Video Prompt Reverse Engineer 视频提示词逆向工程师

中文 | English

---

中文说明

一个用于 AI 视频生成工作流的技能（Skill）：把参考视频逆向拆解成具体、可直接使用的 AI 视频生成提示词，支持 Sora、Kling/可灵、Runway、Vidu、Pika，以及 Seedance 2.0 / 2.5（即梦 / Doubao / 火山引擎）原生格式。

本技能只做分析与提示词文本，不直接提交生成任务。只学习镜头逻辑与视觉方法，不复刻真实人物、品牌、台词与版权音乐。

能做什么

• 快速通用版：短视频、产品展示、Vlog 片段，输出"七要素拆解 + 成品提示词 + 负面提示词"。
• 完整专业版：广告、短剧、MV、多镜头复刻，输出 18 维度逐镜头时间轴 + 分镜成品提示词 + 全局风格锚定。
• Seedance 2.0 / 2.5 原生格式：在标准拆解之上，按 Seedance 官方语法改写成品提示词：
  • @引用 多模态参考（每个参考标注用途 + 排除项）
  • 分时段描述（每段含结束状态，保证镜头衔接）
  • 音频括号语法：音乐 ( )、音效 ``、对白 { }、字幕 【 】
  • 2.0 复杂度按 4–15 秒匹配；2.5 支持最长 30 秒一镜到底、最多 50 个参考素材
• 五类题材补充：人物/剧情、产品/广告、风景/旅行、音乐/MV、美食/手工/教程。

文件结构

video-prompt-reverse-engineer/
├── SKILL.md                                # 技能主入口：原则、模式选择、输出结构
├── references/
│   ├── video-reverse-prompt-template.md    # 通用/专业版拆解模板与题材补充
│   └── seedance-prompt-guide.md            # Seedance 2.0/2.5 原生提示词规范
├── agents/openai.yaml                      # OpenAI 兼容 agent 配置
└── README.md

怎么用

1. 安装到支持 Skill 的客户端（如 ZCode / Claude Code 等兼容 Agent Skills 规范的工具），把本目录放进技能目录（如 ~/.zcode/skills/ 或项目 .codex/skills/），或直接 clone：

   git clone https://github.com/coolxxxx/video-prompt-reverse-engineer.git
   

2. 触发技能：上传或给出一个参考视频，然后说：

   • “帮我反推这个视频的提示词”（通用版）
   • “用完整专业版逐镜头拆解这个广告”（专业版）
   • “反推成 Seedance 2.5 的提示词”（Seedance 原生格式）

3. 产出内容（以 Seedance 目标为例）：

   • 视频概览 + 七要素/逐镜头拆解
   • 参考素材清单（@图片1 人物形象参考、排除项等）
   • 生成设置（模式/时长/比例，不写入提示词正文）
   • Seedance 原生提示词正文（分时段 + 音频括号 + 一致性约束）
   • 负面提示词 + 原创化边界声明

Seedance 提示词示例骨架

[主体与动作] 陶艺师双手在转盘上给陶坯塑形，随坯体升高缓慢收拢指尖
[场景与环境] 日光工作室，木质工作台，背景为虚化的工具架
[视觉风格] 纪录片质感，暖色调，自然光
[镜头运动] 从手部特写缓慢拉远至中景
[分时段] 0–5 秒：双手塑形（结束状态：坯体成型）；5–10 秒：拉远展示工作室全貌
[音频] (轻柔的木吉他背景音乐)  
[保持一致性] 人物服装、发型、工具位置全片不变

友链

• LINUX DO — 本项目经由 LINUX DO 社区分享与交流，感谢社区与佬友的支持和反馈。

---

English

A Skill for AI video generation workflows: reverse-engineers a reference video into concrete, ready-to-use AI video prompts. Supports Sora, Kling, Runway, Vidu, Pika, and the native Seedance 2.0 / 2.5 format (Jimeng / Doubao / Volcano Engine).

This skill only produces analysis and prompt text — it never submits generation jobs. It learns shot logic and visual grammar only; it does not reproduce real people, brands, dialogue, or copyrighted music.

What It Does

• Quick universal mode: short clips, product demos, Vlog moments — seven-element breakdown + finished prompt + negative prompt.
• Full professional mode: ads, short dramas, MVs, multi-shot replication — 18-dimension shot timeline + per-shot prompts + global style anchoring.
• Seedance 2.0 / 2.5 native format: rewrites finished prompts using Seedance’s official syntax:
  • @ multimodal references (each with purpose + exclusions)
  • Timestamped stages (each with an explicit end state for continuity)
  • Bracketed audio notation: music ( ), SFX ``, dialogue { }, captions 【 】
  • 2.0 complexity matched to 4–15 s; 2.5 supports 30 s one-takes and up to 50 reference assets
• Five genre supplements: characters/drama, product/ads, landscape/travel, music/MV, food/craft/tutorial.

File Structure

video-prompt-reverse-engineer/
├── SKILL.md                                # Main entry: principles, modes, output shape
├── references/
│   ├── video-reverse-prompt-template.md    # Universal/professional templates + genre supplements
│   └── seedance-prompt-guide.md            # Seedance 2.0/2.5 native prompt spec
├── agents/openai.yaml                      # OpenAI-compatible agent config
└── README.md

How To Use

1. Install into any Skill-compatible client (e.g. ZCode / Claude Code and other Agent-Skills-compatible tools): place this folder in your skills directory (e.g. ~/.zcode/skills/ or project .codex/skills/), or clone:

   git clone https://github.com/coolxxxx/video-prompt-reverse-engineer.git
   

2. Trigger it: upload or link a reference video, then ask:

   • “Reverse-engineer this video into a prompt” (universal)
   • “Do a full professional shot-by-shot breakdown of this ad” (professional)
   • “Convert this into a Seedance 2.5 prompt” (native format)

3. Output (Seedance target example):

   • Video overview + element/shot breakdown
   • Reference asset list (@Image 1 character look, exclusions, etc.)
   • Generation settings (mode/duration/ratio — kept out of the prompt body)
   • Seedance-native prompt body (stages + audio brackets + consistency constraints)
   • Negative prompt + originality boundary statement

Seedance Prompt Skeleton Example

[Subject & Action] A potter shapes a clay vessel on a spinning wheel, fingertips slowly closing as the form rises
[Scene] Sunlit studio, wooden workbench, blurred tool rack in the background
[Visual Style] Documentary feel, warm tones, natural light
[Camera] Slow pull-back from a hand close-up to a medium shot
[Stages] 0–5 s: hands shaping (end state: vessel formed); 5–10 s: pull back to reveal the studio
[Audio] (Soft acoustic guitar)  
[Consistency] Clothing, hairstyle, and tool positions unchanged throughout

Friend Links

• LINUX DO — This project is shared and discussed on the LINUX DO community. Thanks for the support and feedback.

License

MIT