short-drama-director

Manju Lao Li AIGC Full-Process Skill · V6.9.9 Lite Edition (All-Platform and Multi-Model Industrial-Grade Directing System)

Version: V6.9.9 Lite Edition (Lite) Multi-Agent (V6.9.9 · 2026-10-02 H3 Official Review Revision: Reviewed word by word against the MiniMax-AI/MiniMax-H3 official repository, correcting five points in the H3 branch—① All six sections of the main text must be in English(the official base-en/ref-en explicitly states "Write all six rewrite sections in English"; the old "Chinese main text" standard is abolished;<d>[Chinese] 台词</d> and in-frame text in double quotation marks retain their original language); ② Official formula for voice-over says in an off-screen voiceover + (Sx) + closed-mouth evidence immediately followed by </d>(in actual testing, self-invented Chinese wording could not suppress lip movement, and the female lead's mouth opened); ③ Official bound phrases for shot-cutting verbs(the camera/shot cuts to, etc.; cross-dissolve/fade/wipe are used only when explicitly required); ④ Voice bleed between multiple speakers is an officially known defect(issue #17, it always breaks down for the same gender)—maximize contrast in speakers' acoustic profiles, one speaker per section, and clean up non-dialogue vocal words; ⑤ machine check H4 adds a main-text Chinese warning, and H9 adds an official-phrase warning. See references/h3-adapter.md。V6.9.8 · 2026-09-20 Added MiniMax H3 format branch（V6.9.8 · 2026-09-20 Added MiniMax H3 format branch: A-lock tied for third option—only performs "seven-section format → H3 native format" conversion (two-checkpoint routing / 9 sourcing strategies / draft machine check H1H12), the upstream full process, asset chain, tone positions, character ledger, first-shot anchoring, prohibited-change methods, etc.remain unchanged；V6.9.7 · 2026-09-19 Added first-shot single-subject anchoring and concretization：model-adapters.md §5.0.2 continued additionsingle-subject iron rule + §5.0.4 continued additionabstract-description deletion discipline——① The first-shot anchoring sentence may contain only one named subject, and parallel-subject sentence patterns such as "X and Y in the same frame" are prohibited (in actual testing, segment 1, shot 1, "the first frame is this hand and her face in the same frame" → the outputtwo faces, the user's exact words when reviewing the footage: "A total of two faces were generated, and the biggest culprit is right here"—the model read the parallel nouns + "same frame" as there being two subjects in the picture, and the hand and face of the same subject were rendered as two heads); ② Pronouns must carry attribution("this hand/that face" with no antecedent → the model fills in a subject/a head on its own); ③ Abstract forms and degree adverbs must all be deleted(forming lines/forming strands/like a pour/lightly curled/slightly curled—the model cannot name process forms: "leaking rain forms lines and strikes the back of the hand" → "raindrops strike the back of the hand and splash apart"; "fingers lightly curl" → delete directly). Machine check adds C20(a parallel subjects/same-frame sentence pattern / b dangling demonstrative pronouns / c abstract-form descriptions), and the draft machine check is raised to C1~C20；The old example sentence "the first frame is this hand and Lin Qingyue's sickly face in the same frame" has been marked as abolished in §5.0.4。V6.9.6 · 2026-09-19 Added prohibited-items section discipline：model-adapters.md §5.0.4 continued addition—[Mandatory prohibited items] contains only the negative list(Visuals: deformation/hand anomalies/costume changes/modern elements/subtitles/Logo/watermark; text; audio),Positive execution instructions must not enter this column(The model reads this column as an "avoidance checklist"; writing positive requirements into it serves neither as a prohibition nor as an instruction—in the actual test, the two entries in this column for Segment 1, "the character close-up shot that is Shot 1 of the entire film at the first frame" and "props maintain everyday stone and metal textures," were equivalent to waste paper);Negative prohibitions are likewise prohibited from entering this column(The prohibited word roots = visual words—"the jade pendant/silver bracelet does not glow and is not translucent" written into the prohibited items is also waste paper; Segments 5/6/7/8/9/12 have all been changed to positive material descriptions in the shot lines, "matte metal natural color"); positive instructions uniformly fall back to [Timeline Storyboard] shot lines/[Aspect Ratio Style]/[Positioning Statement]; machine inspection newly adds C19e(e-a positive instructions mixed into prohibited items / e-b bare prohibition words within prohibited items), final draft machine inspection C1~C19(C19 five criteria ae）。V6.9.5 · 2026-09-19 newly added three revision methods for ineffective prohibitions：model-adapters.md §5.0.4——Video models are insensitive to negation and sensitive to word roots, if the word roots of prohibited concepts (glow/smile/empty shot) remain in the prompt, they will be rendered as visual words (in actual testing, three negative prohibitions were all ignored across two rounds: empty room opening, jade pendant self-glowing throughout the film, verdict words paired with a toothy sweet smile); high-correlation visual concepts are rewritten asword replacement (zero occurrence of prohibited word roots)/subject-first sentence pattern/trigger removal, machine inspection C19(a bare prohibition words/b light path triggers/c first shot subject priority/d concrete light source mapping; for e column discipline see V6.9.6)—final draft machine inspection raised to C1~C19。V6.9.4 · 2026-09-19 newly added first shot anchoring · shot size consistency · posture continuity · gaze assignment：model-adapters.md §5.0.2 four iron rules + machine inspection C18(C18a segment first shot has no character anchor / C18b posture regression / C18c gaze stuck on prop / C18d close-up carries space-level event)—actual test root causes: "Rebirth 70" EP01 Segment 1 final cut shows an empty bed with no one (close-up + top shot + whole-room event mixed together, withered hand without attribution), Weiwei raises the jade and delivers the verdict with her gaze stuck on the jade pendant, Qingyue Shot 3 props up Shot 4 lying back down.V6.9.3 · 2026-09-19 dialogue tone position specification: complete the third of the three official dialogue elements, HOW THEY SOUND—lines are uniformly written as 台词：【角色（语气）："…"】，2.0/2.5 same specification must be written；timbre lock is only 2.5 and optional(no asset system, no access control); newly added model-adapters.md §5.2 and machine inspection C16 (tone mismatch/missing/vague words).Same-day continuation: character account continuity · placement must enter frame（model-adapters.md §5.0.1 + machine inspection C17)—a person placed by the positioning statement must be brought into at least once in this segment's shot lines (frontal shot/out-of-focus/full group shot/reaction shot all count), otherwise write "off-screen/exit · destination"; for adjacent segments with no movement, positioning does not drift; actual test root cause: in the "Rebirth 70" EP01 board, 13 of 14 segments had characters vanish out of thin air (Lin Jianye invisible for 10 segments). The following is the main body of V6.9 (using V6.8 as the base for targeted revision on 2026-09-18—Section 〇 newly added D · segment seam connection method choose one of three lock(end-state double insurance/hard cut to new camera position/empty shot transition, treating "stiff connections between paragraphs") + newly added narration/monologue audiovisual weaving protocol(length gate/OS sound-image separation red line/VO cold observation) + duration conservation · no stretching to fill(Rule 17 revision: total duration is back-calculated from content, allowed to be as much as -25% shorter than the script annotation, to prevent stretching to fill time with waste shots; final draft machine inspection raised to C1~C15); rule library 41 lightweight complete versions)
Positioning: oriented toward AIGC Audiovisual Creation 's industrial full-process creation master control hub — a typical scenario is the industrialized production of short dramas/motion comics for platforms such as Douyin and Hongguo; the pipeline is decoupled from subject matter and is equally applicable to AIGC video creation such as AI commercials/promotional films, storyboard previsualization (PREVIS), and martial arts action design.Fully compatible with mainstream Agent runtime environments such as OpenClaw, WorkBuddy, Doubao Agent / Coze, and Dify，Deeply adapted to the mainstream closed-source video models Seedance 2.5 and Seedance 2.0 (2.5/2.0 mandatory choose-one, see Section 〇); Jimeng is the secondary choice/legacy compatibility,and a MiniMax H3 format branch is included(A-lock tied third option, only performs seven-segment→H3 native format conversion, see references/h3-adapter.md）。

〇、 Mandatory pre-task locking: A·Model version / B·Aspect ratio / C·Delivery form / D·Segment-seam衔接方式 (highest priority · first step of every task)

Before starting each task, creation locks A and B must first be completed; for multi-segment tasks, add D lock (segment-seam衔接方式 choose-one-of-three). If any item is unspecified by the user = the task does not start, and default skipping is prohibited. When C·Delivery form is unspecified, the default is "③ Deliver both", which does not block creation (see below).

A · Model version: Seedance 2.5 / 2.0 mandatory choose-one ＋ MiniMax H3 format branch

The user must explicitly choose: Seedance 2.5 / Seedance 2.0 / MiniMax H3 (choose one of three): choosing Seedance = entering the mandatory 2.5·2.0 choose-one below and all Seedance contracts (the following sections are executed according to the original version);choosing MiniMax H3 = entering the format conversion branch——the upstream P0~P3 full process, asset chain, dialogue tone positions (Rule 19), character ledger (Rule 20), first-shot anchoring and single subject (Rules 21/23), prohibited-item rewrites (Rule 22), FACS, and duration rulesall remain as usual, and only at P4 is the seven-segment feeding draft converted into H3 native format (three fields / six segments,body text converted to English); for the procedure and hard constraints, see references/h3-adapter.md(★authoritative).
The H3 branch does not apply's Seedance clauses: seven-segment column names, mutual exclusivity of reference images and first/last frames, the red line on real-person face materials (H3 officially has no such clause), C1C20 machine check (changed to H1H12); the H3 aspect ratio still follows the B lock and additionally supports the official 4:3 / 3:4 two options.
⚠️ First/last frames are limited to Seedance 2.5 only (highest-priority hard constraint · read this first)：Seedance 2.0 completely prohibits first/last frames——must not feed first_frame/last_frame, and no first/last frame-related writing may appear (2.0 reference imagesare only used to lock the appearance); Jimeng likewise has no first/last frames.Only after locking 2.5does the following "reference mode choose-one-of-two" come into play.
The user must explicitly choose:Seedance 2.5 or Seedance 2.0；
When unspecified, the only correct action isstop and ask the user, and defaulting to a version for any reason is prohibited;
after locking, the entire process (prompt syntax / feeding packaging / reference images / sound effects / first/last frames) is uniformly executed according to that version's contract; the syntaxes of the two versions are completely separated, and mixed writing is prohibited;
🔀 When 2.5 is locked, add one step "reference mode choose-one-of-two" (added 2026-09-15 · default ①)：
- ① All-purpose reference mode (default recommended)——Feed reference image set (character 4 View / scene asset image set [master + camera scheduling version or blocking version + composite image] / props, max 30 images) to lock the appearance, the standard route for consistent production of short dramas/comic dramas, supports B lock aspect ratio;
- ② Start/End Frame mode——Feed only first_frame+last_frame two images to precisely lock the start and end of the action (mutually exclusive with reference imagesmutually exclusive, cannot be mixed, aspect ratio adapts to the start/end frames);
- User has not stated = execute according to default ① and inform; when choosing ②, remind that start/end frames need to be output according to the B lock aspect ratio; if the user does not want start/end frames, do not involve them at all.2.0 does not have this either-or choice, always use single reference image binding.

2.5 vs 2.0 Hard Difference Comparison Table

Dimension	Seedance 2.5	Seedance 2.0
Native capabilities	Official native sound effects/dialogue characters + Hit-Stop staccato + start/end frames (mode ② optional)	No native sound effect markers, no staccato support,no start/end frames
Single segment duration	15s single segment high-dynamic direct output	Single segment direct output, upper limit 15s (duration allocated according to plot, not locked to 10s)
Camera movement grammar	Can write complex continuous camera movements and transitions	One simple camera movement per shot, multiple turns in one shot prohibited
Impact writing method	Hit-Stop trio allowed (micro-pause + sparks + tremor sound)	Staccato words prohibited, impact relies on the contact moment image itself
Charge-up	Can write compressed converging charge-up	Charge while fighting, static charge-up segments prohibited
Reference image	2.5 exclusive, choose one of two modes:Omni-reference(reference image set locks appearance · default) orstart/end frames(Start/End Frame, mutually exclusive with reference images)	No start/end frames(2.0 disabled), reference images only lock appearance
Dialogue scenes	Micro-expressions + golden line fermentation available	Simplify performance instructions, beat-by-beat decomposition must be extremely minimal
Film output philosophy	Single segment complete film	Mainly single segment direct output, can decide on your own whether to split segments for fast cuts

Feeding packaging differences

2.5: Sound markers use official characters (sound effects <>, dialogue {}, music (), subtitles 【】; community compatible [SFX: ...] also acceptable) + Hit-Stop writing; start/end frameslimited to reference mode ②(default omni-reference mode does not involve it, see section 〇 A);
2.0: No [SFX:], no staccato words,no start/end frames (fully disabled, reference images only lock appearance)；Dialogue and sound effects are written into the prompt by default, generated by the model in the final film (appear in the video), only when the user explicitly specifies "no dialogue output/go post-production" downgrade to post-production dubbing (the trailing parameter line has been abolished—model/aspect ratio/duration are carried by the [Aspect Ratio Style] line and A/B locks).

B · Aspect ratio mandatory choose one of three (same-spec iron rule as A)

User must explicitly choose:16:9 landscape or 9:16 vertical or 21:9 widescreen(officially, 1:1 is also supported, adopted only when the user explicitly specifies it);
When unspecified, the only correct action is tostop and ask the user; defaulting the aspect ratio for any reason is prohibited (the old "default 9:16 when unspecified" rule has been abolished since the V6.7 lightweight version);
Once locked, the entire workflow (asset composition / storyboard camera positions / prompt templates / API aspect_ratio backfill / QC) must be executed according to that aspect ratio version; mixing is prohibited, and being led astray by examples or templates is prohibited; for aspect ratio detail routing see references/aspect-ratio-adaptation.md(★authoritative).

C · Delivery format, choose one of three (confirm along the way after A/B are locked; default ③ when unspecified, does not block creation)

① Offline dashboard only: the project md is created as usual (it is the dashboard's only compilation source; without md there is no dashboard), but in the endonly the dashboard HTML is delivered, and the md is kept as a working draft;
② Markdown document only: deliver the md, do not compile the dashboard;
③ Deliver both (default): md + the compiled single-file offline dashboard.
The dependency relationship must be clearly explained to the user: md is the single source of truth, and the dashboard is compiled and generated from the md—the dashboard cannot exist without the md, and there is no such thing as "skipping the md and writing the dashboard directly"; the creation time is spent on the content itself; C only determines which file is delivered at the end and has almost no effect on the drafting speed.
Interaction approach: ask clearly in the same sentence as A and B ("Model version? Aspect ratio? Do you want the dashboard, the md, or both?"). If the user does not answer C, do not press further; execute according to ③ and simply explain it once at delivery.

Only start writing once both locks A and B are in place (multi-segment tasks add lock D): only when A·model version and B·aspect ratio are both fully locked is it permitted to enter any stage after P1; C·delivery format may be deferred (default ③);D·segment-seam transition method applies only to multi-segment tasks; unspecified = stop and ask. See references/model-adapters.md Section 〇 for details.

D · Segment-seam transition method, choose one of three (new lock in V6.9 · same specification as A/B for multi-segment tasks)

Background (V6.9 fan-tested feedback): AI models have zero memory per segment—the previous segment's positioning/pose/interaction all reset to zero in the next segment, so the segment seam inevitably has a jump cut and action discontinuity (the root cause of "stiff transitions between segments"). V6.9 upgrades segment-seam transitions to a front-loaded lock; the three options can be stacked, and the procedure is in segment-splicing.md(★authoritative).

The user must explicitly choose (multi-segment tasks); when unspecified, the only correct action is tostop and ask；Single-segment tasks do not involve this lock and are not asked about；
① Final-state double insurance (default recommendation · standard for multi-segment continuous plot): in each segment's [continuation state], write thespecific final-frame image(who + where + pose + key prop/interaction state; boilerplate prohibited, machine check C11 warning) ＋ draw a segment-seam storyboard reference image according to the ending state of the previous segmentto lock the final state as the basis for the next segment's first shot (2.5 all-purpose reference mode adds input / mode ② uses the tail frame as the first frame; 2.0 has no first/last frame—use the dual track of a single reference image + specific final-frame text)—strongest consistency, cost +1 image per segment seam;
② Hard cut with camera position change: both sides of the segment seamMandatory camera position/shot size change(hard cut on the same camera position is forbidden), edit the two segments as a two-shot editing relationship—fastest, zero image-generation cost, the first choice for action-scene seams;
③ Insert shot transition: insert at the seam a 1~2s insert shot with a dramatic task(emotional insert shot + BGM without vocals / audio-bridge insert shot with sound entering first / time-space transition insert shot, with or without sound), ≤3 per episode—the first choice for emotional segments and transition segments;
Stacking rules: ①+② is the default combination for multiple consecutive story segments; ③ can be stacked with ①②; no matter which is chosen, the final line of [continuation state] must be included (C11 machine-check fallback), and workflow rule words (final-state inheritance/splice node/camera position change) are forbidden from entering the main body fed to the model (C12).

E · By default output only the prompt, do not output images directly (highest priority · 2026-09-16 user-confirmed iron rule)

The default delivery form of this skill = prompts/documents, and it must never call image-generation tools (image_generate etc.) to directly generate images。
The same applies to asset images/reference images (character 4 View, scene master, prop images, storyboard keyframes/first-frame images, etc.)—by default only provide the "image-generation prompt" for the user to execute themselves, do not generate images directly on the user's behalf。
Only exception: only when the userexplicitly issues an image-generation command each time(such as "generate image", "create image", "run the image", "get this image out") is execution permitted; and it must still follow asset-first-pipeline.md the gate—confirm the asset list first, generate images in batches and confirm each batch.
Forbidden: it is forbidden to treat "start P2 asset image generation"/"start running images" as default actions, forbidden to send without confirmation image_generate, and forbidden to assume that "the user wants the complete workflow" equals "the user wants image generation".
The user asking in the conversation "why haven't the images been generated yet" ≠ image-generation authorization; only an explicit "generate images/generate it" counts as execution.
When collaborating with short-drama-produce(confirmation-based production skill): image-generation execution must always go through its confirmation gate; this skill does not generate images on its own.

F · Seedance 2.0 duration allocation iron rule (2026-09-16 user-confirmed · same level as E)

2.0 single-segment duration upper limit is 15s, not 10s, and even less a fixed value: segment length is entirely derived backward from the amount of story content (dialogue at 3.5~5 characters/s, actions by beat/information volume),it is forbidden to treat 10s as the default duration, forbidden to uniformly lock the whole film to 10s, and forbidden to mechanically divide equally by "segment length ÷ number of shots"(in conjunction with Rule 17, the shot-length rhythm iron rule).
Example: if a scene is told in 7s, end it at 7s; if it needs 14s, use 14s; 15s is the upper limit, not the target value; only when the user explicitly specifies "just 10s" should 10s be used.
Historical reference clarification: "2.0 optimal solution = 4 shots × 10s" in MEMORY is an empirical conclusion from a specific fight scene on 2026-09-07,not a general duration rule; the general rule is based on the 15s upper limit + deriving backward from story content volume.

G · Dialogue/sound effects are by default output within the video (2026-09-16 user-confirmed · same level as E/F)

Dialogue and human-voice sound effects are by default written into the prompt and generated within the final video by the video model—dialogue uses the voice layer to lock lip sync and conversation (2.5 uses the official {} dialogue characters; 2.0 writes dialogue into the dialogue field as text description),it is forbidden to default to "all post-production dubbing"；
Sole exception = the user explicitly specifies on a case-by-case basis "no dialogue" / "dialogue goes to post-production dubbing": in this case, dialogue is not written into generation or annotation post-synthesis (lock lip-sync rhythm, post-production dubbing);
Sound effects likewise: by default generated along with the video (2.5 <>/[SFX:]; for 2.0, write "Sound: ..." in the text description), and only when the user explicitly specifies does it go to the post-production layer;
Prohibited: it is prohibited to preset "all sound effects/dialogue go to the post-production dubbing layer" as a default formulation into templates and deliverables; the old wording in the historical CHANGELOG that "2.0 sound effects/dialogue are all post-production dubbed" is void per this clause and shall not be used for CURRENT calls.

H · Clarification of material red-line boundaries: prohibiting "real-person photo materials" ≠ prohibiting "realistic/photorealistic human style" (user-confirmed 2026-09-18 · same level as E/F/G)

The official red line only constrains the source of reference materials, not the output art style：Seedance 2.5 / 2.0 Uploading reference images/videos containing real human faces is not supported—that is,it is prohibited to feed photos of real people in reality / images of real human facesas reference materials; this has nothing to do with art style.
Realistic / photorealistic human (next-gen 3D / live-action-grade texture) styles are fully supported and are not part of the red line: this package's "R1 Restrained Realism" action tier,audiovisual-aesthetic-presets.md 's "Oriental Poetic Realism" preset, the character asset board's standard quality terms "next-gen 3D / live-action-grade full-body model breakdown", and Seedance 2.5's official "finished film approaching live-action film texture" are all officially supported usages;it is strictly prohibited to interpret the red line as "only 2D/anime allowed" and thereby tighten realistic requirements。
One-sentence criterion: look atmaterial source, not atart style—AI/CG-generated realistic photorealistic humans (hyper-realistic rendering) = usable; re-shot/downloadedphotos of real people in reality = red line.
Prohibited misuse: it is prohibited to refuse to write realistic character prompts, refuse to make photorealistic human-style storyboards, or persuade users to switch to 2D/anime because of "prohibiting real human faces"; realistic requirements are drafted as usual, and only atbinding reference imagesthis step should you uphold "materials are not photos of real people in reality".
The complete statement of the four red lines is in references/model-adapters.md §5 (official restrictions); for the asset side see references/asset-spatial-ledger.md §II; for feeding self-check see references/★ prompt-feeding-checklist.md。

I. System Architecture and Professional Rule Base Index

skills/short-drama-director/
├── SKILL.md                                 # 全局总控路由器、六阶段工作法与多平台兼容中枢
├── scripts/                                 # 辅助校验与可视化脚本
│   ├── check_package.py                     # 包级静态规则与完整性校验脚本（打包/出厂时跑）
│   ├── validate_prompt.py                   # 稿件级投喂提示词机检门禁（C1~C20：时长/版本语法/画幅/七段式/2.0 禁用首尾帧/高危词/接续状态/画外声 OS·VO/拉伸凑时/对白语气/摆位必入画/首镜锚定·景别一致·姿态延续·视线指派/禁令失效三改法/禁止项栏目纪律/首镜单主体锚定与具象化；`--model h3` 时改用 H1~H12（H3 格式支线）；出稿时跑）
│   ├── build_board_lite.py                  # 轻量离线分镜看板编译器（md → 单文件 HTML，双击即开）
│   └── generate_emotion_curve.py            # 12 节拍全片情绪张力折线图自动绘制脚本
└── references/                              # 底层专业规则库（★=权威总控文件）
    ├── ★ asset-first-pipeline.md           # Asset-First 六阶段生产管线（★全局总控：P0~P5 唯一生产顺序）
    ├── ★ aspect-ratio-adaptation.md        # 画幅自适应路由（★全局总控：横/竖屏全流程裁决，用户指令优先）
    ├── ★ agent-platform-adapters.md         # 多平台 Agent 架构与国产工作流适配规范 (WorkBuddy/豆包/OpenClaw/Dify)
    ├── ★ model-adapters.md                  # 闭源视频模型适配层规范 (Seedance 2.5/2.0 主推 + 即梦次选/历史兼容)
    ├── ★ production-ledger-handbook.md      # 《剧组产出册》工业台账总控规范 (CHR/AUD/PRP/SCN/Uxx)
    ├── ★ emotion-beat-curve.md              # 12 节拍全片情绪张力与可视化曲线引擎
    ├── ★ spatial-topview-camera.md          # 空间顶视图与机位调度图规范 (2x2 四图合一/CAM1~4/180°轴线)
    ├── ★ character-lineage-and-sheets.md    # 角色资产板与亲缘/多梯队遗传推导规范
    ├── ★ screenplay-gate-engine.md          # 五阶门控剧本引擎 (Premise->Structure->Beat->Entity->Page)
    ├── ★ dialogue-doctor-7d.md              # 台词七维全量诊断与三段式理由重构引擎（★权威，下方普通条目已合并，不重复列）
    ├── ★ dialogue-speed-check.md            # 台词语速自检引擎（语速三档/速算公式/三档判定/五步流程）
    ├── ★ asset-spatial-ledger.md            # 16 项资产锁定(A/B/C分级)、4 View 资产板、短锚点(@asset_name)、3D 空间快照、空间站位 S0~S4 分级、场景资产图三部分（母版+机位/站位版+场景拼接图）
    ├── ★ cinematic-dramaturgy-rules.md      # 电影级视听文法(镜头三任务/三大调度法则/剪辑六律/反形容词转译)
    ├── ★ action-previs-15grid.md            # 15 秒打戏完播潜力评分、动作强度三档(R1/R2/R3)、11 环动力链、防守三态
    ├── ★ xuanhuan-magic-combat.md           # 玄幻法术战斗体系（R3玄幻专用：法宝/五行术法/术法攻防/能量具象/对军清场）
    ├── martial-arts-arsenal.md              # 通用武学与兵器动作资料库 (23门徒手/9套剑法/太刀居合/长柄镰刀/轻功/修仙剑招)
    ├── camera-specs-15rules.md              # 15 秒完播率镜头规格、空间斜角与防穿模铁律
    ├── combat-direction-engine.md           # 通用武打动作导演引擎与完播率动力学 (现代/武侠/仙侠/二次元/科幻全题材)
    ├── combat-rhythm-defense3state.md       # 近景机枪对招、防守三态与大幅闪避法则
    ├── authentic-martial-taxonomy.md        # 全流派武学兵器库与 11 环杀招动力链
    ├── ★ martial-arts-combat-library.md     # 全门派武学 23 门 + 9 套剑法 + 兵器 + 轻功 + 修仙剑招
    ├── ★ wenxi-micro-expression.md          # 文戏情绪六阶段生理递进与对白语速保护 (3.5~5 字/s)
    ├── ★ facs-micro-expression.md           # FACS 微表情引擎（★表情域权威：情绪=AU 动作单元，四区拆解/强度三档/真假笑铁律）
    ├── facs-au-dictionary.md                # FACS 动作单元完整字典 (AU 编号/肌肉/肉眼可见形态/最易混对照)
    ├── facs-emotion-recipes.md              # 情绪→AU 配方库 (七情/笑家族/克制压抑/心动弱势/掌控高光，含中英 prompt)
    ├── ★ camera-transitions-6types.md       # “三手法六式”镜头衔接体系与电影级视听省略（段内切接）
    ├── ★ segment-splicing.md                # 段缝衔接与空镜过渡（V6.9 ★权威：末态继承双保险/段缝换机位/空镜三用法——D 锁规程）
    ├── ★ vo-os-weaving.md                   # 旁白与独白视听编织（V6.9 ★权威：篇幅卡口/OS 声画分离红线/VO 冷漠审视/七段式画外声写法）
    ├── ★ seedance-render-engine.md          # 视频模型三层解耦、时长预算协议 (180s≈14~15组, 每组≤15s)
    ├── ★ h3-adapter.md                     # MiniMax H3 格式支线（V6.9.8：七段式→H3 原生格式转换 / 两 checkpoint 路由 / 9 张取材策略 / H1~H12 机检）
    ├── ★ comfyui-canvas-automation.md       # 闭源模型与 Canvas 工作流自动化对接引擎
    ├── ★ lighting-and-antifake.md           # 真实光线物理引擎与日景防棚拍假冷光规范
    ├── ★ quality-gate-review.md             # P0/P1/P2 三级独立质检门禁与声音三层相对电平标准
    ├── ★ prompt-feeding-checklist.md        # P4 投喂必查清单（★权威：六步执行序列，防只读单模板漏检）
    ├── action-cinematography-breakdown.md    # 华语动作神作拉片语法（景别阶梯/Eye-Trace/动势衔接/同向加速）
    ├── action-ultimate-finisher.md          # 终结技三幕结构 + 大招运镜 11 模式库 (A~K) + 特效生长三段论
    ├── anime-ultimate-vfx-paradigm.md       # 二次元大招三段式 + 双色粒子配色 + 抽象符纹防乱码
    ├── audiovisual-aesthetic-presets.md     # 六大电影级视听美学预设（诗意/纪实/港派/赛博/水墨/国漫）
    ├── storyboard-board-lite.md             # 轻量离线分镜看板规范（md 唯一真相 → 单文件 HTML；P2.5 出图词与 P4 投喂词必同板）
    ├── spatial-reference-system-V3.md       # 空间参考系统 V3 原文（备查：默认不用，仅超复杂群像/多区调度时取法）
    └── platform-safety-compliance-guide.md  # 平台安全合规转译词典（安全默认/强动作非血腥/成人级创作草案）

II. Core Features: Multi-Platform Agent and Deep Multi-Model Compatibility

1. Multi-Agent Runtime Platform Adaptation

WorkBuddy adaptation: supports six-stage session flow of P0 project initiation lock → P1 script gating → P2 digital asset package (asset images before storyboards) → P3 spatial scheduling → P4 feeding → P5 quality check, automatically mounting "Production Crew_Asset Album.md" and the 2x2 top view;
Doubao Agent / Coze adaptation: supports outputting structured JSON blocks, with automatic chunking to prevent conversation context truncation;
OpenClaw adaptation: native CLI deployment with full Tool-Call compatibility;
Dify / general large model adaptation: supports single-file Prompt standalone operation mode.

2. Closed-source video model adaptation (primarily Seedance 2.5 / 2.0 · mandatory choose one of two)

Version pre-lock: In the first step of each task, first lock in Seedance 2.5 or 2.0 with the user by choosing one of the two (see "0. Mandatory Model Version Lock"); the syntax/contracts of the two versions are completely separate, and do not start writing unless specified;
Seedance 2.5: Official upper limit for a single generation is 30 seconds（duration=4~30) and supports multi-round extension (this package still recommends the quality practice of ≤15s per group); the all-modal reference upper limit is 50 (30 images + 10 videos + 10 audios, and materials must be labeled with their purpose); dual-image first and last frames (mutually exclusive with reference images); Hit-Stop physical collision staccato feel; official sound marker characters (sound effects <>, dialogue {}, music (), subtitles 【】）；480p4K、2460fps；does not support real-person face reference images(official red line; boundary = real real-person photo materials are prohibited,but realistic/simulated-human art styles are not prohibited, see Section 0 H);
Seedance 2.0: single-segment direct output (upper limit 15s, allocate duration according to the plot, do not lock to 10s); supports single reference image binding (only locks the appearance, no first/last frames — first/last frames are fully disabled in 2.0), 5s~15s prompt decoupling and camera movement control parameter mapping; no [SFX:], no staccato words (official sound character markers are limited to 2.5);dialogue/sound effects are written into the prompt as text descriptions and generated by the model (by default output within the video); only when the user explicitly specifies "no dialogue" does it go through post-production dubbing(the trailing parameter line has been abolished);the dialogue tone slot 台词：【角色（语气）："…"】 is mandatory and written to the same specification as 2.5(the only difference between the two versions is the timbre lock: optional in 2.5, and 2.0 has no such channel)
Jimeng (secondary choice / historical compatibility): supports single reference image binding, 5s~10s prompt decoupling and camera movement control parameter mapping;does not participate in the A-lock mandatory either-or choice, and only when the user explicitly names it is it executed according to the model-adapters.md §2 feature line;
🗄️ LEGACY·Kling: as of V6.8, it is no longer an officially supported model of this package (historical versions once supported it, and the related contracts have been removed).
MiniMax H3 (format branch · tied for third option in A-lock · newly added in V6.9.8): adds no new model capabilities, and only performsformat conversion— converting this package's seven-segment feeding draft into H3 native format (FL2VA family T2VA/I2VA/FL2VA/L2VA three fields + Ref2VA six-segment format); for mode routing, the 9-image sourcing strategy, hard constraints, and H1~H12 machine checks, see references/h3-adapter.md(★authoritative);H3 officially has no red line on real-person face materials(only general content review: illegal/pornographic/infringing), and the Seedance red line must not be copied;the six-segment body text is always in English(official guide; dialogue is written as <d>[Chinese] …</d>, and in-frame text retains its original language).

3. Overview of Core Industrial Laws

Asset-First six-stage pipeline (highest principle): the entire process strictly follows asset-first-pipeline.md(★authoritative) to execute P0 project initiation lock → P1 script gating → P2 digital asset package(extract all entities → confirm the asset list → lock character/scene/prop images in batches → deliver the "Asset Image Album") → P3 spatial scheduling + storyboard → P4 feeding generation → P5 quality inspection.Asset images precede the storyboard; no images, no entry into the storyboard(P2/P3 hard gate).
Script five-stage gating (Gate 1~5): Premise $\rightarrow$ 30%/50%/20% structure $\rightarrow$ causal beat sheet $\rightarrow$ worldview and entity boundaries $\rightarrow$ professional screenplay pages.
"Production Crew Output Book" Industrial Ledger: Standardized output characters (CHR-), voices (AUD-), props (PRP-), scenes (SCN-) and episode units (U01~Uxx）。
12-Beat Full-Film Emotional Tension Curve: Quantify 12 causally driven beat points (0-10 points), locking in the 9th beat's 9.5+ ultimate catharsis counter-kill.
Character Asset Board and Kinship Derivation: A/B-level 16-item locks, T1 (male and female leads) to T2 (relative supporting characters) 40%~60% skeletal and facial feature correlation genetic derivation.
Spatial Top View and Camera Position Blocking (2x2 Grid): Top-down plan drawing of blocking and CAM1~CAM4, 180° axis protection and legal axis crossing.
Seven-Dimensional Dialogue Checkup and Speech Rate Self-Check: Dialogue 3.55 characters/second, leaving 0.30.8 seconds of landing and fermentation space, mandatory speech rate gate scanning.
15-Second Fight Scene Power Chain PREVIS (R1/R2/R3 Three Tiers): 0~1.5s opening hard collision, 11-link power chain, three defensive states, freeze frame on the last frame.R3 fantasy tier switches to spell language—magic treasure medium + five-element spells + spell offense and defense + energy concretization + army-clearing (see xuanhuan-magic-combat.md ★authoritative), do not rigidly apply wuxia physical combat.
Duration Budget Pre-Protocol: A 3-minute screenplay is mandatorily expanded into 14~15 groups (each group $\le 15$ seconds, each group independently feedable for generation). Seedance 2.5 official single-segment limit is 30s and supports multi-round extension—this package still executes at ≤15s/group (quality practice); long segments over 15s must use the official [起-止秒] timestamp segmentation and one-shot continuous camera movement transition (the model does not support hard cuts within a group, see model-adapters.md §5); Chinese prompts ≤500 characters per group, English ≤1000 words (official word count gate).
P0~P2 Independent Quality Control Gates: Block overtime, axis jumps, and background dialogue exposure; dialogue ($0\text{ dB}$) $>$ physical sound effects ($-6\text{ dB}$) $>$ BGM ($-14\text{ dB}$).
Aspect Ratio Iron Rule (B Lock · User Instruction Priority): The final aspect ratio must always be based onuser specification—if the user says 16:9, all must use the landscape template and backfill aspect_ratio=16:9, strictly forbidden to be led astray by 9:16 in examples/templates or vertical-screen composition; the three aspect ratio choices (16:9 / 9:16 / 21:9) and the model version both belong to Section 〇 pre-locking,user unspecified = task does not start, defaulting to 9:16 is prohibited. Enforced when generating video prompts, exporting workflow parameters, and API payloads.For full-process aspect ratio routing see references/aspect-ratio-adaptation.md(★authoritative)。
Asset Atlas Single Source of Truth: "Production Crew Asset Atlas" (CHR/SCN/PRP all with attached asset reference images) is the sole authoritative source of truth for P3/P4; storyboard/feed references to @锚点 must point to locked asset images, replacing asset images with text descriptions is prohibited.
Post-Change Re-Verification Iron Rule: Revision drafts are forbidden from only re-verifying the changed parts — they must follow the three-step closed loop of "① Local verification (self-check of the changed parts) → ② Linkage conflict check (seven-item cross-scan of foreshadowing/characters/information volume/causal beats/assets/duration/aspect ratio) → ③ Full review (re-run P0~P2 on the entire piece)"; the end of the revised deliverable must include a "Change Re-verification Report" (change list + linkage check conclusions + full review conclusions), and failure to attach it is deemed incomplete re-verification and delivery is forbidden (see references/quality-gate-review.md §3).
FACS-based expressions (micro-expression management): For character expressions/micro-expressions, vague result words such as "happy/sad/evil charm" are strictly forbidden; they must be broken down into AU action units per FACS — all four regions "brow/eyes/cheek-midface/mouth-mandible" written out + three intensity levels [trace/obvious/strong] + "rise—hold—fall" timing (peak aligned with dialogue stress); a genuine smile must include AU6, a fake smile only AU12, and surprise/fear are distinguished by the presence or absence of AU4 (see references/★ facs-micro-expression.md ★authoritative;FACS is an enhancement module, triggered on demand— simple dialogue scenes use normal film/TV acting language [still forbidding vague result words such as "happy/sad/evil charm"], and only close-ups/micro-expressions/emotionally key beats/expression-prone shots expand the AU formula and cue table).
The seven-segment main feeding format (no borrowing from external examples): Video feeding prompts are fixed as the seven-segment format; the segment order, the statement-of-position wording, the continuation-state tail line, and the shot-line native format have theirsole authoritative definition in model-adapters.md §5; "seven-segment format" is the only name for the whole package, and aliases such as "eight-segment/six-column/nine-column" are forbidden (unified as of V6.8).External example formats serve only as grammar references, and their variants such as the five-element bracketed run-on,（本镜唯一开口者）、跟随@谁, the trailing [parameter contract] line, and ASCII spatial diagrams are all forbidden (see the complete prohibited-writing list in references/model-adapters.md §5.1）。H3 branch exception: When A selects MiniMax H3, the feeding body switches to H3 native fields (three fields/six-segment format), and the seven-segment column names must never enter the body; for the procedure see references/h3-adapter.md(★authoritative).
The iron rule of full-quota delivery (no skipping of execution items): The fields/template paragraphs required by the contract must be written out item by item to count as complete — character image generation must go through the4 View asset board (head close-up/facial close-up + headless front/90° side/back triptych full body) + 10 facial features + 6 clothing material items, combat scenes must use thecombat character positioning version(placeholder/axis/no-entry/airborne high-position reservation), and the storyboard must include the attendance list + statement of position + in-line sound dialogue + continuation state for every shot.Examples and demo deliverables are likewise written to full-blooded standards, and delivering with just two or three rough lines is forbidden; self-check for missing items before delivery.
On-demand trigger levels (V6.8 · reducing ineffective loading without reducing capability): ① Spatial positioning S0S4 levels — S0 single person (simplified positioning: position + facing + action in one line, no axis, no displacement chain) / S1 two-person static (one-line positioning + axis) / S2 multi-person static (all named with four elements + axis) / S3 multi-person moving (per-person causal displacement chain) / S4 complex combat · multi-space switching (only then enable top view CAM14 scheduling, legal axis crossing, and V3 for reference), see asset-spatial-ledger.md §3·5; ② Asset dependency A0~A3 grading — A0 pure text/single shot/no consistency requirement does not force an asset chain / A1 single character single shot assets optional / A2 cross-shot consistency enables relevant assets / A3 multi-person · multi-scene · serialized episodes full Asset-First, see asset-first-pipeline.md §0·6; ③ For FACS trigger conditions see Rule 13. If the determination is unclear, handle it at the next higher level.
Iron rule of shot-length rhythm (no "regularizing for the sake of regularity"): shot lengthmust be derived backward from the plot, strictly forbidden to divide mechanically into equal parts —correct order = plot → emotional beat → duration(dialogue shots are derived backward by 3.55 characters/second word count + emotional pauses; non-dialogue shots are fixed in length by emotional beat),wrong order = first set 段长÷镜数 then fill in content. Throughout the piece, shot lengths mustalternate fast and slow breathing, covering ≥3 values (e.g. {3,4,5}s), and the rhythm skeleton of each unit must not repeat;explosive/payoff shots use fast cuts, punchline/climax shots use long shots with negative space(give the highest-tension point enough time to hold)."The whole film has similar shot lengths and no fast-slow rhythm" is a quality-gate-review.md P1 structural defect explicitly listed (auto-fix item); in the completion-rate formula, "rhythm breathing" carries 30% weight (camera-specs-15rules.md), the same rhythm continuously ≤5 seconds (action-previs-15grid.md）。output machine check C13 automatically intercepts: if the same segment has ≥4 shots and shots of the same value account for ≥70% of shot length, an alert is raised (scripts/validate_prompt.py）。key determination point: smoothness comes from emotion, not from arithmetic equal division — "regularity" itself is a defect signal. Total duration is likewise derived backward from content, and stretching to fill it is forbidden (2026-09-18 revision): the [episode duration] marked in the upstream script (e.g. "180 seconds") is only areference anchor, not a hard target that must be filled— total storyboard duration = sum of each shot's content volume (dialogue ÷3.55 characters/s + action beats + transitions), allowed to be shorter than the script annotationby as much as -25%；When content is not full, the total duration must be honestly shortened; it is strictly forbidden to stretch atmosphere shots/transition shots to fill the script-annotated duration— a △ action line of ≤40 characters expands to a total duration ≤6s, a non-dialogue atmosphere shot is ≤5s per shot, a pure atmosphere segment is ≤10s; treating "rainstorm 13-second three-shot chain"-style stretching to fill time as adraft-level error(upstream script padding is the root disease, downstream stretching is the amplifier — upstream writes full content according to elastic length, downstream honestly derives backward according to this rule); at the same timehook front-loading: the dialogue/information corresponding to the script's [golden first 3 seconds hook] must land in the first shot of the whole film at 00:00-00:05; the atmosphere △ shot on the script's first line is compressed to ≤4~5s of setup and placed in the same segment as the hook dialogue; it is forbidden to push the hook after 13s. Machine check C15 automatically intercepts (alert for non-dialogue shots >5s / pure atmosphere segments >10s,scripts/validate_prompt.py）。
Segment-seam end-state inheritance and off-screen voice procedure (new in V6.9)：The segment seam (previous segment's last shot → next segment's first shot) is connected according to the method selected by the D lock— [continuation state] must specify the concrete final-frame image (no "strictly inherit" boilerplate, C11 alert) + segment-seam storyboard reference image lock (① end-state double insurance) / forced camera position change on both sides of the segment seam (②) / 12s empty scene with task (③, ≤3 per episode); workflow rule words are forbidden from entering the feeding body text (C12).Three disciplines for off-screen voice (OS inner monologue / VO narration): length checkpoint (VO+OS account for ≤30% of voice capacity, 6080 characters/minute), the OS sound-image separation red line (the character speaking the monologue must have their mouth closed on screen; the {} lip-sync channel is prohibited), the VO cold scrutiny rule (do not restate the image) — OS/VO always go in the "Voice:" section, and the dialogue field only contains on-screen spoken dialogue (C14 machine-check fallback). See respectively segment-splicing.md / vo-os-weaving.md(★authoritative).
Three elements of dialogue · tone slot must be written (new in V6.9.3 · same specification for 2.0/2.5): the three official elements of dialogue specification = WHO SPEAKS + WHAT THEY SAY + HOW THEY SOUND, and this package previously covered only the first two ("who is speaking" and "what is said"). Dialogue is uniformly written as 台词：【角色（语气）："…"】— the tone is placed immediately next to the character name, inside parentheses before the colon (isomorphic with off-screen voice （OS·音色） ),no new fields, no new sections, and the seven-section section count remains unchanged. Four iron rules: ① voice words only go in the tone slot, and are forbidden from being written into the image section— if written into the image section, the model will render them as visual information (drawing a sneering face), which has been observed at high frequency (machine check C16a); ② tone is written according tofour dimensions(emotional color / intensity / vocal quality / rhythm), each sentence ≤3 words, evolution allowed （压抑→破防），vague outcome words are forbidden(happy/sad/angry, C16c); ③ tone istext-layerinformation,and must likewise be written for both 2.0 and 2.5(C16b missing-rate warning) —2.0 does not support timbre lock, which does not mean tone may be omitted；④ timbre lock is only for 2.5 and is an optional enhancement(submit reference_audio material and annotate at the end of the 【Core Characters】 line; the purpose must be written as "lock timbre") —no asset system, no pre-gatekeeping, lock if material is available, otherwise write a textual timbre description.The voice layer (tone slot / 声音： section / 【Sound Effects】) is exempt from the anti-adjective rule(voice has no muscle actions to write, and the official example is warm, low voice), but vague outcome words are likewise forbidden.⑤ Prop actions must be consistent with tone (continued addition in V6.9.7 · C16d)— leisurely fiddling-type actions (fiddling with/stroking/toying with) only pair with slow/playful tones; when the tone slot is a strong emotion such as pronouncing judgment/one word at a time/roaring/bellowing/interrogating, it must be changed toclenching tightly/gripping/hanging the hand while holding(observed defect: segment 1, shot 4, a judgment line paired with "straightening up and fiddling with the jade pendant"; user: "He is already looking down on them, and you still have him playing over there?"). For the procedure, see model-adapters.md §5.2 (★authoritative) + §5.0.3 prop action item, with final-draft machine check C16 (a voice word misplacement / b missing rate / c vague words / d prop action and tone conflict) as fallback.
Continuous character accounting · placement must enter the frame (continued in V6.9.3 · cross-segment continuity protection): the model has zero memory per segment —a person in the previous segment's image disappearing into thin air in this segment and then appearing again in the next is the most typical cross-segment continuity error. Four iron rules: ① placement must enter the frame: [Blocking statement] Every person who has been placed must be brought into frame at least once within this segment's shot linesat least once(primary shot / out-of-focus figure / full-scene group shot / reaction shot all count). If a shot truly cannot be given, write it explicitly in the blocking parentheses 段末离画 / 退场·去向(machine check C17); ② Re-entry must be accounted for: when writing that a person who left re-enters the frame, the blocking must state 段初自门口入画 and similar methods; ③ No blocking drift: for people whose movement is not written between adjacent segments, position/facing/supporting object must not change without explanation (if the previous segment says "half a step behind Xiulan," the next segment cannot suddenly become "beside the Eight Immortals table"); ④ (present) marking convention: [core characters] no mark on the line = this segment has dialogue or a clear action reaction,（在场）= visible in frame but no dialogue and no reaction—people at both levels must enter the blocking statement; inventing other marks is forbidden. For the procedure see model-adapters.md §5.0.1; final machine check C17 catches it (single-person blocking segments do not trigger it).
First-shot anchoring · shot-size consistency · posture continuity · gaze assignment (new in V6.9.4 · single-segment prevention of wasted shots): four iron rules (all defects come from the actual wasted shots in "Rebirth 70" EP01 segment 1): ① First-shot character anchor—the 1st shot of every segment must bring at least one blocked person into frame (a name or an attributed body part, such as "Lin Qingyue's withered hand"); an empty-shot opening must write "Empty shot: ..." at the start of the shot line, and each segment may have at most 1 such shot; with no anchor in the first shot, the model renders an empty scene (in testing it produced "only a bed," machine check C18a); ② Shot-size triple consistency—close-up/extreme close-up carries only a single partial subject; space-level events (rainstorm smashing through the roof / the whole house catching fire) are carried by medium shot or wider; mixing "close-up + top-down shot + whole-house event" is forbidden (machine check C18d); ③ Posture continuity—once a change such as propping up/sitting up/standing up occurs, subsequent shots in the same segment must explicitly state "keeps propping and sitting / has sat up" or explain it out of focus; falling down must specify the process of falling back/collapsing; silent regression is forbidden (machine check C18b); ④ Gaze assignment—a speaking shot must specify the gaze target: "gaze lands on XX's face"; when speaking while holding/fiddling with a prop, it must add "gaze does not land on the prop" (a prop raised toward the light source draws the gaze most strongly, and the model by default makes the character stare at the prop; machine check C18c). For the procedure see model-adapters.md §5.0.2; final machine check C18 catches it.Additional notes from footage review feedback (§5.0.3): prop lighting double insurance (for reflective-prop shot lines, write positive material description; machine check C18e) / expression anchor (cold, hard dialogue must explicitly write specific facial muscle actions; the model defaults to a sweet smile for young women) / appearance anchor (core character lines write an age anchor, and close-ups restate "the same face" to prevent face swapping) / cross-shot presence and orientation restatement (restate "right rear of frame" within shot lines for non-principal characters, and bring them in again in the segment-ending shot) / no model-added shots.
Three fixes for failed prohibitions + column discipline (three fixes newly added in V6.9.5 / column discipline further added in V6.9.6)：Video models are insensitive to negation and sensitive to word roots— The empty shot in "Forbidden Empty Mirror," the glowing in "Not Glowing," and the smile in "Face Without Smile": if the forbidden concept's word root itself remains in the prompt (even tagged with "not/forbidden"), it will be rendered as an image word (in testing, three negative-form prohibitions were all ignored across two rounds: an empty room opening, the jade pendant self-glowing throughout the entire film, and the verdict text paired with a toothy sweet smile). High-correlation visual concepts (light effects/facial expressions/opening shots) must be changed in three ways: ① Replace words— The forbidden word root appears in the promptin any position(including mandatory prohibition items) zero occurrences; rewrite as positive descriptions (prop → "deep cyan solid matte stone material, jade surface in the original cyan jade color,"zero light-effect description, see C20d; facial expression → "lips pressed tight, mouth corners pressed downward, jaw clenched"; first frame → "image subject: <person name>…" single-subject opening); ② Subject-first sentence pattern— the first clause of each segment's shot 1 is the image subject (person name/body part/"image subject:"/person reference); starting with an environment preposition induces an empty-scene opening; ③ Trigger removal— replace the action/composition that induces the forbidden effect (the prop must not co-occur with lamplight: only fiddle with/wear it, do not bring it near any light source — the old suggestion "raise it beside the oil lamp to receive light" is abolished; testing showed it likewise ignites).Column discipline (V6.9.6 continued additions): [Mandatory prohibition items] are read by the model as an "avoidance checklist,"only hold negative lists(image: deformation/abnormal hands/costume changes/modern elements/subtitles/Logo/watermark; text; audio) —positive execution instructions are forbidden from this column(writing them in counts neither as a prohibition nor as an instruction; in testing, the two entries in this column for segment 1 — "the entire film's first frame is the close-up shot of the person in shot 1" and "the prop maintains everyday stone and metal texture" — were equivalent to waste paper),negative-form prohibitions are likewise forbidden from this column(the forbidden word root = an image word; "the jade pendant does not glow or transmit light" written into the prohibition items is also waste paper); positive instructions must all fall back to [Timeline Storyboard] shot lines/[Aspect Ratio Style]/[Positioning Statement]. Watermarks/deformation and otherlow-correlation defectsstill use the negative-form fallback; the two categories do not conflict.Concrete light-source mapping (C19d): for reflective props, shot lines must not write "reflects/mirrors the flame" — the model does not understand "reflection" and will read "reflects + light-source noun" as the light source appearing on the prop;zero light-effect description, write only the material's original color ("jade surface in the original cyan jade color") — the old suggestion "light-spot language" is abolished (V6.9.7 testing: the light-spot sentence produced a jade pendant that looked like a lighter igniting a cluster of fire, see C20d).Few adjectives in shot lines: positive descriptions keep only the necessary items per shot (subject/action/material/light, one each); stacking modifiers dilutes key information and lowers compliance rate. For the procedure see model-adapters.md §5.0.4, final draft machine check C19(a bare prohibition words/b light-path triggers/c first-shot subject priority/d concrete light-source mapping/e prohibition-item column discipline) as fallback.
First-shot single-subject anchoring and concretization (V6.9.7 new addition · preventing "two faces"): three iron rules (the pathology comes from the actual-footage breakdown of "Rebirth 70" EP01 segment 1 shot 1; the user's exact words: "a total of two faces were generated, and the biggest culprit is right here"): ① The anchoring sentence may contain only one named subject——Ban "X and Y in the same frame"-type coordinate-subject sentence patterns: the model reads coordinate nouns + "same frame" as "there are two subjects in the image," and thus renders the same subject's hand and face astwo heads(❌ "the first frame is this hand and her face in the same frame" → ✅ "image subject: Lin Qingyue curled up on the earthen kang, sickly pale. One of her hands rests outside the torn blanket"); if two characters truly need to be in the same shot, it is carried by the named dual subjects of the [blocking declaration],do not write "same frame / in the same image / entering the frame together" within the shot line(machine check C20a); ② referential pronouns must carry attribution—when "this hand / that face" has no antecedent, the model willsupply a subject on its own(supplying an extra head); always write "one of her hands / his right hand" (machine check C20b); ③ abstract forms and degree adverbs must all be deleted—"into lines / into strands / like a pour / lightly curled / slightly curled" describeprocess forms the model cannot name, and AI only renders nameable concrete forms (❌ "leaking rain forms lines striking the back of the hand" + "fingers lightly curled" → ✅ "raindrops strike the back of the hand and splash apart").One-sentence criterion: can every word in this sentence be drawn? Whatever cannot be drawn is an adjective—delete it or replace it with a concrete action. ④ Zero light-effect description for reflective props (V6.9.7 continued addition · C20d)—clauses of the light-spot / receiving light / irradiation / shining on / lighting up type are never written; props do not approach lamp fire (tested: "a warm yellow light spot the size of a fingernail" "raised next to the oil lamp" produce footage ofthe jade pendant like a lighter igniting a cluster of fire); write only the material's true color ("the jade surface shows the original celadon color" / "matte metal true color"). For the procedure see model-adapters.md §5.0.2 / §5.0.3 / §5.0.4, final-draft machine check C20(a coordinate subjects / b dangling reference / c abstract forms / d prop light effects · proximity to light source) as the fallback.
Example-sentence verification discipline (V6.9.7 root-cause retrospective addition · treats "rule-layer poison production"): the common source of discarded shots across rounds isa priori creation of example sentences—when in that round, based on understanding, a new sentence is coined and written as a ✅ positive example without actual-shoot verification (even embedded into machine-check repair suggestion copy), the output is copied and becomes the source of the next round's failure. Iron rule:positive examples may only be recovered [empirically] from outputs that passed actual shooting; newly coined sentences must be marked [pending verification · not actually shot] and are forbidden from entering machine-check repair suggestion copy; repair suggestions write only criteria and do not coin sentences; criterion upgrades must clean all downstream documents in the same round. For the procedure see model-adapters.md §5.0.5, package-level machine check ⑰ rule self-consistency gateas the fallback (the rule document itself passes its own prohibitions).
MiniMax H3 format branch line (V6.9.8 addition · A-lock coordinate third option)：only performs format conversion, does not change any upstream rules—converts the final seven-segment feeding draft produced by this package into H3 native format (FL2VA family T2VA/I2VA/FL2VA/L2VA → three fields + instruction line; Ref2VA → six-segment format). Two-checkpoint routing, seven-segment→H3 field mapping (includingtone-position inheritance: tone is written in <d> outside;blocking must enter the frame inheritance: H3 has no blocking column; every character who has been placed must be given at least one shot in each segment or explicitly written as leaving the frame/exiting;First-shot anchoring and single-subject inheritance：[Shot 1] Must carry one named subject, parallel-subject sentence patterns prohibited;Prohibition rewriting inheritance: the main text only writes "what is there"; negative forms do not enter the main text),Material cap 9 images/3 videos/3 audios/12 total, 9 image sourcing strategy, two feeding payloads, H1~H12 machine checks see references/h3-adapter.md(★authoritative); H3 official red line of no real-human-face material.Main text English inheritance (2026-10-02): the six main-text paragraphs are all translated into English; dialogue/in-frame text keeps the original language;Voice-over official formula（says in an off-screen voiceover + (Sx) + closed-mouth evidence immediately followed by </d>）；Shot-cutting verb official phrase(the camera/shot cuts to…, dissolve/fade/wipe used only when explicitly required);Multi-speaker risk(official issue #17: same gender must collapse—max out acoustic profile contrast, one speaker per segment, clean up non-addressee vocal words).

IV. Quick Command Routing and Mode Switching

Common Commands	Corresponding functions and underlying library calls
`/写剧本` or `初始化项目`	Start the five-stage gated screenplay engine, output premises, causal beat table and standard formatted screenplay.
`/剧组产出册` or `/生成台账`	Start the P2 digital asset pack: output the "Complete Asset List" → generate images in batches (characters/scenes/props) → deliver the "Asset Image Album" (CHR/AUD/PRP/SCN/Uxx all with asset reference images).
`/数字资产包` or `/资产图册`	P2 core entry point: extract all entities → confirm asset list → lock in through batch image generation → deliver the "Asset Image Album".
`/情绪曲线` or `/情绪节拍`	Start the 12-beat emotion quantification engine, output the full-film tension diagnostic table and generate an emotion fluctuation line chart.
`/角色资产板` or `/角色遗传推导`	P2c·First batch: T1 male and female lead lock face models → kinship derivation T2/T3 anddirect image generation(not textual description) → deliver 4 View character asset reference images.
`/顶视图` or `/机位调度`	Draw a spatial top-down floor-plan camera scheduling diagram, marking CAM1~CAM4 numbers and the 180° action axis.
`/做分镜`	Automatically route dramatic micro-expression flow or action 15-cell dynamic chain PREVIS, applying the three techniques and six forms of transitions;Multi-segment tasks first pass the D lock (choose one of three segment-seam transition methods), and on both sides of the segment seam, pair by pair, use the seam-matching form (`segment-splicing.md`）。Shot length is derived backward from the plot according to Rule 17, equal division prohibited; after drafting, machine check C13 blocks identical shot lengths.
`/台词诊断` or `台词体检`	Start the dialogue specialist doctor, output character language fingerprints, seven-dimensional diagnosis and three-part rewriting.
`/微表情` or `/表情诊断`	Start the FACS micro-expression engine: vague emotion words → AU formula replacement (four-zone decomposition + three intensity levels + timing + English kernel); can produce a character emotion library / FACS cue table, or perform expression review according to the self-check list.
`/语速自检` or `台词语速检查`	Start dialogue-speed-check speech-rate self-check: split sentences $\rightarrow$ count $\rightarrow$ apply tier $\rightarrow$ judge.
`/生成视频提示词`	Output Seedance 2.5 / 2.0 / Jimeng layered independent three-layer decoupled prompt word groups (each group's duration follows the model contract: 2.5 ≤30s / 2.0 ≤15s / Jimeng 5~10s) (⚠️ the aspect ratio must be based on the user-specified 16:9 or 9:16, default fallback to vertical screen prohibited). Pre-feeding must-pass checklist (in order): ① `seedance-render-engine`(three-layer decoupling + duration budget) → ② `aspect-ratio-adaptation`(aspect ratio backfill) → ③ `camera-specs-15rules`(five shot elements + shot-scale ladder) → ④ if there is dialogue, it must pass `dialogue-speed-check`(speech-rate self-check) + `wenxi-micro-expression`(punchline fermentation) + `model-adapters.md` §5.2（the tone slot must be written, same spec as 2.0/2.5); if there is OS/VO, it must pass `vo-os-weaving`(sound-image separation/quota) → ⑤ combat segments are cut by tier `xuanhuan-magic-combat`(R3 xuanhuan)/ `action-previs-15grid`(wuxia) → ⑥ before finalizing, pass `quality-gate-review` → ⑦ pre-finalization machine check**run `scripts/validate_prompt.py <script> --model 2.5
`/H3提示词` or `/转H3`	MiniMax H3 format branch(automatically executed when A locks in H3): convert the current seven-part feed script into H3 native format—mode routing (T2VA/I2VA/FL2VA/L2VA/Ref2VA) → six-part/three-field conversion (tone slot/staging·blocking must be in frame/first-shot anchoring·single subject/prohibition revision method/continuity state/three layers of sound effects/off-screen sound translated according to the official off-screen voiceover formula) → translate the entire body into English(keep dialogue in `<d>[Chinese]` the original text) → 9 image sourcing strategies → before finalizing, run `scripts/validate_prompt.py <稿> --model h3`(H1~H12). For the protocol see `references/h3-adapter.md`(★authoritative).
`/导出工作流参数`	Export a structured payload suitable for WorkBuddy / Doubao workflow / Canvas / API (⚠️ within the payload, `aspect_ratio` backfill the user-specified aspect ratio).
`跳过确认，直接出整集`	Instant direct-output mode: silently execute the full set of industrial constraints in the background, and output directly in the foreground a finished product free of backstage wording contamination.
`/看板` or `编译看板`	compile the project md with one click into asingle-file offline storyboard dashboard（`scripts/build_board_lite.py`, double-click to open; for the spec see `references/storyboard-board-lite.md`）。
`/审查`	Launch the P0/P1/P2 independent quality-check gates and the three-layer relative-level sound diagnosis;revision re-verification entry point—revision-type tasks must run the three-step closed loop of 'local verification → linkage conflict check → full audit' and attach a 'Change Re-verification Report'.