← Back to Skills Marketplace
43
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install madstory
Description
电影级影视分镜设计引擎，支持多平台视频生成（Seedance / Runway / Kling / Sora）。 支持 9 种创作模式：电影创意探索、电商产品展示、UGC 原生广告、电影感品牌短片、多镜头叙事、一镜到底、 爆款复刻、Agent 全链路创作、短剧批量生产。 融合 Harness Engineerin...
README (SKILL.md)
MadStory — 电影级影视分镜技能 v3.1 (分层加载架构)
技能定位
MadStory 是面向电影级内容生产的专业分镜设计引擎，基于 Harness Engineering (PPAF 循环 + R.E.S.T 可靠性模型)。
- 触发词:
MadStory,影视分镜,分镜设计,电影分镜,电商视频,UGC广告,品牌短片,多镜头叙事,一镜到底,爆款复刻,短剧创作,从一句话出片,抖音脚本,AI电影 - 目标用户: 电影导演、广告导演、创意总监、影视制作人、电商运营、品牌营销、内容创作者、短剧制作人、零基础创作者
- 默认时长: 15 秒 / 镜头，支持多镜头序列和长视频 (>1 min)
分层加载架构 (Progressive Loading)
核心原则: SKILL.md 始终加载（本文件）。其他资源按交互阶段按需加载，不预加载。
Layer 0: 始终加载 (本文件, ~1,500 tokens)
- 元数据 + 触发词 + 技能定位
- 9 模式概览表 (每模式 1 行)
- Phase 0-7 阶段名称 (无详细说明)
- 输出规范 + 质量门禁
- 分层加载索引
Layer 1: 按需加载 (~2,000 tokens/次)
以下资源在对应触发条件满足时 读取文件内容:
| 触发条件 | 加载文件 | 内容 |
|---|---|---|
| 用户确认/选择模式 | references/modes_detail.md |
该模式的完整说明、约束、护栏 |
| 进入 Phase N 推导 | references/phases_detail.md |
Phase N 的详细推导指引 |
| 生成/校验提示词 | references/prompt_engineering.md |
5层结构、Negative Prompt模板、多镜头/一镜到底/复刻语法 |
| 需要精确镜头术语 | references/terminology.md |
影视专业术语库 |
| 分镜方案输出前检查 | references/pre_flight_checklist.md |
导演级预检清单 |
| 需要平台参数参考 | references/seedance_v2_rules.md |
Seedance 2.0 提示词工程规范 |
Layer 2: 深度加载 (~3,000 tokens/次)
以下资源仅在特定高级场景下加载:
| 触发条件 | 加载文件/模块 | 内容 |
|---|---|---|
| Mode 8 短剧 + 多场景/多集 | references/short_drama_consistency.md |
短剧全流程一致性管控 (A1-D4) |
| Mode 0 创意探索 | assets/cheat_sheet.json → director_style_references + creative_film_prompts_library |
导演风格参考 + 创意提示库 |
| 要求电影级镜头语言 | assets/cheat_sheet.json → cinematic_camera_language + cinematic_lighting_extended |
运镜/构图/布光预设 |
| 要求声音设计 | assets/cheat_sheet.json → cinematic_sound_extended |
声音分层 + 情感映射 |
| CLI/API 调用 | scripts/mad_story_engine.py |
核心引擎代码 |
| 质量校验 | scripts/director_validator.py |
导演级核验工具 |
| 企业集成 | scripts/api_server.py |
REST API 服务 |
| 批量生产 | scripts/batch_runner.py |
批量流水线 |
| LLM 增强 | scripts/llm_router.py |
多意图拆分/Tournament路由 |
| 多平台适配 | scripts/platform_adapter.py |
平台参数映射 |
| HTML 可视化 | assets/storyboard_template.html |
分镜预览模板 |
9 种创作模式概览
| # | 模式 | 场景 | 输入 | 核心约束 |
|---|---|---|---|---|
| 0 | 电影创意探索 | 概念开发、风格实验 | Text | Generate-and-Filter + Tournament |
| 1 | 电商产品 | 商品详情页、主图动效 | Image | 产品不变形、标签可读 |
| 2 | UGC 原生广告 | 信息流、口播、种草 | Reference | 人脸一致、手势自然 |
| 3 | 电影感品牌短片 | 品牌故事、预告 | Text/Image | 镜头有意图、灯光有逻辑 |
| 4 | 多镜头叙事 | 完整叙事弧线 | Reference | ≤3镜头/次，跨镜头一致 |
| 5 | 一镜到底 | 空间巡游、沉浸展示 | Image序列 | 2-10图，空间连续 |
| 6 | 爆款复刻 | 灵感翻拍、竞品复刻 | Reference+Image | 风格还原、主体替换 |
| 7 | Agent 模式 | 一句话到成片 | Text | 自动意图解析+路径规划 |
| 8 | 短剧创作 | AI短剧、漫剧、小说改编 | Reference+Script | 跨集角色一致 |
模式详情见
references/modes_detail.md
分镜推导阶段 (Phase 0-7)
| Phase | 名称 | 产出 |
|---|---|---|
| 0 | 模式选择与意图澄清 | Mode 确认 + 输入策略 |
| 1 | 核心创意锁定 | Subject + Action |
| 2 | 时间轴与节奏 | Timeline + 关键帧 |
| 3 | 视觉构图 | Composition + 画幅 |
| 4 | 镜头运动 | Camera + Motion Strength |
| 5 | 光影与质感 | Lighting + 风格预设 |
| 6 | 声音设计 | Sound + BGM |
| 7 | 最终合成与输出 | 完整提示词 + Negative Prompt |
Phase 详情见
references/phases_detail.md
交互准则
- 逐阶段推进: 每次只推进一个 Phase，不跳过
- 用专业术语但让外行听懂: 提供选项卡片
- 实时展示草稿: 每完成一个 Phase 展示当前草案
- 主动拦截: 用户塞入过多信息时建议拆分
- 模式引导: Phase 0 必须确认模式
- 负向提示词自动注入: 由系统按 Mode 自动生成
- 质量自检: 输出前对关键约束自查
输出规范
每个分镜方案必须包含:
- STANDARD_PROMPT: 符合 5 层结构的完整正向提示词
- NEGATIVE_PROMPT: 按 Mode 自动生成的负向提示词
- TIMELINE: 15 秒时间轴 + 关键帧描述
- CAMERA: 逐秒/逐镜头运动描述
- MOTION_STRENGTH: 建议值 (1-10)
- DURATION: 固定 15s（或按多镜头分配）
- MODE: 模式标签
- MULTI_MODAL_ADVICE: 参考图/视频/音频建议
- SHOT_LIST (多镜头模式): 分镜头脚本表
质量门禁 (Quality Gates)
- 产品标签不可读或变形 → Ecommerce 不合格
- 人脸在两个镜头间不一致 → UGC / Short Drama 不合格
- 一个镜头内超过 1 个主导运动 → 任何 Mode 不合格
- 缺少 Negative Prompt → 任何 Mode 不合格
- 多镜头序列 > 3 个镜头放在一次生成 → Multi-shot 不合格
- 一镜到底出现跳帧或空间断裂 → One-Shot 不合格
- 爆款复刻风格偏离参考 → Viral Replicate 不合格
- Agent 模式意图解析完全偏离 → Agent 不合格
- 短剧角色跨集不一致 → Short Drama 不合格
Usage Guidance
Install only if you are comfortable with optional executable helpers. Avoid running the API server on a public interface without your own authentication, use only trusted LLM endpoints and API keys, avoid pasting confidential scripts or campaign plans unless remote processing is acceptable, and treat the replication mode as inspiration-only for content you have rights to use.
Capability Tags
Capability Assessment
Purpose & Capability
The core purpose is coherent: a cinematic video storyboard and prompt-generation skill with CLI, batch, validation, and optional API helpers. One advertised mode promotes high-fidelity reference-video replication, which is purpose-aligned for that mode but needs lawful-use limits.
Instruction Scope
The trigger language is broad and mandatory for many general video or script requests, which can route user content into this skill even when the user did not clearly opt into its executable helpers or external processing paths.
Install Mechanism
No hidden installer, post-install hook, obfuscated payload, or undeclared package-install behavior was found. The optional API dependencies are disclosed as fastapi, uvicorn, and pydantic.
Credentials
The optional REST server defaults to binding on 0.0.0.0 without authentication, and the optional LLM router can send raw user content to an OpenAI-compatible endpoint selected by environment variables. These are coherent features but under-scoped for potentially sensitive creative or business material.
Persistence & Privilege
Session and output persistence is user-facing, but CLI save/load accepts arbitrary paths and the API save endpoint writes session JSON to a predictable temp-file pattern and returns the raw path. This is not destructive by itself, but it is broader than necessary.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install madstory - After installation, invoke the skill by name or use
/madstory - Provide required inputs per the skill's parameter spec and get structured output
Version History
v3.0.0
MadStory 3.1 introduces a layered architecture with expanded creative modes and deeper modularization.
- Migrated to a progressive, on-demand file loading framework for better performance and extensibility.
- Added support for 9 creative modes, including new "电影创意探索" (creative exploration).
- Split all detailed guides, cheat sheets, and engineering references into individual reference files for phase, mode, prompt, and platform-specific logic.
- Expanded cross-platform support (Seedance, Runway, Kling, Sora), including platform adapter and multi-modal advice output.
- Improved output/quality gate checks and enforced stricter negative prompt automation.
- Added and organized rich sample assets, deep mode/phase guides, and director-level pre-flight checklists.
Metadata
Frequently Asked Questions
What is MadStory?
电影级影视分镜设计引擎，支持多平台视频生成（Seedance / Runway / Kling / Sora）。 支持 9 种创作模式：电影创意探索、电商产品展示、UGC 原生广告、电影感品牌短片、多镜头叙事、一镜到底、 爆款复刻、Agent 全链路创作、短剧批量生产。 融合 Harness Engineerin... It is an AI Agent Skill for Claude Code / OpenClaw, with 43 downloads so far.
How do I install MadStory?
Run "/install madstory" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is MadStory free?
Yes, MadStory is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does MadStory support?
MadStory is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created MadStory?
It is built and maintained by qomob (@qomob); the current version is v3.0.0.
More Skills