[English](./README.md) | [中文](./README_zh.md) # Seedance Prompt Skill for Claude Code A Claude Code custom skill that turns Claude into a professional AI video prompt engineer for ByteDance's **Seedance 2.0** (即梦) video generation platform. Provide a creative idea in natural language, and this skill will generate structured, production-ready Chinese prompts that you can paste directly into the Seedance 2.0 platform to produce cinematic AI videos. ## Features ### Ten Core Prompt Generation Capabilities | # | Capability | Description | |---|-----------|-------------| | 1 | **Pure Text Generation** | Generate videos purely from text descriptions, no reference materials needed | | 2 | **Consistency Control** | Maintain character/product/scene consistency across shots using reference images | | 3 | **Camera & Motion Replication** | Replicate camera movements, complex actions, and pacing from reference videos | | 4 | **Creative Template / VFX Replication** | Reproduce creative transitions, ad templates, and cinematic effects from reference videos | | 5 | **Story Completion** | Auto-generate storylines from storyboard images or scripts | | 6 | **Video Extension** | Extend existing videos forward or backward with smooth continuity | | 7 | **Sound Control** | Voice cloning, dialogue generation, and sound effect design | | 8 | **One-Take Long Shot** | Generate seamless long takes that flow continuously across scenes | | 9 | **Video Editing** | Modify existing videos: swap characters, alter plots, add/remove elements | | 10 | **Music Beat Sync** | Synchronize visual rhythm precisely with music beats | ### Advanced Techniques - **Timestamp Storyboarding** — Precise per-second control for 15-second videos (e.g., `0-3s: ...`, `4-8s: ...`) - **Technical Parameter Specification** — Set aspect ratio, frame rate, color grading, and lens style upfront - **Negative Prompting** — Explicitly exclude unwanted elements (watermarks, subtitles, etc.) - **Multi-Segment Stitching** — Automatically split videos longer than 15 seconds into extension-chained segments with continuity notes ### Multi-Modal Reference System Supports the Seedance 2.0 `@` reference syntax for combining multiple input modalities: - **Images** (`@图片1` ~ `@图片9`) — Character references, scene backgrounds, first/last frame control - **Videos** (`@视频1` ~ `@视频3`) — Camera motion, action, effects, and pacing references - **Audio** (`@音频1` ~ `@音频3`) — Background music, voice tone, sound effect references ### Scenario-Specific Prompt Strategies - **E-commerce / Advertising** — Product showcases, 360-degree spins, 3D exploded views - **Short Drama / Dialogue** — Scripted scenes with character dialogue, emotion tags, and sound design - **Xianxia / Fantasy Animation** — Cinematic martial arts, spell effects, transformation sequences - **Science Education** — 4K medical CGI, transparent anatomy visualizations - **Music Videos / Beat Sync** — Widescreen compositions, multi-image beat matching ### Built-in Prompt Asset Libraries The skill ships with curated vocabularies for: - **Camera Language** — Shot types, camera movements, angles, pacing, focus, transitions - **Visual Styles** — Film grains, color palettes, art movements, lighting setups, animation styles ## Installation ### Prerequisites - [Claude Code CLI](https://docs.anthropic.com/en/docs/claude-code) installed and authenticated - A Seedance 2.0 (即梦) account at [jimeng.jianying.com](https://jimeng.jianying.com) for using the generated prompts ### Steps 1. **Clone the repository** ```bash git clone https://github.com/songguoxs/seedance-prompt-skill.git ``` 2. **Copy the skill into your project's `.claude/skills/` directory** ```bash # Navigate to the project where you want to use the skill cd /path/to/your/project # Create the skills directory if it doesn't exist mkdir -p .claude/skills/seedance # Copy the skill file cp /path/to/seedance-prompt-skill/.claude/skills/seedance/SKILL.md .claude/skills/seedance/SKILL.md ``` 3. **Verify installation** Start Claude Code in your project directory: ```bash claude ``` Then type `/seedance` or simply describe your video idea. Claude should activate the Seedance prompt skill automatically. ### Alternative: Global Installation To make the skill available across all your projects, place it under your home directory: ```bash mkdir -p ~/.claude/skills/seedance cp /path/to/seedance-prompt-skill/.claude/skills/seedance/SKILL.md ~/.claude/skills/seedance/SKILL.md ``` ## Usage ### Quick Start Start Claude Code and describe what video you want to create: ``` You: 帮我生成一段赛博朋克风格的城市夜景视频提示词 ``` Claude will ask you a few clarifying questions (duration, aspect ratio, reference materials), then generate 2-3 prompt versions ready to paste into Seedance 2.0. ### Triggering the Skill The skill activates automatically when you: - Use the `/seedance` command - Mention keywords like: `Seedance`, `即梦`, `视频提示词`, `视频生成`, `AI视频`, `短剧`, `广告视频`, `视频延长` - Ask to "generate video prompts" or "create Seedance prompts" ### Interactive Workflow The skill follows a structured 4-step workflow: ``` Step 1: You describe your idea → "一段古风武侠打斗" Step 2: Claude confirms parameters → duration, aspect ratio, references, style preferences Step 3: Claude generates prompts → 2-3 versions for videos ≤15s, or multi-segment plan for longer videos Step 4: You refine → adjust timing, camera, dialogue, effects, or style ``` ## Examples ### Example 1: Pure Text — Short Action Scene **Input:** ``` 帮我写一段仙侠战斗的视频提示词,15秒,横屏 ``` **Output (one of 2-3 versions):** ``` 15秒仙侠高燃战斗镜头,金红暖色调,0-3秒:低角度特写主角蓝袍衣摆被热浪吹得猎猎飘动, 双手紧握雷纹巨剑,剑刃赤红电光持续爆闪,地面熔岩翻涌冒泡,远处魔兵嘶吼着冲锋逼近, 主角低喝"今日,便以这柄剑,镇尔等邪祟!",伴随剑鸣与熔岩咕嘟声;4-8秒:环绕摇镜 快切,主角旋身挥剑,剑刃撕裂空气迸射红色冲击波,前排魔兵被击飞碎裂成灰烬,伴随剑气 破空声与魔兵惨嚎;9-12秒:仰拍拉远定格慢放,主角跃起腾空,剑刃凝聚巨型雷光电弧劈 向魔兵群;13-15秒:缓推特写主角落地收剑的姿态,衣摆余波微动,冷声道"此界之门, 不容踏越",音效收束为余音震颤与渐弱风声。 ``` ### Example 2: Multi-Modal Reference — Product Ad **Input:** ``` 我有一张可乐产品图,想做一个炫酷的产品展示广告 ``` **Output:** ``` @图片1中的可口可乐饮料,360度高速旋转2圈后,突然停住蓄力分裂成了3个部分进行展示。 随后分解后的可口可乐饮料罐的上中下三部分快速向内旋转合成,一罐完整的可口可乐饮料, 3D渲染产品展示特效,动感产品特效展示 参考素材: - @图片1:可乐产品正面高清图 ``` ### Example 3: Short Drama with Dialogue **Input:** ``` 写一段霸道总裁短剧,15秒,女主撕合同反杀的名场面 ``` **Output:** ``` 画面(0-5秒):特写女主撕契约镜头,纸屑飘落,总裁单膝跪地伸手阻拦,眼神慌乱, 女主侧身躲开,嘴角挂着冷漠笑意 台词1(总裁,卑微慌乱):苏晚!契约还没结束,你不能走!我给你钱,给你地位! 画面(6-10秒):女主抬脚避开他的手,将撕碎的契约纸扔在他脸上,镜头扫过周围宾客的 窃窃私语 台词2(女主,冷漠反杀):契约?顾总,当初是你说,我连给你提鞋都不配,现在求我?晚了! 画面(11-15秒):总裁僵在原地,脸上沾着纸屑,女主转身昂首离开,红裙裙摆飘动 音效:华丽又带张力的背景音,契约撕碎的声响,宾客轻微的窃窃私语声 时长:精准15秒 ``` ### Example 4: Long Video (>15s) — Multi-Segment **Input:** ``` 做一段30秒的仙侠出场短片 ``` **Output:** ``` ## 超长视频提示词(总时长约30秒) **主题**:剑修云海仙山迎敌 **总段数**:2段 **建议比例**:16:9 --- ### 第1段(0-15秒)—— 正常生成 **生成时长**:15秒 #### 提示词 15秒仙侠镜头,0-5秒:俯拍云海翻涌中的仙山全景,镜头缓缓下推穿过云层; 6-10秒:剑修站在山巅悬崖边,背对镜头,衣袍随风飘动,远处魔气升腾; 11-15秒:剑修缓缓转身面向镜头,拔剑出鞘,剑刃金光闪烁,目光坚毅低声道 "来了",定格在剑修持剑面向镜头的画面。 #### 衔接点 本段结尾画面:剑修正面持剑站在山巅,目光前视,远方魔气翻涌。 --- ### 第2段(15-30秒)—— 视频延长 **操作**:将第1段生成的视频上传为 @视频1 **生成时长**:15秒 #### 提示词 将@视频1延长15秒。0-5秒:接上段剑修持剑画面,数十只暗影魔兽从远处魔气中 飞扑而来,剑修纵身跃起迎敌;6-10秒:空中激战,剑气纵横,魔兽被斩成灰烬 粒子消散,镜头环绕快切;11-15秒:剑修落地收剑,身后爆炸的金色粒子缓缓飘散, 镜头缓推特写剑修侧脸,音效渐弱。 ``` ### Example 5: One-Take Long Shot with Multi-Image References **Input:** ``` 我有5张图,想做一个一镜到底的谍战片追踪镜头 ``` **Output:** ``` 谍战片风格,@图片1作为首帧画面,镜头正面跟拍穿着红风衣的女特工向前走, 镜头全景跟随,不断有路人遮挡红衣女子,走到一个拐角处,参考@图片2的拐角建筑, 固定镜头红衣女子离开画面,走在拐角处消失,一个戴面具的女孩在拐角处躲着恶狠狠 的盯着她,面具女孩形象参考@图片3。镜头往前摇向红衣女特工,她走进一座豪宅 消失不见了,豪宅参考@图片4。全程不要切镜头,一镜到底。 参考素材: - @图片1:红风衣女特工正面形象(首帧) - @图片2:旧城区拐角建筑 - @图片3:面具女孩形象 - @图片4:欧式豪宅外观 - @图片5:备用场景(可用于扩展) ``` ## Seedance 2.0 Platform Specs | Dimension | Specification | |-----------|--------------| | Image Input | jpeg/png/webp/bmp/tiff/gif, up to 9 images, each <30MB | | Video Input | mp4/mov, up to 3 videos, total 2-15s, each <50MB, 480p-720p | | Audio Input | mp3/wav, up to 3 files, total ≤15s, each <15MB | | Text Input | Natural language description | | File Limit | Max 12 files total (images + videos + audio combined) | | Duration | 4-15 seconds per generation | | Sound Output | Built-in sound effects and music | | Resolution | Up to 2K output | > **Important:** The platform does not allow uploading images or videos containing realistic human faces. Such materials will be automatically blocked. ## Tips for Best Results - **Be specific and visual** — "a woman in a red trench coat walks through rain-soaked neon streets" works much better than "a woman walking" - **Use timestamp storyboarding** for 13-15 second videos to maintain control over each segment - **Separate dialogue and sound effects** from visual descriptions for cleaner results - **Specify negative constraints** at the end (e.g., "no watermarks, no subtitles, no text overlays") - **Match reference image styles** to your video theme (e.g., use ink-wash style images for Chinese historical themes) - **For videos over 15 seconds**, let the skill generate multi-segment plans with explicit continuity points between segments ## Project Structure ``` . ├── .claude/ │ └── skills/ │ └── seedance/ │ └── SKILL.md # The skill definition (core file) ├── .gitignore └── README.md ``` The entire skill logic is contained in a single file: `.claude/skills/seedance/SKILL.md`. ## Contributing Contributions are welcome. Some areas where help would be appreciated: - Adding more prompt examples for specific scenarios - Translating the skill definition to support English prompt generation - Adding prompt templates for new Seedance features as the platform evolves - Improving prompt quality based on real-world generation results To contribute: 1. Fork this repository 2. Create a feature branch (`git checkout -b feature/your-feature`) 3. Commit your changes (`git commit -m "Add your feature"`) 4. Push to the branch (`git push origin feature/your-feature`) 5. Open a Pull Request ## License MIT ## Acknowledgments - [Seedance 2.0](https://jimeng.jianying.com) by ByteDance for the AI video generation platform - [Claude Code](https://docs.anthropic.com/en/docs/claude-code) by Anthropic for the extensible skill system