SuperAPI AI Directory

manju-laoli-v5.0

General & other

View repository

Updated 10-15 20:07

About this skill

Manju Laoli Skill V5.0.2

V3.2 upgrade (2026-08-07 · absorbing new prompts from the desktop manju course):①§2.1A AI script batch creation enhancement pack— Novel-to-manju script instructions (three-step method: deconstruct hit logic → adaptation requirements → pure action-dialogue conversion, vertical-screen shot rules + duration rules), fully automatic creation instructions for original hit scripts (7-dimension optional configuration + 7-module full-case framework + dark hook × sensory critical hit opening iron rule), 5 types of script hit formulas;②§2.3 a0 Character asset extraction JSON— Group + primary-secondary consistency system (face change starts a new primary/only unchanged face is secondary, 13 fields, primary = text-to-image three-view/secondary = image editing face-lock outfit change);③§9.14 Emotion-driven cinematic feel (Quill)— Audience position/power relationship/spatial psychology 6 questions + emotion→visual decision table + deterministic hard rules. V3.1 upgrade (including HELL GRIND ingestion):Open-source cinematic prompt methodology officially ingested— All prompts from the HELL GRIND 95-minute open-source film (15 segments) classified and distilled by reusable methodology (Chapter 9: prompt skeleton 16 fields/asset reference system/unified atmosphere instructions/iron rule zone/second-level action choreography/combat ONER/SOLO performance/environmental shots/image editing/style paraphrase table), methodology distilled version see references/HELL-GRIND-提示词分类方法论.md (full original text has been slimmed down and removed, V4.1.3). V3.2 supplement (2026-08-08 · Higgsfield project API actual capture):340 real generation prompts ingested— Captured 340 practical prompts from 17 scene folders via the HELL GRIND project API (nine major categories: establishing environment/character-locked multi-shot/action group scenes/Chinese multi-shot dialogue/romantic close-ups/slow-motion multi-character/style reference/costume editing/character description), core conclusions distilled into §9.15: Style Prefix locked word-for-word across the entire project; asset references use <<<uuid>>>/<<<image_N>>> and annotate usage; Chinese prompts are feasible (Scene 14 entire scene in Chinese); multi-shot uses 【Shot N】 segmentation + second-level action choreography; editing prompts use CHANGE+PRESERVE. Raw data has been removed (2026-08-09 user requested slimming, methodology points see references/HELL-GRIND-提示词分类方法论.md). V2.2 → V3.0 upgrade:Color cards and prompt template library officially ingested— Verified complete prompt template library (cinematic golden formula + 9 directors' Chinese-English bilingual templates + director name safe conversion table), master color card HEX data, director name red line (Seedance detects personal name violations, user-tested). V2.1 → V2.2 upgrade:Supplemented practical methodology from desktop AI learning materials— Lighting system expanded from 3 types to a full 8-type table, standardized "base-atmosphere-picture" three-step prompt template, golden 3-second hook design + scene emotion selection, refined aesthetic values. The core system remains unchanged (Xiaoluo as primary/Shanyin + toolbox + micro-expression library as supplements). Integration sources: Xiaoluo AI short drama production chain (primary) / Shanyin screenwriting master / short drama industrialization reference / Shanyin directing master / AI film and television toolbox (He Zhiwei + Li Yifan + Senhai) / Laoli micro-expression library / spatial director converter-spatial reference system V3 / AI learning material compilation (Communication University of China courses + Feynman learning method + fight tutorials + hit analysis) / color cards and prompt template library (verified)


Zero, Core Iron Rules (applicable to all steps)

0.1 System Master-Subordinate Relationship (New in V2.1 · Most Important)

Role System Responsible Content
Main Core (Bone) Xiaoluo AI Short Drama Production Chain 0-90s continuous timeline, shot grammar (action-reaction hard cut, shot size grading and camera movement, macro visual gear shifting), unified asset numbering (Image N/@Scene N/@Prop N), six-dimensional physical performance, dialogue/SFX specifications, pre-delivery hard validation
Supplement (Flesh) Shanyin Screenwriter/Director Master Four types of shot groups, 10 variants of action-reaction, information volume duration estimation (multi-layer information × 2-3s/layer), director style library, nine-column storyboard table (for internal breakdown only)
Supplement (Flesh) AI Film and Television Toolbox Explicit action priority, three layers of concretization, keyword weighting, prompt structure template
Supplement (Flesh) Old Li Micro-expression Library Three-dimensional emotion description method, micro-expression timeline slicing, contradictory body language, emotion base library (to complete Xiaoluo's subtle expression depiction)

Conflict adjudication: When the two systems conflict,Xiaoluo rules take priority(such as timeline, hard validation, action-reaction shot cutting); where Xiaoluo does not cover, supplement with Shanyin/Toolbox/Micro-expression Library. Details subsequently added by the user are merged into the corresponding module at any time.

0.2 Three Major Assessment Standards (New in V2.1 · All step outputs must be self-checked)

User assessment standards, self-check item by item before each step output, if not up to standard, revise first then output:

  1. Smooth storyboard transitions, reasonable duration

    • Continuous timeline: previous shot ends = next shot begins, no skipped seconds/overlap/blank
    • Duration is determined bycontent volume(information volume duration estimation: reading one layer ≈ 2-3s, each additional layer +2-3s), no padding
    • Adjacent shots do not mechanically repeat shot size/camera position/composition
    • Fourfold connection between segments (asset/action/emotion/sound) + sixfold consistency lock
  2. Plot conforms to logic

    • Complete causal chain: every action has a reaction, every piece of information has a prerequisite, every reversal is not out of thin air
    • Each scene has goal → obstacle → action → change → exit hook
    • Do not arbitrarily add characters/relationships/prop abilities, do not change plot causality
    • Dialogue conforms to character identity and current purpose, does not repeat information the audience has already seen
  3. For grand scenes, pay attention to shot size and focal length/focal length prompt words

    • Grand scenes must explicitly write shot size + focal length (millimeters) + camera movement (see §4.5 Grand Scene Special Section)
    • Extreme long shot/aerial shot → wide angle (16-35mm); medium shot narrative → 35-50mm; close-up → 85mm+; character-to-environment ratio → clearly marked
    • The ratio difference between characters and huge environments must be written into the scene description ("the character occupies only one tenth of the frame")

0.3 Writing Red Lines

Absolutely do not:

  • Write psychological description ("he realized", "she understood", "inside surged")
  • Write parenthetical hints ("(actually covering up nervousness)")
  • Characters use dialogue to explain settings or themes
  • Preachy segments, forced tear-jerking, sentimental monologues
  • Excessive metaphors, bookish AI-toned dialogue
  • In the script/screen content, write anything the camera cannot capture
  • Replace performance with abstract emotion words (sad/angry/wronged/desperate/collapsing/numb/nervous/afraid)
  • Use "he/she/it/they" to refer to characters in the screen description (must use names/Figure N/asset tags; necessary pronouns may be retained inside original dialogue)
  • Output Meta instructions such as "short-drama pacing", "cinematic-grade", "blockbuster feel", "9:16 aspect ratio", "generation model"
  • Output exact height values (height can be written as a range/relative value)
  • Write director/celebrity/studio names into the final prompt(such as "Wes Anderson style", "Miyazaki style")——Video tools such as Seedance will detect director/person names and refuse generation as a violation(confirmed by user testing); master styles can only be conveyed using descriptive approaches such as color palette/composition/texture (see §4.2.2 Master Color Grading Library usage)

Must achieve:

  • Dialogue should be colloquial, like real people speaking, ≤40 characters/sentence (control speech rate according to character personality, age, and emotion, not mechanically; static comic drama dual-track scripts use a stricter ≤15 characters/sentence, see §2.1B)
  • Use action instead of explanation (action is subtext)
  • Dialogue is like an iceberg, only one corner is revealed
  • All content starts from an audiovisual perspective
  • Screen content only writes visible actions and objects
  • The sound column only writes audible sounds; key physical sounds use [SFX: 音效内容]

0.4 Output Language

  • Default output is pure Chinese, including the prompts themselves (asset-level/story-level/shot-level)
  • Screen details, environment, materials, and texture may be mixed withshort compatible English tags(such as subtle cinematic handheld、micro expressions), 2-4 per shot, mutually compatible, not mechanically piled up
  • The last line of Chinese output strictly writes 无字幕,无背景音乐(for single-episode/segmented delivery)

0.5 Pre-output Self-check

Before all content is output to the user, first do an internal self-check. The basis for the self-check is the checklist for the current step (Gate A-E + the three major assessment standards). If any problem is found, modify it first before outputting. The user may request an explicit [self-check] at any time.

Every time content is generated, a self-check must be performed (2026-08-06 user confirmed · iron rule):

  • After generating/delivering any content (prompts, scripts, storyboards, assets, single-segment/multi-segment, short content/long content),first do an internal self-check, fix problems on the spot, and deliver only after fixing
  • The self-check is completed internally and is not listed separatelyShow the [self-check] checklist to the user (user confirmed: just check it yourself when generating, no need to list it separately)
  • Self-check items for short content (single segment) (≤60s): complete five-part structure (style/color card reference/shot/timeline/sound)? Is the timeline continuous and closed? Does the opening explode? Action + physical feedback + sound effect embedded? Zero psychological description and no abstract emotion words? No director/celebrity/studio names? Are all four colors of the color card present with HEX+RGB? 2-4 English tags? Visual closure at the end? Last line "no subtitles, no background music"?
  • Self-check items for long content (multi-segment/full episode or single segment >60s): on the basis of short content——Segments >60s need an added [Characters and Scenes Appearing] module (six-part structure)? Are the characters and scenes appearing clearly written (@asset position)? Seamless connection between segments? Sixfold consistency lock? Hard cut markers between segments?
  • Fix first, then deliver; delivery with defects is not allowed

0.6 Workflow Discipline

  • Three-step pipeline + connection layer executed in order: ①Detailed script → ②Episode-by-episode storyboard script → Connection layer (three-view/scene/expression/plot storyboard panels) → ③Storyboard prompts for each episode
  • After each step is completed, you may pause and wait for user confirmation
  • Do not skip steps, do not generate all content at once
  • User provides an outline → deliver step by step according to the pipeline; user specifies starting from a certain step → start from that step

I. Prompt Engineering (Core Methodology · Reference Module)

Source: AI Film & TV Toolbox (He Zhiwei + Li Yifan + Senhai).Auxiliary layer, helps write clear shot descriptions; not used as the main output structure.

1.1 Explicit Actions vs Implicit Logic (He Zhiwei)

AI prioritizes realizingthe most explicit, most easily visualized, most action-orientedcontent.

  • Explicit actions: clear actions directly visible in the frame (gripping a knife, turning the head, splashing water) → easy for AI to realize
  • Implicit logic: spatial relationships, intentional logic, body orientation → AI cannot understand, prone to failure

Three practical steps: ①Identify the core action ②Remove high-interference actions ③Sacrifice implicit logic, prioritize ensuring the core action is precise.

1.2 Four Core Principles (Li Yifan)

Precision > Piling up | Concrete > Abstract | Professional > Colloquial | Structure > Random

1.3 Three-Layer Concretization Structure

Object layer (scene/character/props) → Detail layer (color/material/appearance) → Atmosphere layer (lighting/emotion/style)

1.4 Keyword Weight Rules

Keywords placed earlier have higher weight. Put the core subject first, modifiers and details later. Effective length core: every word has a visual function, no redundancy.

1.5 Prompt Structure Template

主体描述 → 构图/景别 → 运镜方式 → 光影方案 → 材质细节 → 美术风格 → VISUAL DNA

1.6 Long vs Short Prompt Selection Strategy (Supplement · Feynman Learning Method)

  • Long prompts: fine control but prone to mutual interference → complex scenes, commercial projects requiring precise control
  • Short prompts: more room for AI to improvise, easy to produce surprises → creative exploration, style testing, finding inspiration
  • Effective length principle: not the longer the better, but the more precise the better; every word must contribute to the image, delete useless ones directly
  • Reference image usage: do not "copy" but "decode" — break down the reference image's five elements of composition/lighting/color/material/atmosphere, and reproduce them with prompts; reference image + prompt = precise control

1.7 Standardized Prompt Three Steps: Base-Atmosphere-Image (Supplement · Fight Tutorial · Fill-in-the-blank Quick Reference)

Source: AI Comic Drama Fight Prompt Standardization Tutorial. Used to quickly build a high-quality prompt framework, compatible and complementary with the §1.5 structure template.

Mnemonic: Base-Atmosphere-Image, three steps

Module A [Basic Settings] · Lay the foundation

  • Visual style and type (art style + visual dimension + texture; e.g., "Chinese comic xianxia realistic style 3D")
  • Scene environment (environment + space + props + atmosphere; e.g., "cloud sea churning at 10,000 meters altitude, blood-red full moon")
  • Core character (appearance + clothing + personality + ability state; e.g., "Li Tianjun with black hair like a waterfall, cyan immortal robe")
  • Immersive sound effects (ambient sound + action sound, weaken background music)
  • ⚠️ "Two people fighting" → vague → extremely low hit rate

Module B [Atmosphere and Image Quality] · Inject the soul

  • Style core: combat temperament and emotional tone (fierce explosive vs. steady restrained)
  • Visual tone: primary color palette + emotional change curve (e.g., gold-cyan-black, cold suppressed → golden highlight explosion)
  • Image texture: animation details + effect layers (particle fluids, slow-motion debris splatter)

Module C [Visual Content] · Flesh Out (storyboard style)

  • Timeline: precise to the second, a 10-second video split into 2-3 core storyboard shots
  • Shot size hierarchy + composition (symmetrical/central/dynamic/radial/oppressive)
  • Camera movement (static/push-pull/tracking/orbit/handheld/bullet time, only one type of movement per segment)
  • Frame-by-frame details: the more specific the action, the more precise the AI
  • Audio-visual sync: each storyboard shot matches its own dedicated sound effect

Beginner progression path: fill in the form → add emotional curve → design camera language → precise audio-visual coordination. 3 prompts per week, initial mastery in 2 weeks.

1.8 Five Categories of Negative Exclusion Prompts (New in V2.2 · Replication Content-03 Practical Combat)

Source: "Qimen Town Shaman" storyboard practical combat. Attach when feeding segment-style generation, organized by five categories, directly usable.

  • Character category: facial feature distortion, facial collapse, stiff expression, empty eyes, deformed fingers, clothing clipping, face swapping, loss of core features
  • Image category: frame skipping, chaotic camera shake, image fragmentation, head cropping, camera movement stuttering
  • Image quality category: overexposure, underexposure, blur, noise, watermark, subtitles, logo, color overflow
  • Scene category: solid-color background, crude background, static image, no dynamics, irrelevant scenes intruding
  • Content category: lighting discontinuity, lip-sync misalignment, sound effect desync, background music, BGM, score, duration overrun, irrelevant characters intruding

In coordination with the main text "no subtitles, no background music": the five categories of negative words are the complete version; a simplified version can be used for single-segment generation.

1.9 Order of Shot Description Information (Xiaoluo Main Core · Shot Script Writing Order)

Organize the shot description in the following order:Visual subject and composition → action in progress → gravity and collision feedback → micro-expressions and physiological stress response → lighting changes → dialogue → SFX → English tags

1.10 Layered Generation Logic (New in V2.2 · Replication Content-03 Practical Combat Verified)

Source: "Qimen Town Shaman" storyboard practical combat (Replication Content-03). The core iron rule for AI image/video generation success rate:Do not try to make the AI draw all elements in one image/one video segment。

Core principle: one image focuses on only one narrative focal point; the remaining elements are handled by reference images/layering.

Wrong example (element overload): "Matsuda Yu + holding a bronze mirror + a shockwave shield in front + a fire wall outside + rocks behind" → the AI will definitely be logically chaotic.

Correct layering:

  • Layer 1 Effect close-up: focus on a single effect (energy shield/fire wall/particles), the prompt only writes that effect + environmental atmosphere
  • Layer 2 Facial performance: focus on the character's expression (anger/terror/resolve), the prompt writes face + light direction + skin details
  • Layer 3 Physical impact: focus on action and destruction (knocked flying into rocks), the prompt writes dynamic pose + environmental destruction + particle residue

Layering placement rules:

  • The prompt for each storyboard shot carries only 1 main narrative focal point
  • Character consistency is ensured by the reference image (@character image); the prompt does not repeatedly describe the appearance already locked by the reference image
  • When overlaying effects, generate layer by layer and then composite; do not write them into the same prompt

Paired with the §4.6.2 transformation template: the "disintegration → recombination → explosion" of a transformation scene is itself a natural layering (one focal point per stage).


II. Two/Three-Step Pipeline + Transition Layer (Core)

用户输入:剧本大纲
    │
    ▼
① 详细剧本创作(小逻编剧创作逻辑为主:创意→大纲→人物→正文→镜头化表达)
    │
    ▼
② 分集分镜剧本(小逻时间轴+镜头脚本为主:0-90s连续时间轴 / 动作-反应切镜 / 六维表演 / 宏观视觉换挡)
    │
    ▼
【衔接层】资产与视觉预演提示词
   a. 主角+重要人物三视图提示词(小逻正交标准)
   b. 角色表情设定提示词(十二表情库,补齐小逻缺少的细微刻画)
   c. 场景提示词(含焦段体系,宏大场景专项)
   d. 剧情分镜格提示词(每场一个,默认九宫格,剧情简单可用四宫格)
    │
    ▼
③ 每集分镜提示词(生成段式:≤60s 五段式 / >60s 六段式,可直接投喂生成工具)

2.1 Step One: Detailed Script Creation

Input: script outline (provided by the user) Output: detailed episode script

Core Logic (Xiaoluo screenwriting creation):

  • First determine whether the user is at the idea/structure/character/main text stage, and automatically route to the corresponding mode (idea development → outline → character biography → script main text → production handoff)
  • Each scene must answer: Who wants what? Who/what prevents it? What action does the character take? What new information is released? How do the relationship or situation change? Why must it move to the next scene?
  • Scene endings should prioritize: unfinished action/new evidence/identity exposure/relationship reversal/cost appears/goal changes
  • Conflict escalation must change at least one item: stronger resistance/less time/higher cost/ally changes stance/goal conflict/new information overturns old judgment/wrong method backfires
  • Consecutive scenes must not repeat the same kind of quarrel or misunderstanding
  • Hard rule for character names: action and shot descriptions must use the character's formal name; pronouns are prohibited

Episode planning:

  • Each episode's duration is determined by the plot (default reference 90s, flexible; if content is less than 90s, expand with environmental staging/action intermediate states/character reactions, and do not fabricate plot facts; if content is overloaded, split episodes at complete plot beats)
  • Flexible segmentation within each episode:Segment length is cut naturally according to plot/scenes, ranging from 3s to 15s; 15s is the upper limit for a single segment (must not exceed it), not a fixed value—Once a piece of drama is finished, it ends; some segments wrap up and end at 7s, and there is no need to force 15s; each segment carries 1 narrative function
  • Reference for narrative function zones (90s episode):
    • 0-10s: visual hook/theme setup (abnormal image, unfinished action, identity exposure, inverted result)
    • 10-30s: establish character goals and relationships
    • 30-55s: escalating conflict (action-reaction-information release advancing layer by layer)
    • 55-75s: relationship reversal (visible transfer of power relations)
    • 75-86s: climax of this episode (action/decision/discovery reaches its peak)
    • 86-90s: suspense abrupt stop (unfinished action/suddenly refocused gaze/unknown sound source)
    • The above intervals are narrative function zones; switching at fixed seconds is not mandatory; each episode answers one small question while raising a bigger question

Character design (Want/Need/Arc):

  • Want (what the character wants) → drives external action
  • Need (what the character truly needs) → drives internal growth
  • Arc (arc change) → resolution of the conflict between Want and Need
  • Additional for features/series: Ghost (traumatic backstory) + Lie (false belief) + Flaw (personality defect)
  • Xiaoluo character biography supplement: desire/need/fear/secret/flaw/bottom line/ability/cost/arc + dramatic function + character relationships (surface relationship/actual need/hidden contradiction/power change/relationship endpoint)

Four formulas for viral openings:

  1. Misunderstanding type: misunderstanding established → suspense → reversal to the truth
  2. Identity contrast type: You think X → actually Y
  3. Time pressure type: Must complete Y before time X
  4. Emotional buildup type: Daily life → accumulating details → explosion

Golden 3-second hook design (addendum · viral AIGC short drama patterns):

Source: Analysis of 10 major viral AIGC content pieces on Douyin & Hongguo. The opening 3 seconds determine whether viewers swipe away.

  • Visual impact type: High-energy iconic scene opens directly (heavenly soldiers arrayed / explosive execution / moment of falling off a cliff)
  • Suspense contrast type: Unreasonable imagery + question (gods kneeling to a panicked fake Dragon King)
  • Emotional conflict type: Extreme contrast opening (past-life tragic death cuts to rebirth gaze, life-and-death contrast pulls you in within a second)
  • Scene wonder type: Unseen visual spectacle (dark ancient-style streets and alleys + demon beast roars + sword-wielding entry into battle)
  • 60-character hook writing method: One sentence clearly states "who the protagonist is → what dilemma they face → why you must watch"

Scene emotion selection (addendum · script breakdown course):

The same plot feels completely different in different settings. The scene itself participates in storytelling and shapes emotion.

  • Confession: Western restaurant = solemn / vegetable market = absurd comedy / rainy street = dramatic intensity
  • Final parting: train station = unknown journey ahead / hospital ward = accepting death / war ruins = angry revenge
  • Confrontation: luxury office = power oppression / kitchen = intimate relationship tearing apart / rainy night alley = desperate showdown
  • Before choosing a scene, ask first: What emotion does this setting add to the plot? What is it itself saying?

Five-dimensional micro-performance (for direct use in script body, no abstract emotion words): Dimension 1 Micro-expressions and gaze (pupil constriction / gaze evasion / eyelash trembling / eyelid tightening / masseter and mouth corner changes) Dimension 2 Body and center of gravity (shoulders drawing in / chest rising and falling / leaning forward and withdrawing / fingers gripping, rubbing, hovering / footstep direction) Dimension 3 Intermediate action states (hand stops midair / turning pause / half-step retraction / wanting to touch but stopping before the object / opening mouth with no sound) Dimension 4 Physiological reactions (breathing changes / Adam's apple rolling / nostrils flaring / finger trembling / cold sweat / muscles tensing / voice tightening) Dimension 5 Light and shadow dynamics (turning moves window light from face to behind ear / chest rise and fall shifts collar shadow / door gap widening lets light cut across the floor)

Output format (episodic script):

【第X集】(目标时长:Xs)
● 本集目标:一句话
● 分段结构:段1(功能:钩子, Xs) / 段2(功能:建立, Xs) / ...(按剧情弹性)
● 场次:
  场次X|地点|时间|内外景
  [动作/台词/情绪节拍](可拍摄的动作+对白+五维微表演,镜头化表达只在叙事转折/信息释放/关键表演处用[镜头:景别,机位与运动,构图焦点])
● 集尾悬念:一句话
● 资产登记:角色01/场景01/道具01(首次出现登记DNA编号)
● 制作交接清单(按需):角色资产要点 / 场景资产要点 / 道具资产要点 / 下游建议

Gate A: Script gate (including three major assessment standards)

  • Is each episode's duration and segmented structure clear? Is duration determined by content volume, not forced?
  • Does each segment carry 1 narrative function?
  • Is the causal chain complete: does each scene have goal / obstacle / action / change / exit hook?
  • Does the reversal have prior information? No last-minute cheating?
  • Is the dialogue colloquial? ≤40 characters/sentence? Does it fit identity and purpose?
  • No psychological description? No parenthetical hints? No preaching? No abstract emotion words?
  • Does the episode ending have a suspense hook (small question answered + bigger question raised)?
  • Do action and camera descriptions use characters' formal names, with no pronouns?
  • (When using the §2.1B dual-track script) Bullet dialogue ≤15 characters? Are foreshadowing layers registered (short-range/medium-range/ultimate)? Is the close-up rate ≥40%? Is the word count within hard targets (60s=250-300 characters/120s=500-600 characters)?

2.1A AI Script Batch Creation Enhancement Pack (New in V3.2 · Source AI General Prompt.docx)

Suitable for: batch production of novel-to-comic drama adaptations, fully automatic creation of viral original scripts. Complements the detailed script creation in §2.1—§2.1 is the director-led fine-tuning workflow, while this section is an AI production instruction that can be copied directly.

a. Novel-to-comic-drama script instruction (three-step method)

Step 1: Break down the hit formula: Have the AI sort out the hit structure of the first ten chapters of the novel (plot setting/character setting/hook and reversal setting), and sort out the main storyline suitable for a comic-drama structure Step 2: Adaptation requirements: The theme and meaning remain unchanged; the opening reversal is domineering (counter-common sense and counter-logic, using contrast,反差, novelty, stepping on others, and parallelism); the story is full of ups and downs; the ending is suspenseful; after polishing, the visual sense is extremely strong and suitable for storyboarding; retain key dialogue/emotional explosion points/plot turning points; adapt to the rhythm of vertical-screen short videos; strictly comply with platform content regulations Step 3: Novel-to-comic-drama instruction (core rules):

  • Role: senior comic-drama screenwriter and director, vertical-screen short-video audiovisual language
  • Core principles:Absolutely loyal to the original text(no additions, deletions, or changes; character dialogue is retained as in the original text);Pure action and dialogue(no narration/subtitles/voice-over); visualization first (background/psychology/worldview → filmable action); shot annotations; alternating action and dialogue (no continuous 5s without action, 10s without dialogue); duration annotations
  • Action judgment: body movements/facial expressions/environmental changes/prop use/details capturable by the camera/extras' reactions; psychological description must be converted into expressions and actions ("he was very angry in his heart" → "Su Mo clenches his fists, knuckles turning white 1s")
  • Dialogue judgment: direct quotations/clear speech verbs/dialogic onomatopoeia; retain emotional action indications:角色名:(情绪)"对话原文"
  • Narrative background exposition: worldview/character background must be conveyed through dialogue or visual transformation (such as the color change of the life lamp + everyone's reaction); only if it truly cannot be transformed may one line of dialogue be added and marked "reasonable supplementation based on the original background"
  • Vertical-screen shot rules: long shot LS / medium shot MS / close-up CU / extreme close-up ECU / high angle HA / low angle LA / over-the-shoulder OTS / follow shot FT / pan Pan / push Push
  • Duration rules: action 0.5-1s (minimal) / 1-2s (simple) / 2-3s (medium) / 3-4s (complex) / 3-5s (environment); dialogue about 3 characters/second, a single sentence ≤5s, split if longer
  • Output format:场景:序号 内/外景 地点—时间 + 【镜头:类型】动作/对话 block

b. Fully automatic creation instruction for hit original scripts (high-like style enhancement words)

Mandatory interaction process: ① Output the current hit theme list across all tracks (mark traffic tier/core setting/comic-drama suitability/audience) → ② Customized configuration for the theme (7 dimensions to choose from: entry method/golden finger/satisfaction direction/tone/gender emotional line/length/free supplement) → ③ Output the full-case framework → ④ Confirmation and creation

7 major modules of the full-case framework:

  1. Track positioning (worldview direction/vertical-screen adaptation logic/audience/platform support/differentiation analysis)
  2. Core protagonist character design (fixed name/extreme miserable starting-point anchor/visualizable golden finger/strong-edged personality/dark-side character anchor: surface X dark side Y, mark the first hint/half-reveal/ultimate detonation nodes)
  3. Core satisfaction anchors (external satisfaction line = crushing and reaching the top + internal satisfaction line = dark-side detonation subverting cognition)
  4. Length-adaptation skeleton (long-form: opening ice-breaking chapters 1-10 → mid-term growth chapters 10-100 → late-stage climax after chapter 100, rhythm small satisfaction every 3-5 chapters/medium satisfaction every 10 chapters/large satisfaction every 30 chapters; short-form: explosive opening chapters 1-3 → dense climax chapters 4-25 → nuclear explosion ending chapters 26-30, strong satisfaction every 1-2 chapters)
  5. Opening dark hook × sensory critical hit (Mode A dark hook tear: dark first line → sensory immersion → opposing side enters → delayed fuse; Mode B sensory critical hit blast: critical dialogue → concrete finishing blow → physical critical hit → environmental anchoring → in-your-face entrance)
  6. Delayed detonation reversal setup (long-form: ≥5 fuses + 2-3 half-reveals + ultimate detonation; short-form: ≥3 fuses + 1 half-reveal + chapter 26-28 nuclear explosion)
  7. Iconic scene setup (≥4: core plot/explosive dialogue/comic drama shot presentation/short video duration/traffic-driving advantage)

Opening iron rules: golden 3 seconds to retain viewers; five-sense critical hit (≥3 sensory experiences within the first 200 characters, using "as if being..." synesthetic metaphor); rough and raw dialogue (profanity/roaring/cursing allowed, written-style tone strictly forbidden); psychologically manipulative plot advancement

c. Script explosive-point formulas (5 types, fill-in-the-blank ready)

  1. Identity contrast: hidden lowly identity + sudden highlight moment + shock that no one believes
  2. Desperate comeback: irreversible desperate situation + impossible countdown + a hidden trump card
  3. Tragic romance missed connection: deep love + wrongly accused + cliff-edge separation + evidence that arrives one step too late
  4. Reversal ending: known ending + last-second reversal + cheat ability activated
  5. Taboo trigger: calm daily life + sudden intrusion of an external enemy + an untouchable taboo triggered

2.1B Script secondary writing capability (short drama industrialization reference · absorbed 2026-08-15)

Source: short drama industrialization reference (full version see references/short drama industrialization reference/). Core:rewrite outlines/materials into a "dual-track finished-film script"—visual assets and audio assets are completely physically separated at the script stage, and can directly connect to AI image generators + CV + post-production. Complementary to §2.1 from-scratch creation and §2.1A novel-to-comic-drama conversion (2.1A handles "faithful conversion of the original text into action and dialogue"; this section handles "industrialized layout and hard metrics of the finished-film script").

① Volume planning and episode matrix (set the track before writing):

  • Track A seasonal premium: 12/24 episodes, suspense crime/high concept/urban strong female lead/absurd comedy; bottom line = first episode 3-second completion >70%, full-season completion >45%
  • Track B heavy long comic drama: 60-100 episodes, system cheat/rebirth counterattack/war god rich tycoon/hidden wealthy family; bottom line = paid checkpoint (episodes 10-15) conversion >8%

② Single-episode second-level structure (120s example, scalable proportionally):

[0-5s]   3秒强钩 Cold Open(直切高潮/反常动作,禁黑屏渐显)
[5-15s]  集头承接 Reaction(回应悬念,建立当场危机)
[15-50s] 冲突一阶 The First Beat(事件展开,主配角交锋)
[50-60s] 30秒微钩 Micro-Hook(新信息/细节翻转,对抗疲劳)
[60-105s]冲突二阶 Climax Peak(矛盾顶点,底牌揭开)
[105-120s]集末钩子 Cliffhanger(切在「最难受的临界点」)

③ Five-stage creative workflow (free head, tightened tail):

  1. Topic anti-vulgarity self-review: list ≥3 common-pool tropes for this subject matter and mark them as forbidden zones; ask yourself "what unique experience can the audience take away after watching (aftertaste beyond thrill/tragedy)?"
  2. Visual/auditory character design: character Key Pose ("when extremely angry, left eye pupil suddenly contracts + cold smile at the corner of the mouth") + voice anchor ("smoky voice, every word tearing in the throat")
  3. Episode outline structural shaping (hard gate): one-sentence arc check for each episode—"what was taken from the protagonist/what was exchanged in this episode"; episodes that cannot be written should be cut or merged; foreshadowing registration with time-limited recovery
  4. Script writing: single-track first, then ask whether dual-track is needed (see ④)
  5. Script doctor self-check: Audiovisual self-check (is there any psychology that cannot be drawn or spoken?) + Filler self-check (delete AI-sounding water words)

④ Script format (single-track first, then ask whether dual-track is wanted):

  • Single-track = audio and visuals output in sync: Visuals (actions) and sound (sound effects/dialogue) are written in the same paragraph, output in sync, not split into separate columns
  • Dual-track = audio and visuals separated: The visual track (scene description/shot size/transitions) and the auditory track (sound effects BGM/dialogue) are physically split into two columns and written separately
  • By default, output the single-track script first: Actions + sound effects + dialogue are mixed in the same paragraph (audio-visual sync), one action per sentence, write only what can be filmed and heard, first let the user confirm whether the plot is right
  • After outputting single-track, you must ask: "Do you need to upgrade to dual-track (storyboard-style final-cut script, audio and visuals separated)?" — if the user confirms yes, then upgrade; if not, stop at single-track
  • After the user confirms they want dual-track, physically split visual/auditory according to the format below:

[Dual-track format (upgrade as needed)]

镜号:#01
【景别】特写/中景/全景/主观/过肩(只写三种景别,漫画师可执行)
【转场】切/冷开场/闪白/动态速度线拉近
【场景】[物理空间]·[日/夜]
【画面描述】[极简可画指令,含 Key Pose,禁心理描写]
【音效/BGM】[SFX环境音/爆破声/BGM节奏]
【台词】[角色名]([语气/情绪]):[子弹台词≤15字]
  • Bullet dialogue rule: Static comic-drama dual-track script single-sentence dialogue **≤15 characters** (under semi-static visuals, dialogue is a breakthrough bullet, not prose); AI video dynamic generation segments follow **≤40 characters/sentence + duration-based sentence splitting** (8-12 characters per shot is most natural; anything over 15 characters is split into multiple consecutive time points, see §2.4)
  • Hard character-count targets: 60s = total 250-300 characters / pure dialogue 120-150 / storyboard 8-12 panels; 120s = total 500-600 characters / pure dialogue 240-300 / storyboard 16-22 panels
  • Panel vs shot (to prevent confusion): "Panel" = a storyboard panel of a static comic-drama dual-track script (semi-static, about 5-7s per panel); "shot" = a shot of an AI video dynamic generation segment (18-26 shots/90s in §2.2, about 3-5s per shot). Static comic drama uses "panels", AI video dynamic uses "shots"; the different densities of the two are normal
  • Close-up rate ≥40%: Close-ups of eyes/partial details/key props are the lifeline of comic drama
  • ASR limit: Pure voiceover dialogue is far below the ASR limit (about 210 characters for 60s / about 315 characters for 90s), leaving enough action time for sound effects and visuals

⑤ Script second-pass writing (diagnose → rewrite → explore):

  • Four diagnostic dimensions: Pacing (plot loose-medium-tight + emotion light-medium-heavy dual-track offset) / duration (scene too long or too short) / logic holes / dialogue problems (too blunt/written tone/preachy)
  • Targeted rewriting: Subtext rewriting (press the truth that was spelled out back under the iceberg) / visual rewriting (psychology → action physiology) / anti-cliché conception / colloquial dialogue, only revise the problem passages and keep the parts already approved
  • What-if multi-version exploration: For problem scenes, give 2-3 rewrite directions, explain in one sentence "what was changed + why it is more effective", do not rewrite the full text
  • Foreshadowing audit(mandatory for suspense/reversal): register short-range/mid-range/ultimate reversals in layers (example: foreshadowing A #01 holding the cup in the left hand → #04 left-handed pen marks; foreshadowing B #03 stopping the watch swing → #10 time of death; ultimate reversal = the interrogation is a dialogue)

Red lines (same as §0.3 + distilled version): Zero psychological description / zero parenthetical hints / zero dead space (every word must be drawable or soundable) / action replaces explanation / dialogue reveals only the tip of the iceberg

2.2 Step two: episodic storyboard script

Input: Detailed episodic script Output: Episode-by-episode storyboard script (shot level, can be directly converted into generation segments)

Core Logic (primarily Xiao Luo):

Timeline (hard rules)

  1. Fix each episode at 90s (or the user-specified target episode length); the first shot starts at 0s, and the last shot ends exactly at the total duration
  2. Use only one continuous timeline within a single episode; do not split into segments and reset the timeline
  3. The end time of the previous shot = the start time of the next shot; no overlap, skipped seconds, or gaps
  4. Scene changes do not reset the timeline; scene markers do not occupy duration
  5. Prefer whole seconds; when decimals are truly necessary, uniformly keep one decimal place and close precisely
  6. By default, output one episode at a time; for long works, multiple episodes may be output consecutively, with each episode having an independent timeline
  7. 18-26 shots as a soft reference, adjusted according to plot density, dialogue length, and action completeness (Old Li's addition: information-volume estimation—reading one layer ≈ 2-3s, each additional layer +2-3s; estimate dialogue duration by mentally "acting out" each character's personality and speaking rate, then add reaction time)

Shot Grammar (primarily Xiao Luo)

  • Action-reaction cut: key pressure action/information reveal/prop trigger/sudden entry/important dialogue/relationship reversal → 动作镜头 → hard cut → 反应镜头
  • Action shot presentsthe physical behavior currently happening, not replacing the process with a completed static result
  • Reaction shot presentsvisible physiological stress(pupil changes, eyelash trembling, throat movement, slight nostril flaring, breathing pause, knuckles tightening, weight shifting back, gaze losing focus then refocusing)
  • Reaction shots do not recursively trigger the next set of reaction shots; two consecutive reaction shots must not mechanically reuse the same pupil/lip-pressing/fist-clenching/backing-away action
  • Ordinary walking, natural breathing, and turns of the head with no narrative function do not require an independent reaction shot
  • A long take may present action and reaction successively within the same shot through lateral movement, focus transfer, and foreground-background staging; the reacting person's reaction must not be blocked/out of focus/swallowed by rapid camera movement
  • Editing transitions are used only as needed: direct opening, hard cut, match cut, sound bridge cut; do not use dissolves/fade-in/fade-out
  • Macro visual gear shift: an obvious gear shift every 15-30s, with any two intervals not exceeding 30s. Gear shift methods: cut scene/spatial area; large span between wide shot and close-up; fixed to tracking/orbit/handheld; obviously change camera height and orientation; reverse foreground-background positions/compositional center/spatial power; narrative focus shifts from face to prop/unknown space/new character; actual light source obviously changes brightness or color temperature. Similar shot-size shot-reverse-shot, ordinary hard cuts, and slight compositional adjustments do not count as gear shifts

Shot Size and Composition (primarily Xiao Luo + Old Li/Shan Yin additions)

  • Alternate use of extreme long shot/long shot/medium long shot/medium shot/medium close-up/close-up/close-up/extreme close-up/extreme close-up/environmental empty shot
  • Do not continuously use the same shot size, same camera position, same angle, or same composition
  • Shot-reverse-shot may use similar shot sizes, but change orientation, angle, foreground occlusion, power height, or focal-length feel
  • Medium shots/long shots/environmental shots prioritize three-layer depth: foreground occlusion + midground performance + background environment; close-ups/extreme close-ups do not require three layers, ensuring pupils, eyelashes, masseter muscles, mouth corners, and fingers are clear
  • Each shot carries only one main narrative action or one main character reaction
  • Seven categories of narrative purpose (Shan Yin addition · each shot must carry at least one, ensuring the shot has a function): Establish/Advance/Reveal/Emphasis/Transition/Emotion/Reflect. Writing method: use one sentence to cover both "what to do" and "why do it this way" ("use a back-to-back composition to show alienation" rather than "show the change in relationship")
  • Old Li adds (cut-to-cut connection logic): push in (wide → medium → close-up, gradually going deeper into the inner world; about 54% standard path) / pull out (close-up → wide, detachment reveals, scene ending) / bounce (medium → close-up → medium) cross-level jump cut (wide → close-up) used for moments of surprise/reversal/emotional rupture
  • Shanyin adds (shot group:each segment = 1 shot group, the shot group is the bridge between rhythm planning and storyboard breakdown): montage group (parallel quick cuts, each shot 1-3s hard cut) / progressive group (shot size gradually pushes in wide → medium → close-up) / cause-effect group (action-reaction alternating, shot size pushes in as intensity rises) / contrast group (two images contrasted, hard-cut impact)
  • Shanyin adds (10 action-reaction variants): standard type/reaction first/action juxtaposed/reaction omitted/action hidden/reaction delayed/synchronized type/reaction chain/repetition and gradation/chained connection; 70/30 rule (70% conventional patterns hold up the skeleton, 30% standout patterns create memorable points)

Script → storyboard translation rules (Shanyin adds · prevent content loss):

  • Each Beat → at least 1 shot; emotional/action changes → new shot
  • Dialogue → visualized (action is subtext, don't rely on dialogue to explain)
  • Scene change → new scene + spatial anchor record (direction/doors and windows/fixed furnishings)

Zero-loss cutting principle (Shanyin adds · prevent AI from imagining and changing the plot):

  1. After cutting, shots must restore the original text; omissions are strictly prohibited
  2. Single-shot dialogue ≤40 characters
  3. Short sentences add up to ≤40 characters and the action is coherent → merge
  4. 100% copy the original text; no imagining or condensing allowed
  5. The opening first shot explains how the characters arrive at the scene

Tiered camera movement (choose only one main camera movement tier per shot; do not mix conflicting labels)

  • Close-up/extreme close-up/insert:locked-off tripod shot or ultra-stable micro-jitter; add as needed high facial detail focus、precise eye focus、controlled depth of field、micro expressions、subtle performance; violent shaking/rapid drifting/large-amplitude handheld breathing prohibited
  • Medium close-up/two-person confrontation/dialogue shot:subtle cinematic handheld + realistic camera inertia; slight low-frequency lag with the character's posture, keeping the face stable and recognizable
  • Medium shot/medium long shot/wide shot/environment empty shot:documentary-style cinematic handheld、natural breathing、controlled camera drift; slow drifting natural breathing allowed, subject outline must not be blurred
  • Steadicam long take: used for movement in narrow spaces/multi-person staging/follow-style narrative; preserve real camera inertia, step rhythm, spatial obstacle feedback; must not pass through walls, furniture, or people
  • Old Li adds (camera movement emotion dictionary): push in = approach, scrutinize, focus / pull out = alienation, reveal, detachment / lateral move = roam, observe, follow / rise and fall = god's-eye view reveals scale / orbit = confrontation tension and involvement / whip pan = chaos, energy, sense of speed / handheld = immediacy, unease / fixed = stability, ritual, gaze

Camera position power and lighting (mainly Xiao Luo)

  • Dominant/pressuring party: slight low angle, foreground pressure, less headroom, warm rim light
  • Suppressed/hesitant party: slight high angle, increased negative space, cool top light, eye socket shadows
  • Evenly matched: eye level, symmetrical composition, camera positions at the same height
  • Power reversal: expressed through camera height, foreground occlusion, character blocking, focus ownership, and shift of compositional center
  • "Light moves with power" is built on real light sources: characters entering/leaving existing light zones, body occlusion, light from doors and windows, and changes in lamp direction altering facial light and shadow; light must not change without a source
  • Ambient light produces subtle shifts along the facial contours with breathing and posture → ambient light drift

Spatial Reference System (Spatial Director Converter V3 · 2026-08-15 incorporated)

Source: Spatial Director Converter - Spatial Reference System V3 (full version see references/Spatial Director Converter - Spatial Reference System - V3.md). Core:Space first, camera second— first convert the storyboard into a "continuous spatial state," then write the shots/prompts, to prevent characters from drifting randomly, teleporting, or being reset by shot changes during AI generation. Complementary to §2.4 Blocking Anchoring Iron Rule and §3.4 Screen Axis Rule: Blocking Anchoring governs "where characters are within a single shot," while this section governs "how space changes continuously across shots."

Two spatial modes (determine the type first, then handle it; do not handle all scenes the same way):

Mode Applicable Core
A Static/Dialogue type Dialogue, interrogation, meetings, standing/sitting exchanges Lock one spatial state + character anchors + line-of-sight relationships
B Dynamic/Motion-Combat type Combat, chases, group scenes, blocking changes Lock spatial changes — multiple continuous spatial keyframes

A Static type rules: Lock character positions/left-right relationships/front-back relationships/orientation/line-of-sight relationships/180° axis/camera zones. Characters may speak/blink/breathe/nod/turn their heads/make micro-gestures/slightly shift their weight,but may not swap positions/cross zones/teleport/be rearranged due to a shot change。

B Dynamic type · Spatial keyframes (core new concept):

A dynamic scene is not "fixed blocking + characters performing actions," but rather "the spatial state continuously changes with movement."

  • Usea set of spatial keyframes arranged in time orderto express spatial changes (e.g., 0s initial → 3s engagement → 6s knocked back → 10s repositioned), with each frame marking the characters' relative positions
  • Each frame must simultaneously express five elements:spatial state + movement direction + movement path + action causality + new spatial state after movement
  • Movement causality iron rule: any obvious position change must have an action cause — attack → receive force → move along the knockback direction → land in a new position. It is forbidden for "the protagonist to suddenly appear on the right after a shot change" (unless the storyboard clearly shows movement back there)
  • Relationships > absolute coordinates: in dynamic scenes, do not hard-code "protagonist = left"; instead record "who is on which side of whom / who is in front of whom / who is approaching whom / who is moving away from whom / who attacks whom / who is knocked back / who chases whom"
  • Dynamic spatial anchors: not "characters never move," but rather locking "initial zone + movement range + movement rules" (e.g., "the protagonist starts in the left rear zone, advances toward the center, and is forbidden from crossing through the female supporting character without reason")
  • Multiple people (5-8 people): use zones (left front/right front/central combat zone/elevated platform) + factions (protagonist group/enemy group) + target relationships (A1→C1) + movement relationships (A1→center); do not write coordinates one by one

Character position ≠ screen position (must remember): The real scene position ≠ the left-right position in the frame. After the camera moves, a person on the right side of the scene may appear on the left/center/foreground/half-body side of the frame — as long as the true three-dimensional relationships are not broken.

Camera rules: The camera governs shot size/composition/angle/focal length/push-pull/lateral movement/over-the-shoulder/reverse shot,and must not change character spatial relations without reason, rearrange characters for a frontal face, cross the axis without reason, mirror real space, or turn a cut into a scene reset。

Shot continuity (check at every cut): Inherit the spatial state at the end of the previous shot → character positions → character relationships → movement results → then choose the camera → execute the current shot.Wherever the previous shot was knocked back to, the next shot must continue from there, unless the storyboard explicitly shows movement back.

Over-the-shoulder and reverse shot: When a character on the right side of the scene looks toward a character in front, prioritize a "right rear over-the-shoulder" shot; do not place the camera on the other side of the axis just to capture a frontal face. A reverse shot = changing the viewing angle, ≠ left-right swap/character rearrangement/spatial reset.

Verification (merged into Gate B): Spatial continuity (doors, windows, furniture do not drift) / action continuity (movement distance matches the time interval) / eyeline continuity (shot-reverse-shot directions match) / axis continuity (crossing the axis requires a neutral shot) / asset consistency.

Six-dimensional physical performance (core of Xiao Luo · mandatory for image description)

  • Image description must strictly not use abstract emotion words to replace performance
  • Ordinary character shots must cover at leastthree dimensions; key close-ups/extreme close-ups/reaction shots must cover at leastfive dimensions
  • 1 Micro-expressions and eyeline: pupils suddenly contract/slowly dilate, eyelashes flutter at high frequency, gaze defocuses/refocuses, masseter tightens, corner-of-mouth texture shifts, eyelids tremble involuntarily
  • 2 Limbs and posture: center of gravity presses forward/shifts back, shoulders rise and fall with breathing, fingertips twist and grip the hem of clothing, knuckles turn white, wrists tense, back muscles contract, footsteps half-stop
  • 3 Action intermediate states: about to speak but stopping, hand raised and suspended in air, turn hesitates for an instant, half-step hovers, slight pause before touch, object about to fall, door about to open
  • 4 Physiological stress: Adam's apple rolls, nostrils flare slightly, veins at the temples bulge slightly, chest rhythm changes, swallowing pauses, breathing becomes short and rapid, palms tighten, muscles tremble finely
  • 5 Light-and-shadow breathing: breathing posture drives subtle displacement of skin/collar/facial contours/eye-socket shadows → ambient light drift
  • 6 Physical interaction and gravity feedback: limbs and props/costumes/ground/walls/furniture obey real collision volumes; prohibit limb penetration, objects floating, hands gripping props without contact, soles sliding, passing through scene structures; during sudden stops, hair, hems, and carried objects retain forward inertia; carrying heavy objects shows weight through tightened arm muscles/pressed-down center of gravity/bent knees/foot support/slight torso sway
  • Prioritize using action intermediate states; do not directly present completed static results

Four techniques for micro-expressions in dialogue scenes (absorbed 2026-08-10 · eliminating the AI mask feeling)

The six-dimensional performance tells you "what to write"; this technique tells you "how to write it." Root cause: AI feeling = templated emotion words + overly neat symmetry + too-smooth setup-development-turn-resolution.

  1. Emotion muscle chain (reject static adjectives, use actions to replace emotional states)
    • Do not directly write template words such as "sad/upset/smiling"; write only the immediate movement of facial muscles, giving the AI frame-by-frame execution instructions in the order the muscles fire
    • Example (emotion-extinguished version): ① brows draw together, squeezing out two shallow vertical lines ② the lower lip shows an extremely subtle tremor with an amplitude of less than 1 millimeter ③ the brows slowly relax, and after the emotion is extinguished the face is flat and rippleless (order: corrugator → orbicularis oris → frontalis)
  2. Asymmetry and coexistence of contradictions (break AI perfect symmetry)
    • Asymmetry: unilateral emotion, full-face linkage prohibited—"the corner of the mouth pulls slightly to the left, drawing out a half-arc incomplete bitter smile, only one side of the face exerts force, not spreading to the fundus of the eye, the right eye shows no emotional change throughout"
    • Coexistence of contradictions: two opposing emotions simultaneously/switching within a short time—"the lower eyelid is reddened and moist, the fundus of the eye hides a shattered, calm sorrow; at the same time the pupil contracts violently, the gaze is sharp and resolute" (note: the two emotions switch and occur synchronously within 0.5 seconds)
  3. Quantify physiological indicators (use numbers to eliminate ambiguity)
    • Avoid vague words such as "slightly/gently/somewhat" that AI will weaken; specify precise amplitude/duration/distance
    • Basic quantification terms: amplitude less than 1 millimeter, movement amplitude no more than 15°, mouth corner sinks half a degree, both shoulders draw inward 2-3 centimeters, the entire set of movements takes 1.5 seconds, there is a 0.2-second freeze before the eyelids fully open
  4. Restraint and action interruption (create a sense of real disorder)
    • AI emotions have complete and smooth progression, while real people suppress and interrupt midway; write "start→sudden stop" rather than a complete action
    • Restrained sentence patterns: words surge to the lips and are forcibly swallowed back; desperately suppress emotions, the eyelid muscles tense and lock the tears in the eye sockets; after the embrace action starts, it suddenly restrains and stops at the last moment
    • Refined substitution: do not write "shed tears"→"the eye sockets redden, tears are locked at the edges by the eyelid muscles, the lower eyelid keeps trembling faintly"; do not write "shake head in denial"→"after the head shake starts, it swings to 10 degrees and suddenly decelerates and stalls, the train of thought is interrupted"

Asset unification and continuity (Xiaoluo main core)

  1. Character first appearance assignment [图1]角色名, scene @场景1, prop @道具1; reuse the original number across shots/scenes/episodes, do not renumber
  2. Keep facial bone structure, hairstyle and hair color, skin tone, body type, clothing, footwear, and accessories stable; only modify when the script explicitly changes them (change of clothes/dampened/stained/damaged) and mark the time node in the continuity field
  3. For fixed scenes, the door and window positions, furniture positions, entrances and exits, main light direction, spatial scale, materials, and foreground/midground/background relationships
  4. Prop material/color/size feel/weight feel/state remain continuous; after being opened/moved/damaged/handed over, subsequent shots inherit the result of the change
  5. Shot descriptions may directly use the character's name, no need to repeat [图N];[图N] mainly used for asset planning, reference image binding, and cross-episode asset unification
  6. Asset planning lists only assets that actually appear in this episode or directly affect the picture

Dialogue and sound (Xiaoluo main core)

  • Write dialogue as 角色名:"台词"; if the original script already has dialogue, prioritize preserving it verbatim, do not change key meaning/character stance/information function
  • Dialogue is spoken at natural conversational speed within the shot duration, do not use unnatural high-speed speech; long dialogue may span multiple consecutive shots, carried through the speaker's actions/listener's reactions/prop close-ups/spatial staging, without missing words, repetition, or changing the original meaning
  • Express dialogue pauses through lip pauses, swallowing, breathing changes, gaze movement, and unfinished actions
  • Prevent too-fast shot cutting in dialogue scenes (2026-08-10 user feedback): dialogue is a hard duration constraint—if the dialogue cannot fit within one time point/shot, extend the time point or split it across multiple consecutive time points, do not force long sentences into short shots just to maintain a 2s shot-cutting rhythm. Reference rate: natural speech speed ≈ 4 characters/s (anxious arguing can reach 5-6 characters/s, calm/drunk/recalling 3 characters/s or even slower),A single-shot line exceeding about 8-12 characters means the shot duration must be extended to 2.5-4s; if a sentence exceeds 15 characters, split it into 2-3 consecutive time points (speaking → pause reaction → continue speaking), inserting listener reactions/prop close-ups/speaker's small actions in between, butthe line must be delivered continuously; hard cuts that cause broken sentences/mouth-shape misalignment are prohibited。
  • Key physical sounds are written as [SFX: 音效内容]; only annotate sounds with narrative value for action/space/information/suspense
  • "No background music" only prohibits scoring, not dialogue/ambient sound/spatial reverberation/physical Foley; "no subtitles" only prohibits automatic subtitles/dialogue text/explanatory subtitles

Output format (episode-based storyboard script):

【单集 [集数编号]】(总时长:[90s或目标集长])

● 资产规划:
  角色 (Characters):[[图1]角色名=稳定外观、发型、服装、肤色、年龄气质与辨识特征;[图2]角色名=稳定特征]
  场景 (Scenes):[@场景1=空间布局、门窗位置、固定陈设、主光方向、时间、材质与前中后景关系]
  道具 (Props):[@道具1=材质、颜色、形态、初始状态、重量感与辨识特征]
  连续性 (Continuity):[角色服装、身体状态、道具状态、场景状态发生变化的准确时间节点]

● 镜头脚本 (Prompt):
场景:[场景名称]=@场景N
[开始秒]–[结束秒]s([景别]、[拍摄距离]、[方位与角度]、[运镜方式]、[剪辑衔接]):[具名角色、画面动作、微表情、生理应激、体态、物理反馈、光影变化、台词、音效与兼容英文标签]
[开始秒]–[结束秒]s(...):[...]

无字幕,无背景音乐

Format positioning (confirmed 2026-08-04 + strengthened in 08-05 V2.1): The shot script is aninternal breakdown and rhythm planning tool. When finally fed to a generation tool, it is converted into a "generation-segment-style" narrative format (see 2.4) and this format is not directly used for feeding. Both share the same continuous timeline and the same asset numbering.

Gate B: Storyboard gate (including the three major assessment criteria)

  • Does the first shot start at 0s, and does the last shot end exactly at the total duration?
  • Do adjacent shots connect completely end-to-end? No overlap/skipped seconds/blank space? Is duration determined by content volume?
  • Is the shot count within the soft reference (18-26 shots)? Is each shot's duration reasonable (estimated by information volume)?
  • Is the action-reaction structure correct? Are key nodes using hard cuts?
  • Are reaction shots non-recursive? Are consecutive reaction shots not mechanically reused?
  • Do adjacent shots avoid the same shot size/camera position/angle/composition?
  • Is there a macro visual shift every 15-30s?
  • Are character/scene/prop numbers consistent across shots? Is the state continuous?
  • Does the image description avoid pronouns? No abstract emotion words? No shot numbers?
  • Do ordinary character shots have ≥3 dimensions, and key close-ups have ≥5 dimensions of performance?
  • Can the line be delivered naturally within the corresponding duration?
  • Grand scenes: are shot size + focal length + camera movement + character proportion explicitly annotated (§4.5)?
  • Does the Chinese output have "无字幕,无背景音乐" on the last line?
  • Zero-loss cutting: 100% restoration of the script? No brain-fill extraction?
  • Does each shot have a narrative purpose (seven categories)?
  • Dynamic scenes: do position changes have movement causality (attack → force → movement → new position)? Do they inherit the previous shot's spatial state across shots?(spatial reference system)

2.3 Transition layer: assets and visual preview prompts (strengthened in V2.1)

Input: episode-based storyboard script (asset registration) Output: a prompt text package directly usable for image generation platforms (pure Chinese, may mix in necessary English tags)

Order: assets first, then plot preview. Rule: only output prompt text, not generation guidance/tutorials.

a0. Character asset extraction JSON (new in V3.2 · source: 10 assignment comments)

The root of short-drama character consistency:group(same person) + primary-secondary(primary = baseline makeup photo / secondary = each specific look). Input the full script → output standardized character visual asset JSON.

Workflow:

  1. Identify all characters who "need to actually appear on screen" (including those appearing in memories/photos/videos/dream visualizations; also mention those with no dialogue but who form the subject of the image)
  2. Merge all appellations for the same person (formal name/nickname/childhood name/former name/mistaken identity/occupation/kinship term) into one 组别
  3. First create 1 主 asset = baseline appearance (character reference anchor) for each group
  4. If the same person has visual differences (clothing/makeup and styling/hairstyle/age/injuries, blood, dirt, damage), split into multiple 从 assets, sharing the same group
  5. Count the number of appearances + appearing scenes for each 从 asset; 6. Complete the fields and output JSON

Hard rules:

  • Who to extract: people who actually appear on screen; do not mention: those only mentioned, only in background descriptions, or who do not actually appear
  • Split compound appellations: "parents/dad and mom/brother and sister" must be split into individuals; "relatives/passersby" remain generic; it is forbidden to directly use generic terms like "dad/mom" as a group
  • Primary-secondary split:主 = baseline appearance anchor;从 = under the primary'sface and body shape remain unchangedpremise, only change clothing/makeup and styling/state
    • Differences where the face changes → create another "primary"(major age stages/reincarnation/possession/before and after disfigurement), the same group may have multiple primaries
    • Differences where the face does not change → these are "secondary"(clothing/makeup and styling/hairstyle/state: injuries/blood/dirt/damage/soaked/disguise)
    • If the styling is consistent throughout, there is only 1 secondary; if unsure, do not split
  • Naming:组别(阶段/服装/状态), e.g. 林晚(校服)、金妮妈妈(年迈)、陈庆圆(染血衬衫)
  • Scene binding: if consecutive scenes do not mention a styling change = continue the appearance of the previous confirmed asset; if the original text clearly states a change of clothes/injury = create a new "secondary" starting from that scene
  • Field values: gender male/female/unknown; age toddler/youth/young adult/prime adult/middle-aged/elderly/unknown; dialogue yes/no; number of appearances (fill 0 for "primary"); character bio <100 characters; clothing about 20 characters; appearance features ≤50 characters (write only objective appearance, not emotions/temperament/psychology);It is forbidden to fabricate experiences/occupations/relationships/settings with no basis in the original text

Output fields (fixed):姓名 / 组别 / 从属状态 / 性别 / 年龄 / 人物简介 / 对话 / 职业 / 服饰 / 外形特征 / 人设 / 出现次数 / 出现场次

Image generation prompt rules(companion):

  • 主 → Text-to-image: pure white background, front/side/back full-body three-view + one large face close-up, plain clothes (basic neutral clothing without story styling), strictly according to appearance features
  • 从 → Image editing: use the same group's "primary" image as the base, lock the face/facial features/body shape/skin tone/hairstyle base, and only change clothing/makeup and styling/hairstyle/state (injuries/blood/dirt/damage/soaked)
  • White-background asset images do not add negative constraints (no subtitles/watermarks/BGM will appear); defects/scars are depicted truthfully and exposed (to maintain consistency), but stay within review boundaries (no revealing sensitive parts, no gore or shock)
  • Do not write emotions, temperament, or psychological states (these are not appearance)
  • Applicable models: Chinese image generation models (Banana/Nano Banana, GPT), supporting direct input of the primary image as a reference image

a. Three-view prompts for protagonists + important characters (small logic orthogonal standard)

Three-view standard:

  • Orthographic projection, parallel lines remain parallel
  • No perspective foreshortening, wide-angle distortion, near-large-far-small, or lens distortion
  • No top-down shots, low-angle shots, three-quarter views, or dynamic tilts
  • Front/Left/Right/Back four views (optionally Top/Bottom), same skeleton pose (default A-Pose: standing upright, head straight, gaze level, neutral expression, arms naturally apart, legs naturally apart, feet parallel, soles fully on the ground)
  • All views at the same projection scale, soles aligned to the same horizontal baseline, key joint points (eyes/shoulders/elbows/wrists/waist/hips/knees/ankles) horizontally aligned
  • No perspective ground, vanishing point, depth of field, or spatial haze; pure white background, uniform diffusion, no shadows, no vignetting, no glow
  • Left/Right are defined according to the character's own orientation, preserving real asymmetries (hair part/ear ornaments/protective gear/waist ornaments/scars); simple mirroring is prohibited
  • Props such as weapons are moved out of the character and displayed separately, not held in hand; when long hair/cloaks cause occlusion, add an "unoccluded structure auxiliary diagram"

Asset consistency lock:

  • Same character identity, sense of age, facial bone structure, facial feature proportions, head-to-body ratio
  • Same hairline, hairstyle structure, hair color, hair length, hair ornaments
  • Same skin tone, makeup, body proportions, physique
  • Same clothing cut, color, material, patterns, footwear, accessories
  • Same character temperament
  • Strictly prohibited: face swapping, age drift, physique drift, head-to-body ratio changes, clothing style changes, hairstyle changes, accessory loss, art style drift

Output format (three-view prompt):

【角色三视图】角色01|角色名
● 不可变DNA:面部骨相/瞳色/肤色/发际线/发色发质/身体比例/胎记疤痕(一次定稿,全片复用)
● 当前变体:[角色01-A] 服装/妆发/状态(变体先复述不可变DNA,再描述可变部分)
● 五官细节(V5.0.1 新增·必写):脸型(瓜子/方圆/棱角/鹅蛋)+ 眉形(剑眉/平直眉/柳叶眉/粗短眉)+ 眼型(丹凤眼/桃花眼/细长眼/圆眼)+ 瞳色瞳形(墨黑/琥珀/灰蓝 + 瞳仁大小)+ 鼻型(高挺/直鼻/微塌/鹰钩)+ 唇形(薄唇/厚唇/唇珠/嘴角微垂)+ 下颌线(清晰/圆润/棱角分明)+ 耳廓(贴耳/招风)+ 骨相(颧骨/眉骨/腮骨突出度)+ 面部特征(痣/疤/法令纹/卧蚕)——五官是角色辨识度根基,跨段逐字复用
● 服装材质细节(V5.0.1 新增·必写):面料(丝绸/棉麻/皮革/金属甲片/毛呢/纱)+ 纹理(刺绣/暗纹/织纹/鳞甲)+ 工艺(滚边/盘扣/铆钉/绳结/绲边)+ 层数结构(外袍→中衣→内衬逐层写)+ 颜色分区(主色/辅色/点缀色)+ 旧化状态(磨损/破口/污渍/湿透/血染/破损程度)——材质细节防"衣服像塑料/换面料"漂移
● 视图规格:正面/左侧面/右侧面/背面 正交四视图(按需Top/Bottom),A-Pose标准姿势,脚底基线对齐,无透视无阴影
● 风格渲染:渲染质感 + 材质 + 主辅色 + 画风(参考图唯一标准时可附参考图绑定)
● 负向提示词:卡通、Q版、低幼、模糊、畸形、水印、文字、杂物、塑料质感、透视变形、动态姿势、额外人物、剧情背景、干净数码锐度、CG感、海报构图、居中肖像

b. Character expression setting prompt (new in V2.1 · fills in the subtle depiction missing from Xiaoluo)

Xiaoluo's existing expression setting isimage generation instruction(twelve-grid expression board); Lao Li incorporates it into the connection layer as a standard asset step;subtle depiction of shot-level performanceis additionally supplied by §5 Lao Li's micro-expression library + Xiaoluo's six-dimensional physical performance.

Twelve-expression library (fixed order): Calm / Smile / Happy / Sad / Angry / Afraid / Surprised / Confused / Shy / Indifferent / Crying / Laughing

Expression intensity and differentiation lock:

  • Smile < Happy < Laughing (eyelid narrowing, cheek raising, mouth opening extent, orbicularis oculi involvement increase progressively)
  • Calm vs Indifferent: lower eyelids, more distant gaze, more restrained perioral muscles
  • Sad vs Crying: tears appear + stress intensity of eyelids/nostrils/perioral area
  • Surprised vs Afraid: brow furrowing/lower eyelid tension/mouth corners stretched laterally (Afraid) vs open tension/lower eyelids not tense/mouth shape round or oval (Surprised)
  • The twelve grids must not repeat, be near-duplicates, be mirror copies, or change only the mouth shape while ignoring eyebrow and eye muscles

Identity consistency lock: Each grid shows the same character (skull/cheekbones/jaw/interocular distance/brow peaks/nasal bridge/lip shape/skin tone/moles, spots, scars/hairline and hairstyle/clothing and accessories); expressions may only change facial performance and necessary slight head dynamics, and must not redesign the character. Strictly prohibited: face swapping, facial feature reshaping, age drift, lighting drift, art style drift

Output format (expression setting prompt):

【角色表情设定】角色01|角色名
● 表情板规格:4列×3行十二宫格,统一头肩像/胸像裁切,统一正面视角(或统一四分之三),统一眼线/肩线/留白,纯白或中性灰背景,同一光源方向/光比/色温/眼神光
● 十二表情(按序):平静(眼睑正常开合、眉间无挤压、唇自然闭合)/ 微笑(眼睑轻微收窄、嘴角小幅上扬、低于开心)/ 开心(眼睑明显收窄、面颊抬升、可轻微露齿)/ 悲伤(眉头内侧上提、眼睑沉重、嘴角下坠、克制的下唇紧张)/ 愤怒(双眉向内向下压、眉间纵褶、鼻翼轻张、咬肌下颌收紧、双唇紧抿)/ 害怕(双眉上提聚拢、上眼睑大幅抬起、下眼睑紧绷、嘴角横向拉开、下颌后收)/ 惊讶(双眉整体高抬不聚拢、双眼圆睁、下颌下落、嘴型圆或椭圆)/ 疑惑(单眉上挑单眉下压、单眼微眯、单侧嘴角收紧、可极轻微偏头)/ 羞涩(视线下垂或侧移、眼睑柔和收窄、克制小幅上扬、面颊自然微暖)/ 冷漠(上眼睑比平静更低、眉平直、双唇水平闭合、口周低活动)/ 哭泣(眉头内侧明显上提、眼眶积泪自然泪痕、鼻尖微红、下唇颤动)/ 大笑(眼睑明显收窄、眼角自然纹理、面颊高抬、张口露齿符合人体结构))
● 表情差异锁定:微笑<开心<大笑;平静≠冷漠;悲伤≠哭泣;惊讶≠害怕
● 禁止:手势/手部/全身姿势、剧情背景、道具、夸张腮红、性化、重复表情、错字、拼贴错位、英文标签、水印签名

c. Scene prompt (including focal length system · strengthened in V2.1)

Scene standard:

  • Spatial structure (walls/doors and windows/columns and beams/stairs/level differences)
  • Layout (positions of furniture/lighting/props)
  • Material/color/light and shadow/time/weather/style
  • Focal length system (Xiaoluo scene layout alignment): C1 Main visual 24mm or 35mm (space establishment) / C2 Side 35mm / C3 Detail 50mm / C4 Functional area 24mm or 35mm
  • For grand scenes, see §4.5 Special (extreme long shot/aerial focal lengths + layered depth + character proportion)

Output format (scene prompt):

【场景】场景01|场景名
● 空间结构:墙体/门窗/柱梁/楼梯/高差(写死,跨镜不得漂移)
● 布局陈设:家具/灯具/道具位置(写死)
● 材质色彩:主辅色 + 材质谱系 + 旧化程度
● 光影方案:主光方向/光源类型/色温/时间/天气
● 美术风格:建筑风格 + 时代背景 + 渲染方式
● 焦段方案(按需):空间建立 24-35mm广角 / 关系叙事 35mm / 材质细节 50mm / 宏大场景按§4.5
● 负向提示词:卡通、Q版、低幼、模糊、畸形、水印、文字、杂物、塑料质感、多余人物、CAD线稿、施工图、干净数码锐度、CG感、海报构图

d. Storyboard panel prompt (one per scene)

Rules (V2 confirmed + Xiao Luo nine-grid alignment):

  • One per scenestoryboard panel
  • Defaultnine-grid (3×3), for simple plots you may usefour-grid (2×2), depending on the plot
  • The nine panels belong to the same story world and occur continuously from left to right and top to bottom
  • Each panel is an independently valid film still
  • Characters/scenes/props/style remain stable throughout (identity lock/scene lock/style lock)
  • Camera language coverage: at least 1 extreme long shot or wide-angle space establishment, 1 full body, 2 medium shots or half-body shots, 1 side or three-quarter side view, 1 over-the-shoulder for multi-person scenes (single-person scenes must not add a second person just to fit the form), 1 close-up or facial close-up, 1 close-up of eyes/hands/key prop, at least one each of eye-level/low angle/high angle, at least 3 of front/side/back/three-quarter side
  • No three consecutive panels with the same shot size/same camera height/similar composition
  • Continuity: temporal continuity (actions do not return to the starting point without reason), action continuity (action matching, props do not return to their original position, keep the same hand), spatial continuity (the first panel establishes the spatial axis, no teleporting, doors/windows/furniture do not drift), eyeline continuity (shot-reverse-shot eyeline matching), screen axis (by default stay on the same side of the axis; if crossing the axis, first add a neutral shot or movement transition)
  • Performance continuity: expression intensity develops panel by panel, no unexplained jumps; close-ups and facial close-ups preserve original facial features; hand close-ups have the correct number of fingers
  • Cinematography optics: wide angle establishes space but avoids extreme stretching; medium shots compress naturally; close-ups and facial close-ups control depth of field to keep eyes and expressions clear; Cinematic lighting is implemented as explicit key light direction/fill light ratio/rim light source/environmental reflection; Realistic material reflects the specific reflection roughness and texture of wood, stone, metal, fabric, leather, and glass; Natural depth of field changes with shot size; Film color grading is unified across the nine panels

Default nine-panel narrative structure (adjustable):

  • First row: establish space and trigger event (panel 1 extreme long shot or wide angle establishes space + character position → panel 2 character enters or full-body positioning → panel 3 action trigger medium shot or side view)
  • Second row: develop relationships and advance action (panel 4 character narrative medium shot/half-body/over-the-shoulder → panel 5 core action intermediate state → panel 6 reaction or spatial change close-up/side/back)
  • Third row: details, outcome, and suspense (panel 7 close-up of eyes/hands/prop details → panel 8 action consequence or character decision close-up/medium close-up → panel 9 ending suspense: open-ended close-up/back view/long shot/key detail)

Output format (storyboard panel prompt):

【剧情分镜格】第X集·场次X|场景名
● 格位规划(3×3或2×2,按剧情):
  格1:[景别+机位高度+角度] 画面内容(动作中间态优先)
  格2:[景别+机位高度+角度] 画面内容
  ...(按格位顺序,共九格或四格)
● 连续性锁定:同一人物身份/服装/场景结构/道具位置/光源方向/时间天气/风格
● 镜头语言覆盖清单:大远景/全身/中景/侧面/过肩/近景特写/道具特写/平视低机位高机位(核对)
● 风格渲染:渲染质感 + 色彩体系 + 光影 + 画风
● 负向提示词:卡通、Q版、低幼、模糊、畸形、水印、文字、杂物、多余人物、画面割裂、镜头编号、红框箭头、参数说明、干净数码锐度、CG感、海报构图

Gate C: Asset prompt gate (including three major assessment criteria)

  • Three-view: orthographic projection? No perspective? A-Pose? Immutable DNA lock? Left and right not mirrored?
  • Expression settings: twelve-grid? Same character? Clear physical differences among four easily confused expressions?
  • Scene: complete spatial structure/material/lighting/style? Clear focal length plan?
  • Storyboard panels: one per scene? Continuous narrative? Complementary shot sizes? Camera language coverage meets standard? Characters/scenes/style locked?
  • Grand scene: shot size + focal length + character proportion (§4.5)?

2.4 Step 3: Per-episode storyboard prompt (generation-segment format)

Input: per-episode storyboard script + transition-layer asset prompts Output: complete prompt package for each episode (can be fed directly into the generation tool)

Core format (user-confirmed on 2026-08-04 + five-segment format on 08-06 + six-segment format on 08-09 (only for >60s long content) + V4.0 returns to primarily the five-segment format): It is not a tabular storyboard, but a narrative-style timeline that can be fed directly into the generation tool. One episode = several "generation segments" spliced together, each segment carrying a completefive-segment format:【风格】+【色卡参考】+【镜头】+【建议时间轴】+【声音】。Segment length is cut naturally according to plot/scene (ranging from 3s to 15s); 15s is the upper limit for a single segment, not a fixed value—once a scene is finished, it ends; it can end at 7s/10s/12s, and each segment does not have to be padded to a full 15s。 Content ≤60s (including single-segment/short content): use the five-segment format(V3 has proven effective in practice; default for action/dialogue scenes);Only >60s long content (such as a long segment for an entire episode) is upgraded to the six-segment format(five-segment format + the opening [Characters and Scene] module, used to anchor character/scene assets over long durations without drift). Short content within 10s (when the user requests a single segment): do not split into segments; directly output one complete five-segment generation segment, with the timeline narrated continuously from 0-Xs to the end; only long content (≥15s/full episode) is split into multiple segments. Shares with the storyboard scriptthe same continuous timeline: starts at 0s, the final shot ends exactly at the episode length, with no skipped seconds and no overlaps.

Five-segment structure (none may be missing for each segment ≤60s):

## 【段N · 段名】X-Xs

【风格】渲染质感 + 色调 + 对比度 + 氛围 + 高亮区/色彩体系声明
【色卡参考】本段主色相 + 辅助色 + 高亮色 + 暗部色,每色附 HEX + RGB 参考值(锁死色彩,防跨段色彩漂移;色卡参考图垫首帧时提示词只写动作不写光影,画面自动咬住色卡色彩)
【镜头】焦段 + 画幅 + 景别 + 运镜方式 + 晃动/呼吸感 + 镜头跟随逻辑

【建议时间轴】(秒数为节奏建议,AI 可微调,非硬命令)

约0-Xs:连续叙述式画面描述。动作+物理反馈+括号音效嵌在描述里(如"削断三根竹子,竹干从中断裂的声音")。
约X-Xs:下个时间点。
...(每个时间点约 1.5-3s 一段,连续叙事)

【声音】该段全部声音按序列出

Embedded camera-movement writing (new in V4.0.6 · AI cinematic camera movement, verified by user examples):

Each time point can embed an independent camera movement, format:[时间码] 运镜:... 画面:.... Camera movement evolves along the timeline, strung into an emotional arc (establish → introduce → focus → explode → conclude), no longer one global camera movement. Key writing points:

  • Write the complete trajectory + shot-size changes + camera-position switches for camera movement: "AI cinematic aerial dive push-in shot, large depth of field extreme wide shot slowly descending, the camera passes through layers of clouds and mist from dome height and drops vertically, switching to a low-angle upward shot at medium shot", do not just write the two words "push-in shot"
  • Each camera movement has a narrative function: descent = establish environment / orbit = introduce characters / push in = focus on details / pull back and slow motion = explode and release / static gaze = conclude
  • Motion texture words can include: "silky camera movement without shake" "dynamic tension camera movement" "slow-motion high-speed shot"
  • Visual staging hierarchy large-medium-small-large: full environment → full body of character → facial close-up → explosive wide shot, forming a scale contrast for the ending ("she is very small, the giant door is very large")
  • Information planting: objective visuals carry narrative hooks ("what is crushed underfoot is dried black bloodstains (not hers)" "she is not looking at the camera—she is looking above and behind the camera")
  • Overall parameters can be placed up front:【整体参数】时长,画幅,风格,负向提示词 at the beginning of a block (for direct single-segment feeding); at this time the remaining modules of the five-segment format are still retained in order

Six-segment format (only for >60s long content): insert after [Style] and before [Color Card Reference] in the five-segment format 【出场人物与场景】本段出场人物 + 场景位置(生成时在此处 @ 人物/场景资产图,是固定的资产引用位,每段必写), the other five segments remain unchanged.

Color card reference writing rules (user-confirmed on 2026-08-06 · must be written for every segment):

  • Located after [Style] and before [Shot], it is an independent fixed module and cannot be omitted
  • Write the main hue + secondary color + highlight color + shadow color for this segment, with each color accompanied by HEX + RGB reference values(e.g., 千禧粉 #F1BB7B 241,187,123, share the same color card across segments to prevent drift; when the §4.2.1 color card is placed on the first frame, the prompt only writes action and not lighting, and the image automatically locks onto the color card's colors)
  • Share the same color card across segments: the main hue of adjacent segments stays unchanged, only the highlights/shadows may be fine-tuned according to the emotion curve
  • How to use: place the color card reference image on the first frame (§4.2.1 three steps) → the prompt only writes the subject's action and not lighting → the image automatically reuses the color card's colors, achieving unified color grading

Recommended timeline writing rules (confirmed 2026-08-04 + reinforced 08-05 + camera movement embedded 08-12 + made advisory 08-15):

  • Positioning: the timeline is a "suggested rhythm", not a hard command. The seconds are a rhythm reference for the AI—during generation, the AI may naturally fine-tune based on the visual performance, action completeness, and dialogue speed; a deviation of 1-3s forward or backward is normal, and it does not need to be precise to the second. Our estimated seconds are occasionally inaccurate,leave room for the AI to perform, don't lock it down. The only two hard bottom lines that cannot be compromised: ①dialogue must be fully spoken(no cutting off half a sentence/mouth-shape misalignment); ②the action cause-and-effect chain must be continuous(no character teleporting/jumping/position changes without cause)
  • Each time point is a continuous passage of narration, duration about 1.5-3s (Old Li reference: 1-1.5s flash-cut impact / 1.5-2s hook climax quick cut / 2-2.5s regular action dialogue / 2.5-3s narrative establishment / 3-4s emotional settling reaction / 4-6s slow-motion freeze ending), with seamless connection between passages (end of previous passage = start of next passage);Each time point may embed independent camera movement [时间码] 运镜:... 画面:...(V4.0.6), camera movement evolves along the emotion arc, see the dedicated section below the five-part structure
  • Open by exploding directly: do not explain the backstory, start from the middle of the conflict ("the image explodes directly—already fighting")
  • Action + physical feedback + parenthetical sound effectsembedded in the description ("cuts three bamboo stalks, the sound of bamboo trunks breaking from the middle")
  • Zero psychological description: only write what the eyes can see and the ears can hear
  • One sentence one action, active subject-verb-object sentence pattern, specific to physical details (firelight/water splash/sparks/breakage)
  • Time point segmentation: use 1.5-2s quick cuts for hooks/climaxes, 2.5-3s for emotional settling
  • Dialoguewritten as 角色名:"台词", dialogue time estimated by "acting" it once at natural speaking speed + reaction time
  • Hard rules for dialogue scene duration (2026-08-10 user feedback):do not cut the shot before the dialogue is finished. Within a time point, the dialogue must be able to be fully spoken at natural speaking speed (≈4 characters/s, impatient 5-6 characters/s, calm/drunk/recalling ≈3 characters/s); single-shot dialogue of 8-12 characters → extend that time point to 2.5-4s; 15 characters or more → split into 2-3 continuous time points (speaking → pause/listener reaction/prop close-up → continue speaking), dialogue across time points must be continuous, and it is forbidden to break sentences/misalign mouth shapes just to preserve a 2s quick-cut rhythm. The overall rhythm of a dialogue segment is determined by the amount of dialogue, not by 2s shot cuts
  • Grand scenesmust clearly write within the time point: shot size + focal length + proportional difference between characters and environment + camera movement (§4.5)
  • Iron rule for position anchoring (2026-08-11 user feedback: AI-generated video positions are chaotic): at each time pointthe first sentenceCharacter spatial relationships must be hard-coded—[角色A]在画面[左/右/中/前景/后景],[距谁多远/面向哪],[角色B]在[相对位置]; for multi-person shots, add foreground/background or three-by-three grid coordinates; key props must also be position-locked (sword/bell/door/table). Use specific numbers for distance (two steps/two meters/three steps), and use left/right/center/foreground/background for direction.Not writing blocking clearly = characters drift randomly during generation. Changes in character position across time points must specify movement trajectories ("walks toward the cliff edge on the left side of the frame, passing by the master's side"), not just the action
  • Gaze Anchoring Iron Rule (2026-08-11 user feedback: dialogue scenes had no gaze constraints): Every shot in dialogue/confrontation scenes must specifywho looks at whom, where they look, and when they look away——[角色A]视线[低垂/直视/追随/望向X]; both the listener and the speaker must be written ("Lin Che keeps his eyes lowered and does not look up" "his gaze meets the master's for the first time"). For receiving instruction/submission/guilt, use averted gaze; for taking a stance/confrontation/realization, use direct eye contact; for parting/nostalgia, use following the retreating figure.No gaze instructions = AI eyes wander randomly and dialogue has no tension. Must be checked together with blocking anchoring as a gatekeeping item.Expanded rules (2026-08-11 incorporating Xiaoluo's three-by-three grid): ① When a character looks at an off-screen target, the target shown in subsequent shots must be in a reasonable opposite position; ② shot-reverse-shot must maintain matching gaze height and direction (speaker's gaze height = listener's gaze height); ③ eye close-ups must be consistent with the observed object in the preceding and following shots; ④ expression intensity develops shot by shot with the plot, and adjacent shots must not jump without reason (show changes through eyes/eyebrows/ mouth corners/breathing/fingers/shoulders and neck/body center of gravity)
  • Key physical sounds are marked in parentheses within the description and summarized in the 【Sound】 section

Complete segmented generation example (user-confirmed template · five-segment format):

【风格】胶片动作片质感,冷墨绿色调,高对比度,湿竹叶反光,高速运动时有轻微运动模糊
【色卡参考】主色冷墨绿(竹林深处 #2F4F4F 47,79,79),辅助色湿竹叶青(雨湿竹叶 #2C9678 44,150,120),高亮色雨珠白(雨珠反光 #F3F9F1 243,249,241),暗部色竹影深青(竹丛阴影 #1C2C5B 28,44,91)
【镜头】24mm广角,16:9宽幅,全程手持跟拍近景,轻微呼吸晃动,镜头跟随每一次出刀和闪避的节奏摇摆

【建议时间轴】

0-1.5s:画面直接炸开——已经在打。白衣武者从上往下重劈,剑锋垂直斩向黑衣刀客头顶。黑衣刀客双手举刀格挡,双刃相击火花溅开。湿竹叶从两人脚下被气浪推飞。

1.5-3s:黑衣刀客格挡住这一击后立刻反击,压低身体横斩白衣武者腰部。白衣武者没有后退——直接腾空跃起,剑砍空了从脚下掠过(削断三根竹子,竹干从中断裂的声音)。白衣武者在空中转体90度,落地时面向黑衣刀客侧背。

3-4.5s:落地瞬间白衣武者剑已先到——直刺黑衣刀客右侧肩。黑衣刀客迅速回身以刀背斜挡,剑尖在刀面上刮出一串火星(尖锐金属摩擦声)。白衣武者的剑被挡开,但剑身弹回时顺势沿着刀面滑向刀柄方向,将黑衣刀客的刀挑脱手。

4.5-5.5s:短刀旋转着飞出去,一片竹叶被劈成两半。刀还没落地,黑衣刀客已经空手抓向白衣武者握剑的手腕——他要夺剑。白衣武者后退两步,手腕翻转躲过,剑气带起的风声里他的白袖翻飞。

5.5-6.5s:白衣武者借后退之势调整重心,双手握剑柄从右侧向左侧全力横扫。角度极低——从膝盖高度扫去——逼黑衣刀客不得不起跳躲避。黑衣刀客在起跳的瞬间左手抓住头顶一根横竹,把自己吊上竹干。

6.5-7.5s:白衣武者不让他停——挥剑砍向那根横竹。竹子应声断开,黑衣刀客连同断竹一起从空中坠落。白衣武者调整剑尖方向,剑锋停在坠落的黑衣刀客胸口前三厘米处。黑衣刀客伸手抓住剑身侧面——手指划破流血——但稳住了自己没有再下落。

7.5-9s:两人同时着地,距离只有一剑之长。都在喘气,胸口剧烈起伏。雨水混着汗水从两人的额头流下。黑衣刀客低头看了一眼流血的手指,然后抬头看着白衣武者——嘴角缓缓向上扬了一点。不是嘲讽,是某种确认。

9-10s:白衣武者收剑——干净利落地从左下往右上挑入剑鞘。剑入鞘的瞬间,一声脆响定住整个画面。两人的呼吸声在安静的竹林中持续。一根断竹从高处坠落,轰的一声砸在两人侧面的地面。

【声音】金属刃击高频撞击、竹干断裂的脆裂声、剑尖刮金属的火星尖啸、湿竹叶被踩碎的低沉声响、断竹从高空坠落地面的轰响、两人粗重的喘息

Combination consistency (must be checked for each segment):

  • Use hard cut → markers between segments (or a directly continuous timeline)
  • Cross-segment generation continuity: adjacent segments share the same keyframe/asset ID reference to ensure prop and particle continuity
  • Iron rule for independent segmented generation (2026-08-11 user confirmed): When generating separately, each segment's prompt must befully self-contained—【Style】【Color card reference】【Shot】must be written out completely for each segment,it is forbidden to write any inherited expression such as "continues from the previous segment/inherits/carries over from the previous shot/same as above"(each segment is fed separately, and the model has no cross-segment memory); unity across segments in the same scene relies onthe four color card colors being completely identical+ style descriptions being verbatim identical, not on textual references
  • Sixfold consistency of asset lock/timeline lock/spatial lock/lighting lock/style lock/final-state inheritance lock (see section three)
  • The ending returns to a spatial scale contrast to form a visual closed loop (e.g., opening explosion firelight → ending skylight outside the window)

Time allocation principles:

  • Each time point 1.5-3s; hooks/climaxes 1.5-2s quick cuts; emotional settling 2.5-3s
  • Each segment's duration is determined by the amount of plot content (dialogue segments are estimated by line volume at natural speech rate ≈4 characters/s, action segments by action density; reading one layer of information ≈2-3s, and each additional layer +2-3s)
  • Hard rules for dialogue scene duration (2026-08-10): Do not cut shots before the dialogue line is finished. Single shot 8-12 characters → 2.5-4s at that time point; 15+ characters → split into 2-3 consecutive time points (speaking → listener reaction/prop close-up → continues speaking), with the dialogue spoken continuously; do not break sentences/mismatch lip sync just to keep a 2s quick cut. The pacing of dialogue segments is determined by the amount of dialogue; 2s cuts serve only as a reference for action segments without dialogue
  • One segment has only one camera movement, no compositing
  • Visual content is bound to camera movement but they do not compete with each other
  • Physical details anchor realism (specific external physical events)

Gate D: Prompt gate (generation segment style, including three major assessment criteria)

  • Does each segment have a complete five-part structure (style/color card reference/camera/timeline/sound)? (Only >60s long content uses six parts: + [Characters and Scene])

  • For segments >60s: characters and scene are clearly written (@asset slots)? For segments ≤60s: no such module (return to the V3-tested five-part structure)?

  • Is the color card reference located after [Style] and before [Camera]? Are all four colors present—primary hue + secondary color + highlight color + shadow color—with HEX+RGB attached?

  • Is the timeline continuously narrated? Are segments seamlessly connected? Is there a total duration reference target (suggestive, flexible)?

  • Embedded camera movement (V4.0.6): does the camera movement at each time point evolve along the emotional arc (establish → introduce → focus → explode → converge)? Does the camera movement have a complete trajectory + shot size changes? Does the visual staging have a large-medium-small-large hierarchy or a scale-contrast ending?

  • Does the opening explode directly? No explaining of origins?

  • Action + physical feedback + parenthetical sound effects embedded in the description?

  • Zero psychological description? Only write what is visible and audible? No abstract emotion words?

  • Can the dialogue be finished at a natural speaking rate? Character name: "dialogue" format?

  • Dialogue scene duration: no shot cuts before the dialogue line is finished? Single shot 8-12 characters stretched to 2.5-4s? 15+ characters split into consecutive time points with continuous dialogue?

  • Grand scenes: shot size + focal length + character proportion (§4.5)?

  • Staging anchor: does the first sentence of each shot clearly state the characters' spatial relationships (position/distance/facing/foreground-background)? Are position changes across time points written with movement trajectories?(added based on user feedback on 2026-08-11)

  • Gaze anchor: in dialogue/two-hander scenes, does each shot clearly state who looks at whom, where they look, and when they look away? Is the listener's gaze constrained?(added based on user feedback on 2026-08-11)

  • Independent generation per segment: is each segment's [Style], [Color Card Reference], and [Camera] fully self-contained? Are there no inherited expressions like "continue/use the same/previous segment"? Are color cards for the same scene identical word-for-word across segments?(added with user confirmation on 2026-08-11)

  • Fight segments (§4.6.8): is the action power chain complete (intent → wind-up → center of gravity → force generation → contact → feedback → recovery → carry into next panel)? Are attacks and defenses paired (every attack must have a defense/dodge/block/counter)? No teleporting/no reasonless side switching/clipping/spatial reset? Do P15-type endings avoid static pose-holding?(added when the martial arts system was incorporated on 2026-08-16)

  • Is each segment mainly in pure Chinese? Are there 2-4 English tags per shot and are they compatible?

  • Are the six-fold consistency locks (assets/timeline/space/lighting/style/final-state inheritance) satisfied?

  • Are hard cut markers or timeline continuity used between segments?

  • Is there a visual closed loop at the ending (imagery callback)?

  • Does the Chinese output have "No subtitles, no background music" on the last line?


III. Storyboard Combination Consistency (V2 core upgrade · retained)

Goal: Multiple video segments are stitched into one episode, with seamless transitions between segments and high overall storyboard requirements.

3.1 Dual-track timeline model

  • Segment-level timeline: Each segment is an independent shot group;segment length is cut naturally according to plot/scene, ranging from 3s to 15s; 15s is the upper limit, not a fixed value(a scene ends when it is finished; it can end at 7s/10s/12s, without forcing it to 15s)
  • Episode-level timeline: Multiple segments connect end to start = continuous total duration (the end of the previous shot = the start of the next shot, with no skipped seconds/overlap/blank space)
  • The two tracks coexist: use segment-level when generating in segments, and use episode-level when stitching the final film

3.2 Sixfold consistency lock

Lock Content Check item
Asset lock Character/scene/prop IDs must not be reset across segments Costume/bone structure/prop state continuity
Timeline lock Time continuity between segments End of previous segment = start of next segment, with no skipped seconds
Space lock Continuity of spatial direction/position Eyeline axis/blocking direction/door and window positions do not drift
Lighting lock Consistency of light source direction/color temperature/time Lighting and shadow in the same scene do not change abruptly
Style lock Consistency of rendering/color/texture VISUAL DNA declared for each segment
End-state inheritance lock(newly added in V4.0.2) Variable data is synchronously updated for each segment The later segment inherits the character blocking/pose/interaction/composition at the end of the previous segment, locked down with a storyboard reference image

Detailed explanation of the end-state inheritance lock (the core of AI video multi-segment stitching): AI video 2.0 has a single-segment limit of 15s, and each segment has no memory of the previous segment when generated (character blocking/pose/interaction reset to zero every 15s).It is forbidden to directly stitch by using the last frame of the previous segment as the first frame of the next segment(there is no dynamic inheritance, and the jump cut is the most severe). The correct approach: before generating each segment, draw a storyboard reference image according to the character state at the end of the previous segment, fix the pose/position/interaction relationships, and synchronously update the changing data in the prompt.

3.3 Fourfold rules for transitions between segments

  1. Asset transition: Adjacent segments share the same keyframe/asset ID, and props and particles are continuous
  2. Action transition: The action at the end of the segment and the action at the start of the segment are coherent (action-match transitions preferred)
  3. Emotional transition: Emotion at the end of the segment → emotion at the start of the segment progresses or reverses (logically)
  4. Sound transition: SFX continuity across segments (ambient sound is not interrupted)

3.4 Screen axis rules (absorbing Xiaoluo nine-grid on 2026-08-11)

  • Establishing the main axis: The main axis is formed by person-to-person, person-to-prop, or the direction of character movement; by default, keep the camera on the same side of the axis (180-degree spatial relationship)
  • Shot/reverse shot: Maintain matching eyeline height and direction; do not mirror character costumes, scene structure, or left-right accessories because of the reverse shot
  • Crossing the axis: When the axis must be crossed, first add a neutral frontal shot, an on-axis shot, or a clear movement transition; do not cross hard directly
  • Reverse position: When a character looks at an off-screen target, the target shown in the subsequent shot must be in a reasonable reverse position
  • Spatial stability: When using reverse angles across the axis, keep the left-right spatial relationship stable; do not mirror the room
  • Character movement: When a character moves from one side of the frame, subsequent shots must maintain a consistent screen direction; do not teleport to the other side of the room without a transitional shot; doors, windows, furniture, lighting fixtures, and key props must not drift due to a change in camera position
  • For the complete nine-grid storyboard specification, see references/小逻-视觉分镜九宫格生成.md

3.5 Four seamless transitions

  1. Wipe transition: Foreground full-screen occlusion → cut. Occluding objects: trees/walls/passersby
  2. Whip pan transition: High-speed whip pan motion blur cut point
  3. Action match transition: Cut at the same action apex before and after
  4. Similar composition transition: Before and after composition/shape highly matched

3.6 Combination verification

  • Total duration after multi-segment assembly = target episode length ± 0.5s
  • No black screen between segments/no transition break
  • Check each episode: whether the hook is at the beginning and the suspense is at the end
  • Assessment standard ① review: smooth connection (fourfold connection + sixfold lock) + reasonable duration (determined by content volume)

3.7 AI video multi-segment assembly methodology (new in V4.0.2 · practical verification)

AI video 2.0 has a single-segment limit of 15s. When assembling multiple segments into a full episode, the most common issues are: visual jump cuts, broken character action continuity, and dialogue with no breathing room. Root cause: Each segment is generated without memory of the previous segment; a new clip does not inherit dynamic information such as the previous segment's character blocking, posture, and interaction, and the data resets to zero every 15s. Incorrect approach: Using the last frame of the previous segment as the first frame of the next segment for assembly—there is no dynamic inheritance at all, and the assembly traces are the most obvious.

Two major data categories (core understanding):

Category Content Handling method
Invariant technical data Character asset images, scene images Fixed import throughout, invariant
Variable changing data Character blocking, posture, spatial interaction between characters, image composition Must be synchronously updated for each segment——the key to smooth assembly

The two approaches must be used together; neither can be omitted:

Approach A · Technical means (adapted to AI generation characteristics):

  1. Improve the variable data (top priority): Variable data = prompts + reference images/storyboard images + audio/video materials. Use text + reference images together (relying only on prompts easily leads to repeated gacha pulls). Before generating each segment, draw a storyboard reference image based on the character state at the end of the previous segment, giving the AI clear character posture, position, and interaction relationships, and synchronously update the prompts.
  2. Use "cut-in (camera position change)" at the assembly point: Do not force the same continuous long take; change the camera position at the assembly node and split it into two different shots, avoiding the jump-cut stutter of directly connecting the same camera position.

Approach B · Directorial thinking (optimize the viewing experience from the audience's perspective): Avoid stiff assembly from the viewing experience level. Only by cooperating with technical means can it be thoroughly solved—a single method alone cannot handle all plot scenarios.

Practical workflow:

找拼接问题 → 分析根源(AI 无长时序记忆 + 可变数据未同步)
→ 双方案处理 → 核心操作:① 绘制分镜参考图  ② 拼接节点换机位

Connection with existing rules:

  • When generating across segments: first determine the character state at the previous segment's tail frame → draw the storyboard reference image (fixed pose/position/interaction) → update the changed data in the prompt accordingly → change the camera position at the splice node
  • Hard cut shots are recommended to be generated separately and then edited together
  • Three techniques and six forms for inter-segment transitions (V5.0.1 enhanced · must be used at every segment splice node, splicing first and last frames is prohibited):

Core:splice node = camera position change node. Reserve transition nodes at the end of each segment and the start of each segment—the last shot of the segment tail should not come to a complete closure (leave the action in an intermediate state), the first shot of the segment head should begin from the action continuation point, and the splice node must satisfy one of the three options.

Three techniques (choose at least one; camera position change is the basic mandatory choice):

  1. Camera position change: change angle, camera height, or axis orientation for the same action/state (front → side / eye level → high or low angle / far → near)—avoids the jump-cut stutter of directly connecting from the same camera position, highest priority
  2. Shot size cut: cross-level jump cut (wide shot ↔ close-up / medium shot ↔ extreme close-up); the larger the shot size span, the weaker the splice trace
  3. Cross-segment reaction: segment tail A's action → segment head B's reaction shot (action-reaction paired across segments)

Six forms (specific transition techniques):

Form Technique Usage Applicable to
① Action continuity + camera position change Segment tail starts the action, segment head completes the same action from a different camera position (match on action across camera positions) Between battle/action segments · best
② Reaction shot transition Segment tail A's action → segment head B's visible physiological stress reaction After pressure/information reveal
③ Shot size jump cut Segment tail wide shot → segment head close-up (or vice versa), emotional rupture/emphasis moment Emotional turning point/transition
④ Sound advance Segment tail sound continues into segment head image (rain sound/sword ring/dialogue breath across segments) Atmosphere continuity segments
⑤ Prop/detail transition Segment tail lands on a prop close-up, segment head shows the same prop from a new angle Key prop running through
⑥ Empty shot transition Segment tail action ends, segment head environmental empty shot (rain curtain/candle flame/dust) then connects to the new segment Rhythm breathing point

Splice node design rules:

  • Segment tail last shot: leave the action in an intermediate state/unfinished closure (sword swing halfway/turn halfway/dialogue not finished), do not land on a static pose
  • Segment head first shot: cut in from the action continuation point + camera position change (mandatory) + optionally overlay shot size jump cut/cross-segment reaction/sound advance
  • Prohibited: direct continuation from the same camera position / using the tail frame as the first frame / connecting a static image to a static image / cutting after the segment tail has fully closed
  • Every splice node must be marked at the storyboard script stage (see §2.2 storyboard table "inter-segment transition" column)

For the complete methodology see references/AI视频拼接问题总结-多段拼接方法论.md


IV. Director's Audiovisual Control

4.1 Camera Movement Dictionary (Lao Li's supplement · camera movement emotion)

Camera movement Keyword Emotion
Dolly In slow dolly in Approaching scrutiny, focus
Dolly Out slow dolly out Alienation reveal, detachment
Tracking side tracking Roving observation, following
Crane crane up/down God's-eye view, revealing scale
Orbit counter-clockwise orbit Confrontational tension, involvement
Whip Pan fast whip pan Chaotic energy, sense of speed
Handheld subtle handheld Immediacy, unease
Locked-off tripod shot, locked-off Stability, ritual feel, gaze

Xiaoluo's graded camera movement as the foundation (stable close-ups/slight handheld for medium close-ups/breathing feel for wide shots/Steadicam long takes), with the above table supplementing emotional semantics.

Three real handheld moves (new in V2.2 · Hedgehog Planet · eliminating staged feel):

  1. Movement path not overly smooth: brief deceleration while pushing in, slight pause before approaching the subject (traces of human operation)
  2. Subject and camera not synchronized: add tiny lag to camera movement, simulating real-shot reaction delay
  3. Low angle with strong perspective: lower the camera for an upward shot, amplify foreground scale contrast, wide-angle perspective distortion makes lines converge toward the vanishing point (epic feel/oppressive feel/e-commerce tension)

4.1.1 Shot Selection Decision Table (new in V2.2 · Source: Cinematic World Generator)

Don't memorize a camera movement checklist; directly choose the camera position based on "what emotion do I want to express." Shot size/height/occlusion are all narrative decisions, not random switches.

Narrative intent Camera position choice Image composition
Giant objects/awe/sacred scale Low angle, ground-level camera, foreground scale, tiny figures Giant buildings/colossal beasts filling the upper part of the frame
Isolation/ritual/social order High angle, overhead shot, negative space Figures like chess pieces within the environment
Intimacy/labor/emotion/tension Close-up, partial (hand/half face), shallow depth of field Only part of the body shown, focusing on details
Secrets/peeping/hierarchy/witnessing Over-the-shoulder, door-frame composition, frame within a frame, occlusion Viewing the subject through a foreground frame
Travel/pressure/mystery/memory Depth, silhouette, reflection, smoke/dust/rain, firelight Partially visible, environment participates in the narrative

Usage: before writing a shot, first ask "what do I want the audience to feel in this shot?", consult the table to determine the camera position, then pair it with the §4.1 camera movement dictionary to choose the movement method.

4.1.2 Six-Dimensional Control of One-Shot Long Takes (absorbed 2026-08-10 · running/chase/follow narrative)

For running chases and follow-style one-shot long takes. Core: replace all abstract atmosphere words with executable physical processes.

  1. Shot narrative goal: define the camera movement purpose before writing (example: a 15s one-shot long take laterally tracking the running female lead, creating an oppressive tension of a tense chase; the camera briefly loses focus and shifts to the pursuer, then quickly snaps back to the female lead, building spatial layers)
  2. Three-layer spatial layering (mandatory): foreground layer / subject layer (running protagonist) / midground layer (street tables and chairs, passersby, pursuers, carrying the conflict) / background layer (street vanishing point, exit at the end). Layering prevents AI spatial confusion
  3. Physical inertia + imperfect handheld: Realism comes from imperfect camera shake and inertial lag—handheld cameras have a natural breathing tremor, and after an impact the camera settles
  4. Focus narrative design (simulating the photographer's on-the-spot mistakes): Brief defocus enhances the documentary feel—"the focus briefly shifts to the pursuer's face, drifts past the crowd to the female lead in the distance, and after 0.3 seconds quickly refocuses and pulls back to the subject," replicating the on-the-spot operation of a real person re-locking onto the target
  5. Layered sound space: Sound changes in sync with the camera's shot scale to build a three-dimensional sound field—when the camera turns to the pursuer, the female lead's panting recedes to the far end of the sound field, while the pursuer's footsteps/heavy breathing occupy the center of the sound field; when the camera cuts back to the female lead, her panting once again becomes the main audio subject
  6. Negative constraints must be added: No mid-shot cuts, no instant camera teleport zooms, no character limb sliding/body merging, no objects losing their sense of weight

4.2 Complete Table of Eight Cinematic Lighting Setups (Supplement · Director's Thinking Course + Feynman Learning Method)

V2.1's original three (split lighting/side backlight/top light) → V2.2 expanded to a complete eight standard lighting setups.

No. Lighting method Light source position Visual effect Keywords Applicable scenes
1 Butterfly light (beauty light) Front-top Beautiful, slimming, three-dimensional face butterfly lighting, paramount lighting Character beautification, advertising, female close-ups
2 Top light (skull light) Directly overhead Eerie, oppressive, deep shadows in the eye sockets hard top light, overhead, dramatic shadow Interrogation, thriller, oppressive, horror
3 Rembrandt lighting Front-side Classic portrait, three-dimensional, triangle of light Rembrandt lighting, triangle light Character portraits, authority figures, classic narrative
4 Split lighting (division light) Direct side Half light half dark, dramatic tension split lighting, dramatic, hard light from side Suspense, good-versus-evil opposition, duality
5 Rim light (edge light) Diagonal rear Outlines contours, separates from background rim lighting, backlight, golden hour, edge light Hero entrance, silhouette, atmospheric feel
6 Backlight Direct rear Silhouette, mysterious backlight, silhouette Silhouette, suspense, epic feel
7 Bottom light Below Horror, eerie bottom light, under lighting, horror Horror, villain, nightmare
8 Front light Direct front Bright, flat, few shadows front light, flat lighting News, beauty, product display

Note: ① Complete aperture range f/1.0-f/64, thirteen stops in total (larger aperture → shallower depth of field → more blurred background); ② longer focal length → shallower depth of field (16mm deep depth of field → 200mm shallow depth of field); ③ light intensity and light source distance follow an inverse-square relationship

Supporting subsections: §4.2.1 Color system control (three-step HEX color card method) → §4.2.2 Master color grading style reference library (color system quick lookup + safe conversion table for director names) → §4.2.3 Verified complete prompt template library (cinematic golden formula + 9 director templates). The lighting table governs a single shot's light source, while the color card system governs color consistency across shots; the two are used in combination.

4.2.1 Color System Control (New in V2.2 · Hedgehog Planet HEX Color Grading Method)

Source: Hedgehog Planet super-i. Solves random colors in AI-generated visuals and cross-shot color drift (a color-layer supplement for the XiaoLuo timeline consistency).

Underlying logic: move color control ahead of generation, using color values instead of vague words ("I want blue" → lock down the HEX color gamut).

Three-step workflow:

  1. Get a dedicated color card: upload 10 reference images in the same style → scan and extract the core HEX color values → generate a dedicated color card image
  2. Lock colors when generating images: feed the color card as a reference image to the image generation tool, write only the subject in the prompt → the image automatically reuses the color system
  3. Continue colors into video: use the high-quality static image generated from the color card as thefirst-frame reference imageand place it into the video tool → write only the action in the prompt, not the lighting → the video holds onto the first frame's colors, achieving cinema-grade unified color grading

Usage suggestions:

  • When a comic drama is greenlit, first make a dedicated color card (10 same-style references → color card), unifying the color system for the whole series
  • Each scene variant (day/night/battle) can derive a sub-color card, but the main hue stays unchanged
  • Cross-episode color consistency: the same character/scene continues to use the same color card

4.2.2 Master Color Grading Style Reference Library (New in V2.2 · V3.0 HEX Enhanced · Color Card Quick Reference)

Paired with §4.2.1: when determining the style tone, directly consult the table to choose a master color system → use that color system's reference images/keywords to generate a color card.Director names are not written into the final prompt(a content safety red line), only the color scheme is written.

Master/School Core color system Keywords (can be included in prompts) Suitable subjects
Nolan Cool gray-blue + single point of warm color (lake blue #3B9AB2, dark cyan-gray #243940) desaturated cool blue, teal & orange, gritty realism Psychological suspense, sci-fi, realistic subjects
Villeneuve Desert gold + cyan-gray (deep sand brown #4A3728, earth yellow #D6A01D) golden hour, sand tones, muted teal, vast negative space Epic, sci-fi, road movies
Wes Anderson Candy colors (millennial pink #F1BB7B, coral red #FD6467, cream yellow #F3DF6C, mustard gold #CEAB07) pastel color palette, symmetrical, candy colors Comedy, absurdist, healing genres
Wong Kar-wai Neon cyan-green + dark red (qipao red #7C1010, dark blood red #8B0000, ink green #2F4F4F, dark golden brown #B8860B) neon teal, moody red, film noir, long exposure Urban romance, nostalgia, mood pieces
Tim Burton Gothic dark purple + cyan-black gothic purple, dark teal, eerie, high contrast Fantasy, dark fairy tales, thriller
Kubrick Cool white + vivid red stark white, bold red, symmetrical, sterile Sci-fi, thriller, cold and severe subjects
Roger Deakins Natural light + single point of highlight natural light, single strong light source, deep shadows Realistic, western, epic
Wes Craven horror style Cyan-green dark light + bloody red sickly green, desaturated, blood red accent Horror, thriller
Hou Hsiao-hsien Low-saturation earth yellow + cyan-gray muted earth tones, dusty, naturalistic, long take Artistic, period, realistic
Zhang Yimou High-saturation red + gold (vermilion #FF4C00, Forbidden City vermilion #C83C23) vivid red, imperial gold, bold color blocking Costume drama, epic, ethnic
Makoto Shinkai Translucent blue + highlight white + orange vibrant blue sky, lens flare, glowing highlights, clean anime Youth, healing, animation
Cyberpunk Neon pink-purple + electric blue neon pink, electric blue, rain reflections, haze Sci-fi, post-apocalyptic, urban night
Chinese horror Dark cyan + cinnabar red + gray-white (azurite #1685A9 cinnabar #FF461F earth yellow #D6A01D) dark teal, cinnabar red, pale grey, lantern light Folk horror, Republic-era mystery
Steampunk Brass brown + teal steam brass brown, copper, steam green, warm gears Alternate history, mechanical fantasy
Spider-Man: Across the Spider-Verse style (logged 2026-08-10) High-saturation pop warm-cold clash (main golden warm sunlight + cool skylight coexisting) Halftone dots shading, light-colored hand-drawn black ink outline, silkscreen printing texture, cyan/magenta chromatic aberration in highlights Animation, superhero, trendy illustration style

Three steps to use:

  1. Define the theme → check the table and choose 1 main color system (+ optionally 1 auxiliary color system)
  2. Use color system keywords to generate reference images / extract color palettes (§4.2.1 three steps)
  3. Insert the color palette into the video tool to carry through the entire film

Example: "Qimen Town Evil Suppressor" Republic-era mystery → Chinese horror color system (dark cyan + cinnabar red + gray-white) + Nolan realistic texture = the cold, stern base tone of folk mystery.

Master color grading key points quick notes (from the color palette document · can be written directly into the color grading description):

  • Wong Kar-wai: ①Overall offset method (shift the image toward the lower left, with cyan-green in the shadows) ②Lifted blacks (darkest brightness 3-4%, soft film texture) ③Low brightness, high saturation ④Red-green complementary contrast
  • Nolan: ①Cool tone, low saturation (steel blue/gray, desaturated) ②High contrast (heavy shadows + bright highlights) ③IMAX widescreen epic scale ④Live-action practical effects texture
  • Dunhuang mineral colors: azurite #1685A9 + malachite #2C9678 + cinnabar #FF461F + earth yellow #D6A01D (three schemes: layered green, brilliant vermilion-gold, and antique brown-black)
  • Forbidden City court colors: Forbidden City vermilion #C83C23 + glazed yellow #F0C75E + lapis blue #1C2C5B + jade white #F3F9F1

Spider-Man: Across the Spider-Verse three-in-one art style (logged 2026-08-10 · reproducible formula):

All three are indispensable: printed comic texture + pop color clash + halftone/chromatic aberration lighting.Do not write film titles/director names(content safety red line), only write the visual grammar.

  • Art style: shadow areas are densely covered with fine halftone dots; object edges have thin, light-colored hand-drawn black ink outlines; the entire frame is covered with ultra-fine digital noise + light silkscreen printing texture, presenting the tactile feel of high-end comic printing
  • Color: the same frame simultaneously contains two strongly clashing groups of warm and cold light sources — the main light is golden warm sunlight, plating a warm golden rim light on the right contour of the character; cool skylight enters from the left window, laying pale cyan reflected light on the left shoulder/left edge of objects, with warm and cold coexisting on the same object surface
  • Lighting: shadows have a light halftone dot texture (not ordinary smooth CG gradients); highlight edges have extremely faint cyan/magenta chromatic aberration, simulating printing misregistration defects

✅ Verified prompt (2026-08-06 user-tested): cinematic golden formula

[镜头类型与角度],[主体正在进行的明确动作],[具有层次深度的具体场景],[灯光设计],[命名色彩分级],[氛围/类型] --ar 16:9 --stylize 200 --no text, watermark, brand logos
  • Quick fixes: Flat depth of field → add foreground steam/rain + midground props + backlight in fog; blurred subject → emphasize verbs + negative fill light + lower stylize; text ghosting → --no text, watermark, brand logos
  • 6 color grading styles: teal-and-orange blockbuster / bleach bypass (desaturated high contrast) / silver-retained / pastel film / film noir / Wong Kar-wai tones
  • See Obsidian for the complete bilingual templates for 9 directors 03-素材/色卡与提示词模板库/AI影视创作提示词模板库.md(✅Verified)

Director name → descriptive phrasing conversion (to prevent Seedance violations · user-tested):

Director Safe phrasing
Wes Anderson perfectly symmetrical composition, pastel candy color palette, flat lay aesthetic
Wong Kar-wai highly saturated emotional color grading, urban isolation, neon signs, dreamy bokeh
Zhang Yimou bold monochromatic color scheme, Chinese cultural symbolism, ink wash aesthetic
Nolan cool desaturated palette, epic architectural scale, dramatic chiaroscuro lighting
Quentin violence aesthetics, vivid retro palette, dramatic angles, extreme close-ups
Ridley epic widescreen, industrial design, volumetric lighting, desaturated earth tones
Kurosawa dynamic symmetrical composition, dramatic weather elements, samurai aesthetic
Tim Burton gothic aesthetic, exaggerated character design, twisted surreal architecture
Ang Lee East-meets-West fusion, poetic natural lighting, restrained elegant composition

4.2.3 Verified complete prompt template library (new in V3.0 · ✅ user-tested)

Source: desktop《AI影视创作色卡与提示词模板库.docx》, all verified as usable by the user on 2026-08-06. Complete bilingual templates for 9 directors + color grading see Obsidian 03-素材/色卡与提示词模板库/AI影视创作提示词模板库.md。

① Cinematic golden formula (verified):

[镜头类型与角度],[主体正在进行的明确动作],[具有层次深度的具体场景],[灯光设计],[命名色彩分级],[氛围/类型] --ar 16:9 --stylize 200 --no text, watermark, brand logos
  • Quick fixes: Flat depth of field → add foreground steam/rain + midground props + backlight in fog; blurred subject → emphasize verbs + negative fill light + lower stylize; text ghosting →--no text, watermark, brand logos
  • 6 color grading styles: teal-and-orange blockbuster (Teal&Orange) / bleach bypass (desaturated high contrast) / silver-retained / pastel film / film noir / Wong Kar-wai tones
  • Lighting verbs: hard key light/soft key light/negative fill light/practical light source/backlight in fog/rim light/specular highlight
  • Atmospheric elements: volumetric fog/rain droplets/dust particles/lens flare/light wrap

② Chinese templates for 9 directors (verified · when feeding, it is recommended to remove director names and use the §4.2.2 conversion table):

  • Wes Anderson: perfectly symmetrical composition, centered framing, candy-colored palette (pink/yellow/mint green), exquisite set design, flat-lay aesthetic, retro props, whimsical atmosphere, eccentric character styling, extremely precise storyboarding
  • Wong Kar-wai: highly saturated emotional colors (deep red/emerald green/electric blue), slow motion, handheld cinematography aesthetic, urban alienation, motion blur, Hong Kong neon signs, nostalgic romance, dreamy bokeh, cigarette atmosphere
  • Zhang Yimou: bold monochrome palette (dominant red/black/white), Chinese cultural symbols, epic grand scenes, traditional Chinese architecture, ink-wash painting aesthetic, dramatic composition, wuxia elegance
  • Nolan: cool-toned low saturation (steel blue/gray), epic architectural scale, IMAX widescreen, practical-effects aesthetic, time manipulation/fragmented narrative visuals, chiaroscuro lighting, rational atmosphere
  • Quentin: violence aesthetic, vivid saturated retro colors (yellow/orange/red), dramatic low-angle/worm's-eye views, stylized action choreography, 1970s B-movie aesthetic, extreme close-ups, trunk shot, dramatic blood effects
  • Ridley: epic widescreen composition, industrial design aesthetic, atmospheric volumetric lighting, Tyndall light, low-saturation earth tones, smoke atmospheric effects, gritty realism
  • Kurosawa: dynamic symmetrical composition, samurai warrior aesthetic, dramatic weather (rain/wind/fog), black-and-white or low saturation, classical Japanese architecture, epic action choreography, humanist storytelling
  • Tim Burton: Gothic aesthetics, dark fantasy, exaggerated character design (big eyes, slender limbs), stop-motion handmade texture, distorted surreal architecture, German Expressionist lighting
  • Ang Lee: Fusion of Eastern and Western aesthetics, poetic natural lighting, delicate emotional expression, restrained and elegant composition, cultural symbolism, contemplative and introspective mood, soft colors

Example: Wong Kar-wai bar scene (verified example) → A woman in a red cheongsam leans against a bar, neon lights reflecting off wet surfaces, slow motion, handheld camera, dreamy bokeh, cigarette smoke atmosphere

4.3 Material Emotion Reference

Metal → cold and technological | Wood → retro and warm | Stone → solemn and heavy | Glass → clean and premium | Fabric → comfortable and gentle

Modifiers: rough / smooth / polished / matte / glossy / worn / scratched / weathered

4.3.1 Imperfect Realism (New in V2.2 · Anti CG Plastic Feel)

Source: Cinematic World Generator. AI images that are too clean = plastic feel/CG feel; use the sense of everyday wear to make the world "lived-in."

Texture elements (write directly into the image description):

  • Film grain / haze / weather traces (rain stains, dust)
  • Worn fabric / cracked stone / hand-shaped wood
  • Mud / skin texture / animal breath (white vapor in winter)
  • Practical light (torches, oil lamps, candlelight rather than studio lighting)
  • Environmental wear: old tools, footprints, worn door frames, cooking smoke

Usage: after writing ornate material descriptions, add 1-2 wear elements to break the "brand-new feel." Epic feel ≠ clean; realism = traces of use.

4.3.2 Two Great Techniques for Foreground Occlusion (New in V2.2 · Hedgehog Planet)

Source: Hedgehog Planet super-i. AI by default generates unobstructed "omniscient perspective" images; foreground occlusion tells the AI "where the camera is hidden," using perspective relationships to build three-dimensional space.

① Physical foreground occlusion (establishes camera position and spatial layering)

  • Principle: In real shooting, the camera has a physical position; foreground occlusion = telling the AI where the lens is hiding (behind a pillar/outside a door/in front of a window)
  • Universal formula:[工业级摄影参数] + [极度前景虚化(叙事介质)] + [中景主体与动作] + [背景环境] + [胶片与色彩质感]
  • Example: blurred marble pillar foreground framing the left edge + midground heroine on the phone + hotel lobby + teal & orange color grading = the narrative feel of "hiding behind a pillar and eavesdropping"
  • Pitfalls to avoid: ① Do not write a "passport photo"-style flat list (no story feel); ② Do not add occluders that clash with the environment (adding cherry blossoms to an underground garage → breaks logic)

② Atmospheric foreground occlusion (breaks the vacuum feel, makes the air perceptible)

  • Principle: Use aerial perspective so that light travels through a medium (Makoto Shinkai = floating dust motes/rain streaks/thin mist giving light a path)
  • Medium words go directly into the prompt:atmospheric haze(thin mist)dust motes(dust motes)particles(grain)
  • Pitfalls to avoid: ① Do not overuse "extreme detail, 8k, sharp focus" to erase airborne impurities (increases plastic feel); ② Writing only "Tyndall light" without a medium → the light shafts look hard-painted
  • Example: golden sunlight pouring in through the window + thin mist filling the room + Tyndall effect + dust motes glowing in the light beam = the emotional feel of time passing

Complementary to §4.3.1 Imperfect Realism: wear handles materials, occlusion handles space and air.

4.4 Shot Duration Reference (Generation Segment Time Points)

Duration Purpose
1-1.5s Flash cut/impact (high-speed collision, moment of fright)
1.5-2s Hook/climax quick cut
2-2.5s Regular action/dialogue
2.5-3s Narrative narration/establishment
3-4s Emotional settling/reaction
4-6s Slow-motion atmosphere/freeze-frame ending

Information volume duration estimation (supplement by Shan Yin, master director · core duration rule):

  • How many layers of information does the audience need to "read" within one shot? Reading one layer ≈ 2-3s, each additional layer adds 2-3s.
  • Single information 2-3s (a close-up of an object, a simple action) / Double information 4-5s (action + environmental relationship, dialogue + reaction) / Multi-layer information 6-8s (complex scene, multi-person interaction) / Comprehensive passage 10-15s (long-take staging, emotional negative space)
  • Dialogue duration: mentally "act out" using the character's personality/age/emotional state — an impatient young person arguing: about 1.5s for 10 characters; a calm elderly person reminiscing: 10 characters may take 5s (including pauses, breathing, silence); drunken slurring: about 4s for 10 characters; in addition to the dialogue time, add reaction time (the other party's pause after listening, facial expression changes).

4.5 Grand Scene Special Section (V2.1 new addition · core of assessment standard ③)

Triggering scenes: battlefields, city panoramas, palaces/sects/giant structures, mountains and great marshes, starry sky and cosmos, disaster scenes, thousands of troops and horses, giant fantastical objects, etc., anyscene where characters are juxtaposed with a huge environment/giant object. Requirements:Shot type + focal length (in millimeters) + camera movement + proportional relationship between characters and environmentAll four items must be explicitly written into the prompt; none can be missing.

Shot type-focal length reference table (grand scenes):

Shot type Recommended focal length Purpose Prompt writing example
Extreme long shot/God's-eye view 16-24mm ultra-wide angle Establish the scale of the world, the full view of mountains, rivers, cities "16mm ultra-wide angle extreme long shot, characters occupy only one-twentieth of the frame"
Full shot (ground level) 24-35mm wide angle Army formations, crowds, full view of giant objects "24mm wide-angle full shot, the formation extends from the near end of the frame to the horizon"
Medium full shot/character-scene relationship 35-50mm Characters conversing with the environment, walking among giant objects "35mm medium full shot, characters standing beneath the city gate, the gate opening about eight times the height of the characters"
Medium shot narration 50mm Character actions and dialogue within grand scenes "50mm medium shot, the general rides a horse through burning streets"
Close shot/close-up 85mm+ telephoto Individual emotions and key details within grand scenes "85mm telephoto close shot, soot and sweat on the soldier's face, the background army formation compressed and blurred"

Grand scene prompts must contain four elements:

  1. Shot type: extreme long shot/full shot/medium full shot, etc. (see table above)
  2. Focal length: explicit millimeter number (16-24mm ultra-wide angle / 24-35mm wide angle / 35-50mm standard / 85mm+ telephoto)
  3. Camera movement: aerial shot/crane up to reveal scale (crane up from a character close-up to a full city view); slow lateral tracking in extreme long shot (establishing spatial breadth); push-in (approaching key characters from a full shot); for epic scenes, "God's-eye view vertical overhead + slow push-in" can be used
  4. Proportional difference between characters and environment: must be written clearly ("characters occupy only one-tenth of the frame" "the city gate is about eight times the height of the characters" "the formation spreads across the plain like a swarm of ants")

Grand Scene Composition Rules (Shanyin/Xiaoluo combined):

  • Three-Layer Depth: foreground (characters/occluders) + midground (main action/crowd) + background (massive environment/sky/horizon) must all be present simultaneously
  • Power Composition: characters' insignificance within a massive environment = sense of oppression (miniaturization); character centered + symmetrical environment = sense of ritual; giant object pressing in from above the frame = sense of threat
  • Scale Anchors: first establish scale with reference objects that have a clear sense of size (city gate/tree/flag/steps), then place the characters; only then will AI not draw the giant object as small
  • Aerial/High-Altitude Camera Position: vertical top-down or high-angle overhead shot, used for military formations, cities, terrain; combined with slow push-in or lateral tracking
  • Particles and Aerial Perspective: add aerial perspective to the background (fog/dust/light attenuation), sharp foreground and soft background to reinforce depth; windblown sand/flags/smoke density must be written into the prompt
  • Light and Shadow: grand scene light follows scale — dawn/dusk golden side-backlight emphasizes volume; overcast top light emphasizes oppression; light source direction and time are already locked in the scene assets and do not change abruptly within a shot
  • VFX Prompt Specifications: for thousands of troops/giant objects/disaster scenes, write "crowd density, formation order, ratio of giant object to architecture" rather than "a lot of people"; do not write unrenderable quantifiers ("tens of thousands" → "formation extending to the horizon")
  • Macro Visual Gear Shift: grand scenes are naturally gear-shift points — the first shot entering a grand scene uses an extreme long shot, and exiting a grand scene uses a close-up to close it out, forming a "macro → micro" visual loop

Example of a grand scene shot description (can go directly into the generation segment):

4-6s:16mm超广角大远景,垂直航拍缓慢推进——千军阵列如蚁群铺满原野,战旗在风沙中翻卷,阵列从画面近端一直延伸到地平线的城墙之下。人物在画面中仅占极小比例。空气透视让远端的城墙隐入烟尘。(远处战鼓声、风沙呼啸声)

Gate E: Grand Scene Gate (dedicated to Assessment Criterion ③)

  • Shot size explicitly labeled (extreme long shot/extreme wide shot/medium wide shot...)?
  • Focal length explicitly labeled in millimeters (16-24/24-35/35-50/85+)?
  • Camera movement explicitly labeled (aerial lift/lateral tracking/push-in...)?
  • Is the scale difference between characters and the massive environment/giant object clearly written?
  • Are the three layers of depth (foreground/midground/background) present?
  • Are aerial perspective/particle density/light direction clear?
  • No unrenderable quantifiers such as "tens of thousands" or "a lot of people"?

4.5.1 Giant Object Aesthetics Special Template (V3.2 addition · user-verified and confirmed 2026-08-09)

Source: 4 user-provided, verified-usable giant object prompts (mythical heavenly dragon/ancient deity statue/sky giant whale/polar giant ice beast) + same-day consecutive test verification. Used for "giant object aesthetics verification" shots such as giant objects/divine beasts/giant ruins.

Two writing styles (user-verified and confirmed):

Style A · Single-segment direct delivery (the giant object prompt itself, suitable for static/single-image):

【神话XXX】巨物美学,【远古传说中的XXX,体貌描述——身躯/头部/材质/部分隐没("仿佛无尽延伸")】XXX巨大到超出人类理解并遮蔽天空,主体占据画面绝大部分空间,压迫感极强。画面前景或角落出现一个极其渺小的【人类】,如米粒般微小,用于形成强烈尺度对比。低角度仰拍,史诗级纵深构图,电影级光影,真实自然环境,超现实照片级真实,空气透视,翻涌云海,逆光,体积光与漂浮颗粒,低饱和高级色彩,窄画幅,极致细节,营造神圣、恐惧、孤独与敬畏感。【XXX】以具有压迫感的姿态凝视【镜头方向】

Style B · Six-segment format (user confirmed "this is our video prompt format" · only use this for >60s long video generation segments): For giant object aesthetics segments, only use the six-segment format if >60s ([Style] → [Characters and Scene] → [Color Card Reference] → [Shot] → [Timeline] → [Sound]); break the giant object structure from Style A into each time point of the [Timeline], write the shot segment as low-angle upward shot + slow push-in, and within the timeline keep the subject filling the frame + rice-grain human scale anchor + gaze toward the camera direction.Do not mix: generation segments ≤60s always use the five-segment format (no [Characters and Scene]), and only >60s uses the six-segment format.

Core iron rules (user-verified):

  1. The subject occupies the vast majority of the frame's space— the sense of oppression comes from filling the frame, not from displaying it at a distance
  2. A tiny human as a scale anchor— place a rice-grain-sized human in the foreground/corner of the frame to create a stark scale contrast (write the proportion difference as "like a grain of rice," "as tiny as a speck of dust")
  3. Low-angle shot looking up— epic depth composition, the giant looming from the frame toward the viewer
  4. Volumetric light / aerial perspective / floating particles— the holy trinity of sacredness: backlight + low-saturation high-end colors
  5. End with an emotional word— sacredness, fear, loneliness, awe (choose according to the subject matter)
  6. The giant gazes toward the camera direction— drooping / looking down / gazing at [camera direction], closing the loop of oppression
  7. ❌ God's-eye overhead view is forbidden— an overhead view makes the giant look small and dilutes the oppression (user-tested: "the God's-eye view felt weird, and actually made the giant look smaller")
  8. The subject is partially hidden— "part of its body is hidden in clouds/dense fog/glaciers," suggesting endless extension, which is more oppressive than showing the whole
  9. ❌ No human speech during the giant segment(2026-08-09 user-tested) — during the giant segment, once a human speaks, the model tends to mess up the lines, so write no dialogue at all; keep only gasps/breathing for the human voice, pure visuals + ambient sound effects

Eight techniques to strengthen the sense of a giant (confirmed through user discussion on 2026-08-09 · stackable):

  1. Local close-up contrast (strongest)— don't shoot the whole body; shoot one part of the giant in the same frame as a human ("one scale of the giant towers on the ground, a human stands at the base of the scale, their height reaching only one tenth of the scale's edge"), and the brain automatically fills in the whole body—what the brain imagines will always be bigger than what is drawn
  2. Out-of-frame composition— the giant pushes beyond the frame boundary ("the giant's body extends infinitely toward the top of the frame, its top exceeding the frame boundary—you can't see its head"), the frame can't contain it
  3. Human-perspective limitation— write from a human's upward-looking perspective, looking up to the limit without seeing the whole ("the human tilts their head back to the limit of their neck, and can only see the lower edge of the giant's chin, everything above is blocked")
  4. Multi-layered scale anchors (two-level steps)— a three-tier comparison of human → reference object → giant ("the human stands beside a broken stone pillar, the pillar is about five times the human's height; a single scale of the giant is taller than the entire pillar")
  5. Speed contrast— the giant is extremely slow while the human is extremely fast ("each blink of the giant takes three seconds; the human runs desperately at its feet, taking dozens of steps before the giant slowly turns its head"), the bigger it is the slower it moves, and the speed difference itself implies scale
  6. Physical feedback— when the giant moves, the environment reacts ("the giant takes a step, cracks spread across the ground like a spiderweb toward the horizon, gravel bounces at the human's feet, and distant stone pillars topple in rows")
  7. Sound comes first + low frequency— sound arrives before the image, low frequencies swallow the human voice ("the giant's heartbeat is like muffled thunder, coming from deep within the earth, and the gravel beneath the human's feet bounces to the rhythm of the heartbeat")
  8. Silhouette + concealment— most of the giant is hidden in fog/clouds/darkness with only its outline showing ("most of the giant's body is submerged in the thick dusk fog, leaving only a mountain-ridge-like outline undulating in the fog, as if it is still growing")

Combination suggestions: The most ruthless combination = ①Local close-up + ③Perspective restriction + ⑧Obscuration (humans only see a part, looking up they cannot see the whole, and most of it is still in the fog).

Subject replacement slot:

  • Mythical heavenly dragon: dragon body coiled around the sea of clouds, long horns piercing through the cloud layer, scales heavy and ancient, lowered dragon head gazing at the camera
  • Ancient deity statue: standing like a mountain range, sacred light beams cast down from the rift in the sky/dome, the giant stares in the direction of sight with an oppressive posture
  • Sky giant whale: floating in the sea of clouds like a moving continent, whale fins cutting through the cloud layer, tail fin submerged into the skyline, slowly looking down and gazing at the camera
  • Polar giant ice beast: ten-thousand-year ice armor with angular ice spikes, crouching on the snowy plain, beast head breaking through the blizzard, coldly lowering its head and gazing at the camera
  • Xuanming giant serpent: serpent body coiled like city walls, scales layered, serpent head emerging from the abyss, dark-gold vertical pupils gazing at the camera (user 2026-08-09 test)
  • Ruin stone spirit giant god: stone skin with rune patterns, rising from cracks in the earth, runes spreading from the eye sockets across the whole body, lowering its giant head and gazing at the camera (user 2026-08-09 test)
  • General replacement: just replace the [Mythical XXX] name, appearance description, and subject emotion words to apply it

4.6 Action scene special section (new in V2.2 · referencing 3D action scene video generation notes + user verification)

Source: desktop "[1.24]--3D Action Scene Video Generation Notes" (Jimeng Seedance practical course) + user-verified transformation scene template. Action scenes/transformation scenes/high-speed fight scenes can directly apply the framework in this section.

⚠️ Primary-secondary relationship (established in V4.1.1): Within this section, 4.6.1-4.6.7 aresupplementary items(transformation scenes/three levels of destruction/materials/multi-person constraints/multi-character splitting—content not covered by the combat special section);fight actions and prompts shall always follow §4.6.8 (action correctness) + §4.6.9 (intensity), and in case of conflict, §4.6.8/4.6.9 and the two reference files shall prevail.Special section takes priority over general rules (supplemented in V5.0.2): fight segments are also constrained by the general rules for generated segments in §2.4, and when the two conflict (such as §2.4 "only one camera movement per segment" vs §4.6.9 richness of fight camera movement, §2.4 camera writing vs §4.6.9 five elements of camera specifications)the fight special section §4.6.8/4.6.9 shall prevail, but the five-part skeleton ([Style] → [Color Card Reference] → [Camera] → [Suggested Timeline] → [Sound]) remains unchanged.

4.6.1 Three-level hierarchy of destruction effects (essential for upgrade/destruction scenes)

From slight to complete, describe progressively according to the degree of destruction so that AI can present a perceptible destruction process:

  • Beginner: cracks appear on the surface (deep cracks in the crystal platform, wall cracks spreading)
  • Intermediate: structural disintegration (corridor pillars turning to sand, wall peeling, beams and columns breaking)
  • Ultimate: total annihilation (buildings turning to dust, objects disappearing completely)

Writing method: first write the starting point of destruction (cracks), then write structural disintegration (sanding/peeling), and finally write total annihilation (dust), escalating level by level without skipping levels.

4.6.2 Transformation scene template (user-verified usable · 7 stages)

Standardized timeline structure for transformation/awakening/evolution scenes:

Stage Time proportion Content
1 Buildup 0-15% Static close-up of the original form, material details clear (rough fabric/old armor texture), surrounding environment slightly moving
2 Disintegration 15-30% The original clothing/armor begins to disintegrate from the shoulders/chest, tearing into glowing fragments, with the fragments suspended in midair
3 Energy Gathering 30-45% The fragments and flowing light rise while circling along the body, the air distorts, and strands of hair rise without wind
4 Pupil Mutation 40-55% Close-up of the pupils: phantom visions and divine radiance emerge, marking the activation of the transformation
5 Reassembly 50-70% New form elements such as petals/scales/light wings gather from all around, embedding and taking shape along the body's curves, with dense clicking sounds
6 Burst 65-85% At the moment the new form takes shape, energy explodes (sea of flowers tornado/shockwave/light burst), and the environment is damaged (see 4.6.1)
7 Settling 85-100% The protagonist stands amid ruins/radiance, with lingering aftershocks slightly trembling, dust settling, and a slow push-in close-up of the eyes to finish

Five hard rules for transformation scenes:

  1. The forms before and after awakening must be uploaded as two separate design images (tool layer: platforms such as Jimeng require @respective form images)
  2. The disintegration -> reassembly must clearly write the transformation process of "old material -> fragments -> new material," without directly jumping
  3. The destructive effects in the burst segment progress in three levels (4.6.1), matching the intensity of the transformation
  4. The settling uses a slow push-in + eye close-up freeze-frame to complete the emotional loop
  5. Material contrast competition: rough texture of the old form vs. sharp semi-transparency of the new form (petals/scales/crystal)

4.6.3 Action Scene Rhythm Combinations

Slow-fast-static rhythm (transformation/awakening scenes): Opening medium shot circling slow-motion high-speed ramp -> high-speed push-in burst -> ending environmental silence Fast-hold-shake-follow-settle rhythm (multi-person fight scenes): Rapid lateral tracking shot (speed vision) -> freeze key frame -> impact/explosion shake -> follow action/projectiles -> ending freeze-frame

The three-piece speed expression set (mandatory for high-speed motion):

  1. Afterimage trails (high-speed movement leaves afterimages)
  2. Air distortion (air deformation around breaking the sound barrier/high-speed objects)
  3. Motion blur (blurred edges of moving objects)

4.6.4 Action Scene Material Precision Comparison (write into the image description)

Object Precision description
Clothing fabric Texture + motion wrinkles (changes in wrinkles when sleeves and skirts flutter)
Plants/vines Rough surface texture + contrast with the fresh tenderness of new buds
Armor/metal Reflection + cold metallic feel of chains
Energy body Light-transmitting property + glowing edges
Building stone Heavy feel + cross-sections/dust after destruction

4.6.5 Action Scene Special Effects Checklist (can be written directly into prompts)

  • Dash air-wave particles (air waves around the body during high-speed dashes)
  • Trailing light points on seeds/projectiles
  • Mud clumps splashing when vines grow wildly
  • Green leaves/petal light particles dancing
  • Shockwave particle diffusion
  • Destruction dust settling

4.6.6 Multi-Person Fight Scene Constraints (prevent proportion imbalance)

  • Write an explicit number or range for the number of crowd members/soldiers ("several dozen faceless silver-armored dust purifiers"), not "many people"
  • Combat proportion balance: our side's/enemy's numbers and combat power match, using text constraints to prevent one-sided imbalance
  • New characters joining are marked with @character name and their setting images uploaded; characters that do not appear are deleted from the prompt
  • Multi-character coordination: each character's action direction/rhythm is aligned to avoid each fighting their own battle

4.6.7 Multi-character action splitting (new in V2.2 · Hedgehog Planet)

Source: Hedgehog Planet super-i. Solves the "chaotic actions, changing A ruins B" problem in multi-character scenes — AI attention and compute power shift: when multiple actions are crammed into the same sentence, the AI preserves only one character while the others become background boards, and it does not understand the timing of "simultaneously/then".

① Time-segment splitting (video): Do not use a running account ("the man on the left drinks coffee, the woman on the right dances, then the man claps" → AI confuses the timing). Change to:

  • Clarify the compute focus for each time segment:0-3s 焦点:左男喝咖啡;4-8s 焦点:右女跳舞;8-10s 左男鼓掌
  • Keep only one main character action per time segment

② Spatial placement (frame composition): Clearly mark character positions:

  • 画面左侧:男人;画面右侧:女人 → Avoid the AI grabbing the wrong subject

③ Semantic editing lock modification: Do not use mask repainting (edge flickering + brushing away satisfactory characters); use the tool's video editing function:

  • 保留右侧的女人,把左侧男人的动作改成抬手打招呼 → AI precisely identifies the character and changes only the target part

④ Storyboard splitting to break posed shots: Do not cram all actions into one long shot; split into three segments (using "station farewell" as an example):

  • Stage 1 Close-up - emotional buildup: the heroine's slightly reddened eyes, the hand clutching the ticket
  • Stage 2 Medium shot - action explosion: letting go, turning and running
  • Stage 3 Wide shot - climax freeze: slow-motion embrace, train roaring
  • Pitfall avoidance: there is no need to force the background to be completely consistent; the audience's attention is on the character actions, and subtle background differences are automatically ignored

Tool adaptation (Jimeng Seedance): upload scene reference image + character setting image (full body/three-view) → use @character name in the prompt to associate → select model (2.0/2.0fast) → all-purpose reference type → 16:9 or 9:16 → default 15s per segment (can be split into 7 groups of shots). To batch-generate shots in the same style, only change the varied parts.

4.6.8 Martial arts action system (new in V4.1 · integrating the industrial-grade 15-cell PREVIS system)

Source: desktop "Vocational School Recommendation skill/6.txt" industrial-grade 15-cell martial arts PREVIS system (archived 2026-08-16). Supplements what 4.6.1—4.6.7 lack about "the action itself": power chain / attack-defense interplay / continuity / axis / QA.Fight scene storyboards and prompts can directly apply this section. For the complete version (20 chapters, including the 15-cell PREVIS storyboard generation and full rules for converting to prompts), see references/武打PREVIS-15格动作预演系统.md。

① Action power chain (11 links must be written for every move): intent → starting preparation → center-of-gravity change → footwork displacement → core power generation → limb transmission → attack trajectory → contact point → hit/block feedback → recovery inertia → inheritance by the next cell. It is forbidden to draw only the result (punching/kicking/explosion/beam clash/static posing) — the audience must be able to understand how the action happens.

② 15-cell rhythm structure (default skeleton, subordinate to the plot): P01 Establish space and distance → P02 Starting probe → P03 Footwork starts, center of gravity shifts forward → P04 First attack enters → P05 Dodge and block → P06 First collision → P07 Loss of balance under force → P08 Counterattack and seize initiative → P09 Displacement into close quarters → P10 Combo/weapon interplay → P11 Key hit/disarm and break defense → P12 Body feedback inertia → P13 Space re-establishment → P14 Stabilize center of gravity → P15 Counterattack starting move.P15 does not end the fight(unless the plot explicitly requires it) — a fight is a high-energy cross-section of the process, not the ending.

③ Attack-Defense Game: Rhythm 快→快→停→爆→反馈→反制→续; one rhythm change every 3-5 panels, one change of dominance every 6-8 panels; no unilateral combo throughout, no same action throughout.

④ Fists, Feet, and Weapons: For punches, always specify which hand/ground push/center of gravity/waist and hips/shoulder drive/knuckle landing point/what the opponent reacts to first; for kicks, always specify support leg/knee lift/hip rotation/trajectory/landing point of instep or palm-heel/center of gravity after impact; for throws and takedowns, break down into six steps (entry → control → lock center of gravity → exert force → off-ground imbalance → landing feedback); complex ground techniques are not completed in a single panel. For weapons, always specify who holds it/which hand/length/blade direction/trajectory/contact point/sliding/blade press/hand shock;blade collisions must not float in the air, at least one of: blade press/blade slide/blade lift/deflection/hand shock/position lock/disarm/aftershock; no weapon clipping/hand switching/disappearing/length drift/no contact point.

⑤ Five Continuity Locks (each panel inherits from the previous one): state inheritance (stance/orientation/center of gravity/foot position/hands/injuries/stamina) · 180° axis (A left, B right; axis jumps can only be rebuilt via orbit/top-down/frontal/pan, and after the jump maintain the new relationship) · inertia (forward dash throws clothes back, sudden stop throws clothes and hair forward, rotation in the same direction; no mutual conflict) · injuries and damage persist and do not automatically recover · special effects do not replace physical contact. Prohibited: teleportation/groundless side switching/space reset/multiple hands and feet/limb clipping/body fusion.

⑥ Action QA Quick Check (self-check before writing): 15 panels complete? Attack and defense paired? Power chain present? Contact points clear? Body feedback after impact? Weapons not clipping? Left-right stance continuous? Axis established? Inertia direction correct? Injuries continuous? No realistic skin tones/colored materials?

⑦ Storyboard to Seedance Prompt (when the user requests panel-by-panel output): the storyboard locks order/shot size and camera position/stance/action direction/weapon position/contact points/transitions; character images lock face shape and clothing, scene images lock environment lighting and weather, weapon images lock appearance and grip. Prompts must ignore white-model materials/numbers/arrows/annotations. Each entry has 14 elements: shot number/suggested duration/corresponding grid panel/shot size and camera position/camera movement/action starting point/footwork and center of gravity/power chain/weapon action/contact feedback/recovery inertia/transition to next shot/environmental sound/prohibited items. Sound defaults to no BGM, keeping only environment + breathing + footsteps + fabric + wind break + collision + landing; realistic combat prohibits colored energy effects; the end of the previous shot naturally connects to the beginning of the next shot, and each shot must not restart from stillness.

4.6.9 Anime Combat Prompt System (New in V4.1 · Distilled Anime Combat Methodology)

Source: desktop 《职业学校推荐skill/武打skill.md》 (353KB original text, distilled and archived on 2026-08-16). Division of labor with §4.6.8:PREVIS governs "whether the action is correct"; this section governs "whether the fight is hype". For the complete version (including the full table of 32 iron rules, 15 camera movements + combination formulas, 9-character energy quick reference, technique overview, and three close-up templates), see references/二次元打戏提示词体系.md. After writing a fight segment, it must pass: §4.6.8 Action QA + this section's hype-oriented iron rules + Gate D.

① Completion Rate Formula: opening impact 20% × rhythm and breathing 30% × camera movement richness 25% × emotional turning point 25%. Within the first 1.5s of the opening there must be attack-defense contact (no panoramic setup); upgraded version: within 0.5s, two extreme moves collide head-on + micro slow-motion freeze-frame (2-second bounce rate drops 38.98%).

② 15-Second Skeleton: 0-1.5s opening hard collision → 1.5-5s close-up exchange → 5-8s spatial change (dimensions ≥2 times) → 8-9s fast attack escalation → 9-12s emotional turning point → 12-14.5s finishing move (breath frame → breakthrough → bloom → aftershock) → end on the climaxno ending frame(The shockwave/dust from shattered stone is the final frame).

③ Dual watchability: Long shots show big moves (spectacle type), close shots show killing intent (density type).Within 15s, at least 1 segment of 2-3s upper-body close-up exchange of moves, pure long-shot big moves ≠ a complete fight scene. Close-up rhythm = machine-gun rhythm: 3 light 1 heavy / alternating left-right acceleration / seal-break-kill.

④ Move serialization (Iron Rule 25): AI execution deviation on qualitative words like "machine-gun rhythm/fast" is 30-40%—must write move sequence + landing point + defensive feedback:右摆拳打脸→左勾拳打下巴→右直拳打鼻梁→左横肘切眉骨→右标指刺咽喉,防守方依次侧偏、下潜、拍开、架桥、后撤. Writing "attacks per second ≥ how many times" is prohibited. Close-up hand watchability ranking: smashing elbow > horizontal elbow > inch punch > uppercut > piercing finger > palm push > straight punch > palm slap.

⑤ Five elements of shot specifications (must be written for every shot):[景别]·[拍摄视角+角度]·[运镜方式]·[跟随目标]—Shot size choose one of four (full/medium/close/extreme), camera angle choose one of five (front/side/oblique/high/low), specific angle value (30° low/90° side/45° oblique high), follow target must be written. "Camera pulls back/moves closer/follow shot" is prohibited.⚠️Iron rule of spatial sense: medium shots and above prohibit 90° side-on frontal shots, must use 30-60° oblique angle to create depth。

⑥ Ten commandments of hype-direction iron rules (selected): Five elements of visual peak triggered in the same frame (large movements + fast rhythm + full shot + special effects + camera movement, missing one discounts it) · Breathing frame requires ≥1.5s high-speed buildup (freeze-frame without buildup = boring) · Every hit must have air feedback (punch wind + hit air burst, form changes with the attack) · Differentiated hit reactions (face→lean back/abdomen→curl up like a shrimp/lower body→kneel) · Large-scale dodges (prohibit "sidestep to let it pass", low duck to the ground/lean back ≥45°/backflip) · Three states of defense (after guard break, cannot restore full defense for 0.3-1s) · Special effects obey the character's main color (delete colors that cannot be attributed) · Prohibit 360° surround/spiral-rise camera movement · Zero repetition of moves in the same fight · Full text ≤2000 characters (if over, delete negative prompts first).

⑦ Quick reference for energy auras: Taiji=black and white, White Dove=purple-black, Black Tea=orange-red, Lin Yu=ice blue silver-white, Lightning=pure purple silver-purple, Guava=pink gradient, Ice Woman=blue-white hard ice, Sword God=red-purple instant flash, Feixue=fire-red blazing blade arc. Writing method:右腿如钢鞭甩出,紫黑气焰沿小腿弧线炸裂成新月弧刃(prohibit "character releases purple special effects").

⑧ Character combat structure (must be used when creating fight characters): 1-2 main combat techniques make up 70% + 1-2 counter techniques make up 20-30% (guard break/throw/lock control) + 1 finishing technique (visual highlight); distance coverage ≥2 types (close/mid-range/spatial change); action dimensions 2-3 types (standing/aerial flip/ground control). Even use of techniques is prohibited; stacking all techniques of the same type is prohibited; must have a character action signature (one sentence explaining how it differs from others) and special-effect personality (burn/freeze/bloom/arc, color is only the foundation). Creation process: show technique overview (recommend 4-5) → assign structure → output character card → confirm.

⑨ Weapon rules (must be used for armed combat): Medium/full shot weapon attack and defense prohibit 90° side-on frontal shots (like a side-scrolling fighting game); high-speed weapon moves cannot be all close-ups (only used for openings/hits/block moments 0.2-0.5s); large circular sword flowers/full-moon slashes prohibit close-ups (must use medium/full shots to see the complete trajectory);weapon collision cannot only be written as explosion—must write physical contact of blade body/crossguard/blade edge. Prompt conversion formula:[角色]+[步法/重心]+[身体发力]+[兵器部位]+[攻击轨迹]+[命中/格挡目标]+[兵器反馈]+[镜头目的]. When the user specifies a specific weapon system, first read references/武学招式库-经典招式速查.md the corresponding specialty (swordplay/long-handled scythe/tachi·Dream of a Single Heart/tachi·Feixue Cherry Blaze Style/advanced qinggong/immortal-cultivation advanced sword moves); generating from move names alone is prohibited; unarmed combat does not invoke the weapon library.

⑩ Ultimate Move Material Quick Reference: 1 of 8 breathing frames (extremely slow freeze / sweat droplets drifting / hair strands still / gaze lock freeze / wind stops / aura withdrawal); minimum spatial variation requirement ≥2 types within 15s (airborne + ground-hugging / inverted + side roll, no standing still throughout); two types of emotional turning points — suppression type (rage awakening / combo acceleration / ultimate move release / full-power crushing) or reversal type (force redirection counter / opening capture / movement reversal / retreat to advance / desperate outburst / entering-body reversal); ultimate move three acts = breaking the deadlock (breaking the offense-defense balance) → blossoming (full-power output) → aftershock (shockwave / dust finish). Detailed material library (20+ spatial variation types, 14 reversal techniques, full table of 8 breathing frames) see references.

⑪ Martial Arts Move Library (use moves when writing fight scenes): Fist systems (Wing Chun straight chain punch / biu jee, Bajiquan iron mountain lean / elbow strike, Taiji cloud hands / grasp sparrow's tail, Baguazhang circle walking / sparrowhawk piercing forest, Xingyi beng quan), Leg systems (Tan Tui twelve routes snap kick stomp intercept heel kick hook lift sweep whip kick knee swing, Cai Quan side-door rapid strike, Brazilian capoeira inverted ground-hugging, Taekwondo 540° tornado kick), Throw systems (Judo seoi nage / osoto gari / uchi mata, Jujitsu armbar / triangle choke, grappling joint locks), Movement systems (Lingbo Weibu / shrink the earth into inches / swallow three dips in water / advanced qinggong · Lingbo Nine Turns / cliff-edge moonbreak / sheer-cliff hanging thread), Weapon systems (Huashan / Dugu Nine Swords / Jade Maiden Sword / Ziwu Sword, long-handled scythe, tachi · Dream One Mind 12 forms, tachi · Crimson Snow Cherry Blaze Style 12 forms, advanced cultivation sword moves · Ten Thousand Swords Return to Origin / flying sword long-range kill / sword formation / sword wall).Usage Rules: Move names are written directly into the prompt ("phoenix eye piercing finger straight thrust at the throat"), pair moves use typical combos from each martial art, moves must pass the power chain (§4.6.8) + serialization (§4.6.9), zero repetition within the same fight, special effects follow the character's primary color. Full table (including tachi / qinggong / cultivation sword move camera iron rules) see references/武学招式库-经典招式速查.md。

4.7 Nine-Aspect Worldbuilding Method (new in V2.2 · optional reference)

Source: Cinematic World Generator (OAK). Used forbefore designing new scenes/new worldviewsquick world reference, or cultural setting before grand scenes. Core idea:turn settings into "things happening right now in the frame," each aspect is a cinematic still, not a setting book. Regular storyboards for already-finalized scripts may skip this section.

4.7.1 Nine-Aspect Checklist (each aspect = one specific scene, no repeating scene types)

# Aspect Core Question Scene Content Example
1 Inhabitants Who lives here? Identity/posture/clothing/role/body language, living in the environment rather than standing still
2 Animals What other life shares this world? Labor/migration/being raised/being hunted/being worshipped, or merely hinted through traces/shadows/bones
3 Architecture Where do they shelter/gather/rule/remember? Spaces in use: smoking rooftops, people entering doors, guards beneath stone carvings
4 Landscapes What terrain shaped the civilization? Environment as force: weathering, ruins, crop lines, cooking smoke, footprints, small figures to contrast scale
5 Daily Life What repeated daily actions? Labor/food/crafts/repair/trade/family rhythm, the ordinary over the ritual
6 Travel How do they move? Walking/Cycling/Rowing/Dragging/Climbing/Migration, showing distance, terrain, weather, weight, effort
7 Sound/Culture What does it sound like? Chanting/Listening/Instruments/Tool strikes/Dance/Silence/Oral tradition/Funerals, making sound visible
8 Power/Intensity What tests this world? Storm/Flood/Ritual authority/Social silence/Scarcity/Manhunt/Confrontation/The moment before change
9 Portrait What private moment carries the soul of the world? An intimate emotional moment: reaction/gesture/a glance/wound/breath/dirty hands/a face half in shadow

4.7.2 Checklist of World Continuity Elements (Same Cinematic Feel)

All scenes share: climate / materials / clothing / architecture / tools / symbols / animals / behavior / terrain / lighting logic / color logic / texture.

Complementary to the Sixfold Consistency Lock in §3.2 (within the same episode): this section isat the worldview levelcontinuity (across episodes and scenes); once a world is established, all scenes inherit it automatically.

4.7.3 Principle of Space in Use (Architecture/Scenes Not Displayed)

  • Architecture: smoking roof > empty building; people entering the city gate > panoramic view of the city gate; stone carving under guard > close-up of stone carving
  • Scenes: only space that is being used has life; write "function," not "appearance"
  • Criterion: If the people are removed from this shot, is there still a story? No → qualified; Yes → change it to in use

Consistent with the existing "no display-style composition," this section gives positive phrasing.

4.7.4 Five Essential Questions for Scenes/Shots (Ask Yourself Before Writing Each Shot)

  1. What is happening in this frame?
  2. What are the foreground/midground/background respectively?
  3. Why did the shot choose this camera position (see the decision table in §4.1.1)?
  4. Which element in the scene is "being used"?
  5. Does this image have the on-the-spot feel of something "discovered" (not staged)?

4.8 Director Style Reference Library (Genre Routing · Added by Old Li)

Emotional tone (A): Everyday realism/emotional aesthetics/poetic meditation/absurdism/tragedy/comedy/black humor/romance Genre film (B): Thriller/psychological thriller/horror/crime/film noir/mystery whodunit/science fiction/fantasy and magic Action and confrontation (C): Modern action/wuxia/war/disaster Subject matter and relationships (D): Romance/family bonds/youth coming-of-age/biography/sports competition Form and narrative (E): Mockumentary/road movie/musical/animation/experimental avant-garde Social perspective (F): Social realism/historical costume drama/political satire/main melody

Key references for wuxia action:

  • King Hu: fixed camera positions + precise blocking, still → still → burst rhythm, one move decides victory or defeat, negative space
  • Tsui Hark: creative camerawork unrestrained by physics, whip pans + fast cuts to create chaotic energy, extremely high particle density, slow motion for the key finishing kill
  • Ang Lee: action choreographed as dance, long takes showing continuous action sequences, landscape-painting-style composition, action as inner state

Key references for modern action:

  • Chad Stahelski: wide shots showing full-body action, multiple people fighting in long takes, neon color palette, action clarity first
  • John Woo: slow-motion core, multi-angle coverage of the same action, dual-gun symmetrical composition, violence as aesthetics
  • George Miller: center-frame rule (key information always at the exact center of the frame), two-color high contrast, extremely fast but readable

Key references for epic/grand scenes (new in V2.1):

  • Peter Jackson: alternating epic long shots + intimate character close-ups; large-scale staging (battle lines of thousands of characters); sweeping aerial landscapes
  • Hayao Miyazaki: dynamic composition of clouds, wind, and water; three independent layers of motion—foreground grass and flowers / midground characters / background clouds and mountains; sense of weightlessness in flight
  • Kubrick: symmetrical composition + one-point perspective (grand sense of symmetrical oppression); god's-eye overhead view; wide-angle and telephoto juxtaposition

Reference styles are used only for shot and prompt creation,do not write director/celebrity names into the final prompt(content safety red line).


V. Performance Control (Old Li Micro-expression Library · Supplementary Layer)

The core is Xiaoluo six-dimensional physical performance (§2.2); this module covers areas not covered by Xiaoluosubtle expression depictionto complete it, and the two are used in combination: key medium close-ups/close-ups = Xiaoluo six-dimensional (≥5 dimensions) + Old Li micro-expression library (restraint is tension).

5.1 Three-dimensional emotion description method

Dimension Incorrect Correct
Expression Happy corners of the mouth raised, outer eye corners curved upward, natural smile lines at the eye corners
Action None hands slightly raised, body gently swaying a little
Tone None light and brisk tone, with a slight tremolo of surprise

Emotion must be paired with a triggering scene: it is only natural with cause and effect.

5.2 Micro-expression timeline slicing method

Core rule:restraint is tension.Emotion contracts in the gaze, tiny facial muscle movements, and breathing.

Emotion decomposition formula:动机 + 眼神 + 呼吸 + 脸部 + 身体 + 声音 + 收尾

Timeline format (2-3 seconds/segment):

0-2s: [角色外貌前缀],[初始状态],[微表情细节]
2-4s: [第一个变化],[动作幅度],[情绪过渡]
6-9s: [情绪深化],[面部变化]
12-15s: [最终状态],定格收尾,无黑屏无转场

5.3 Contradictory body language (advanced technique)

Describe contradictory physiological signals rather than stating the emotion directly.

  • "the corners of the mouth are still smiling, but the light in the eyes slowly weakens"
  • "nostrils move slightly, throat swallows, eyes rim slightly wet but no tears fall"
  • mouth smiling while eyes are tired > simply crying

5.4 Emotion baseline library

Emotion Micro-expression formula
Disappointment/heartbreak light in the eyes weakens, eye rims slightly red without tears falling, gaze shifts downward, extremely slight sigh
Guilt/lying gaze wanders within an extremely small range, blink rate increases slightly, throat moves slightly, breathing becomes shallow
Letting go/smiling through it gaze shifts from tense to soft, slight long exhale, shoulder line relaxes, extremely faint smile at the corners of the mouth
Anger/suppression gaze fixed and dead, breathing slows and becomes heavy, jawline extremely slightly tense, shoulders completely still

5.4.1 Gaze-emotion quick reference (2026-08-11 incorporating AI Film & TV Toolbox + Xiaoluo expression settings)

Gaze state Corresponding emotion/intent English label
Gaze lowered to the ground/desktop listening to a reprimand, submission, endurance, avoiding conflict averted gaze
Direct gaze into the other person's pupils Stating a position, standoff, questioning, realization, threat direct eye contact
Gaze follows the receding figure/moving target Parting, reminiscence, reluctance, covert attention tracking gaze
Gaze wanders within a very small range Guilty conscience, lying, scheming shifty eyes
Gaze fixed without blinking Suppressed anger, killing intent, interrogation pressure dead stare
Gaze lowered + slightly reddened eye rims Disappointment, heartbreak, feeling wronged (without tears) downcast eyes
Gaze looks toward a direction off-screen Suspense, memory, target lock (next shot gives the reverse-position target) off-screen gaze
Gaze shifts from tense to soft Letting go, forgiveness, warmth softening gaze
One eyelid narrows slightly + gaze reassesses Suspicion, reassessment, putting up defenses appraising gaze
Gaze distant and level (eyelids heavier) Coldness, disregard (distinguished from "calm" by amount of movement) detached gaze

Usage: When writing scene partner interactions, select the gaze state directly from the table (including English tags), and use it together with the §2.4 gaze anchoring iron rule; expression intensity develops frame by frame with the plot, and adjacent shots do not jump without reason.

⚠️ Tag responsibility division (2026-08-11 user test: forcibly applying the table made the effect worse): The quota of 2-4 English tags per shot,prioritize shots/performance textures that cannot be described in Chinese(cinematic handheld, depth of field, motion blur, precise eye focus); once the gaze relationship is written clearly in Chinese, do not repeatedly occupy English tag slots, and only use averted gaze / direct eye contact to emphasize key performance nodes (lowering the eyes to avoid/direct eye contact to state a position).Verified tag combinations take priority; do not replace them just to fit the table——replacing camera words with gaze words loses image texture (tested: after changing subtle cinematic handheld to tracking gaze, the effect got worse).

5.5 Four red lines (self-check before performance output)

  1. No environmental characters or spatial displacement: Replace interaction targets with camera/frame direction; in fight scenes, change opponents to passive frame descriptions ("blade strikes armor," not "opponent blocks")
  2. No abstract psychological description: Translate everything into facial muscle and breathing actions
  3. No action overload: Within a 2-3 second slice, at most 1-2 tiny actions, using "slowly, very slightly, gradually"; fight scenes may be moderately relaxed but each shot is still controlled to 2-3 core actions
  4. Mandatory self-check: Are there personal pronouns other than the protagonist? → Delete. Is there inner mental activity? → Translate into muscle actions.

5.6 Hollywood acting system (ACTING SYSTEM upgrade · V3.1)

Source: HELL GRIND project ACTING SKILL (Seedance 2.0 character performance system), full text see references/ACTING-SKILL-角色表演系统.md。 Core axiom: acting is behavior under pressure (BEHAVIOR under pressure), not emotional display. What the character wants, what is blocking him, how he goes to get it—emotion is only a byproduct of this struggle, and emotion is never written directly.

5.6.1 Five pillars of a scene (must be broken down for each character in each scene)

Pillar Content Incorrect writing Correct writing
Objective What is wanted here and now,verbs directed at the opponent Angry Make him pay back the money
Obstacle and stakes What's in the way + cost of failure None Two-hour deadline / pride won't allow begging
Tactics The specific method to achieve the goal (verbs) One move for the whole scene Pressure → seduce → humiliate → bargain → threaten, switch if it fails
Beats Smallest unit of action, 2-4 changes No change throughout the whole scene Pause / change posture / change speed / shift gaze, beat changes must be visible
Subtext True thoughts vs what's said aloud Say it directly A question that isn't a question / repeated question / sudden topic change / ill-timed joke / overly short answer

5.6.2 Character Performance Master Profile Template (150-220 words, single paragraph, one per character)

Character acting as [NAME]. [年龄/体型/体态——身体是履历书]. [心理引擎一句]。Vocal profile: [音色/口音/语速/情绪下如何破音]. Key physical habits and tics: [招牌小动作+触发器;压力小动作+触发器;隐藏行为;面具及其破裂条件]. Walking style: [命名步态+重量/节奏/落脚]. However, when [情绪触发器], [面具崩溃后的转变]. [可选: 软化对象——唯一让脸真软下来的人/物].

Rules:

  1. Write only observable behavior: don't write "he is nervous", write "lower lip trembles, heavy swallow, mouth opens for a deep inhale, lips press together and exhale sharply"
  2. Every small action has a trigger: not "he cracks his knuckles", but "he cracks his knuckles during boring conversation to feign confidence" (format: action + when/why)
  3. Name the gait: "power-walk" "peacock strut" "battering-ram stride", then break down weight/stride/torso/head
  4. Mask + crack: at least one "However, when X——" clause (the dominant gait shatters into a childish slump after rejection)——a person performing two truths at once is a person, not a puppet
  5. Give the softening object only one: one person/animal/object, no dilution
  6. No costume: costume goes in the scene/appearance block, the performance profile must survive across costumes
  7. No camera, no color: performance only handles performance
  8. The body carries biography: "a physique worn down by past struggles" "shoulders that are always tense"——profession/old injuries/self-perception written into the body

5.6.3 Eye life (must be written for every profile, every scene)

Dead eyes are the number one flaw in AI performance. Must write:

  • Micro-saccades + gaze target: gaze keeps moving (drift/glance away to think/scan for details/fall back), never frozen and locked on one point
  • Realistic blink rate: rapid consecutive blinks under pressure; slow and calm when in control; "blink-defocus" when distracted
  • Living eye light: eyes must be moist, have light, be alive
  • Controlled stillness is a choice, not death: predator-style infrequent blinking must be slow and deliberate, gaze still moving slowly
  • Eyes before thought/head: eyes reach the target first, then the head turns; thought is readable before the words come out

5.6.4 Voice permanently locked (not adapted per scene)

  • One permanent voiceprint per character via Voice prompt:"A [age]-year-old [origin/accent]. [音色]; [语速]; [情绪特征——压力下如何变]."
  • When the character speakspaste verbatiminto the audio field, never modify; if present but not speaking, omit
  • Vocal descriptions in the performance paragraph (beat speed changes/voice cracks/whispers) must be consistent with the Voice prompt

5.6.5 States, not transitions (iron law of video generation)

Models are good at states and mess up transitions. WriteAlready in action: "mid-throw, mid-punch, mid-pace," do not write the process of "reaching into the bag, pulling out, winding up." Chain states beat by beat, do not narrate a continuous process.

5.6.6 Group Portraits and Space

  • Group reactions are waves, not synchronized: one person laughs first, the second half a beat late, the third does not laugh; simultaneous synchronization = fake
  • Reactions are worth more than actions: the most precious image after an event is the witness's face
  • Freeze before threat: the group portrait keeps micro-moving → at the key threat everything goes still (the motion→stillness contrast is the punctuation)
  • Movement = motive: no one crosses the scene for no reason (approach = escalation / turn away = expulsion / standing up = seizing power / sitting down during conflict = paradoxical power / stopping at the doorway = decision point / starting to pack up = the body's final ultimatum)
  • The strong are still while the weak are restless: danger does not come from the dangerous person themselves, it comes from the tension of those around them; threat has no warning (no deliberate pause, no slow turn)
  • Accumulated depletion: a character worn down, the body accumulating grayness/heaviness/slower reactions, never reset between scenes

5.6.7 0-5 Performance Rating Scale (Self-Check)

Level Name Characteristics
0 Puppet Recites lines, no behavior
1 Reciter "Facial-expression-ized" text + indicated emotion + diagrammatic body
2 Diligent Can guess the objective, but single tactic / slow reactions / empty pauses
3 Craftsman Has objective and beats, listens to the partner, body is reasonable. Missing: subtext / unexpected tactics / state inertia
4 Alive Continuous behavior + contrasting tactics + subtext diverging from the text + status in the body + reaction before lines + at least one unexpected but true choice
5 Magnet All of 4 + paradox: simultaneously funny and terrifying / cruel and fascinating / lying and moving——dual truth, a single clean emotion in close-up feels synthetic

Aim for 4+ for the protagonist's shots; if self-check is below 2, revise before submitting.

5.6.8 15 Kinds of Bad Performance Atlas (Quick-Reference Fixes)

# Symptom Prompt-level fix
1 Indication (emoji) Take the face out of the task, write the objective + give the hands something to do
2 Playing the result Only write what the character knows at this moment
3 Waiting for lines Reaction starts midway through the partner's line
4 Single tactic Mark the beats, a new tactic verb for each beat
5 Diagrammatic gestures Gestures either precede thought, contradict the words, or are absent
6 Free emotion Build a ladder: restraint→breakdown, emotion must pay a price
7 Body contradicts biography Set a body file: center of gravity / rhythm / wear
8 Lines too clean Add dirt: interruptions / swallowed endings / repetition
9 Threat signals Threat is everyday, violence has no warning
10 Synchronized group portrait Wave-style staggered reactions, varying in strength
11 Dead pause Fill pauses with evaluation or business—or cut them
12 Emotional reset States have inertia; aftereffects carry into the next beat
13 Commentary character Fully believe in the situation; comedy must be performed seriously
14 Close-up overload The tighter the shot, the less movement: only the eyes and the thoughts
15 Dead eyes Fully apply 5.6.3; eye life is never omitted

5.6.9 Pre-publish checklist (performance layer)

  • In every frame, the character's objective = a verb directed at the opponent
  • Obstacles and stakes exist; the cost of failure is real
  • 2-4 beat changes, each visible (pause/posture/speech rate/gaze)
  • Reactions precede the end of the opponent's line; there are moments of evaluation
  • Every character has business; accents that interrupt actions are used deliberately
  • Changes in distance are motivated; status is in the body, not in the words
  • Every small action has a trigger; the mask has its cracks (However, when X)
  • Eye life is explicitly written in: saccades/blink quality/eye light/eye leading
  • If the character speaks, the Voice prompt is pasted verbatim; if silent, it is omitted
  • States are not transitions; the performance layer has no costume/camera/color
  • Group reactions are staggered; the strong are still, the weak are restless
  • Self-check on the rating scale ≥4

VI. Film and television aesthetics and imagery (Old Li's supplementary layer)

6.1 Core values of film and television aesthetics (supplement · aesthetic appreciation + directing class)

Unity of form and content (文质彬彬)

When substance prevails over adornment, one is crude; when adornment prevails over substance, one is pedantic. Only when adornment and substance are in harmony is one a gentleman. — The Analects · Yong Ye

  • Substance (content) = story plot, characters, theme
  • Adornment (form) = visual style, music style, editing style, performance style
  • The pseudo-refinement trap: the more gorgeous the visuals and the emptier the story, the more form fails; form must provide support for content
  • The most common mistake in AI creation: piling on special effects but lacking a solid storyline, reasonable visual continuity, and matching music and sound rhythm

The creator's three levels of cognition

  1. Low eye, low hand: cannot tell good from bad, and cannot make it (beginner, can only do elementary execution)
  2. High eye, low hand: can tell good from bad, but still cannot make it (the key stage of progress—wash the eyes first, then train the hand)
  3. High eye, high hand: both aesthetic judgment and creative execution (the ultimate goal)

Core conclusion: to make good work, first watch good work. AI creation = using AI tools to recreate the aesthetic accumulation from live-action/animated works.

6.2 The seven major traditions of film and television aesthetics

Realism / Formalism / Classicism / Expressionism / Surrealism / Poetic aesthetics / Epic aesthetics

6.3 Imagery system design

  • Visual motif: a visual element that recurs throughout the entire film
  • The only highlight zone: the core visual anchor that makes the audience's gaze follow throughout the entire film
  • Imagery closure: the opening and ending use different scales/states of the same image to form a payoff (especially applicable to grand scenes: a giant object in a wide shot at the opening → a fragment of the same giant object in a close-up at the ending)

6.4 Dual-track rhythm system

External plot rhythm(loose/medium/tight) + Internal Emotional Rhythm(Light/Medium/Heavy)

Misalignment between the two tracks is an advanced narrative technique—misalignment of loose plot + heavy emotion (the calmer the visuals, the heavier the emotion).

6.5 The Director's Three Questions

  1. What do the audience know that the characters do not? → Suspense/information asymmetry → Determines camera perspective
  2. What emotion do you want the audience to feel? → Emotional goal → Visual means → Determines rhythm, lighting, and shadow
  3. What does cinematic feel come from? → Pure cinematic expression distinct from novels → Determines camera movement/composition/sound

Seven, Industrialized Workflow

7.1 Three-Stage Structure

Pre-production 30%: Concept → Storyboard → Material preparation (character design sheets/scene images/keyframes/style references) Mid-production 40%: AI-generated visuals → Manual iteration Post-production 30%: Editing and compositing → Music and sound effects → Final cut

7.2 Three Rounds of Iteration

Round 1 sets the style → Round 2 adjusts composition → Round 3 refines materials, lighting, and shadow. Do not fuss over details from the start.

7.3 Pitfall Avoidance Checklist

  • Resolution Trap: Image quality is determined by the prompt; resolution only changes the canvas size
  • Macro Before Micro: Minimal prompts set the direction → Gradually refine
  • Single-Variable Debugging: Change only one variable at a time, so the source of the effect can be traced
  • Reference Image Strategy: Break down elements (composition/lighting/color/material/atmosphere), do not copy directly
  • Forcing Duration: Do not stretch an explosion into 12s just to fill 90s; content volume determines duration (user correction record)
  • Forcing Segment Count: Do not force 6 segments per episode; let the plot drive it (user correction record)
  • Feeding Storyboards as Tables: What the user wants is generated segment-style narration, not tables (user correction record)
  • Missing Performance: Do not write "happy/sad"; use micro-expressions + physiological stress responses (Xiaoluo Six Dimensions + Old Li Micro-expression Library)
  • Painting Grand Scenes Small: Do not write empty words like "many people/big scene"; use proportional anchors + focal length + layered depth (§4.5)

Eight, Output Quality Gates (Must Pass Before All Outputs)

This set of gates is thesole authoritative version(2026-08-09 fix: unified with the gates embedded in each step of §2, eliminating dual-version desynchronization).

Gate A: Script Gate (After ① Is Completed)

  • Is the duration and segment structure of each episode clear? Is duration determined by content volume, not forced?
  • Does each segment carry 1 narrative function?
  • Is the causal chain complete: does each scene have a goal/obstacle/action/change/exit hook?
  • Does the twist have prior information? No last-minute cheating?
  • Is the dialogue colloquial? ≤40 characters/sentence? Fits identity and purpose?
  • No psychological description? No parenthetical hints? No preaching? No abstract emotion words?
  • Does the episode ending have a suspense hook (small question answered + bigger question raised)?
  • Do action and camera descriptions use characters' formal names, with no pronouns?
  • (When using the §2.1B dual-track script) Are bullet lines ≤15 characters? Are foreshadowing layers registered (short-range/mid-range/ultimate)? Is the close-up rate ≥40%? Is the word count within the hard targets (60s=250-300 characters/120s=500-600 characters)?

Gate B: Storyboard Gate (After ② Is Completed)

  • Does the first shot start at 0s, and the last shot end exactly at the total duration?
  • Do adjacent shots connect seamlessly at head and tail? No overlap/skipped seconds/blank gaps? Is duration determined by content volume?
  • Is the shot count within the soft reference (18-26 shots)? Is each shot's duration reasonable (estimated by information volume)?
  • Is the action-reaction structure correct? Are hard cuts used at key nodes?
  • Are reaction shots non-recursive? Are consecutive reaction shots not mechanically reused?
  • Do adjacent shots avoid the same shot size/camera position/angle/composition?
  • Is there a macro visual gear shift every 15-30s?
  • Are character/scene/prop numbers consistent across shots? Is the state continuous?
  • Does the image description contain no pronouns? No abstract emotion words? No shot numbers?
  • Do ordinary character shots have ≥3 dimensions, and key close-ups/insert shots have ≥5 dimensions of performance?
  • Can the dialogue be spoken naturally within the corresponding duration?
  • Grand scene: are shot size + focal length + camera movement + character proportion explicitly marked (§4.5)?
  • Does the Chinese output end with the last line "无字幕,无背景音乐"?
  • Zero-loss cutting: 100% restoration of the script? No brain-completion condensation?
  • Does every shot have a narrative purpose (seven categories)?
  • Dynamic scene: do position changes have运动 causality (attack→force→movement→new position)? Does it inherit the previous shot's spatial state across shots?(spatial reference system)

Gate C: Asset Prompt Gate (after the衔接 layer is completed)

  • Three-view: orthographic projection? No perspective? A-Pose? Immutable DNA locked? Left and right not mirrored?
  • Expression settings: twelve-grid? Same character? Are the physical differences of the four easily confused expressions clear?
  • Scene: are spatial structure/materials/light and shadow/style complete? Is the focal length plan clear?
  • Storyboard panels: one per scene? Continuous narrative? Complementary shot sizes? Does the shot language coverage meet the standard? Are characters/scenes/style locked?
  • Grand scene: shot size + focal length + character proportion (§4.5)?

Gate D: Prompt Gate (after ③ is completed)

  • Is each segment a complete five-part format (style/color card reference/shot/timeline/sound)? (Only >60s long content uses the six-part format: +[Characters and Scenes Appearing])

  • For segments >60s: are the characters and scenes appearing clearly written (@asset slots)? For segments ≤60s: is this module absent (returning to the V3-tested five-part format)?

  • Is the color card reference located after [Style] and before [Shot]? Are all four colors present—primary hue + auxiliary color + highlight color + shadow color—with HEX+RGB attached?

  • Is the timeline narrated continuously? Do segments connect seamlessly? Is the total duration referenced to the target (suggestive, flexible)?

  • Embedded camera movement (V4.0.6): does the camera movement at each time point evolve along the emotional arc (establish→introduce→focus→explode→converge)? Does the camera movement have a complete trajectory + shot size changes? Does the visual scheduling have a large-medium-small-large hierarchy or scale-contrast ending?

  • Does the opening explode directly? No explanation of its origin?

  • Action + physical feedback + parenthetical sound effects embedded in the description?

  • Zero psychological description? Only write what can be seen and heard? No abstract emotion words?

  • Can the dialogue be finished at a natural speaking rate? Character name: "dialogue" format?

  • Dialogue scene duration: do not cut the shot before the dialogue is finished? Has a single shot of 8-12 characters been stretched to 2.5-4s? Have 15+ characters been split into continuous time points with continuous dialogue?

  • Grand scene: shot size + focal length + character proportion (§4.5)?

  • Blocking anchoring: Does the first sentence of each shot clearly state the characters' spatial relationships (orientation/distance/facing/foreground-background)? Are movement trajectories written out for position changes across time points?(Added from user feedback on 2026-08-11)

  • Gaze anchoring: In dialogue/opposing scenes, does each shot clearly state who looks at whom, where they look, and when they look away? Are the listener's gaze constraints specified?(Added from user feedback on 2026-08-11)

  • Fight segments (§4.6.8): Is the action power chain complete (intent → wind-up → center of gravity → force generation → contact → feedback → recovery → next-panel inheritance)? Are offense and defense paired (every attack must have defense/dodge/block/counter)? No teleportation/no unmotivated side-switching/clipping/spatial resets? Does the P15-type ending avoid a static pose?(Added when the martial arts system was incorporated on 2026-08-16)

  • Independent generation by segment: Is each segment's [style] [color card reference] [shot] fully self-contained? No inherited phrasing such as "continue/use the same as above/previous segment"? Are the color cards for the same scene verbatim identical across segments?(Added with user confirmation on 2026-08-11)

  • Is each segment primarily in pure Chinese? Are there 2-4 English tags per shot and are they compatible?

  • Are the sixfold consistency locks (assets/timeline/space/lighting/style/final-state inheritance) satisfied?

  • Are segments marked with hard cuts or is the timeline continuous?

  • Is there an ending visual closure (imagery callback)?

  • Does the Chinese output end with the line "无字幕,无背景音乐"?

Gate E: Grand-scene gate (Assessment Standard ③ · must pass whenever a grand scene is encountered)

  • Is the shot size explicitly labeled (extreme long shot/extreme wide shot/medium wide shot...)?
  • Is the focal length explicitly labeled in millimeters (16-24/24-35/35-50/85+)?
  • Is the camera movement explicitly labeled (aerial/dolly/crane/truck-in...)?
  • Is the scale difference between characters and the huge environment/giant object clearly written?
  • Are there three layers of depth (foreground/midground/background)?
  • Are atmospheric perspective/particle density/light direction clear?
  • Are there no unrenderable quantity words such as "tens of thousands" or "many people"?

Nine, Open-source Cinematic Prompt Methodology (HELL GRIND · Added in V3.1)

Source: Higgsfield open-source AI film "HELL GRIND" (95 minutes, all prompts/assets open-sourced, incorporated on 2026-08-07). User-validated professional prompts—classified and incorporated by reusable methodology. Full classified methodology text:references/HELL-GRIND-提示词分类方法论.md(Since V4.1.3, this is the only retained item; the full 252KB original set has been removed as requested). Usage:For long-segment/multi-shot segment prompts, directly apply this chapter's skeleton; for short segments, still use the five-part format in §2.4 (≤60s); only long content >60s is upgraded to the six-part format.

9.1 Prompt Framework—Top-level structure for long content

Long prompts are organized in this order, with none omitted (for detailed explanation, see Methodology File §1):

SCENE CONTEXT        场景上下文(承接上一镜,1-2句)
STRICT / WORLD LOCK  铁律区(本段所有硬约束集中声明)
ACTIVE REFERENCES    本段激活的资产引用清单(角色/场景/道具 @引用)
LOCATION MAP         空间布局锚点(方位/距离/几何关系)
FIRST FRAME          第0帧画面状态(开局即定)
FORMAT MODE          格式(时长/画幅/帧率/慢动作策略)
OPTICS               光学(胶片/镜头/颗粒/快门)
CAMERA               摄影机(运镜/机位/性格)
ACTION TIMING        动作时间轴(逐beat秒级编排)
PHYSICS              物理(重量/惯性/接触阴影)
LIGHTING             光线(光源/色温/方向)
AUDIO                声音(环境SFX/台词/音乐规则)
CHARACTER ACTING     表演(微表情/反应节奏)
STYLE                风格(导演风格转述/质感)
QUALITY              质量(8K/毛孔级皮肤/实拍感)
POSITIVE CONSTRAINTS 正约束(明确"画面里必须有"什么)

Key design: The skeleton is ordered by "AI consumption priority" (iron rules/space first, action in the middle, style and quality at the end); each field is written only once, and the action section references previously locked content ("Camera rules continue"); the ending always writes References = visual identity / costume / environment matching only. Generate 100% original continuous live-action cinematography from scratch.(reference images are used only for identity matching; dynamic footage is 100% original).

9.2 Asset Reference System (Foundation of Feature-length Consistency)

  • Asset numbering:<<<uuid>>> + [imageN] dual identifiers, with textual descriptions as fallback
  • Create assets by state: the same character is split by state (normal/wet/injured), and only the needed states are activated in the same segment
  • IDENTITY SWITCH: explicitly declare during transformation/injury <<<A>>> → <<<B>>>, from a certain frame onward use only B — eliminate mid-way drift
  • Reference image annotation purpose:CHARACTER REF — appearance only / SCENE REF — space only / ELEMENT REF — do not copy framing; scenes are split into two: FRONT SIDE / BACK SIDE
  • Environment reference images only inherit style/color tone/atmosphere; copying composition and camera position is prohibited

9.3 Unified atmosphere directive (the killer move for cross-shot consistency)

Put the atmosphere parameters that run through all shotsinto a separate block, requiring each segment to be pasted verbatim— this is the core secret to a feature film not drifting. Locked parameters: color temperature (e.g., 4500K cold overcast day) / snowstorm level (100-200 snowflakes per frame) / visibility (30-50 meters) / color palette (60:30:10) / lighting (overcast cold diffuse light) / horizon (no visible horizon) / wind direction.

Infinite legion mode(battlefield type): every 2-3 seconds 3-5 new enemies rush out of the fog; underground regeneration 5-8 times per shot; at frame 0 already at the center of the battlefield (no start transition); background enemy motion vectors always point toward the protagonist; for epic scale write "a single frame of 500-1000 people implies 10000+" not "tens of thousands"; for crowds write "a mix of 3 styles" not clones.

9.4 How to write the ironclad rules zone (STRICT / HARD LOCK)

  • Positive constraints: write "the frame must contain" —EXACT 3 CHARACTERS / No duplicates / 第0帧长矛必须在骑兵手中 / 盾牌开拍前已长好
  • Hard lock list: slow motion / position anchors / character anatomy / action choreography / hard cut transitions / lighting, each as a separate line ⚠️ XXX (HARD LOCK)
  • Geometry lock: compass orientation (NORTH=skylight wall...) + locked distance (the two people EXACTLY 2.5 meters apart) + anchors do not move (the pillar is fixed and does not follow) + scale anchors (the monster is 360cm, its head is twice the height of a person)

9.5 Action choreography (second-level beat writing)

  • Single-shot beat:动作(0-2秒): [核心动作] + [SLOW-MOTION段落,放慢4-6倍] + 慢动作细节 + [回到正常速度]
  • Frame rate ramp(within one shot):60fps → 240fps(重慢动作)→ ⚡10000fps(子弹时间,只给3个峰值瞬间)— the ramp is all done in-camera, absolutely not an edit cut
  • Multi-shot segment:multishot sequence — 15 SECONDS · 21:9 · 8K · 60fps · hard cuts only · exactly 6 numbered shots; between each shot HARD CUT — [大气参数锁定摘要](condense the parameters that must continue into one line to prevent drift)
  • Multi-character timing: each beat keeps only one main character's action, other characters only have their presence status written

9.6 Combat scene special section

  • Unified death rules: undead-type death = black dust mist/black sand;切开即崩解,两半分离后变黑色尘雾爆发
  • Organic weapon vs armed:整条手臂变成水晶武器——不是握在手中的武器 / 骨刃是身体的一部分,没有手没有握柄(write into AVOID: do not draw hands)
  • Combat camera personality:非常活跃手持剧烈摇晃反应性相机,像战地摄影师冲入战场; reactive jolt; insert a dynamic CU every 0.5-1 second (0.2-0.5 seconds), connected by whip pan
  • One continuous shot (ONER):SINGLE CONTINUOUS HANDHELD ONER, NO editorial cuts; whip-pan is in-camera movement not an edit; in PHASE A mid-section opening → B block and counterattack → C chain-kill frenzy → D continuous combat with no ending; horse rules written in hard (马永不停留/永不撞盾/从旁全速掠过); the ending does not resolve (take ends mid-action)
  • Stamina fatigue progression: for long battles write the protagonist's fatigue (white breath/breathing, turning slower, one blade grazes the shoulder armor) — a sense of real exertion

9.7 Performance and emotion (SOLO PORTRAIT template)

  • Emotional arc timeline(10-second single-person close-up): 0-3s continuous realization (eyes move down/pupils dilate/shallow shocked breathing) → 3-6s realization deepens into sorrow (inner brows lift/lips pressed tight/a tear slides over the bloodstain) → 6-10s sorrow turns to fury (eyes lift/brows press down/jaw clenched tight/slight facial trembling/temple veins bulge)
  • Performance hard rules: Leave 1 second of silence before and after the line (first show the face, then speak; after finishing, continue to hold); never visualize the line content (no flashbacks, no phantoms); prioritize dialogue in the mix (duck ambient sound to a low volume while speaking); act only the scripted lines (no improvisation, no mumbling, no voice-over; if there is no line, it is silence)
  • pore-level skin:pore-level realism — vellus hair(毳毛), asymmetric moles(不对称痣), capillary flush(毛细血管泛红), pore-shadow matching on-set light

9.8 Environmental Shot Template

  • City Aerial Empty Shot(8 seconds): complete parameter lock (600 m altitude + 15-degree forward tilt + horizon y=35% + 2.39:1 + constant-speed rightward movement for the entire segment); motion hierarchy (camera constant-speed lateral move / diagonal rain in the environment / two cars as subjects / lightning within the clouds); lightning timetable (0.8s + 3.2s + 6.1s, each followed several seconds later by muffled thunder); color palette 60:30:10; continuity lock (always only two cars / elevated subway empty / clouds never disperse)
  • Natural Intimate Empty Shot(5 seconds): starts as a handheld medium shot -> slowly and steadily pulls back -> the two people become smaller and merge into the vast nature; the negative list is extremely detailed (no music, no kissing, no slow motion, no other people, no beards, no dialogue, no text)

9.9 Image Editing Prompts (Damage Correction Template)

  • Structure:Edit this image. [总目标] Apply these corrections consistently across all views. -> CHANGE 1 deletion category / CHANGE 2 reduction category / PRESERVE retention category / CHANGE 3 proportion category / CHANGE 4 readability -> Final goal: [最终画面一句话]
  • Principles: write the three categories of deletion/reduction/retention separately; for reduction, provide a "complete list" (这就是面部损伤的完整清单) to prevent the AI from adding extra drama; for the negative list, provide specific reference objects (NOT horror movie victim); write retention items as fixed (identity/structure/body type/hairstyle/clothing/pose/layout/background)

9.10 Unified Style Prefix + Director Style Paraphrase Table

Reusable base prefix for each segment (isomorphic to §1.7 base-atmosphere-image):Style: 8K IMAX. Photorealistic — no 3D render, no game engine → Lighting: Natural light only — contre-jour backlight, camera on shadow side, atmospheric haze → Color: 60:30:10 → Camera: Physical cine lens. 180° shutter motion blur → Skin: Pore-level realism → Acting: Hollywood — micro-pauses → Physics: Gravity and inertia respected → Composition: Rule of thirds + golden ratio → Continuity: No identity drift → Technical: 24fps, 8K → Audio: Environmental SFX only. No music. No subtitles.

Director Style Paraphrase Table(actually used in HELL GRIND, following the §0.3 red line of not writing names):

Original reference Prompt paraphrase wording
Lubezki × Deakins natural light + backlight (contre-jour) + atmospheric haze + handheld
Kurosawa Ran dynamic symmetrical composition + epic battlefield + dramatic weather + low saturation
John Wick / The Raid one-take fight + clean decisive kill beats + camera lives inside the violence
Terrence Malick intimate connection dissolving into vast nature + slow pullback
Cuaron / Villeneuve world density + handheld long take / surveillance-level restraint + low-illumination practical light
anime final battle (Bleach/Demon Slayer) explosive action + keyframe freeze + "manga still frame" beats

9.11 Project-Level Engineering Secrets (Summary)

  1. Consistency does not rely on "remembering," it relies on "copying" - the unified atmosphere directive / infinite legion directive / frame 0 directive must be pasted verbatim in every segment
  2. Assets are the lifeblood of a feature - each character and each state needs an independent asset; use IDENTITY SWITCH for transformations
  3. Positive constraints are better than negative constraints - "EXACT 3 CHARACTERS" is more effective than "do not have a 4th person"
  4. Geometry locked down - compass + distance + anchor points, so space does not drift
  5. Hard cut = state summary - HARD CUT is followed by a one-line atmospheric parameter summary
  6. Feature = multi-segment skeleton - each segment has its own complete skeleton, and segments are connected by "Scene context: continue from the previous shot"
  7. End with a negative list - at the end of each segment, list AVOID/NEGATIVES, covering all the points most likely to go wrong

9.12 CINEDANCE Director System (V3.1 upgrade · see references for the detailed version)

📄 Full version:references/CINEDANCE-HIGGSFIELD-提示词导演系统.md. Positioning: the "master template" of the §9.1 skeleton (4-D methodology + silent pre-output QA, suited to high-demand scenarios where "first generation must be usable"). Key points for quick reference below:

4-D internal methodology: D1 Deconstruct (keep only what is visible and audible in this shot, delete stale @ tags/previous-scene wording) → D2 Diagnose (pre-diagnose failure risks such as empty first frame/late character arrival/gaze reversal/prop held in wrong hand, and add a short lock for each) → D3 Develop (order: scene context → output settings → activate references → location map → first-frame placeholder → spatial blocking → character anchors → format → optical lens → camera → action timing → physical materials → light exposure → audio → positive locks → local failure locks;spatial rules before lens style, light is a priority lock not decoration) → D4 Deliver (output only the final prompt, without QA/reasoning/notes).

Lens control: field of view angle instead of millimeters(Seedance responds better to observable lens results): 107° wide-angle close to face/84° classic wide-angle/47° standard human eye/29° short telephoto medium shot/18° classic telephoto tight close-up/8° super telephoto surveillance (must add 30-45% foreground occlusion). Face portrait 84→29→18; environmental action 47→84→107; distant observation 8°.mixing content types within the same beat = lens drift。

telephoto/wide-angle observable result stack(≥4/3 items per shot): telephoto = background fully softened/subject sharp/creamy bokeh/background compressed/haze suspended; wide-angle = foreground body presence/environment visible all around/edge-to-edge focus/straight lines stay straight/camera physically close. Anti-drift locks:本镜任何部分不得变成[广角/长焦]。

write the camera as a physical operator behavior:camera fixed at X/lens at hip height/subject occupies screen-left third/profile preferred; for handheld, write physical description (operator breathing/weight shift/organic imperfect correction), forbid digital jitter/random shake/gimbal smoothness.

first-frame placeholder lock:The first visible frame already contains all required characters in their correct positions. No empty establishing frame. No delayed character reveal.(empty opening allowed only when the user explicitly requests it).

spatial blocking lock: for each subject write screen position + world position + distance to landmark + body orientation + gaze direction + movement direction + foreground/midground/background;write gaze and body orientation separately(torso faces X/eyes stay locked on X);landmark proximity lock forbids weak words(forbid near/around/beside, replace with within 1 meter/back against the wall/hand on the handle)。

multi-shot continuity lock: default to a single continuous shot; multiple shots must explicitly define every cut (Shot A duration/camera/first frame/blocking/action + cut type + full Shot B set), never let the model invent cuts; after a cut do not reset action or teleport characters; HARD/SMASH/MATCH/INSERT/REVERSE/WHIP CUT allowed, fade/crossfade/dissolve forbidden.

lighting priority lock: backlight lock Subject stays between camera and the brighter background. Camera stays on the shadow side. Faces remain in deep shadow unless explicitly lit. Only rim light, edge light, wet speculars, eye glints, and environmental bounce reveal detail.; light direction must always be written (key light source/light direction/camera side relative to light/shadow contour side/background brightness/exposure priority/allowed highlights/forbidden light failures).

context isolation: the final prompt is a sealed document of the current shot; forbid scene numbers/episode labels/script headers/previous-scene summaries/unused character or location tags/previously/again/same as before/continues.

reference control hierarchy: identity reference governs face/body/proportions/clothing; location reference governs architecture/materials/geography/light direction (must not override the required camera position); prop reference governs shape/scale/state; style reference must not override identity/blocking/action/optics/lighting.give high detail only where control is needed, low detail for general aesthetic description—improvement comes from stronger signal, not more bloat。

silent pre-output QA (20 items): Are all @ tags actually used? Deleted stale @ tags? Correct first frame? Is each character's position/line of sight/body orientation clear? Are landmarks physically anchored at close range? Is the camera side clear? Does the lens character match the content type? Is the camera drift-proof? Is the lighting flat-proof? Are props in the correct hands? Are the actions physically feasible? Are the time blocks consistent? Is the dialogue clean with only lines? Avoid scene numbers and context leakage? Is QA hidden and not exposed?

9.13 LIRA Image Asset System (New in V3.1 · see references for the detailed version)

📄 Full version:references/LIRA-SKILL-图像提示词优化.md. Positioning: advanced specification for bridging-layer asset prompts (three-view/scene/prop/editing).

Model routing: character sheets/portraits/consistency → Soul 2.0 (Soul ID face lock); scenes/environments/cinematic stills → Soul Cinema (21:9); prop sheets → NBP/GPT Image 2;Frame editing always chooses NBP first(minimal changes + exhaustive preservation) → Seedream 4.5 (texture channel only) → GPT Image 2 as last resort. Scenes that need rebuilding go back to Soul for regeneration. Aspect ratio/resolution = platform UI parameters, not included in prompt text; no model has negative prompt parameters, everything relies on positive description.

10 anti-failure iron rules(5 core ones): ① natural prose, not keyword stacking (do not use ALL-CAPS section headers, CAPS only for editing CHANGE/PRESERVE); ② no bloat (80-150 words optimal); ③positive > negative(generation prompts must never write "what not to do"; NOT stacking injects the concept itself; in editing, "remove" is only legal and must be paired with "fill"); ④ technical light and materials, not vague atmosphere (single overhead key light, soft 2:1 ratio is better than dramatic cinematic lighting;board-formed concrete is better than "concrete"); ⑤ character consistency = Soul ID, not prose; ⑥ prevent illustration drift (write studio photographs/film character sheet/cinematic film still, forbid character reference sheet/painterly); ⑦ text gets exact copy + font weight and color; tattoos get specific designs; never write real names/IP/brands; ⑧ 60/30/10 palette derived from instructions/context/references, never invented out of thin air; ⑨ editing NBP first + minimal CHANGE + exhaustive PRESERVE; ⑩ no bloat ≤1500-2000 characters.

Technical block (camera + film): film stock Photorealistic ARRI Alexa LF anamorphic Cooke S4 lens at T2.0, organic 35mm Kodak Vision3 250D film grain, soft cinematic falloff; modern digital stock Shot on ARRI Alexa Mini LF with ARRI Signature Prime lens, clean modern digital cinematic capture. Palette wrapper:Refined desaturated palette: [cool] dominating, [warm] as only warm contrast, deep crushed blacks(remove "painterly" for photorealistic character work).

9.14 Emotion-driven cinematic prompts (New in V3.2 · source Quill free version)

📄 Full version see Obsidian 07-课件内容/AI导演思维10天课程/提炼内容/. Reverse-derive camera angle/focal length/composition/depth of field/lighting/color/material/atmosphere/film texture from emotion/narrative intent/audience psychology.Never stack terms as generic quality boosters。

First think through six questions: audience position (victim/witness/pursuer/God's-eye view/intimate participant) → power relationship (who dominates whom) → spatial psychology (open/closed/compressed/empty/maze/vertical/collapsed) → narrative focus (person/threat/object/environment/light/absence) → visual temperature (alive/warm/sterile/dead/sacred/polluted) → degree of realism (documentary/mythic reality/disaster reality/dream logic).

Emotion → visual decision table quick reference: Fear/oppression = low-angle shot + telephoto compression + side backlight through dust + low-saturation cool tone; giant object making people tiny = eye level + person pressed to bottom edge + giant object clearer than person; awe = low-angle frontal symmetry + small figure + monument as main subject + realistic atmospheric backlight; claustrophobia with no way out = telephoto flattening/tight wide-angle distortion + remove sky and horizon + low brightness + no light in exit direction; loneliness and abandonment = distant static observation + large amount of negative space + cold gray weak light; power and authority = low angle or eye-level symmetry + centered vertical axis + hard side light with controlled shadows; chase panic = handheld shoulder-height tracking + tilted lines/partial occlusion/foreground blur + practical light strong contrast; intimacy and vulnerability = close eye level/over-the-shoulder + standard to short telephoto + shallow focus locking eyes/hands + warm skin soft contrast.

Shot shorthand: Low angle = power shifts upward (weakness/oppression); high angle = vulnerability/being watched; eye level = documentary presence; wide angle = approach/distortion/panic; telephoto = compressed depth/inescapable; shallow depth of field = locks narrative focus; deep focus = all relationships readable.

: ① Certainty—do not write or/or/slash options, choose only one for each variable; ② Default cinema camera texture—write real cinema camera + film (: ① Certainty—do not write or/or/slash options, choose only one for each variable; ② Default cinema camera texture—write real cinema camera + film (电影机拍摄/35mm 胶片扫描质感/fine grain/soft dynamic range/slight halation), never write only "cinematic feel"; ③ Every choice has an emotional function—break the effect into lens/exposure/grain/highlight/color behavior, do not write abstract praise. When the user provides a field template, strictly preserve the headings/punctuation/order.

9.15 Practical Prompt Breakdown Library (V3.2 supplement · quick-reference version see references)

📄 The core conclusions have been distilled into the quick reference below (nine major categories + eight practical iron rules); the raw data (340-item JSON/selected md) was removed on request on 2026-08-09, methodology points see references/HELL-GRIND-提示词分类方法论.md。

Nine practical categories: ① Establishing environment (segmented uppercase title THE FIELD/THE ALTAR + 60:30:10 + technical parameters, prohibited words listed at the end); ② Character lock multi-shot (EXACT N CHARACTERS locks number of people + <<<uuid>>> references assets and marks usage); ③ Action group scene (Style Prefix fixed opening word-for-word + Camera/Environment as independent paragraphs + ending technical tags); ④ Chinese multi-shot dialogue (the entire scene written in Chinese, Seedance supports; ⚠️ one line of dialogue = one shot, dialogue must not cross shots); ⑤ Romantic close-up (CAMERA/LIGHT/ACTION/SOUND/NEGATIVE sections + ACTION at second-level 0.0s/1.5s/3.0s/4.5s/5.0s); ⑥ Slow-motion multiple characters (CUT tag segmentation + frame rate conversion 96→24fps 4x SLOW-MO + dual annotation of screen time/real time); ⑦ Style reference ((StyleReference) declaration + oil painting terms brush textures/thick paint instead of "oil painting style"); ⑧ Costume editing (CHANGE exactly what to change + PRESERVE lock what stays unchanged + material details + REMOVE explicit deletion items); ⑨ Character description (face + body type + hairstyle + clothing all stated positively, do not write emotion, three views in one frame).

Eight practical iron rules: ① Style Prefix locked word-for-word across the entire project (8K IMAX...60:30:10...180° shutter...); ② Asset references use <<<uuid>>>/<<<image_N>>> and mark usage (CHARACTER REF/SPATIAL ref/StyleReference); ③ Chinese prompts are feasible (Scene 14 entire scene in Chinese); ④ Multi-shot = [Shot N] segmentation, each shot has independent camera position/shot size/action/dialogue; ⑤ Second-level action choreography (within [ACTION] 0.0s/1.5s/3.0s/4.5s/5.0s broken down second by second); ⑥ Negative list [NEGATIVE] locks details (no music/no slow motion/no extra fingers); ⑦ Editing prompt = CHANGE+PRESERVE; ⑧ Establishing shots use uppercase segmented titles.

Chinese multi-shot dialogue template structure(Scene 14 evidence): asset references (<<<image_1>>> Role/Location, marked MAIN VIEW/REVERSE VIEW/layout) → Character positions (A stands at X facing Y, B is 1.5 meters directly behind A) → ⚠️Rules upfront (screen lighting rules/dialogue rules) → [Shot N] each shot's camera position + focal length + shot size + duration + action + dialogue → Ending Style section unified style (8K IMAX film/lighting only from actual light sources within the scene/60:30:10/brief pause before performance reaction/physical gravity and inertia/60fps/ambient SFX only/no music no subtitles/duration and aspect ratio).

X. Usage Recommendations

Complete Workflow

剧本大纲 → ①详细剧本 → ②分集分镜脚本 → 衔接层(三视图/表情/场景/剧情分镜格) → ③每集分镜提示词 → 生成验证

①Detailed Script: XiaoLuo screenwriting creative logic (stage routing/scene dramatic units/conflict escalation/five-dimensional micro-performance/hard rules for character names) → episode planning → segmented structure → asset registration ②Episode Storyboard Script: XiaoLuo timeline (0-90s continuous/18-26 shots/action-reaction hard cut/macro visual gear shifting/six-dimensional performance/asset consistency/hard validation) + Shanyin supplements (shot groups/action-reaction variants/information-density duration) Transition Layer: Character three-view prompt (XiaoLuo orthographic standard) → character expression setting prompt (twelve-grid) → scene prompt (focal length system) → plot storyboard grid prompt (one per scene) ③Storyboard Prompts for Each Episode: Output pure Chinese prompts according to the generation segment format (≤60s five-segment format / >60s six-segment format) → self-check gates D+E Generation Validation: AI video generation → character skeletal consistency check → prop continuity check → lighting source check → inter-segment transition check (assessment standard ①)

Model Selection

  • Extreme close-up/macro shots → Kling 1.6 or Runway Gen-4
  • Subtle character performance (micro-expressions/physiological stress) → Pika 2.0 or Kling high-detail mode
  • Particle-effect-intensive scenes → Kling 1.6 (particle physics consistency)
  • Orbiting camera movement + physical interaction → Kling keyframe first/last frame function
  • Grand scenes/large long shots/aerial shots → can first output a still keyframe to verify proportions before generating motion (prevent giant objects from being drawn small)

Cross-segment Generation Transitions

  • The first/last frame solution has been abandoned (V4.0.2): no longer use the previous segment's last frame as the next segment's first frame for direct stitching—AI has no long-term temporal memory, dynamic information resets to zero every 15s, and direct stitching will inevitably cause jump cuts.
  • Correct approach: before generating each segmentdraw a storyboard reference image(fix the character pose/stance/interaction at the end of the previous segment) → prompts synchronously update changed data (stance/pose/composition) → stitching nodeschange camera position(cut in/reverse shot/change shot size)
  • Unchanged data (character asset images/scene images) is fixed and imported throughout; variable data (stance/pose/interaction/composition) is synchronously updated for each segment
  • Hard-cut shots are recommended to be generated separately and then edited together
  • If lighting does not change, lock the light source direction description during generation
  • Fourfold inter-segment transition rules (assets/action/emotion/sound) checked segment by segment + final-state inheritance lock check (see §3.7)

XI. Version Record

  • V5.0.2(2026-08-20):Move library completion + combat-specific priority statement + version number alignment——①references/Martial Arts Move Library - Classic Moves Quick Reference.md newly addedTachi · Dream of Oneness (12 forms)、Tachi · Crimson Snow Cherry Blaze Style (12 forms)、Advanced Qinggong (25 selected forms)、Advanced Immortal Cultivation Sword Moves (25 Selected Forms) Four Major Systems (Source: Desktop Martial Arts skill.md Positioning Quick Reference + Camera Iron Rules), each system includes a move table + camera iron rules + prohibited items; ②§4.6.9 ⑨Weapon rule reference correction: removed the reference to the nonexistent "corresponding specialty library (sword techniques/scythe/tachi/lightness skill/immortal cultivation sword moves)", changed to point to the corresponding specialty in references/武学招式库-经典招式速查.md; ③§4.6.9 ⑪Move library list synchronized with added new systems; ④§4.6 master-subordinate declaration newly addedspecialty takes precedence over generalRuling: when a fight segment conflicts with §2.4 general rules (e.g., one shot per segment vs. fight camera movement richness), §4.6.8/4.6.9 shall prevail, the five-segment skeleton remains unchanged; ⑤Version number alignment: title/frontmatter updated from V5.0 to V5.0.2(V5.0.1 content was already in the main text, this time unified identification); ⑥§2.4 five-segment complete example title [Timeline] changed to [Suggested Timeline], aligned with V4.0.9 rule naming.

  • V5.0.1(2026-08-16):Character asset refinement + inter-segment transition systematization(User feedback: ①Character assets need more detail—add facial feature descriptions, clothing material details; ②each segment transition should not use first/last frames, but rather cut shots/change camera positions/change shot sizes)—①§2.3 three-view prompt format newly addedfacial feature detailsmodule (face shape/brow shape/eye shape/pupil color/nose shape/lip shape/jawline/ear shape/bone structure/facial features, ten items mandatory) +clothing material detailsmodule (fabric/texture/craftsmanship/layer structure/color zoning/aging state, six items mandatory), to prevent facial feature drift and "clothes looking like plastic"; ②§3.7 newly addedinter-segment transition three methods six forms: three methods = change camera position (basic mandatory)/cut shot size/cross-segment reaction, six forms = action continuity + change camera position/reaction shot transition/shot size jump cut/sound lead-in/prop detail transition/empty shot transition; splice node design rules = segment end leaves action in intermediate state + segment start enters with changed camera position + prohibit same camera position continuation/last frame as first frame/static to static; ③At the storyboard script stage, each splice node must be annotated with its transition method.

  • V5.0(2026-08-16):Major version number upgrade + full logic check(Continuing V4.1.1-V4.1.3 fight system trilogy and slimming)—①Major version number V4.1.3 → V5.0: the martial arts system (§4.6.8 PREVIS action correctness + §4.6.9 anime fight excitement + ⑪martial arts move library) upgraded from "patch level" to "major version level" capability, frontmatter/title/directory name synchronized (manju-laoli-v5.0); ②full logic check: chapter numbering complete (0.1-9.15), two gate D copies consistent, references citation cleanup (the only missing full-collection reference has been fixed), §0.5 self-check consistent with §2.4 five-segment/six-segment rules; ③HELL-GRIND original full collection (252KB) slimmed down to retain only the methodology (V4.1.3), entire skill reduced from 878KB to 626KB.

  • V4.1.3(2026-08-16):HELL-GRIND original full collection removed(User request: keep only the core framework, no need to fully integrate)—①references/HELL-GRIND-开源95分钟电影提示词全集.md (252.7KB original text) moved to recycle bin,only retain references/HELL-GRIND-提示词分类方法论.md (16.8KB methodology distillation, including skeleton/assets/iron rules/action/editing/style full classification); ②§9 introduction citation synchronized to point to the methodology file; ③methodology file header annotated that the full collection has been removed.

  • V4.1.2(2026-08-16):martial arts move library added to the repository(User request: select classic moves from the martial arts skill move library into the skill) — ① Add references/武学招式库-经典招式速查.md: a curated selection of 12 sets of high-frequency martial arts for comic dramas (Wing Chun/Baji/Taiji/Bagua/Xingyi/Tan Tui/Choy Li Fut/Capoeira/Taekwondo/Judo/Jujutsu grappling/footwork/weapons), each set containing a one-sentence positioning + key points of core move actions + typical combos, with Choy Li Fut and Tan Tui given execution-level descriptions; ② §4.6.9 add ⑪ Martial Arts Move LibraryUsage pointers: write move names directly into prompts, use typical combos for move pairing, must pass the power chain + serialization double check, zero repetition within the same fight, special effects subordinate to the main color.

  • V4.1.1(2026-08-16):Combat System Reinforcement · Establishment of Principal-Subordinate Relationship(User feedback: the original version 4.6 martial arts provisions are weaker than the martial arts skill level) — ① At the top of §4.6 addPrincipal-Subordinate Relationship Statement: 4.6.1-4.6.7 are downgraded to supplementary components (transformation/destruction/materials/multi-person constraints); fight actions and prompts shall all be based on §4.6.8+§4.6.9 and the two references files; ② §4.6.9 add ⑧ Character Combat Structure Allocation(main combat 70%/counter 20-30%/1 finisher/distance coverage/action dimension/action signature/special effect personality/character card process),⑨ Weapon Rules(no 90° side shots/no full-duration close-ups/no large arcs in close shots/weapon collisions must have physical contact/prompt conversion formula/read the specialized library first),⑩ Finisher Material Quick Reference(breathing frames choose 1 of 8/spatial changes ≥2 types/two types of emotional turning points/three acts of the finisher); ③ references/二次元打戏提示词体系 add three chapters: character combat structure allocation, weapon invocation rules (camera ironclad rules table + 9 taboos + formula + invocation priority), emotional turning points/breathing frames/spatial changes/finisher material library (suppression 4 types + reversal 14 types + breathing frames 8 kinds + spatial changes 20+ types + three-act structure).

  • V4.1(2026-08-16):Martial Arts System Dual-Pack Added to Library(User named: comic drama Old Li most lacks a combat system; source desktop 《职业学校推荐skill/》) — ① §4.6 add 4.6.8 Martial Arts Action System(industrial-grade 15-cell martial arts PREVIS system): action power chain 11 links, 15-cell rhythm structure skeleton, attack-defense game rules, fist-and-foot and weapon rules (six steps of throws/knife and sword collisions must not float), five continuity locks (state inheritance/180° axis/inertia/injury persistence/special effects do not replace contact), action QA quick reference, storyboard-to-Seedance cell-by-cell prompt 14 elements; ② §4.6 add 4.6.9 二次元打戏提示词体系(distilled from the 353KB 二次元打戏 original text): completion rate formula, 15-second skeleton, dual watchability + close-range exchange must be selected, move serialization (no abstract frequency words), five elements of camera specifications + spatial sense ironclad rules (no 90° side shots at medium shot or above), ten commandments of the fiery-direction ironclad rules, 9-character energy quick reference; ③ Gate D two copies add mandatory combat segment check items; ④ full version included in references/武打PREVIS-15格动作预演系统.md(20 chapters) + references/二次元打戏提示词体系.md(distilled version). Division of labor: PREVIS manages action correctness, the 二次元 system manages fiery intensity.

  • V4.0.10(2026-08-15):Segment length "upper-limitization"(User feedback: 15s is not mandatory, it's that it must not exceed 15s; some plotlines wrap up in 7s) — ① In three places, §2.1 Episode Planning, §2.4 Core Format Description, and §3.1 Segment-Level Timeline, uniformly change to "Segment length is cut naturally according to plot/scene, ranging from 3s to 15s; 15s is the upper limit for a single segment, not a fixed value; once a scene is finished, it ends (can wrap up at 7s/10s/12s), no forcing to fill 15s"; ② At the same time, update [Timeline] in the five-segment module references to [Suggested Timeline].

  • V4.0.9(2026-08-15):Make the generated segment-style timeline "suggested"(User feedback: the seconds we give are sometimes inaccurate; leave room for AI to improvise) — ① In §2.4, change the five-segment [Timeline] module to [Suggested Timeline] (add "approximately" to the seconds, and label it "rhythm suggestion, AI may fine-tune, not a hard command"); ② Add "Suggested Timeline Positioning" to the timeline writing rules: the seconds are a rhythm reference for the AI, may shift forward or backward by 1-3s, and the two hard bottom lines remain unchanged (① Dialogue must be fully spoken, no cutting a sentence in half/lip-sync misalignment; ② The action causal chain must be continuous, no character teleportation/jumps/position changes without cause); ③ In Gate D, change the two instances of "total duration precisely closed" to "total duration reference target (suggestive, flexible)". The storyboard script (§2.2) timeline still retains precise requirements (for internal breakdown use); only the generated segment-style (§2.4 for feeding AI) is made suggestive.

  • V4.0.8(2026-08-15):Comprehensive inspection and repair(5-round cyclic self-check) — ① Unify dialogue word count: sentence-splitting threshold "15-20 characters" → "15 characters" (6 places globally, aligned with bullet dialogue ≤15 characters); ② §0.3/§2.1B clarify the two standards: "static comic drama dual-track script ≤15 characters/sentence" vs "AI video dynamic ≤40 characters/sentence + duration-based sentence splitting"; ③ Gate C deletes the misplaced stance/line-of-sight anchoring check items (the rule belongs to §2.4 generated segment-style and should only be in Gate D); ④ Gate A (§2.1+§8) adds §2.1B dual-track script check items (bullet dialogue ≤15 characters/foreshadowing layered registration/close-up rate ≥40%/hard word-count metrics); ⑤ §2.1B adds a distinction explanation of "panel vs shot" (static comic drama uses panels, AI video dynamic uses shots); ⑥ Version number V4.1→V4.0.1b, bottom version records rearranged (V3.0→V2.2→V2.1→V2→V1 in version/time order); ⑦ "Four-step pipeline" corrected to "Three-step pipeline + connection layer" (4 places; the connection layer has no number and does not count as a step).

  • V4.0.7(2026-08-15):Absorb the short-drama industrialized reference script secondary writing + spatial reference system— ① §2.1 adds 2.1B Script Secondary Writing Capability(short-drama industrialization reference): volume matrix (Track A seasonal broadcast/Track B long comic drama), single-episode second-level structure (3-second strong hook/30-second micro hook/end-of-episode hook), five-stage workflow (topic selection anti-vulgarity self-review → audiovisual character design → outline shaping → writing → script doctor), dual-track finished-film script format + bullet dialogue ≤15 characters + close-up rate ≥40% + hard word-count metrics, diagnosis → rewrite → What-if exploration + foreshadowing audit, full version (25 files) in references/短剧工业化参考/; ② §2.2 adds Spatial Reference System(Spatial Director Converter V3): two spatial modes, static/dialogue type vs dynamic/motion-combat type, dynamic space keyframes (five elements: spatial state + motion direction + path + action causality + new state), iron law of motion causality (attack → force → movement → new position), relationships > absolute coordinates, character position ≠ frame position, the camera does not change the real space, shot continuity (inherit the previous shot's spatial state), full version in references/空间导演转换器-空间参考系统-V3.md; ③ Gate B (§2.2+§8) adds mandatory check items for motion causality/cross-shot spatial inheritance; ④ The integration source line adds "空间导演转换器-空间参考系统V3".

  • V4.0.6(2026-08-12):AI cinematic camera movement embedded library(User-provided excellent example: 15s fantasy opening, proven effective in testing) — ① §2.4 Five-Segment [Timeline] newly addedinline camera movement notation: each time point can be written as [时间码] 运镜:... 画面:..., camera movement evolves along the timeline into an emotional arc (establishment→introduction→focus→outburst→resolution), no longer a single global camera movement; camera movement should be written with full trajectory + shot size changes + camera position switches ("descending vertically from the dome through layers of clouds and mist, switching to a low-angle upward shot at medium shot"); ② newly addedoverall parameters upfrontnotation:【整体参数】时长,画幅,风格,负向提示词 single block at the beginning (for single-segment direct submission); ③ newly addedinformation embeddingnotation (objective visuals carrying narrative hooks: "the crushed dried blood is not hers" "looking at the upper rear behind the camera") andscale contrast ending("she is small, the giant gate is large"); ④ Gate D newly added mandatory checks for inline camera movement/emotional arc/scale ending; ⑤ in the source example, position anchoring (left/right of frame / foreground-background), color card HEX, three sound categories, and dialogue rules are increments of this skill, retained when absorbing.

  • V4.0.5(2026-08-11):Iron rule of independent segment-by-segment generation entered into the repository(user confirmed: when generating each segment separately, do not write inheritance expressions) — ① §2.4 combination consistency newly addediron rule of independent segment-by-segment generation: when generating separately, each segment's prompt must be completely self-contained, with [Style] [Color Card Reference] [Shot] fully written out for each segment,it is forbidden to write any inheritance expression such as "continuing from the previous segment / following / carrying on from the previous shot / same as above"(the model has no cross-segment memory); uniformity across segments in the same scene is achieved by fully identical four colors of the color card + word-for-word identical style description, not by textual reference; ② Gate D (§2.4 + §8) newly added corresponding mandatory checks.

  • V4.0.4(2026-08-11):Absorbing the Obsidian gaze/position/axis system(source: local Obsidian notes) — ① §2.4 iron rule of gaze anchoring expanded: reverse position of off-screen target, matching of gaze height and direction in shot-reverse-shot, eye close-ups consistent with the observed object in adjacent shots, expression intensity developing frame by frame; ② §3 newly added 3.4 Screen axis rules(180° principal axis / crossing the axis handling / reverse position / spatial stability / screen direction of character movement), the original 3.4/3.5/3.6 are renumbered as 3.5/3.6/3.7; ③ §5 newly added 5.4.1 Gaze-emotion quick reference(10 gaze states × emotions × English labels); ④ references newly added Xiaoluo-Visual Storyboard Nine-Grid Generation.md, Xiaoluo-AI Short Drama Scene Layout Design.md for reference.

  • V4.0.3b(2026-08-11):Iron rule of gaze anchoring entered into the repository(user feedback: in dialogue scenes "when the master speaks, the disciple does not look at him, there is no gaze constraint") — ① §2.4 newly addediron rule of gaze anchoring: in dialogue/two-person scenes, each shot must specify who looks at whom, where they look, and when they look away ([角色A]视线[低垂/直视/追随/望向X]), both listener and speaker must be written; looking down and avoiding = listening to a reprimand/submission/guilt, direct gaze = taking a stance/confrontation/realization, following = parting/longing for the past, with English labels averted gaze / direct eye contact; ② mandatory gaze anchoring checks newly added to 4 checklists including Gate D; ③ simultaneously fixed: the main text of the V4.0.3 iron rule of position anchoring had previously only entered the gate and version records, and was missing from the §2.4 main text, and is now added in this update.

  • V4.0.3(2026-08-11):Iron rule of position anchoring entered into the repository(user feedback: character positions are chaotic when generating segments and feeding them to AI video) — ① §2.4 timeline writing rules newly addediron rule of position anchoring: The first sentence of each time point must explicitly state the characters' spatial relationships ([角色A]在画面[左/右/中/前景/后景],[距谁多远/面向哪],[角色B]在[相对位置]), for multi-person shots add foreground/background/rule-of-thirds grid coordinates, key props must also have locked positions, distances use specific numbers, directions use left/right/center/foreground/background, and any positional changes across time points must clearly state the movement trajectory; ② The checklists at 4 locations including Access Control D add a mandatory item for position anchoring.

  • V4.0.2(2026-08-11):AI Video Multi-Segment Stitching Methodology added to the library(user practical experience) — solves the cross-segment stitching discontinuity caused by the 15s single-segment limit of AI Video 2.0. ①Added §3.6 Multi-Segment Stitching Methodology: data divided into two categories (invariant technical data = asset images/scene images imported as fixed throughout; variable changing data = position/pose/interaction/composition updated synchronously per segment), dual approaches (technical means: changing data = prompt words + storyboard reference image pairing + camera position change at the stitch point; director's mindset: coordination with audience perception);② Deprecated the first-and-last frame stitching approach: no longer use the previous segment's last frame as the next segment's first frame (AI has no long-term temporal memory, dynamic information resets to zero every 15s, and direct stitching will inevitably cause jump cuts), instead draw a storyboard reference image before generating each segment to fix the previous segment's ending state + synchronously update prompt words + change camera position at the stitch node; ③ §3.2's five-fold consistency lock upgraded tosix-fold: added final-state inheritance lock; ④ §10 cross-segment generation continuity synchronously updated; ⑤ synchronously updated all five-fold lock references such as Access Control D/Access Control E; ⑥ added references/AI视频拼接问题总结-多段拼接方法论.md.

  • V4.0.1b(2026-08-10):Drama micro-expressions/single-take long shot/special art style three sections added to the library(user-provided materials) — ① Added §2.2 "Four Techniques for Drama Micro-Expressions": emotional muscle chain (reject adjectives, execute frame by frame following corrugator supercilii → orbicularis oris → frontalis), asymmetry and coexistence of contradictions (unilateral exertion/0.5s dual-emotion transition), quantified physiological indicators (amplitude <1mm/range ≤15°/precise duration), restraint and action interruption (words rushing to the lips forcibly swallowed/head shake initiates then abruptly stops at 10 degrees) — supplementing the six-dimensional performance's "how to write" beyond "what to write"; ② Added §4.1.2 "Six-Dimensional Control for Single-Take Long Shots": shot narrative goal/three-layer spatial layering/physical inertia imperfect handheld/focus narrative design (0.3s defocus pull-back)/layered sound space/negative constraints — supplementing §4.1's missing focus design and audio layering; ③ §4.2.2 Master Color Grading Library added "Spider-Man: Across the Spider-Verse style" three-in-one formula (halftone dots + pop art warm-cold confrontation + highlight dispersion, without writing the film title to avoid red lines).

  • V4.0.1(2026-08-10):Dialogue scene duration hard rules added to the library(user feedback: 2s shot cuts cause broken sentences/mouth shape misalignment with long lines) — dialogue is a hard duration constraint,no shot cut allowed before the line is finished: single shot 8-12 characters → time point extended to 2.5-4s; 15 characters or more → split into 2-3 consecutive time points (speaking→pause/listener reaction/prop close-up→continue speaking), dialogue across time points must be continuous; dialogue segment rhythm determined by dialogue volume (≈4 characters/s, impatient 5-6 characters/s, calm/drunk/reminiscing ≈3 characters/s), 2s shot cuts serve only as reference for dialogue-free action segments. Synchronously updated: §2.2 Dialogue and Sound, §2.4 Timeline Writing Rules/Time Allocation Principles/Access Control D, §8 Access Control D.

  • V4.0(2026-08-09):Return to five-segment format + SKILL.md slimming(user testing: six-segment format performs worse than V3 five-segment format in ordinary fight scenes/dialogue scenes) — ①Approach A · Five-Segment Format Return: generation segment format primarily usesfive-segment format(【风格】→【色卡参考】→【镜头】→【时间轴】→【声音】); the [Characters and Scene] module is downgraded toonly used for long content >60s(upgraded to six-segment format), ≤60s always uses five-segment format (V3 tested well in practice). §0.5 self-check, §2.4 core format/examples, Gate D (§2.4+§8), §4.5.1 Giant Object Writing Method B, §9 usage instructions are all updated in sync. ②Plan B · SKILL.md slimming: §9.12 CINEDANCE (79 lines), §9.13 LIRA (99 lines), §9.14 Quill (56 lines), §9.15 340-item breakdown library (56 lines) compressed from full text topointer-style quick-reference summary(full versions in references/ and Obsidian), main SKILL.md slimmed from 1985 lines to 1696 lines (approx. 85k→73k tokens), to avoid English film methodology grabbing attention from the Chinese-language context of comic drama during every generation. ③Update version record historical notes (six-segment format = V3.2 temporary solution, has been corrected by V4.0 regression).

  • V3.2 Giant Object Aesthetics Special Upgrade (2026-08-09): §4.5.1 Giant Object Aesthetics special template expanded——①AddedWriting Method B · Six-Segment Format(user confirmed "this is our video prompt format"): the giant object segment must follow the standard six-segment format ([Style]→[Characters and Scene]→[Color Card Reference]→[Camera]→[Timeline]→[Sound]), no mixing with single-segment direct format; ②Added Iron Rule 9:❌ No human speech in giant object segments(user tested: once a person speaks, the model randomly assigns dialogue, only retain panting/breathing); ③AddedEight Techniques for Giant Object Sense Enhancement(partial close-up contrast/out-of-frame composition/human perspective limitation/multi-layer scale anchors/speed contrast/physical feedback/sound-first + low frequency/silhouette + fading out) + strongest combination (①+③+⑧); ④New subject replacement slots added Xuanming Giant Serpent, Ruins Stone Spirit Giant God (user tested 2026-08-09).

  • V3.2 Six-Segment Format Upgrade (2026-08-09): five-segment format upgraded tosix-segment format——a fixed module [Characters and Scene] is added between [Style] and [Color Card Reference] (characters appearing in this segment + scene location, at generation time @ character/scene asset images here, as a fixed asset reference slot, must be written in every segment). §2.4 structure description, color card reference position rules, Gate D updated in sync.

⚠️ V4.0 Correction (2026-08-09): six-segment format tested less effective than V3 five-segment format in ordinary fight scenes/dialogue scenes, has been reverted——[Characters and Scene] only retained for long content >60s (see V4.0 entry), ≤60s always uses five-segment format.

  • V3.2 Color Card Reference Value Upgrade (2026-08-09): color card reference writing rules updated——four colors (primary hue/auxiliary color/highlight color/shadow color) each color accompanied by HEX + RGB reference values(format: color name #HEX R,G,B, e.g. 千禧粉 #F1BB7B 241,187,123). §0.3 writing red lines removed "do not output HEX color codes" (changed to retain "do not output exact height values"), §2.4 color card reference writing rules and Gate D updated in sync. User tested and confirmed reference values = HEX+RGB.
  • V3.2 Supplement (2026-08-08):340-item practical prompt breakdown library added to repository——Scraped 17 scene folders totaling 340 real generated prompts from the HELL GRIND project API, and added §9.15 Practical Prompt Breakdown Library (nine-category quick-reference table + eight implementation iron rules + Chinese multi-shot dialogue template structure). Core conclusions: The Style Prefix is locked word-for-word across the entire project; asset references use <<<uuid>>>/<<<image_N>>> and are annotated with their purpose; Chinese prompts are viable (Scene 14 is entirely in Chinese); multi-shot uses 【Shot N】 segmentation + second-level action choreography; editing prompts use CHANGE+PRESERVE. ⚠️ The original data (340 raw JSON entries + curated md) was removed on request on 2026-08-09; the core methodology is retained in the §9.15 quick reference and references/HELL-GRIND-Prompt-Classification-Methodology.md.
  • V3.2(2026-08-07): Absorbed the new prompts from the desktop 漫剧课程/ training courseware—①Added §2.1A "AI Script Batch Creation Enhancement Pack": novel-to-manga-drama script instructions (three-step method: break down hit logic → adaptation requirements → pure action-dialogue conversion rules, vertical-screen shots LS/MS/CU/ECU/HA/LA/OTS/FT/Pan/Push + duration rules) + fully automated original hit script creation instructions (7-dimension configuration + full-project framework 7 modules + dark hook × sensory overload opening iron rule) + 5 script hit formulas (identity contrast/desperate comeback/tragic love missed/plot-twist ending/taboo trigger); ②The connection layer added §2.3 a0 "Character Asset Extraction JSON": group + primary-secondary consistency system (a face change starts a new primary/a face that does not change is secondary, naming rules, scene binding, 13-field output, primary = text-to-image three-view/secondary = image editing with locked face and outfit swap); ③Added §9.14 "Emotion-Driven Cinematic Prompts (Quill Free Version)": 6 questions including audience position/power relationships/spatial psychology + emotion→visual decision table + shot shorthand + spatial closure + deterministic hard rules + user field discipline. Sources: desktop manga-drama course (AI General Prompts.docx / 10 assignment reviews md / Quill Free Version), knowledge base mirror in Obsidian 07-Courseware Content/AI Director Thinking 10-Day Course/Refined Content/

⚠️ Historical Note (2026-08-09 Fix + V4.0 Update): In this entry, "five-segment format" and "color-system keywords do not use HEX" were the rules at that time (V3.1); HEX+RGB has been established; after the generation segment format briefly went through a six-segment format in V3.2,V4.0 returned to the five-segment format as the primary one(【Characters and Scene】 is used only for long content >60s). During execution, always follow the current §2.4 five-segment/six-segment rules and Gate D; do not execute according to this historical record.

  • V3.1 (2026-08-06 five-segment format · 08-07 HELL GRIND ingestion · 08-07 three-piece set upgrade): ①The generation segment format was upgraded from four-segment tofive-segment(user confirmation: the format is approved, but it lacks a color card reference; required to embed a fixed 【Color Card Reference】 module between 【Style】 and 【Shot】): Added §2.4 five-segment structure 【风格】→【色卡参考】→【镜头】→【时间轴】→【声音】 + color card reference writing rules (primary hue/secondary color/highlight color/shadow color, color-system keywords do not use HEX, share the same color card across segments to prevent drift) + two new checks added to Gate D + complete examples updated accordingly.Same-day user addendum: Short content within 10s is not segmented; directly output one complete five-segment generation segment (only long content ≥15s/full episode is split into multiple segments);After each generation, the content must be self-checked(§0.5 Iron Rule). ② New Chapter 9 "Open-Source Cinematic-Grade Prompt Methodology (HELL GRIND)": sourced from all prompts of the Higgsfield open-source 95-minute AI film "HELL GRIND" (15 segments, user-verified), classified and extracted into the repository by reusable methodology: prompt skeleton (16-field top-level structure), asset reference system (uuid + state assets + IDENTITY SWITCH), unified atmospheric directives (cross-shot consistency lock), iron-rule section writing (positive constraints/hard locks/geometry locks), second-level action choreography (frame-rate gradients/multi-shot segments), combat scene special (ONER/death rules/camera personality), SOLO PORTRAIT performance template, environment shot template, image editing prompts, unified style prefix + director style paraphrase table; the complete segment-by-segment original text goes into references/HELL-GRIND-开源95分钟电影提示词全集.md, and the full methodology text goes into references/HELL-GRIND-提示词分类方法论.md; director names written directly in the original text (Lubezki/Deakins, etc.) have all been paraphrased into descriptive wording, complying with the §0.3 person-name red line.③ Three-piece set merged in (2026-08-07): The three major skills of the HELL GRIND project are merged in by category—CINEDANCE director system → §9.12 (4-D methodology/field-of-view system replacing millimeters/shot decision tree/telephoto-wide-angle result stack/anti-drift lock/physical handheld/first-frame lock/spatial blocking lock/multi-shot continuity lock/lighting priority lock/context isolation/reference hierarchy/20-item QA); ACTING performance system → §5.6 (scene five pillars/main profile template/eye life/voice permanent lock/state not transition/ensemble wave/0-5 scale/15 bad performance patterns/pre-publish checklist); LIRA image assets → §9.13 (model routing/10 anti-failure iron rules/character sheet·scene·prop·surgical editing four major templates/technical block/color palette wrapping/pre-publish checklist). All three original texts are fully included in references/ACTING-SKILL-角色表演系统.md, references/CINEDANCE-HIGGSFIELD-提示词导演系统.md, references/LIRA-SKILL-图像提示词优化.md
  • V3.0(2026-08-06): Color cards and prompt template library officially added to the repository (source: desktop "AI影视创作色卡与提示词模板库.docx", prompts all verified usable by the user). ① §4.2.2 Master color grading library enhanced with HEX color values (Nolan lake blue #3B9AB2/Wong Kar-wai qipao red #7C1010/Anderson candy colors/Dunhuang mineral colors/Forbidden City palace colors, master color grading key points quick notes: Wong Kar-wai offset method + black level lift, Nolan cool colors low saturation + high contrast); ② New §4.2.3 verified complete prompt template library (cinematic golden formula + 6 color grading types + lighting verbs + atmospheric elements + 9 director Chinese templates); ③ Director name red line strengthened (§0.3: Seedance detects person-name violations, user-verified; new director name → descriptive wording conversion table §4.2.2)
  • V2.2(2026-08-06): Supplementary updates compiled from the desktop "AI Learning Materials" — the lighting system has been expanded from 3 types to a full table of 8 types (butterfly lighting/top lighting/Rembrandt lighting/chiaroscuro lighting/rim lighting/backlighting/bottom lighting/front lighting); added the standardized prompt "base-atmosphere-image" three-step template (§1.7) and long/short prompt selection strategy (§1.6); added the golden 3-second hook design + scene emotion selection methodology (§2.1 script creation); added the core value of film and television aesthetics "substance and refinement in harmony" and the creator's three levels of cognition (§6.1); supplemented the physical relationship of aperture/focal length/luminosity (§4.2). After user confirmation, the reference table of the five major fight types was removed (the old version was unusable). Added the action scene special section (§4.6: three-level destruction hierarchy/seven-stage transformation scene template/slow-fast-static and fast-fix-shock-follow-collect rhythm/speed trio/material precision comparison/special effects checklist/multi-person fight scene constraints, referencing the desktop "[1.24]--3D Action Scene Video Generation Notes", the transformation template verified by the user, see examples/Transformation Scene 10s Prompt.md for an example). Added the worldview construction module (§4.7 nine-facet method + continuity elements + space principles in use, §4.1.1 shot selection decision table, §4.3.1 imperfect realism), with anti-CG terms appended to three negative prompts (clean digital sharpness/CG feel/poster composition), referencing Cinematic World Generator (OAK). Further additions: Hedgehog Planet four modules (§4.2.1 color system control HEX color card + first-frame image bedding, §4.3.2 two great techniques of foreground occlusion, §4.6.7 multi-character action breakdown, §4.1 three tricks for realistic handheld) and reproduction content-03 practical application (§1.10 layered generation logic, §1.8 five categories of negative exclusion). Further additions: §4.2.2 master color grading style reference library (quick reference for 14 master/school color systems: Nolan/Villeneuve/Wes Anderson/Wong Kar-wai/Tim Burton/Kubrick/Deakins/Makoto Shinkai/Zhang Yimou/Chinese horror/cyberpunk, etc., including keywords + suitable themes, paired with §4.2.1 color card generation) → enhanced with HEX color values from the desktop "AI Film and Television Creation Color Card and Prompt Template Library" (specific color values such as Nolan lake blue/Wong Kar-wai cheongsam red/Anderson candy colors/Dunhuang mineral colors/Palace Museum court colors, etc., master color grading key points quick notes: Wong Kar-wai offset method + black level lift, Nolan cool colors low saturation + high contrast), for the complete color card see Obsidian 03-Materials/Color Card and Prompt Template Library/AI Film and Television Creation Color Card Library.md. All prompt templates verified by the user as usable (§4.2.2 newly added cinematic golden formula + 9-director descriptive conversion table), director name red line strengthened (§0.3: Seedance detects personal name violations, user-tested)
  • V2.1(2026-08-05): The core logic switches to the Xiaoluo AI short drama production chain as the main focus (timeline/shot grammar/asset numbering/six-dimensional performance/hard validation), with Shanyin/toolbox/micro-expression library downgraded to reference supplements; added three major assessment standards (smooth transitions + reasonable duration / plot logic / grand scene shot size + focal length); added the grand scene special section (§4.5 + Gate E: shot size-focal length comparison table, ratio anchors, three-layer depth, VFX prompt specifications); the transition layer adds character expression settings (twelve-grid chart); strengthened hard rules for timeline/assets/performance/dialogue and pre-delivery hard validation; subsequent user-supplied details will continue to be incorporated
  • V2(2026-08-04): three-step mandatory pipeline + transition layer + three-level prompt system + storyboard combination consistency (fivefold consistency lock + fourfold inter-segment transition rules) + phased quality gates + pure Chinese output
  • V1(2026-08-04): full-process integration package (screenwriting→assets→storyboarding→prompts)