seedance-prompt-skill

---
name: seedance-prompt-skill
description: "Use when building ready-to-paste video prompts for ByteDance's Seedance 2.0 (即梦) — turns a one-line brief into cinematic, multi-shot Chinese prompts with camera language, timing, and @-referenced assets."
source: https://github.com/songguoxs/seedance-prompt-skill
generated: 2026-07-02T18:43:22.267Z
category: tool
audience: creators
---

## When to use

- Turning a short creative brief ("a wuxia fight", "a milk-tea ad", "a 30s suspense short") into a complete Seedance 2.0 video prompt
- Writing multi-modal prompts that reference uploaded images, videos, and audio via the @图片1 / @视频1 / @音频1 system
- Controlling 4-15s shots with timestamp-based shot lists, camera moves, dialogue, and sound design
- Producing videos longer than 15s by chaining segments with Seedance's video-extension (视频延长) feature
- Maintaining character, product, or scene consistency across shots using first/last-frame reference images

## Key concepts

### Prompt formula

The base text-to-video pattern is (主体描述) + (动作序列) + (环境/光影) + (镜头语言) + (风格关键词) — subject, action sequence, environment/lighting, camera language, then style keywords.

### @reference system

Uploaded assets are cited inline with official names: @图片1–@图片9 for images, @视频1–@视频3 for video, @音频1–@音频3 for audio. Each reference must state its purpose, e.g. @图片1为首帧 (as first frame) or 参考@视频1的运镜 (reference its camera work). Up to 12 files total.

### Timestamp shot list

For 13-15s narrative clips, split the prompt into timed beats (0-3秒 / 4-8秒 / 9-12秒 / 13-15秒), each with its own shot description and camera language — the most-used advanced technique.

### Video extension for >15s

Single generations cap at 15s. For longer pieces, generate segment 1 (≤15s), then feed it back as @视频1 and "将@视频1延长Xs" with a matching hand-off frame between segments. Extension length = the duration of the new portion only.

### Natural-language Chinese, no negative prompts

Seedance 2.0 prefers fluent, grammatical Chinese describing a scene to a director — not keyword tags. All prompts (video and image-gen) are written in Chinese. Unwanted elements are stated in a 禁止 (forbidden) block at the end rather than via --no negatives.

## API reference

```
Install: copy .claude/skills/seedance into your project
```

Drop the SKILL.md skill into a Claude Code / Cursor / Cline project so the agent activates on Seedance, 即梦, 视频提示词, or video-prompt requests.

```
git clone https://github.com/songguoxs/seedance-prompt-skill
cp -r seedance-prompt-skill/.claude/skills/seedance .claude/skills/
```

```
Text-to-video: (主体) + (动作序列) + (环境/光影) + (镜头语言) + (风格)
```

Generate a base prompt with no reference assets.

```
镜头跟随黑衣男子快速逃亡，后面一群人在追，镜头转为侧面跟拍，人物惊慌撞倒路边的水果摊爬起来继续逃，人群慌乱的声音。
```

```
Consistency: [角色]@图片N + [动作/剧情] + [场景]@图片N + [运镜/光影]
```

Keep a character, product, or scene consistent by anchoring it to reference images.

```
男人@图片1下班后疲惫地走在走廊，脚步变缓，停在家门口，脸部特写深呼吸调整情绪，特写翻找钥匙插入门锁，进门后小女儿和宠物狗欢快迎接拥抱，室内温馨，全程自然对话
```

```
Timestamp shot list: 0-3秒:[...] 4-8秒:[...] 9-12秒:[...] 13-15秒:[...]
```

Precisely control each beat of a 13-15s narrative clip.

```
15秒仙侠高燃战斗，金红暖色调，0-3秒：低角度特写主角紧握雷纹巨剑，剑刃赤红电光爆闪；4-8秒：环绕摇镜快切旋身挥剑迸射红色冲击波；9-12秒：仰拍拉远慢放跃起劈向魔兵群；13-15秒：缓推特写落地收剑，余波微动。
```

```
Extend >15s: 将@视频1延长Xs + [新增内容的时间戳分镜]
```

Chain segments past the 15s cap, each picking up from the prior segment's final frame.

```
将@视频1延长15秒。0-5秒：接上段剑修持剑画面，暗影魔兽飞扑而来，剑修纵身迎敌；6-10秒：空中激战剑气纵横，魔兽斩成灰烬；11-15秒：落地收剑，金色粒子飘散，镜头缓推特写侧脸，音效渐弱。
```

## Gotchas

- Photoreal human faces in uploaded images/videos are auto-blocked by the platform — do not use real-face reference material.
- Seedance does not support negative prompts; declare unwanted elements (字幕/LOGO/水印) in a 禁止 block at the end instead.
- Write all prompts in fluent Chinese — Seedance 2.0 is tuned for natural-language Chinese, not English keyword tags.
- Use the official @ names (@图片1, @视频1, @音频1), never @img1/@video1, and label what each reference is for so images, video, and characters don't get mixed up.
- When extending a video, the chosen generation length is the NEW portion only (extend by 5s → set length to 5s), and each segment needs an explicit hand-off frame to the next.
- Distinguish 参考 (reference style/motion) from 编辑 (edit the source asset) — they trigger different behaviors.
- Single generation caps at 15s; anything longer must be split into chained extension segments.

---
Generated by SkillMake from https://github.com/songguoxs/seedance-prompt-skill on 2026-07-02T18:43:22.267Z.
Verify against source before relying on details.

File: ~/.claude/skills/seedance-prompt-skill/SKILL.md