🎬 超级导演系统 · Super Director System

把导演的判断力装进 Claude。 光影、分镜、调度、节奏、美学——数百集 AI 短剧实战打磨出的全套导演工艺,打包成 Claude Skills 插件:丢进一个剧本,走出带调度图的完整分镜和可直接投喂视频模型的提示词。

A director's judgment, installed into Claude. Forged across hundreds of episodes of real AI short-drama production — feed it a script, get back blocking diagrams, a full shot list, and production-ready prompts for Sora / Kling / Veo / Seedance. Jump to English ↓

中文用户从这里开始 ↓ ｜ English speakers: full English docs below ↓

一、这是什么

给 AI 装上"导演脑"的知识库+工作流。它不是提示词模板集——它先像导演一样想(这场戏观众该感到什么?权力关系怎么用站位讲?光从哪来?),再像制片厂一样交付(调度图→分镜表→六段式提示词,层层过检)。

13 个技能,一次安装:

二、工作流程(四步交付)

        剧本 / 一个点子
             │
   ┌─────────▼─────────┐
   │ 第0步 · 美学圣经    │  没有圣经不开机:七步法锁死全剧的看相
   │ (每个项目做一次)    │  (照 91示例圣经 抄结构,10分钟粗版也行)
   └─────────┬─────────┘
             │
   ① 🎬 导演分析          节奏弧 / 镜头段落总量 / 最难一点
             │
   ② 🗺️ 场面调度图        每场一张文字版俯视图:站位+走位线+机位+灯位+轴线
             │            → 跳轴/无源光/空间失忆在图纸期就死掉
   ③ 📋 分镜表            每段≤15s,景别交替,动静匹配,段尾=下段开头
             │
   ④ ✍️ 六段式提示词      每段一条,六大板块焊死,负面词兜底
             │
   🖼️ 图先行生成          先出首帧图审过 → 再图生视频 → 过交付快检六问

三、五分钟上手教程

第 1 步 · 安装(二选一)

# 方式A:Claude Code 插件
/plugin marketplace add fanchengchen1-collab/super-director
/plugin install super-director

方式B:把 skills/ 下的文件夹复制到 ~/.claude/skills/,重启 Claude Code。

推荐运行环境(按体验排序):

为什么要 dist/:技能版是"懒加载"架构(路由表按需读文件),只有 Claude Code 的技能机制能保证执行;其他 AI 拿到多文件包不会主动全读。dist/ 是"饿加载"编译产物——链路内嵌成单文件,每阶段带强制检查清单,任何 AI 照读就能跑完整流程。知识库更新后运行 tools/build_bundle.ps1 一键重编。

第 2 步 · 给你的项目建一本美学圣经(每个项目一次,10-30分钟)

对 Claude 说:

「用七步法帮我的项目《XXX》建一本美学圣经。题材是____,平台是____(竖屏短剧/横屏)。照示例圣经的结构来。」

你会得到:渲染域声明、命门句、视觉根脉、hex 锁死的色彩配方、差异化签名、分区应用表——之后所有产出都长在这本圣经上,保证 60 集像同一部剧。

第 3 步 · 跑一集

「这是 EP01 剧本(粘贴/拖入文件)。按四步流程跑:先出导演分析和全部调度图,然后逐场出分镜表和六段式提示词。圣经用我们刚建的。」

大体量剧集会分轮交付(先分析+调度图,再逐场分镜),这是刻意设计——每轮你都能插手改方向。

第 4 步 · 生成

按"图先行"纪律:先让图片模型出每镜首帧 → 你审(像不像、光对不对)→ 首帧过了再图生视频。角色一致性靠"身份描述块逐字复用+参考图锁脸",手册里有完整流程。

第 5 步 · 翻车了?

把生成结果的问题描述给 Claude:

「这段画面人物像贴在背景上/动作僵硬/太干净像渲染图,帮我诊断改提示词。」

会按"去AI感三方"给出病灶和修正版提示词,改一处测一处。

四、输出长什么样(真实格式示例)

「视频段落 P3」(镜头 S3-01 至 S3-02,总时长 7s)

【核心风格基调】
中式武侠+风雪孤驿杀气临门+北宋雪景寒林美学+沉浸式氛围感+电影级镜头语言
+实拍电影质感,真实光影物理+9:16竖屏构图

【画面质感参数】
老木柜台包浆、粗陶碗豁口、门外雪粒被风卷入,灶火橘暖光为全帧唯一暖锚
(光源=北墙灶膛,向南衰减,近亮远暗)+黛蓝雪夜冷底+光比6:1+呼出白气
+接触阴影清晰,营造《绣春刀》式风雪意境

【角色/主体设定】
店小二:三十岁上下,粗布短打油渍围裙,擦碗的手稳而慢,眼皮低垂神情温吞;
赏金头目:披斗篷带雪进门,按刀而立;保持五官发型服装不变,同一角色,面部清晰不变形

【镜头与分镜脚本】
场景总设定:夜,风雪客栈大堂,灶火为主光源,9:16竖屏
0-4s [S3-01]:中景,固定,大门被推开,头目披雪而入,门外风雪灌进,桌上油灯齐齐一矮,
(门轴呻吟|风雪嘶吼|大堂骤静)
4-7s [S3-02]:特写,固定,柜台后店小二擦碗的手没有停,把碗轻轻放下,布叠得方正,
(布擦陶碗的沙沙声|风声被门板隔弱)
⚡ 景别交替检查:MS-CU ✅
⚡ 动静匹配检查:动-静(骤静收口,形状故意为之) ✅
⚡ 段尾状态:店小二双手按在叠好的布上,抬眼 → 下段从抬眼开始

【技术与输出参数】
4K,30fps,9:16,时长7s,运镜:S3-01固定/S3-02固定

【负面提示词】
避免:横屏构图,背景音乐,配乐,动画感,塑料皮肤,影楼打光,均匀无方向的全局照明,
人物滑步,出厂全新材质

五、为什么它不一样

1. L1/L2 分离:"电影是这样"(永恒规律)与"AI 目前只能这样"(临时约束)严格分层——模型换十代,工艺层一字不改。混写的提示词库,模型一升级整库作废。
2. 调度画成图再拆镜:文字版俯视调度图(站位+走位线+机位+灯位+轴线)前置于分镜——跳轴、无源光、空间失忆在图纸期就死掉。全赛道没有第二家做这个。
3. 三渲染域正交轴:写实/2D动漫/3DCG 各有光影法则,开工先拨域,规则不串味。
4. 先情绪后技术:每颗镜头先答"观众此刻该感到什么"。这是系统第一铁律,也是它产出不像"AI 味模板"的根本原因。

六、开源版 vs Pro 版

开源版=方法论全量+产线够用版:全部导演工艺与美学知识+调度图圣经+实战手册(生成红线/去AI感/角色一致性/硬性规则/交付快检)——照着做,能稳定产出制作精良的成片。

Pro 版=工业化量产线:逐模型脾气红线库与翻车病历本、素材QA评分器(六维打分+自动重生红线)、表情表演库(40情绪×三档+FACS防变脸)、剧组流水线(5工种智能体+双质检打回)、完整首帧锚定与角色资产台账、后期质感SOP、进化引擎。一句话:开源版让你拍出好片,Pro 版让你批量稳定地拍出更好的片。 文件中指向这些编号的引用属有意剥离,非死链。

七、FAQ

Q:支持哪些视频模型? 红线是跨模型通用的(单镜单向/禁焦距数值/5-15s/光线详写);中文提示词适配即梦 Seedance/海螺,英文提示词适配 Kling/Veo/Sora/Runway。用 cinema-director-en 直接输出英文。 Q:必须会用 Claude Code 吗? 不必须。推荐顺序:Claude Code(自动触发)> Codex(已带 openai.yaml 适配)> Kimi K2 Agent(挂载为知识库)。任何能读文件的 AI 对话都能用,见上方"推荐运行环境"。 Q:文档都是中文,英语用户怎么办? 用 cinema-director-en:交互和输出全英文,中文知识库由 Claude 原生阅读,对你透明。 Q:能商用吗? 创作你自己的片子=可以;把本内容做成课程/付费资料卖=不可以(License 见下)。

License

CC BY-NC-ND 4.0(署名-非商业性使用-禁止演绎)。可自由学习、使用、分享;不得商用(含卖课/付费社群),不得改编后再分发。商业合作请开 issue。

🎬 English

A director's judgment, installed into Claude. Lighting, storyboarding, blocking, pacing, aesthetics — a complete film-craft system distilled from hundreds of episodes of real AI short-drama production. Feed it a script; get back blocking diagrams, a full shot list, and production-ready prompts for Sora · Kling · Veo · Seedance · Runway.

What is this?

A knowledge base + director workflow that makes Claude think like a director before it writes a single prompt. It is not a prompt-template pack: it first thinks (what should the audience feel? how does blocking tell the power story? where does the light come from?), then delivers like a studio (blocking diagram → shot list → Six-Block prompts, each stage gated by checks).

13 skills in one install — the English master system cinema-director-en (English interaction, English prompt output, English handbook/blocking-diagram/example-bible), the Chinese master system, and 11 master-director style skills (Zhang Yimou, Wong Kar-wai, Ang Lee, Johnnie To, Tsui Hark, Stephen Chow, Peter Pau, a 71-director deep-profile library, and more).

The deep craft files are written in Chinese — Claude reads them natively and serves you in English. You never touch the file language.

The Workflow (4-step delivery)

        your script / your idea
              │
   ┌──────────▼──────────┐
   │ Step 0 · Style Bible │  no bible, no camera: the 7-step method locks
   │ (once per project)   │  the show's entire look (copy example-style-bible.md)
   └──────────┬──────────┘
              │
   ① 🎬 Director Analysis     pacing arc / shot & segment counts / hardest problem
              │
   ② 🗺️ Blocking Diagram      one text-based overhead map per scene: positions,
              │               movement paths, cameras, lights, axis
              │               → axis jumps, unmotivated light, spatial amnesia
              │                 die on paper, before generation
   ③ 📋 Shot List             segments ≤15s, shot sizes alternate, motion matches,
              │               each segment's end = next segment's start
   ④ ✍️ Six-Block Prompts     one per segment, six fixed blocks, negatives included
              │
   🖼️ Image-first generation  first frames as stills → approve → image-to-video
                              → pre-delivery checklist

Five-Minute Tutorial

Step 1 · Install (either way)

# Option A: Claude Code plugin
/plugin marketplace add fanchengchen1-collab/super-director
/plugin install super-director

Option B: copy the folders under skills/ into ~/.claude/skills/ and restart Claude Code.

Recommended runtimes (ranked by experience):

Why dist/ exists: the skill edition is a lazy-loading architecture (the routing table reads files on demand) — only Claude Code's skill mechanics guarantee that actually happens. Other AIs won't voluntarily read a multi-file package. dist/ is the eager-loaded build artifact: the whole chain compiled into one file, each stage gated by a mandatory checklist any AI can follow. Re-build after updates with tools/build_bundle.ps1.

Step 2 · Build your project's style bible (once per project, 10–30 min)

Say to Claude:

"Build a style bible for my project «Title» using the 7-step method. Genre: ____, format: ____ (vertical short drama / 16:9). Follow the structure of the example style bible."

You get: render-domain declaration, a lifeline sentence, visual roots, a hex-locked color recipe, a differentiating signature, and a zone table — every later output grows from this bible, which is what keeps 60 episodes looking like one show.

Step 3 · Run an episode

"Here is the EP01 script (paste / attach). Run the 4-step delivery: director analysis and all blocking diagrams first, then scene by scene the shot list and Six-Block prompts. Use the bible we just built."

Big episodes deliver in rounds by design — you can steer between rounds.

Step 4 · Generate

Follow the image-first discipline: generate each shot's first frame as a still → approve it (likeness, light) → then image-to-video. Character consistency comes from verbatim identity blocks + reference-image face lock — full flow in the production handbook.

Step 5 · Something looks wrong?

"This segment looks AI-generated — the character floats on the background / motion is stiff / it's too clean. Diagnose and fix the prompt."

Claude applies the three de-AI fixes and returns a corrected prompt — change one thing, test one thing.

What the Output Looks Like

"Video Segment P3" (shots S3-01 to S3-02, total 7s)

[Core Style & Tone]
Chinese wuxia + killing intent at the door of a snowbound inn + Northern-Song
snow-forest aesthetics + immersive atmosphere + cinematic camera language
+ live-action cinematic realism, physically-grounded light + 9:16 vertical composition

[Visual Texture Parameters]
Aged counter wood with hand-worn patina, chipped coarse ceramic bowls, snow grains
blown in through the door; hearth-orange firelight as the frame's ONLY warm anchor
(source = north-wall stove, falling off southward, near-bright far-dark) + indigo
snow-night cold base + contrast ratio 6:1 + visible breath + grounded contact
shadows, evoking the mood of "Brotherhood of Blades"

[Character/Subject Setup]
The innkeeper: about thirty, coarse cloth shirt and oil-stained apron, wiping a bowl
with slow steady hands, eyes lowered, mild expression; Bounty leader: cloaked, snow
on his shoulders, hand resting on his sword hilt. Keep facial structure, hairstyle
and wardrobe unchanged; same characters throughout; faces clear, no morphing

[Shot-by-Shot Script]
Scene master setup: night, snowstorm inn hall, stove fire as key light, 9:16 vertical
0-4s [S3-01]: medium shot, locked camera, the door bursts open, the cloaked leader
steps in trailing snow, wind pours in and every oil lamp flame dips at once,
(door hinge groan | storm howl | the hall goes dead silent)
4-7s [S3-02]: close-up, locked camera, behind the counter the innkeeper's wiping
hand never pauses — he sets the bowl down gently, folds the cloth flat,
(soft rasp of cloth on ceramic | the storm muffled by the door)
⚡ Shot-size alternation: MS-CU ✅
⚡ Motion-energy match: motion → stillness (deliberate hard-stop shape) ✅
⚡ End state: innkeeper's hands rest on the folded cloth, eyes lifting →
   next segment starts on the lifted gaze

[Technical & Output Specs]
4K, 30fps, 9:16, duration 7s, camera: S3-01 locked / S3-02 locked

[Negative Prompts]
avoid: horizontal composition, background music, soundtrack, cartoon look, plastic
skin, studio-portrait lighting, uniform directionless lighting, foot sliding,
factory-new materials

Why It's Different

1. L1/L2 separation — "how cinema works" (permanent law) is strictly separated from "what AI can do today" (temporary constraints). Ten model generations from now, the craft layer won't change a word. Mixed-up prompt libraries die at every model upgrade.
2. Blocking is drawn before shots are broken — a text-based overhead map (positions, paths, cameras, lights, axis) precedes the shot list, so axis jumps, unmotivated light and spatial amnesia die on paper. Nobody else in this category does this.
3. Three orthogonal render domains — live-action, 2D anime and 3D CG each get their own lighting law; you set the domain switch before anything else, and the rules never bleed.
4. Emotion before technique — every shot first answers "what should the audience feel right now." This is Iron Rule #1, and the reason the output doesn't read like AI-flavored templates.

Open vs Pro

Open edition = the full methodology + a workable production line: all craft & aesthetics knowledge, the blocking-diagram bible, and the production handbook (universal red lines / de-AI fixes / image-first consistency / ten hard rules / pre-delivery checklist). Follow it and you ship polished films.

Pro edition = the industrial line: per-model quirk & red-line libraries with tested failure casebooks, a 6-dimension footage QA scorer with auto-regenerate thresholds, a facial performance library (40 emotions × 3 intensities + FACS anti-face-drift), a 5-agent crew pipeline with dual QA rejection, full first-frame anchoring & character asset ledgers, post-production SOP, and the evolution engine. In one line: open gets you a great film; Pro gets you great films at scale, repeatably. In-file references to those module numbers are deliberate separation, not dead links.

FAQ

Q: Which video models? The red lines are model-agnostic (one move per shot / no lens numbers / 5–15s / detailed light). English prompts target Kling, Veo, Sora, Runway; Chinese prompts target Seedance/Jimeng and Hailuo. Use cinema-director-en for English output. Q: Do I need Claude Code? No. Preference order: Claude Code (auto-triggering) > Codex (openai.yaml adapters included) > Kimi K2 Agent (mount as knowledge base). Any file-reading AI chat works — see "Recommended runtimes" above. Q: The deep files are in Chinese — is that a problem? No. Use cinema-director-en: interaction and output are fully English; Claude reads the Chinese knowledge base natively, invisibly to you. Q: Commercial use? Making your own films with it: yes. Selling this content as courses/paid material: no (see License).

License

CC BY-NC-ND 4.0 — free to learn from, use, and share with attribution; no commercial redistribution (including paid courses), no derivative redistribution. For commercial licensing, open an issue.