Video Prompt

Video Prompt

The prompt layer of the video-prompt-craft toolkit: turn a shot (or a full shot list from video-storyboard) into structured, platform-ready prompt text. This skill owns the platform syntax, the universal formula, and optimization — it does NOT own narrative structure or shot design (that is video-storyboard's job).

三件套协作 / Three coordinated skills:

• video-storyboard — idea → narrative shot list (story structure, shot language, visual intensity)
• video-prompt (this skill) — shot list → platform prompts
• video-review — multi-angle audit of storyboard & prompts

If a user hands you a rough idea with no shot plan and it needs multiple shots, suggest running video-storyboard first, then return here to render each shot. For a single shot or a direct prompt request, proceed directly.

Core Principle

You are not describing the entire world — you are describing one shot through a camera lens. For Seedance 2.0: think like an AI director — spatial layer (what's in frame) + temporal layer (how it changes over time).

Every effective video prompt answers five questions:

Step 1: Classify the Task

Determine the generation mode before writing the prompt:

Step 2: Construct the Prompt

Basic Formula (Universal)

[Subject description] + [Scene/Environment] + [Primary action] + [Shot type] + [Camera movement] + [Lighting/Style] + [Quality/Constraint tags]

Advanced Formula (Seedance 2.0)

精准主体 + 动作细节 + 场景环境 + 光影色调 + 镜头运镜 + 视觉风格 + 画质 + 约束条件

This maps to two internal model dimensions:

• Spatial layer: who/what, where, appearance, environment
• Temporal layer: action sequence, camera movement, timing, transitions

Rules

1. One shot, one action — keep the primary motion singular and clear.
2. Be concrete, not abstract — "a woman in a white sundress" beats "a beautiful person"; "golden hour warm sunlight" beats "nice lighting".
3. Order matters — Subject first, then scene, then motion, then camera, then aesthetics.
4. For I2V: skip what the image already shows — only describe motion and camera work.
5. For multi-shot: use 镜头1/镜头2/镜头3 labels — do NOT force precise timestamps; let the model pace naturally.
6. For sound: be explicit about silence — write "无台词" / "No dialogue." and "无背景音乐" / "No background music." if needed.
7. Keep it under 300 words per shot — most models perform best with 50–200 words per shot.
8. One camera move per shot — mixing push+orbit+pan in one shot increases instability.

Step 3: Seedance 2.0 — Subject Definition & Reference System

This section is critical for Seedance 2.0 multimodal reference tasks.

3a. Subject Definition

When referencing specific subjects in uploaded assets, always define them explicitly.

Syntax:

将<图片/视频N>中的[2-3个清晰静态特征]定义为<主体N>

Example:

将图片1中穿红色连衣裙、戴草帽的女人定义为主体1

Rules:

• Use 2–3 stable static features (clothing, hairstyle, appearance, type) for identification
• For multi-character scenes: name each character and link to their reference image

  将视频1中的高个子男人定义为警察，矮个子男人定义为小偷
  

• For unnamed subjects: use 主体1@图片1 inline every time the subject is referenced
• Do NOT use Asset ID as a subject reference — always use <图片/视频N>
• Most important references go first in the prompt

3b. Reference Syntax (Multimodal Tasks)

Input limits (Seedance 2.0):

• Images: 0–9
• Videos: 0–3
• Audio: 0–3
• Recommended configuration: 1–2 character photos + 1 scene photo + 1 camera reference video + 1 audio = 4–5 assets total
• Do not max out inputs — too many assets confuse the model's priority weighting

3c. Task Type Sentence Patterns

Multimodal Reference (生成新视频):

参考<图片/视频N>中的<动作/运镜/风格/音效>，生成...

Video Edit (修改已有视频):

严格编辑<视频N>，将其中的<原特征>修改为<新特征>
增加元素：清晰描述<元素特征> + <出现时机> + <出现位置>
删除元素：点明需要删除的元素，并在提示词中强调保持不变的部分

⚠️ For edit/extend tasks: use <视频N> directly — do NOT write 参考<视频N> or the model will treat it as reference-only.

Video Extension (延长视频):

向前/向后延长<视频N>，生成...

Track Completion (轨道补全/多片段串联):

<视频1> + <过渡画面描述> + 接 <视频2> + <过渡画面描述> + 接 <视频3>

Combined tasks:

参考<图片/视频N>的[参考维度]，严格编辑<视频X>，[具体编辑内容]

Step 4: Multi-Shot Storyboard Structure

Break complex videos into numbered shots. The model's internal modeling decouples space and time, so explicit shot ordering helps.

Structure per shot:

1. Camera move / transition type (全景缓慢推近 / 固定机位 / 镜头切至...)
2. Subject action and expression (core character's key motion + emotional state)
3. Location / spatial relationship
4. Audio (SFX, voice, BGM for this shot)

Example:

镜头1：街巷侧拍，男人缓慢起跑，带有急促的呼吸感。
镜头2：男人撞翻水果摊，镜头快速摇动并给到男人惊恐的特写。
镜头3：男人翻过矮墙消失，镜头缓慢拉远定格在空荡的街道。

Note: Precise time segments (e.g. 0–3s) are not recommended for Seedance 2.0 — the model's timing support is unstable. Use shot labels instead and let the model pace naturally.

Step 5: Action Description

Motion Guidelines

• Specify body parts: hand, leg, head, shoulders, back
• Add intensity/speed: 缓慢抬手 / 快速转头 / 用力蹬地 / 微微低头
• Prefer smooth, continuous micro-motions over high-intensity actions (running full speed, big jumps, violent rolls)
• Describe motion continuity: 借着转身惯性顺势抬手 / 从停顿状态自然过渡到举手

Emotion → Physical Expression

Avoid abstract emotion words. Externalize emotion as body language:

Step 6: Apply Aesthetic Controls

Consult references/aesthetics.md for the complete keyword dictionary. Key categories:

• Lighting: natural light, artificial light, special/effect light, light quality
• Shot size: extreme close-up → close-up → medium close-up → medium → medium full → full → extreme wide
• Camera angle: eye level, high angle, low angle, bird's eye, worm's eye, over-the-shoulder, Dutch angle
• Lens: telephoto, wide angle, fisheye, macro, anamorphic, tilt-shift
• Camera movement: push in, pull back, pan, tilt, dolly, orbit, crane, tracking, handheld, Steadicam
• Style: cinematic, photorealistic, anime, cyberpunk, vintage film, minimalist, stop-motion, clay, felt, pixel art

Step 7: Quality, Style & Constraint Tags

Always close your prompt with these three types of tags:

Quality Tags

高清，细节丰富，电影质感，色彩自然，光影柔和

Style Tags

赛博朋克冷蓝紫色调 / 复古胶片 / 日系清新 / 写实纪录片风

Constraint Tags (Critical — reduces artifacts)

Step 8: Special Features — Text Generation in Video

Seedance 2.0 can render text (slogans, subtitles, speech bubbles) in video.

Text generation formula:

「文字内容」+「出现时机」+「出现位置」+「出现方式」，「文字特征（颜色、风格）」

Symbol conventions for special content:

Text rendering tips:

• Use common characters; avoid rare characters (生僻字) and special symbols
• Supported: advertising slogans, subtitles (字幕), speech bubbles (气泡台词)
• For logo precision: upload the logo as a reference image

Step 9: Review Checklist

Before delivering the prompt, verify:

• Subject is concretely described (appearance, clothing, features)
• For multimodal: subject defined explicitly with <图片N> binding
• Scene/environment is specified
• There is exactly one primary action per shot
• Camera shot type is specified (close-up, medium, wide, etc.)
• Camera movement is specified (or explicitly "固定镜头 / static camera")
• Lighting and style are specified
• Constraint tags added (no subtitles, no watermark, etc.)
• For I2V: prompt does NOT redundantly describe what the image already shows
• For multi-shot: 镜头1/2/3 labels used (not forced timestamps)
• For edit/extend: no 参考 prefix before the video reference
• For sound-synced: sound explicitly described or silenced
• Prompt is under 300 words per shot
• Assets ≤ recommended count (4–5 total, not maxed out)

Output Format

When generating prompts, output in this structure:

## 模式 / Mode: [T2V / I2V / 首尾帧 / Multi-shot / Multimodal-Reference / Video-Edit / Video-Extension / Sound-synced]
## 平台 / Platform: [Any / Sora / Runway / Kling / Seedance / Wan / etc.]
## 宽高比 / Aspect ratio: [16:9 / 9:16 / 1:1 / 21:9 / 4:3]
## 时长 / Duration: [5s / 10s / 15s / etc.]
## 分辨率 / Resolution: [480p / 720p / 1080p / 4K]

### 提示词 / Prompt

[The actual prompt text here]

### Prompt (English)

[English translation — bilingual output for cross-platform portability]

If the user does not specify a platform, write in a generic style that works across all major platforms.

Tips for Common Problems

Language

• Write prompts in the same language the user uses.
• Always offer a bilingual version (Chinese + English) since different platforms prefer different languages.
• For Chinese: use 镜头推进 not push in; for English: use push in not 镜头推进.
• For Chinese dialogue in prompts: keep all dialogue in one language — avoid mixing Chinese and English within a single character's lines (proper nouns excepted).

Platform Notes

Video Extension Strategy

For videos longer than 15 seconds, chain multiple segments:

Option A — Video Extension (连续长镜头):

• Best for: dialogue scenes, emotion progression, single-location continuous takes
• Use 向后延长<视频N> for each extension
• Fix seam artifacts: trim last 6 frames of prior clip + first 1 frame of next clip

Option B — Segment + Edit (分段拼接):

• Best for: action sequences, scene changes, montage, complex motion
• Generate each shot separately with consistent style/lighting tags
• Combine in editing software

Option C — Track Completion (轨道补全):

• Chain up to 3 video clips with transition descriptions:

  <视频1> + <过渡画面描述> + 接 <视频2> + <过渡画面描述> + 接 <视频3>
  

Asset Configuration Strategy

Categorize your reference materials by function:

Recommended: 4–5 assets total. Do not max out the input limit — too many assets cause the model to lose feature priority.

Examples

See references/examples.md for a library of examples across modes, styles, and platforms.

Optimization

See references/prompt-optimization.md for iterative refinement strategies and known failure modes.

Subagent 协作(可选)

若环境已注册 prompt-engineer 子 agent(插件 agents/ 目录自动加载),可委派逐镜提示词渲染:

Agent(subagent_type: "prompt-engineer", prompt: "把这份分镜表渲染成 Seedance 2.0 提示词:[shot list]")

降级规则:Agent 不可用时,主线程直接按本文件执行。

完成之后

• 提示词写好后,可用 video-review 做多视角审查(结构/镜头语言/视觉/一致性),按报告返回本 skill 或 video-storyboard 修改。
• 若提示词是从 video-storyboard 的分镜表来的,审查时对照分镜表的 +/-、><、屏幕方向、视觉强度是否被如实翻译。

安装与整体架构见仓库根 README.md。