🎬 漫剧老李 AIGC 创作全流程 Skill
V6.9.8 轻量版 · Short-Drama Director Suite
🎨 面向 AIGC 视听创作 的工业化全流程技能包 从创意立项到视频模型投喂提示词 —— 给 AI 助手 / LLM 直接调用
🎥 生产管线
📋 立项锁定 → ✍️ 门控编剧 → 🎙️ 台词诊断 → 🖼️ 资产锁定 → 🎬 文武双模分镜 → 🚀 提示词封装 → ✅ 独立质检 → 🖥️ 离线看板
🏆 赞助商
感谢以下赞助商对本项目的支持：
| LOGO | 类型 | 赞助商名称 | 说明 | 网站 |
|---|---|---|---|---|
| （LOGO 占位） | — | 虚位以待 | 赞助位开放中——让你的产品出现在众多 AI 短剧创作者面前 | 联系我们 |
💡 招商说明：赞助位对模型渠道、算力平台、API 服务商及创作团队开放。 合作请联系：邮箱
lixiaoxiao9888@gmail.com· 微信lisenlin998
🎬 分镜引擎专项赞助用户
感谢以下用户赞助分镜引擎的开发（开放中）：
你的头像将出现在这里。
| 赞助用户 | 联系方式 |
|---|---|
| 汤圆（用老李 skills 出剧 · 包算力 · 包发行） | 微信：fush989 |
🧭 这是什么
一套 给 AI 助手执行的 AIGC 创作规则库（OpenClaw Skill / Agent 技能包），不是面向消费者的成品剧本——把 SKILL.md 与 references/ 放进 Agent 技能目录即可调用。
它把「AIGC 视听创作」拆成一条可复用的工业化管线，核心解决三件事：
- 不漂移 —— 角色 / 场景 / 道具先出图锁定，再进分镜（Asset-First：资产图先于分镜），跨镜跨集一致性有据可依；
- 不返工 —— 前置强制锁定（模型版本 / 画幅）+ 投喂前必查清单 + P0~P2 独立门禁 + 稿件级机检，把错误拦在生成之前；
- 能直接投喂 —— 输出七段式结构化提示词，整组复制即可喂给视频模型，不需要二次翻译；走 MiniMax H3 时自动转成 H3 原生六段式。
🎯 适用场景
管线与题材解耦，凡「剧本 → 分镜 → 视频模型投喂」的 AIGC 创作都可套用：
- AI 短剧 / 漫剧 / 短视频剧集（本包最典型的高密度场景）
- AI 广告片 / 品牌宣传片 / 概念片
- 分镜预演（PREVIS）与动态分镜
- 武打 / 动作场面设计（现代格斗、武侠、仙侠、二次元、科幻）
- 情绪戏 / 微表情专项（FACS 表情工程）
- 任何需要跨镜头角色一致性的 AI 视频生产
题材覆盖：现代都市 / 古装武侠 / 玄幻仙侠 / 悬疑 / 科幻 / 赛博 / 水墨国漫 —— 含六大电影级视听美学预设（诗意 / 纪实 / 港派 / 赛博 / 水墨 / 国漫）。
🔥 核心能力
| 阶段 | 能力 | 说明 |
|---|---|---|
| 📋 立项 | 前置强制锁定 | 模型版本（三选一：Seedance 2.5 / Seedance 2.0 / MiniMax H3 格式支线）+ 画幅（16:9 / 9:16 / 21:9 三选一）+ 段缝衔接方式（多段任务）——未指定不动笔，禁止默认 |
| ✍️ 剧本 | 五阶门控引擎 | Premise → Structure → Beat → Entity → Page；含三幕结构、因果节拍表与专业剧本页排版 |
| 📈 情绪 | 12 节拍曲线 | 全片张力 0~10 量化 + 折线图脚本；波峰/波谷与第 9 节拍反杀峰值校验 |
| 🎙️ 台词 | 七维诊断 + 语速自检 + 语气位 | 三维语速档（3.5~5 字/秒基准）、超长句拆镜、金句发酵气口；对白三要素之三（HOW THEY SOUND）——台词统一 台词：【角色（语气）："…"】，2.0/2.5 同规格必写 |
| 🖼️ 资产 | Asset-First 数字资产包 | 角色 4 View 资产板（五官 10 项 + 服装材质 6 项）/ 场景空间图三件套 / 道具图；《资产图册》CHR/AUD/PRP/SCN/Uxx 台账；A/B/C 分级、依赖 A0~A3 按需触发 |
| 🎥 空间 | 一行式站位 + 分级 | 【站位声明】四要素（位置/依托物/朝向/姿态）+ 轴线锁 + 位移链；空间 S0~S4 分级；180° 轴线；人物账连续·摆位必入画（摆了位的人本段必须被镜头带到，否则写明离画/退场） |
| 🎬 分镜 | 文武双模 + 首镜锚定 | 文戏微表情六阶段递进 / 武戏 15 秒 PREVIS 完播评分（R1 写实 · R2 商业高燃 · R3 玄幻）；镜头规格 15 铁律 + 三手法六式衔接；首镜锚定·景别一致·姿态延续·视线指派（首镜必带具名主体）+ 首镜单主体（禁「X 与 Y 同框」并列句式） |
| 😀 表演 | FACS 微表情引擎 | 情绪 AU 化：四区拆解 + 强度三档 + 真假笑铁律 + 情绪→AU 配方库与 AU 字典（增强模块 · 按需触发） |
| ⚔️ 资料库 | 武学 / 玄幻 | 23 门武学 + 9 套剑法 + 兵器 + 轻功 + 11 环杀招；玄幻法术 R3（法宝 / 五行术法 / 能量具象 / 对军清场） |
| 🚀 提示词 | 七段式主格式 + H3 格式支线 | 【画幅风格】→【场景资产】→【核心人物】→【站位声明】→【时间轴分镜】→【音效】→【强制禁止项】；2.5 走官方声音字符（<> {} () 【】）+ Hit-Stop 顿挫 + 任务模式二选一（全能参考 / 首尾帧）；选 H3 时自动转 Ref2VA 六段式（subject_definitions / summary / retention_analysis / detailed_description / overall_soundscape / non_diegetic_music） |
| 🚫 禁令 | 禁令失效三改法 | 视频模型对否定不敏感、对词根敏感——被禁概念的词根哪怕挂着"不/禁"也会被渲染；禁止项栏目纪律（只装负面清单，禁混正向执行指令） |
| ✅ 质检 | P0~P2 门禁 + 稿件级机检 | 时长 / 台词 / 跳轴 / 画幅 / 资产锚点一致性 + 声音三层相对电平 + 改动后三步复验铁律；validate_prompt.py 出稿机检 C1~C20（Seedance 侧）/ H1~H12（H3 侧，--model h3） |
| 🖥️ 看板 | 离线分镜看板 | 项目 md（唯一真相）→ 单文件 HTML，双击即开；P2.5 出图词与 P4 投喂词必同板 |
| 🛡️ 合规 | 安全转译词典 | 把血腥/断肢类动作转译为「气浪震散 / 火花消散 / 重创定格」，降低平台风险（不承诺 100% 通过） |
🖥️ 适用平台与模型
| 支持 | |
|---|---|
| Agent 平台 | OpenClaw · WorkBuddy · 豆包智能体 / 扣子(Coze) · Dify 及通用 LLM（支持豆包 JSON 分块，防上下文截断） |
| 视频模型 | Seedance 2.5 / 2.0（强制二选一，语法契约彻底分流）+ MiniMax H3（格式支线：只做七段式→H3 原生格式转换，单段 4–15s）；即梦为次选 / 历史兼容；可灵自 V6.8 起移出正式支持（LEGACY） |
| 输出 | 七段式投喂提示词组（Seedance）· H3 Ref2VA 六段式稿 · Canvas / API 结构化载荷 · 单文件离线分镜看板 · 情绪曲线图 |
⚙️ 快速开始
# 安装到 OpenClaw
openclaw skills install ./short-drama-director --as short-drama-director
# 自检包完整性（41 模块静态 + 语义回归检查 ①~⑱ 组）
python3 short-drama-director/scripts/check_package.py
# 出稿机检（FAIL 清零才交付）
python3 short-drama-director/scripts/validate_prompt.py <稿>.md --model 2.5|2.0|h3
# 项目 md → 单文件离线看板
python3 short-drama-director/scripts/build_board_lite.py <项目.md>
# 12 节拍情绪曲线
python3 short-drama-director/scripts/generate_emotion_curve.py
给 AI 助手的最小执行序列：读 SKILL.md → 完成 A/B（+多段任务 D）锁 → 定动作强度档（R1/R2/R3）→ 有对白先过语速自检 + 语气位 → 资产出图锁定 → 分镜 → 按 ★ prompt-feeding-checklist.md 生成投喂词 → 跑 validate_prompt.py 机检 → 过 quality-gate-review
🧮 常用指令
/写剧本 · /拆资产 · /资产图册 · /角色资产板 · /顶视图 · /做分镜 · /台词诊断 · /语速自检 · /微表情 · /情绪曲线 · /生成视频提示词 · /H3提示词 · /导出工作流参数 · /看板 · /审查 · 跳过确认，直接出整集
完整规则与指令路由见 short-drama-director/SKILL.md
📦 项目结构
lixiaoxiao9888-create/manju-laoli-skill/
└── short-drama-director/ # V6.9.8 轻量版主包（49 文件 · 41 规则库）
├── SKILL.md # 总控路由 + 工作法 + 模块索引（法则 0~25）
├── references/ # 41 部专业规则库
│ ├── asset-first-pipeline.md # Asset-First 六阶段（★流程权威）
│ ├── model-adapters.md # 七段式主格式唯一权威定义 + 语气位/摆位/首镜锚定
│ ├── ★ h3-adapter.md # MiniMax H3 格式支线（★：两 checkpoint 路由/字段映射/9 张取材策略/H1~H12）
│ ├── aspect-ratio-adaptation.md # 画幅三选一锁定与全流程路由（★权威）
│ ├── segment-splicing.md # 段缝衔接（D 锁：末态双保险/换机位/空镜）
│ ├── vo-os-weaving.md # 旁白与独白视听编织（OS 声画分离红线）
│ ├── seedance-render-engine.md # 时长预算 + 三层解耦
│ ├── ★ facs-micro-expression.md # FACS 微表情引擎（★表情域权威）
│ ├── storyboard-board-lite.md # 轻量离线分镜看板规范
│ ├── ★ prompt-feeding-checklist.md # P4 投喂必查清单（★执行序列权威）
│ └── ... 共 41 模块
├── scripts/
│ ├── check_package.py # 静态 + 语义回归自检（①~⑱ 组）
│ ├── validate_prompt.py # 稿件级投喂机检（C1~C20 / --model h3 → H1~H12）
│ ├── build_board_lite.py # 分镜看板编译器（md → 单文件 HTML）
│ └── generate_emotion_curve.py # 12 节拍情绪曲线绘图
└── README.md / LICENSE / CHANGELOG.md
📜 更新历史
- V6.9.8 轻量版（2026-09-20）— 新增 MiniMax H3 格式支线：A 锁并列第三选项，只做「七段式 → H3 原生格式」转换（
h3-adapter.md：两 checkpoint 路由 / 六段式映射 / 素材上限 9 图·3 视频·3 音频·总 12 / 9 张取材策略 / 两套投喂载荷）；出稿机检新增 H1~H12（--model h3） - V6.9.3 ~ V6.9.7 轻量版（2026-09-19）— 五轮实拍复盘驱动的定向修订：对白语气位（C16）· 人物账连续·摆位必入画（C17）· 首镜锚定·景别一致·姿态延续·视线指派（C18）· 禁令失效三改法 + 栏目纪律（C19 / C19e）· 首镜单主体锚定与具象化（C20）· 示范句验证纪律（⑰ 规则自洽门）；机检升至 C1~C20
- V6.9 轻量版（2026-09-18）— 新增 〇节 D·段缝衔接方式三选一锁（末态双保险 / 换机位硬切 / 空镜过渡，治"段落间衔接生硬"）/ 旁白·独白视听编织规程（篇幅卡口 · OS 声画分离红线 · VO 冷漠审视）/ 时长守恒·禁拉伸凑满（法则 17 修订 + C15 门）
- V6.8 轻量版（2026-09-15）— 可灵移出正式支持（降级 LEGACY）/ 角色资产统一 4 View / 「七段式」命名统一、末尾参数行废止 / 空间 S0~S4 · 资产 A0~A3 · FACS 三级按需触发 / 2.5 新增任务模式二选一（全能参考 · 首尾帧）
- V6.7 轻量版（2026-09-14）— 双锁同规格（模型版本 + 画幅三选一强制锁）/ 按官方文档校准 Seedance 2.5 / 吸收 FACS 微表情引擎三件套 / 吸收轻量离线分镜看板
- V6.5 Multi-Agent（2026-09-01）— Asset-First 六阶段 + 多 Agent 平台适配 + Seedance 2.5 深度适配 + 静止段禁用铁律（AI 视频无静止表演能力，情绪蓄力必须在运动中完成）
- 旧版归档：manju-laoli-skill-legacy
🇬🇧 English Overview
V6.9.8 Lite · Short-Drama Director Suite
An industrial, full-pipeline creation skill package for AIGC audiovisual works — from concept lock-in to video-model feed prompts — built for AI agents / LLMs to call directly.
🎥 Production Pipeline
📋 Lock-in → ✍️ Gated Screenwriting → 🎙️ Dialogue → 🖼️ Asset Locking → 🎬 Dual-mode Storyboard → 🚀 Prompt Packaging → ✅ QC → 🖥️ Offline Board
🧭 What This Is
An AIGC creation rulebase for AI agents (OpenClaw Skill / Agent skill package) — not a finished screenplay. Drop SKILL.md plus references/ into your agent's skill directory.
It turns AIGC audiovisual creation into a reusable industrial pipeline, solving three things:
- No drift — characters / scenes / props are generated and locked as reference images first, before any storyboard (Asset-First: asset images precede storyboards);
- No rework — mandatory front-end locks (model version / aspect ratio) + a pre-feed checklist + independent P0–P2 gates + shot-level machine QC stop errors before generation;
- Feed-ready — outputs a structured 7-section prompt you can copy straight into the video model; when MiniMax H3 is selected, it converts automatically into H3's native 6-section format.
🎯 Use Cases
The pipeline is genre-agnostic: anything that goes script → storyboard → video-model feed fits.
- AI short dramas / animated shorts / short-form series (the densest use case)
- AI commercials / brand films / concept films
- Storyboard previsualization (PREVIS)
- Action & fight choreography design (modern combat, wuxia, xianxia, anime, sci-fi)
- Emotional & micro-expression work (FACS expression engineering)
- Any AI video production needing cross-shot character consistency
Genres: modern urban / period wuxia / fantasy xianxia / mystery / sci-fi / cyberpunk / ink-wash guoman — including six cinematic audiovisual aesthetic presets.
🔥 Capabilities
| Stage | Capability | Description |
|---|---|---|
| 📋 Lock-in | Mandatory front-end locks | Model version (choose one: Seedance 2.5 / Seedance 2.0 / MiniMax H3 format branch) + aspect ratio (16:9 / 9:16 / 21:9) + seam-joining mode for multi-segment work — no work starts without them |
| ✍️ Script | 5-stage gated engine | Premise → Structure → Beat → Entity → Page; three-act structure, causal beat sheet, professional script page |
| 📈 Emotion | 12-beat curve | Full-episode tension on a 0–10 scale + chart script; peak/valley and 9th-beat reversal validation |
| 🎙️ Dialogue | 7-dimension diagnosis + speed check + tone slot | Three speed tiers (3.5–5 chars/s baseline), long-line shot splitting, punchline breathing room; the third element of dialogue (HOW THEY SOUND) — every line is written as 台词：【角色（语气）："…"】, mandatory for both 2.0 and 2.5 |
| 🖼️ Assets | Asset-First digital asset pack | Character 4-View sheets (10 facial + 6 material items) / 3-piece scene maps / prop sheets; 《Asset Ledger》 CHR/AUD/PRP/SCN/Uxx; A/B/C grading, A0–A3 dependency triggering |
| 🎥 Space | One-line stance + grading | Stance declaration (position / anchor / facing / pose) + axis lock + movement chains; S0–S4 grading; 180° rule; placement must appear on screen (anyone placed in the stance block must be covered by at least one shot, else marked as exiting) |
| 🎬 Storyboard | Dual-mode + first-shot anchoring | Drama: 6-stage micro-expression progression / Action: 15s PREVIS completion scoring (R1 realistic · R2 commercial · R3 fantasy) + 15 camera rules + 3-technique / 6-form transitions; first-shot anchoring / consistent shot size / pose carry-over / gaze assignment, plus single-subject anchoring (no "X and Y in the same frame" phrasing) |
| 😀 Performance | FACS micro-expression engine | Emotions as Action Units: 4-zone breakdown + 3 intensity tiers + real-vs-fake smile rule + emotion→AU recipes & AU dictionary (on-demand module) |
| ⚔️ Libraries | Martial arts / fantasy | 23 schools + 9 sword forms + weapons + lightness skills + 11 finisher links; fantasy spell combat R3 |
| 🚀 Prompts | 7-section main format + H3 branch | 【Style】→【Scene Asset】→【Characters】→【Stance】→【Timeline Storyboard】→【SFX】→【Hard Negatives】; 2.5 uses official sound chars + Hit-Stop + task-mode choice; selecting H3 converts it into Ref2VA's six sections |
| 🚫 Negatives | Three rewrites for failed bans | Video models are insensitive to negation but sensitive to word roots — a banned concept's root still renders even with "no/never" attached; the negatives block holds only negative lists, never positive execution instructions |
| ✅ QC | P0–P2 gates + shot-level lint | Duration / dialogue / axis / aspect / anchor consistency + 3-layer audio levels + mandatory 3-step re-verification after edits; validate_prompt.py runs C1–C20 (Seedance) / H1–H12 (H3, --model h3) |
| 🖥️ Board | Offline storyboard board | Project md (single source of truth) → single-file HTML, double-click to open |
| 🛡️ Compliance | Safety translation dictionary | Rewrites gore/amputation into shockwave / spark / freeze-frame language to lower platform risk (no 100% guarantee) |
🖥️ Platforms & Models
| Support | |
|---|---|
| Agent platforms | OpenClaw · WorkBuddy · Doubao Agent / Coze · Dify, plus generic LLMs (Doubao JSON chunking supported) |
| Video models | Seedance 2.5 / 2.0 (mandatory choice, fully split syntax contracts) + MiniMax H3 (format branch: 7-section → H3 native format, 4–15s per segment); Jimeng = secondary / legacy; Kling dropped to LEGACY as of V6.8 |
| Outputs | 7-section feed prompt sets (Seedance) · H3 Ref2VA six-section scripts · Canvas / API payloads · single-file offline storyboard board · emotion curve chart |
⚙️ Quick Start
# Install into OpenClaw
openclaw skills install ./short-drama-director --as short-drama-director
# Lint package integrity (41-module static + semantic regression, groups ①~⑱)
python3 short-drama-director/scripts/check_package.py
# Shot-level feed lint (zero FAIL to ship)
python3 short-drama-director/scripts/validate_prompt.py .md --model 2.5|2.0|h3
# Project md → single-file offline board
python3 short-drama-director/scripts/build_board_lite.py
# 12-beat emotion curve
python3 short-drama-director/scripts/generate_emotion_curve.py
Minimal agent sequence: read SKILL.md → complete A/B (and D for multi-segment) locks → pick action tier (R1/R2/R3) → run the dialogue speed check + tone slot if there is dialogue → lock assets → storyboard → generate feed prompts via ★ prompt-feeding-checklist.md → pass validate_prompt.py → pass quality-gate-review
🧮 Commands
/写剧本 · /拆资产 · /资产图册 · /角色资产板 · /顶视图 · /做分镜 · /台词诊断 · /语速自检 · /微表情 · /情绪曲线 · /生成视频提示词 · /H3提示词 · /导出工作流参数 · /看板 · /审查 · 跳过确认，直接出整集
Full rules & routing: short-drama-director/SKILL.md
📦 Structure
lixiaoxiao9888-create/manju-laoli-skill/
└── short-drama-director/ # V6.9.8 Lite main package (49 files · 41 rulebooks)
├── SKILL.md # Router + working method + module index (laws 0–25)
├── references/ # 41 professional rulebooks
│ ├── asset-first-pipeline.md # Asset-First 6 stages (★flow authority)
│ ├── model-adapters.md # 7-section format authority + tone/placement/first-shot anchoring
│ ├── ★ h3-adapter.md # MiniMax H3 format branch (★: 2-checkpoint routing / field mapping / 9-image sourcing / H1–H12)
│ ├── aspect-ratio-adaptation.md # Aspect-ratio locking & routing (★)
│ ├── segment-splicing.md # Segment-seam joining (D-lock: end-state twin insurance / camera swap / empty shot)
│ ├── vo-os-weaving.md # Narration & inner-monologue weaving (OS audio-visual split rule)
│ ├── seedance-render-engine.md # Duration budget + 3-layer decoupling
│ ├── ★ facs-micro-expression.md # FACS engine (★expression authority)
│ ├── storyboard-board-lite.md # Offline board spec
│ ├── ★ prompt-feeding-checklist.md # P4 pre-feed checklist (★sequence authority)
│ └── ... 41 modules total
├── scripts/
│ ├── check_package.py # Static + semantic regression lint (groups ①~⑱)
│ ├── validate_prompt.py # Shot-level feed lint (C1–C20 / --model h3 → H1–H12)
│ ├── build_board_lite.py # Board compiler (md → single-file HTML)
│ └── generate_emotion_curve.py # 12-beat emotion curve chart
└── README.md / LICENSE / CHANGELOG.md
📜 Changelog
- V6.9.8 Lite (2026-09-20) — adds the MiniMax H3 format branch: a third option at the A-lock that only converts the 7-section script into H3's native format (
h3-adapter.md: 2-checkpoint routing / six-section mapping / material caps of 9 images · 3 videos · 3 audios · 12 files total / 9-image sourcing strategy / two feed payloads); shot-level lint gains H1–H12 (--model h3) - V6.9.3 – V6.9.7 Lite (2026-09-19) — five rounds of revisions driven by real shoot retros: dialogue tone slot (C16) · placement must appear on screen (C17) · first-shot anchoring / shot-size consistency / pose carry-over / gaze assignment (C18) · three rewrites for failed bans + negatives-block discipline (C19 / C19e) · first-shot single-subject anchoring and concreteness (C20) · exemplar validation discipline (⑰ self-consistency gate); lint raised to C1–C20
- V6.9 Lite (2026-09-18) — adds the D-lock: three ways to join segments (end-state twin insurance / camera swap / empty shot), the narration & inner-monologue weaving spec (length quota · OS audio-visual split · detached VO gaze) and duration conservation / no stretching to fill (law 17 revised + gate C15)
- V6.8 Lite (2026-09-15) — Kling demoted to LEGACY / character assets unified to 4 Views / "7-section" naming unified, end-of-prompt parameter line abolished / space S0–S4 · asset A0–A3 · FACS three-level on-demand triggering / 2.5 gains task-mode choice
- V6.7 Lite (2026-09-14) — dual locks at equal weight / Seedance 2.5 calibrated against official docs / FACS engine absorbed / lightweight offline board absorbed
- V6.5 Multi-Agent (2026-09-01) — Asset-First 6 stages + multi-agent platform adapters + Seedance 2.5 deep adaptation + no-static-scene iron rule
- Legacy archive: manju-laoli-skill-legacy
📄 License
MIT · For creative & learning use; comply with platform content policies
Made with ❤️ for AIGC Creators
🙌 贡献者与团队
感谢参与产品设计、开发、测试、内容与社区建设的成员：
老李
项目发起者 · 总编导 · 微信：lisenlin998
用作品说话。
evo
H3 支线贡献者 · GitHub：coco38622
让 H3 支线与官方文档逐字对齐。
JY
海外剧剧本 SKILL 贡献者
好故事，不分国界。
虚位以待 —— 期待你加入。