ai-short-drama

AI Short Drama Refined IP Creation v0.6.0

v0.6.0 changelog（2026-05-16）

SD-002 "City Dating Show" EP01 36 grid × 3 rounds (v2.5/v2.6/v2.7) run + full review reverse-engineering. Fixed from "16/36 catastrophic-level consistency breakage" to "0 catastrophic / 60% ≥4 points" + distilled 2 industrial-grade SOPs.

Key Upgrades

⭐ references/storyboard-frame-industrial.md v0.3.0: Storyboard Image Craft SOP
- 8-segment prompt template (CHARACTER/BACKGROUND/ACTION/SCENE/CAMERA/LIGHT/TEXT/STYLE)
- NOT humans clause must be added: each protagonist separately Xxx body IS [arch] architecture NOT a human wearing X NOT a human with headpiece
- ref library craft bias iron rule: modern skyscrapers easily anthropomorphized / ancient architecture inevitably produces chibi human bodies
- Multi-character SOP: 1 protagonist ref + N text silhouette + bokeh blur
- Dramatic lighting vocabulary (25+ keywords) + Hongguo-grade first frame 10-item checklist
- 11-section complete document (including EP01 v2.7 measured verification grid13/21/22 all 5/5)

⭐ references/jimeng-failure-modes.md v0.1.0 (new): Jimeng 5.0 failure modes + full sensitive word list
- 3 types of fail fully distinguished: InvalidNode (character count exceeded) / generation failed (moderation) / no image_url (network)
- prompt character hard limit: 1500 characters(measured distribution table)
- Villain word replacement table：sinister/menacingly/dark intent → moody/leaning/quiet
- Ambiguous word list：blush/cheeks/extreme close-up → rose tint/medium shot
- Concurrency vs serial rate limiting: long prompt + concurrency = fail, long prompt must run alone
- 4 grid real-world diagnostic cases + craft verification checklist
⭐ scripts/generate_storyboard_images.py added --no-fingerprint flag
- Prevents 02_IP_brief.md's ref style visual fingerprint (neutral / flat / NOT cinematic) from polluting storyboard images' dramatic lighting
- max_workers 8 → 4 (lower concurrency to prevent rate limiting)
- timeout 300 → 420s + retry 2 times (prevent long prompt timeout)

Biggest Lesson

ref 工艺是分镜图工艺的上游。
ref 不纯（chibi 人体），分镜图加 NOT humans 也救不回来。
前期 IP 设计阶段就要避坑：选玻璃塔 / 简单几何，避复杂古建筑。

v0.3.0 changelog（2026-05-08）

SD-001 real-world practice + industrial-grade methodology research distillation. Added Phase 1.5 storyboard image layer + 4-mode video generation.

Key Upgrades

⭐ Phase 1.5 Storyboard Image Layer(industrial-grade 5-step process core)
- scripts/生成分镜图.py: 1-4 candidate keyframes per grid (text2image)
- Storyboard image = video first frame (image2video / frames2video input)
- Modification cost 3 credits/image vs video 55 credits/segment = 18 times cheaper
- Video composition = static image (no longer random)
⭐ 4-Mode Video Generation(auto selects automatically based on file existence)
- image2video: storyboard image first frame + action prompt (default primary use 80%)
- frames2video: first frame + last frame locked (key climax 10%)
- multiframe2video: 2-20 keyframes one-shot continuous (complex action 5%)
- multimodal2video：9 image + 3 video + 3 audio fallback（5%）
⭐ ref 5-8 optimal(don't stuff 9 → attention dispersion)
⭐ @ instruction explicit responsibility syntax(@image1 as character reference / @video1 reference camera movement)
⭐ Key grid first and last frames(climax/villain/inner voice/beat must lock)

Document Additions

references/storyboard-frames-craft.md(Storyboard image workflow + 4-mode selection)

Script additions / changes

scripts/生成分镜图.py(new, Step 3)
scripts/生成分集视频.py(changed, supports 4-mode auto selection)
assets/templates/分镜.json(added video_mode / first_frame_path / last_frame_path / multiframe_paths fields)

v0.2.0 changelog（2026-05-08）

Practical takeaways from SD-001 "Reborn as Xiaotian Quan". 13 failure modes → 13 fixes + industrial-grade standards.

Key fixes

⭐ scripts/生成分集视频.py(replaces bash): subprocess args list directly calls dreamina, avoiding bash eval quoting errors under Chinese/special-character prompts (the old bug caused multimodal2video tasks to be silently dropped by the server, with list_task falsely showing alive for 30+ minutes)
⭐ scripts/ref完备性检查.py: before generating video, forcibly scans all (@xxx_ref.png) references, exits 1 if any are missing locally
⭐ scripts/派生分集文件.py: one-click derivation from 分镜.json of 即梦批量包.md / 镜头清单.md
regex fix: containing full-width Chinese ） stop character (r'@([^\s),）]+_ref\.png)'）
multimodal2video stuck auto-detection: after submission check queue_status == 'Generating', otherwise fall back to list_task to recover submit_id

Pacing upgrade

36 grid × 4-10s variations replaces 60 grid × 3s rigid (multimodal2video minimum is 4s, 3s is invalid)
Hongguo must-explode 7 moves (details references/v21-97-percent-rules.md）
Duration distribution: 4s × 11 + 5s × 14 + 6s × 5 + 7s × 3 + 8-10s × 3 = 36 segments / ~190s

Asset standard upgrade

ref library target 80-150 images(vs v0.1.x 22 images)
single prompt 300-500 characters(vs v0.1.x 100-200 characters)
ref layering: ① master style ② character base multi-angle ③ character expression library ④ key actions ⑤ scene multi-angle ⑥ prop details

Documentation

Added references/troubleshooting.md(13 failure modes + solutions)
Added references/v21-97-percent-rules.md(7 moves + 24-episode calibration template)
Changed references/storyboard-craft.md to 36 grid 4-10s variations
Changed references/jimeng-cli-guide.md to add v0.2.0 upgrade + web-side visibility

Core architecture: 2 Phase

┌─ Phase 1：创作（成本低，反复迭代）─────────────────┐
│  阶段 0：市场情报扫描                              │
│  阶段 1：IP 简报（5 子步）                         │
│  阶段 2：故事架构（4 子步）                        │
│  阶段 3：设定圣经（4 子步）                        │
│  阶段 3.5：ref 图自动生成（dreamina CLI）          │
│  阶段 5：批判 refine                              │
└──────────────────────────────────────────────────┘
                      ↓
┌─ Phase 1.5：分镜图（v0.3.0 工业级核心）⭐ 新增 ─────┐
│  ⚠️ 写 prompt 前必读：                            │
│   - storyboard-frame-industrial.md §5（8 段模板）  │
│   - jimeng-failure-modes.md §1+§2（字数+敏感词）   │
│  对每一集独立运行：                               │
│  3A：写该集分镜.json（v0.3.0 字段：video_mode +    │
│       first_frame_path + last_frame_path）         │
│  3B：python scripts/ref完备性检查.py              │
│  3C：python scripts/生成分镜图.py（每 grid 1-4    │
│       张候选关键帧 = 视频首帧）                    │
│  3D：人工选图（grid01_候选3.png → grid01.png）    │
│  3E：关键 grid（爆点/反派/心声）生首尾帧           │
│       （gridXX_首.png + gridXX_尾.png）           │
└──────────────────────────────────────────────────┘
                      ↓
┌─ Phase 2：出片（成本高，按集解锁）────────────────┐
│  4.A：python scripts/生成分集视频.py              │
│       auto 模式按文件存在自动选 4 模：             │
│       - frames2video（首尾帧锁定）                │
│       - image2video（分镜图首帧 + 动作）          │
│       - multiframe2video（2-20 帧一镜到底）       │
│       - multimodal2video（fallback 多 ref）       │
│  4.B：用户在剪映拼接 + 配音 + BGM + 字幕           │
└──────────────────────────────────────────────────┘

Why split it this way：

Script / ref unsatisfactory → Phase 1 iterate repeatedly, without burning video money
Storyboard image = video first frame → Phase 1.5 confirm the static image first (3 credits/image), if unsatisfactory change prompt and regenerate → video composition fixed (no longer random)
Modification cost: storyboard image 3 credits vs video segment 55 credits = 18 times cheaper, saving 50%+ on video trial and error
Video production cost is high → Phase 2 unlocks per episode, letting the user accept each episode
The real cost of AI short dramas is not creation, it is production. Use Phase 1 + 1.5 to polish the composition into place, then Phase 2 is worth spending money on

Trigger conditions

Any one of the following triggers:

User says "make/write/help me make ... a short drama"
Mentions Hongguo/Fanqie/Douyin/Kuaishou IP adaptation
Mentions Jimeng/Seedance/Kling/Vidu production, and talks about script
Mentions storyboard, storyboard, shot list, shot script
Mentions satisfying drama, angst drama, chasing wife, revenge, rebirth, time travel, live-in son-in-law, grandma awakening, miracle doctor physiognomist, AI comic drama spectacle

Counterexamples that do not trigger:

User already has a complete script and only wants to change one line of dialogue → change it directly
User asks "What are the prospects for the short drama industry" → ordinary Q&A
User asks "What is an AI short drama" → ordinary Q&A

Default parameters (v0.6.0)

Item	Default value	Source
Single-episode duration	180-195s	Hongguo measured optimum (not strictly capped at 180)
Number of episodes	24 episodes(essence) / 72 episodes (regular)	24 episodes saves cost + tight climaxes
Storyboard shots per episode	36 grid × 4-10s variation	multimodal2video minimum 4s, variation prevents mechanical feel
Duration variation distribution	4s×11 + 5s×14 + 6s×5 + 7s×3 + 8-10s×3	v2.1 Hongguo guaranteed-hit iron rule
Video aspect ratio	9:16 vertical	Unified across Hongguo/Douyin/Kuaishou
Video mode	multimodal2video (omni-reference)	Seedance 2.0 Fast VIP, most stable with multiple ref inputs
Subtitles	On(wrapped in Chinese brackets + AIGC label)	Mandatory on Hongguo
ref library target	80-150 images(industrial-grade)	Includes multi-angle action/expression/scene/prop details
Single prompt word count	300-500 characters	Includes action physics + expression frame breakdown + lighting + subtitle specs
Platform	Hongguo (default) + Douyin (second choice)	300 million MAU + traffic funneling
Genre	Interactive selection	Hongguo pure satisfaction / premium suspense / comic drama spectacle / silly light comedy / period satisfaction drama
Payment model	Interactive selection	Free (ads) / paid (unlock)
ref image generation	dreamina text2image 4-grid (multi-angle)	Phase 1 stage 3.5
Video output	dreamina multimodal2video (with automatic visual fingerprint append)	Phase 2 stage 4.C
Pre-check before running video	Must run `python scripts/ref完备性检查.py`	Prevents missing ref → AI free generation → consistency break

Every time it starts, ask the user whether to adjust these.

About AI advantages(user hint, don't push hard but point it out):

The real moat of AI-generated video = hard for live action to shootspectacles (cultivation / apocalypse / Journey to the West / fantasy / superpowers). If the user's subject matter is grandma awakening / urban wife-chasing, which iseasy for live action to shoot, tell the user: "Although AI can also do it, AI's real advantage is in the comic drama spectacle genre. Would you consider switching genres?" — give 1 hint, the user decides.

Phase 1: Creation (5 stages + stage 3.5)

Stage 0: Market intelligence scan (4 substeps)

0.1 Search online for the current Top 5 drama titles

搜索 1：红果短剧 排行榜 <YYYY-MM>
搜索 2：番茄小说 改编 IP <近 30 天>
搜索 3：抖音短剧 <用户题材关键词> 热度
搜索 4：DataEye 短剧观察 / QuestMobile 月报
搜索 5：短剧 平台 审核 <当前年> 红线

0.2 Actually read industry breakdowns of 1-2 hit dramas (core)

WebFetch Plot explanations / industry breakdown articles of 1-2 hit dramas. Preferred sources: woshipm / 36Kr / Tencent News / 21 Economic Network.

If all 403 → fallback to built-in references/red-fruit-data.md。

0.3 Give the user a 3-option menu (do not output the report directly)

Let the user pick a differentiation angle.

0.4 Write to `01_市场情报.md`

See references/market-pulse.md。

Phase 1: IP Positioning (5 substeps)

1.1 Genre selection (key)

5 options for the user to choose: A Hongguo pure-satisfaction style / B premium suspense style / C comic-drama spectacle style / D silly light-comedy style / E period satisfaction-drama style.

New in v0.1.2: Add a note next to option C: "Hard to shoot with real people = AI advantage. If you plan to use AI for production, C is the most cost-effective choice, because real-person crews for cultivation/apocalypse/fantasy genres are extremely expensive, while AI generates directly." But do not push it hard; let the user decide.

See references/genre-flavors.md。

1.2 Theme archetype (1 primary + 1 secondary)

See references/archetype-catalog.md。

1.3 Core hook (draft 3 versions for the user to choose)

1.4 Repetitive satisfaction-point formula (draft 2 versions for the user to choose)

1.5 Write to `02_IP简报.md`

See references/ip-strategy.md。

Phase 2: Story Architecture (4 substeps)

2.1 Three-act structure skeleton

2.2 Key reversal nodes (let the user choose)

2.3 72-episode outline (first give the key 25 episodes for the user to confirm)

2.4 Write `03_完整剧本.md` + `04_节奏地图.json`

See references/story-architecture.md。

Phase 3: Setting Bible (4 substeps)

3.1 Character list (≤ 6 people)

3.2 5 close-up shots for personality anchoring

3.3 Visual motif + contrast points + costume table + Jimeng ref prompt

3.4 Write 3 bibles

05_角色圣经.md + 06_场景圣经.md + 07_道具圣经.md

See references/character-design.md。

Phase 3.5: Automatic ref image generation (new in v0.1.2, key)

This is the last step of Phase 1 and also a prerequisite dependency for Phase 2。

3.5.1 Detect dreamina CLI

which dreamina && dreamina --version

If not installed → guide installation:

curl -fsSL https://jimeng.jianying.com/cli | bash

For detailed installation instructions, see references/jimeng-cli-guide.md。

3.5.2 Detect login status

dreamina user_credit

If not logged in → guide:dreamina login(browser authorization)

If balance is insufficient → prompt the user to top up (cannot continue)

3.5.3 Extract ref prompts

Read the "Jimeng character/scene/prop reference image generation package" section at the end of the 3 bibles, and extract all prompts:

N characters → N character refs (9:16 half-body portraits)
M scenes → M scene four-grids
K props → K prop three-view images

Estimated N+M+K images, each at 2k resolution about ¥0.5-1,total cost about ¥5-15。

3.5.4 Batch call dreamina text2image

Submit each image independently:

dreamina text2image \
  --prompt="<从 05_角色圣经.md 提取>" \
  --ratio=9:16 \
  --resolution_type=2k \
  --download_dir="<项目>/ref图/角色/" \
  --poll=60

Save to the project subdirectory:

SD-XXX_<拼音>/
├── ref图/
│   ├── 角色/
│   │   ├── 周翠英_ref.png
│   │   └── ...
│   ├── 场景/
│   │   └── scene_01_四宫格.png
│   └── 道具/
│       └── prop_01_三视图.png

Or use the project's built-in script:scripts/生成参考图.sh

3.5.5 Verify + user confirmation

After ref images are generatedthe user must take a look：

Are the character face shapes correct?
Are the costume colors correct?
Is the overall scene atmosphere correct?

Incorrect → modify prompt → rerun (low cost, rerunning a single image is about ¥1)

OK → mark 项目元数据.json of ref_done = true, proceed to Phase 5 critique.

Phase 5: Critique refine (4 substeps)

5.1 6-dimension self-evaluation + anti-cliché detection + platform review scan

5.2 If hit → must fix → patch the original file → 09_Revision_Log.md

5.3 Major revision → Auto-archive `归档/v1.X_<日期>/`

Phase 1 Completion Criteria：

6-dimension total score ≥ 50/60
Anti-cliché Tier-1 blacklist: 0 hits
All platform reviews passed
All ref images OK
User confirms "ready to produce"

See references/critic-checklist.md。

Phase 2: Production (unlocked per episode)

Prerequisites: Phase 1 complete + ref images ready + user explicitly says "start production".

Each episode runs 4.A → 4.D independently.

Stage 4.A: Select the episode to produce

Phase 1 已完成 ✅。准备进 Phase 2 按集出片。

成本提醒：每集 60 段视频，约 ¥X-XX（按当前 dreamina 单价计算）。

要先出哪一集？建议顺序：

1. 第 1 集（开篇，最关键，先出测试 ref 图效果是否符合预期）
2. 第 8 集（一卡 / 一爆，付费短剧首付费点）
3. 第 27 集（二卡）
4. 第 50 集（三卡 / 终极反转）
5. 第 72 集（结局）
6. ...

挑 1 集开始。出完该集，确认效果再决定下一集。

Default recommendation: Produce Episode 1 first as a trial to see how the ref images perform in the video. If not good → go back to Stage 3.5 to revise ref → re-produce Episode 1. Once Episode 1 is OK, batch-produce the rest.

Stage 4.B: Write the storyboard + script + Jimeng batch package for that episode

If that episode's 分集/第XX集_<集名>/ does not yet exist → follow references/scripting-craft.md(script) + references/storyboard-frame-industrial.md(storyboard image 8-segment prompt + ref process bias) + references/jimeng-failure-modes.md(sensitive words / character limit / concurrency throttling) to write:

剧本.md
分镜.json（60 grid）
镜头清单.md
即梦批量包.md(60-segment prompt)

Stage 4.C: dreamina image2video automatic production

Read 分集/第XX集_<集名>/即梦批量包.md 's 60-segment prompt,call segment by segment dreamina image2video：

dreamina image2video \
  --image="ref图/角色/周翠英_ref.png" \
  --prompt="<段 N 的 prompt>" \
  --duration=3 \
  --ratio=9:16 \
  --download_dir="分集/第XX集/视频段/" \
  --poll=60

Each segment takes about 5-15 seconds to generate (dreamina server-side queue). 60 segments take about 30-90 minutes.

Or use the project's built-in script:scripts/生成分集视频.sh <集编号>

Video segments are saved to:

分集/第01集_<集名>/
├── 剧本.md
├── 分镜.json
├── 镜头清单.md
├── 即梦批量包.md
└── 视频段/
    ├── 段01.mp4
    ├── 段02.mp4
    ├── ...
    └── 段60.mp4

Stage 4.D: CapCut assembly (manual)

The skill does not call CapCut, but provides complete guidance:

Create a new 9:16 / 1080×1920 / 30fps project in CapCut
Import in segment order 1-60
Transitions between segments: "Dissolve 0.3s" (use "White flash 0.2s" for hype segments)
Voiceover: CapCut TTS or real person
BGM: Search the CapCut music library using keywords from 02_IP简报.md
Subtitles: Add subtitles for all dialogue, font size 60, white text with black outline
AIGC label (mandatory for Hongguo 2026): 5s each at the beginning and end
Output MP4 H.264 / 1080P

Phase 2 Single-Episode Completion Criteria

All 60 video segments ready
User has watched the rough cut in CapCut
User says "OK this episode is good"
Mark 项目元数据.json 's episodes_video_done.append(集号)

Then go to 4.A to select the next episode.

⭐ v0.6.0 Visual Consistency SOP (required reading)

v0.6.0 update: The original v0.1.4 8-item SOP has been overturned by three rounds of SD-002 real-world testing / split into 3 industrial-grade documents. Beginners only need to read this section + the 3 references.

Storyboard image prompt process → references/storyboard-frame-industrial.md

8-segment template (CHARACTER/BACKGROUND/ACTION/SCENE/CAMERA/LIGHT/TEXT/STYLE)
1 protagonist ref + N silhouette multi-character SOP
Dramatic lighting vocabulary + Hongguo-grade first-frame 10-item checklist
ref library process bias (modern skyscrapers tend to be objectified / ancient architecture must produce chibi human bodies)

ref image prompt process → references/ref-prompt-industrial.md

6-segment identity block (IDENTITY/BODY/FACE/ATTIRE/LAYOUT/STYLE)
Place the visual fingerprint in the first 200 tokens at the beginning of the prompt
Objectified IP "IS the object NOT humanoid" must be added
4-view label watermark trap

Jimeng 5.0 troubleshooting + sensitive words → references/jimeng-failure-modes.md

Distinguish 3 types of failures: InvalidNode / generation failed / no image_url
prompt hard character limit: 1500 characters (under v0.6.0 measured conditions)
Villain word / ambiguous word / Chinese sensitive word replacement table
Long prompts must be run individually (do not run 4+ concurrently)

3 items from the old v0.1.4 SOP that were overturned(must know):

❌ Old SOP "scripts automatically append visual fingerprint" → v0.6.0 uses --no-fingerprint to disable (ref style words will pollute the storyboard image's dramatic lighting)
❌ Old SOP "4 distinct angles" → only applies at the ref stage, do not write it into the storyboard image prompt
❌ Old SOP "after changing the prompt, must delete the old image and regenerate" → still holds, but the script will automatically skip ones that already exist; after changing the prompt you must manually delete the old files before running

See references/visual-consistency-sop.md(the old v0.1.4 document has been greatly slimmed down, keeping only the cost / process parts).

Project directory location rules (v0.1.2)

When the skill creates a short drama project,the project root directoryis determined in the following order:

User explicitly specifiesa path → use the specified path
Current cwd is a short drama workspace(contains aiForShortDrama / short-drama / 短剧 and other keywords, or already has a SD-XXX_*/ subdirectory) → create directly under cwd SD-XXX_<拼音>/
Current cwd is the user's home directory or a general workspace → default suggestion D:\hrdai\aiForShortDrama\SD-XXX_<拼音>/(the user's standard short drama workspace, already created on this machine)
None of the above → ask the user for the path, do not decide on your own

Default creation workspace：D:\hrdai\aiForShortDrama\

Every newly created short drama project goes under this directory, for centralized management + reuse of ref images / character assets.

Project file structure (v0.1.2)

D:\hrdai\aiForShortDrama\                       ← 用户标准短剧工作区
├── 即梦 CLI 体验指南.pdf                       ← 用户参考文档
├── README.md                                    ← 工作区 README
│
├── SD-001_<拼音>/                              ← 第 1 部短剧
│   ├── README.md                                ← 项目首页
│   ├── 项目元数据.json
│   │
│   ├── 01_市场情报.md
│   ├── 02_IP简报.md
│   ├── 03_完整剧本.md
│   ├── 04_节奏地图.json
│   ├── 05_角色圣经.md
│   ├── 06_场景圣经.md
│   ├── 07_道具圣经.md
│   │
│   ├── ref图/                                   ← 阶段 3.5
│   │   ├── 角色/
│   │   │   ├── <角色>_ref.png
│   │   │   └── ...
│   │   ├── 场景/scene_01_四宫格.png
│   │   └── 道具/prop_01_三视图.png
│   │
│   ├── 分集/                                    ← Phase 2 按集解锁
│   │   ├── 第01集_<集名>/
│   │   │   ├── 剧本.md
│   │   │   ├── 分镜.json
│   │   │   ├── 镜头清单.md
│   │   │   ├── 即梦批量包.md
│   │   │   └── 视频段/                          ← 阶段 4.C 产出
│   │   │       ├── 段01.mp4
│   │   │       └── ...
│   │   └── 第08集_<集名>/...
│   │
│   ├── 08_自我批判.md
│   ├── 09_修订记录.md
│   │
│   └── 归档/                                    ← 大改前 snapshot
│       └── v1.0_<日期>/
│
├── SD-002_<拼音>/                              ← 第 2 部短剧
│   └── ...
│
└── SD-003_<拼音>/                              ← 第 N 部

`项目元数据.json` key fields

{
  "project_id": "SD-001",
  "project_name": "<>",
  "skill_version": "v0.1.2",
  "version": "v1.0",
  "phase": "Phase1_创作 | Phase1_完成 | Phase2_出片 | 完成",
  "stage_progress": {
    "0_市场": "completed",
    "1_IP": "completed",
    "2_故事": "completed",
    "3_设定": "completed",
    "3.5_ref图": "completed",
    "5_批判": "completed",
    "Phase2": "in_progress"
  },
  "ref_done": true,
  "episodes_video_done": [1, 8],
  "next_action": "出第 27 集（二卡）"
}

Rules for interacting with the user

Break each stage into 3-5 substeps, and for each substep give anoption menufor the user to choose from (A/B/C), not a complete document for the user to read through
Proactively do a market intelligence scan at the start of each session + WebFetch to actually read hit shows
If the user's subject matter falls in a declining period / red ocean, honestly tell the user and give alternative directions
If the user has not specified genre / episode count / payment, ask
AI production tip: if the user's subject matter falls within the range that is easy to shoot with real people (grandma / mother-in-law and daughter-in-law / chasing the wife), give a one-time hint that "anime-drama spectacle style is AI's advantage", but do not push it hard
Phase 1 → Phase 2 must have an explicit switch confirmation: only when the user says "start production" can you enter Phase 2
Phase 2 must be unlocked episode by episode, you cannot produce multiple episodes at once
Automatically archive before every major change

dreamina CLI integration

See references/jimeng-cli-guide.md. Key points:

Install:curl -fsSL https://jimeng.jianying.com/cli | bash（Mac/Linux/Windows Git Bash）
Login:dreamina login(browser authorization)
Check balance:dreamina user_credit
Text-to-image:dreamina text2image --prompt=... --ratio=9:16 --resolution_type=2k --poll=60
Image-to-video:dreamina image2video --image=... --prompt=... --duration=3 --poll=60
Async query:dreamina query_result --submit_id=<ID>

The skill calls the dreamina command directly through the Bash tool, no API key management needed.

Startup command

The user might trigger it like this:

"Make a ... short drama"
"Help me write a ... themed feel-good drama"
"I have an idea: ..."
"Adapt ... this novel into a short drama"

Response template for the first sentence after startup:

收到。本技能 2 阶段架构：

Phase 1 创作（成本低，反复迭代）
  - 市场情报 + IP 简报 + 故事架构 + 设定圣经 + ref 图自动生成 + 批判
  - dreamina text2image 生 ref 图，成本约 ¥5-15

Phase 2 出片（成本高，按集解锁）
  - dreamina image2video 单集 60 段，按集生成

先 Phase 1，跟你 5 阶段互动确认。3 个最关键问题：

1. 流派：A 红果纯爽 / B 精品悬疑 / C 漫剧奇观（AI 优势）/ D 沙雕轻喜 / E 年代爽剧
2. 集数：默认 70，要改吗？
3. 付费：免费（广告）/ 付费（解锁，第 8 集首付费）

回答后启动阶段 0。

Final delivery

When Phase 1 is completed, provide a "Creation Completion Checklist":

✅ 《XXX》Phase 1 创作完成

📋 IP 信息
- 流派、集数、平台、对标、钩子、爽点公式

📊 创作产出
- 9 个编号文件（市场情报 → 自我批判）
- N 张 ref 图（角色 / 场景 / 道具）
- 72 集大纲 + 节奏地图

🎬 下一步
准备进 Phase 2？告诉我先出哪一集（建议第 1 集试水）。

When each episode of Phase 2 is completed, provide an "Episode Production Completion Checklist".

About this skill

v0.6.0 changelog（2026-05-16）

Key Upgrades

Biggest Lesson

v0.3.0 changelog（2026-05-08）

Key Upgrades

Document Additions

Script additions / changes

v0.2.0 changelog（2026-05-08）

Key fixes

Pacing upgrade

Asset standard upgrade

Documentation

Core architecture: 2 Phase

Trigger conditions

Default parameters (v0.6.0)

Stage 0: Market intelligence scan (4 substeps)

0.1 Search online for the current Top 5 drama titles

0.2 Actually read industry breakdowns of 1-2 hit dramas (core)

0.3 Give the user a 3-option menu (do not output the report directly)

0.4 Write to 01_市场情报.md

Phase 1: IP Positioning (5 substeps)

1.1 Genre selection (key)

1.2 Theme archetype (1 primary + 1 secondary)

1.3 Core hook (draft 3 versions for the user to choose)

1.4 Repetitive satisfaction-point formula (draft 2 versions for the user to choose)

1.5 Write to 02_IP简报.md

Phase 2: Story Architecture (4 substeps)

2.1 Three-act structure skeleton

2.2 Key reversal nodes (let the user choose)

2.3 72-episode outline (first give the key 25 episodes for the user to confirm)

2.4 Write 03_完整剧本.md + 04_节奏地图.json

Phase 3: Setting Bible (4 substeps)

3.1 Character list (≤ 6 people)

3.2 5 close-up shots for personality anchoring

3.3 Visual motif + contrast points + costume table + Jimeng ref prompt

3.4 Write 3 bibles

Phase 3.5: Automatic ref image generation (new in v0.1.2, key)

3.5.1 Detect dreamina CLI

3.5.2 Detect login status

3.5.3 Extract ref prompts

3.5.4 Batch call dreamina text2image

3.5.5 Verify + user confirmation

Phase 5: Critique refine (4 substeps)

5.1 6-dimension self-evaluation + anti-cliché detection + platform review scan

5.2 If hit → must fix → patch the original file → 09_Revision_Log.md

5.3 Major revision → Auto-archive 归档/v1.X_<日期>/

Stage 4.A: Select the episode to produce

Stage 4.B: Write the storyboard + script + Jimeng batch package for that episode

Stage 4.C: dreamina image2video automatic production

Stage 4.D: CapCut assembly (manual)

Phase 2 Single-Episode Completion Criteria

⭐ v0.6.0 Visual Consistency SOP (required reading)

Project directory location rules (v0.1.2)

Project file structure (v0.1.2)

项目元数据.json key fields

Rules for interacting with the user

dreamina CLI integration

Startup command

Final delivery

0.4 Write to `01_市场情报.md`

1.5 Write to `02_IP简报.md`

2.4 Write `03_完整剧本.md` + `04_节奏地图.json`

5.3 Major revision → Auto-archive `归档/v1.X_<日期>/`

`项目元数据.json` key fields