byted-ark-seedance-pe

Seedance PE - Original Video Prompt Optimization Skill

Core Functions

This skill receives the original, informal video prompt description provided by the user, analyzes it, and optimizes and organizes it into a standard structured prompt that meets Seedance 2.0 engineering requirements.

Use Cases

• The user provides a rough video description and requests "help me optimize this prompt"
• The user shares an unstructured original prompt and hopes to polish it to obtain better video generation results
• Key elements need to be extracted from messy input (subject, action, style, atmosphere, composition, lighting and shadow, camera movement, etc.) and correctly organized
• The user provides complete request body content that can be publicly shared and hopes to extract the text prompt from it and optimize it

⚠️ Important Prerequisites (Must Be Followed)

Each time a user's prompt optimization request is received, before starting optimization, the contents of the following six reference documents must be read completely, ensuring that all rules, best practices, and typical effect issue handling cases are mastered before carrying out optimization work. This is the core prerequisite for ensuring output quality and cannot be skipped.

Reference Materials

This skill relies on the following reference materials. During optimization, the best practices in all materials must be combined:

• references/prompt-guide.md - Seedance 2.0 official prompt guide, including basic sentence patterns for various scenarios
• references/api-doc-rules.md - Summary of rules in the API documentation
• references/business-guide-examples.md - A collection of excellent prompt examples from various industries
• references/engineering-prompt-methodology.md - Engineering prompt methodology
• references/seedance-2-troubleshooting-guide.md - Seedance 2.0 reference-to-video common issues and handling guide
• references/typical-effect-cases.md - Seedance 2.0 typical effect issue tuning case reference, including the causes and optimization suggestions for typical issues such as human heads, audio dialogue, overall image quality, multiple people, style, subtitles, video editing, video extension, props and animals, aspect ratio parameters, etc.

Hard Requirement: The six documents are the knowledge foundation of this skill. They must be reread before each optimization to ensure rules are not forgotten, examples can be referenced, typical issue handling strategies can be reused, and output quality remains stable.

Workflow

Step 1: Determine the Input Type

First determine whether the user input is an original prompt, a complete request body, or only a task ID:

• Original prompt: Proceed directly to Step 2 for parsing.
• Complete request body: Extract the text prompt from the request body, and retain only the general fields required by the public API (such as model / content / ratio / duration / resolution / generate_audio) for subsequent generation; do not retain or rely on internal fields.
• Only a Task ID (such as cgt-xxx）: Do not attempt to query the original prompt or request body through the task id. Please inform the user that you do not have task query capability, and ask the user to directly provide the original prompt or the complete request body content that can be publicly shared.

Step 2: Receive and Parse the Original Prompt

Carefully read the user's original input and identify all mentioned key elements:

• Subject: The main object being filmed in the video
• Action/Plot: What is happening in the video
• Scene/Location: Where the event takes place
• Style: Art style, video type (cinematic, anime, realistic, etc.)
• Mood/Atmosphere: Overall emotional tone (dramatic, calm, energetic, etc.)
• Lighting: Lighting conditions (golden hour, soft light, dramatic, neon, etc.)
• Camera Movement: Shot type, movement method (wide shot, close-up, push-in, etc.)
• Duration/Pacing: Time allocation, pacing speed
• Other Details: Any other specific requirements

Step 3: Analyze gaps and ambiguities

Check what missing information would help generate a better video:

• Unclear descriptions
• Missing key elements important to the video
• Vague words that could be more specific
• Conflicting instructions

If important information is missing, ask the user for clarification. For minor gaps, make reasonable assumptions based on context and state them in the output.

Step 4: Build the optimized prompt

Organize the information according to the Seedance official standard three-part structure, integrating best practices from the engineering methodology:

Standard structure (strictly follow):

1. Global basic settings: Lock characters, environment, and core assets, use @图N syntax to explicitly declare mapping relationships
   • If there are first and last frame constraints, declare them here
   • Must clearly specify @图N the corresponding character/scene name
2. Time-slice storyboard script: By time slices (e.g., 0-3s, 3-10s), each slice contains:
   • Who does what where
   • One camera movement method
   • Must use strong visual references with @图N , followed immediately by the name to prevent ambiguity
3. Editing instructions(only needed for video editing scenarios):
   • Add/delete/modify: specify the time period and spatial position
   • Extend/concatenate: use standard syntax
   • Text generation: specify content, timing, position
4. Image quality, style, and constraints: Auto-mount:
   • Image quality enhancement:4K 高清，细节丰富
   • Anti-collapse fallback constraints:面部稳定不变形、五官清晰、人体结构正常、动作自然流畅、不僵硬、画面无卡顿、无闪烁

Writing guidelines:

• Use clear, concise descriptive language
• Order keywords by importance
• Avoid redundancy
• Strictly preserve the user's original intent and all information, including dialogue and audio requirements, which must not be discarded
• Use the style vocabulary and camera movement vocabulary accumulated in the reference materials to make the description more professional and accurate
• ⚠️[New mandatory rule] The optimized prompt itselfAbsolutely prohibited from using any special icon symbols such as ⚠️✅❌, only standard Chinese, English, and necessary punctuation; icons may only be used in the user explanation section outside the prompt, and must not appear in the final generated prompt code block

Mandatory constraints (from reference materials):

• All @图N / @视频N references must be immediately followed by the character name or noun explanation to prevent word segmentation ambiguity
• Only one camera movement method exists within a time-slice shot
• Must use @图N to establish a textual bridge for the Asset ID; it is forbidden to directly use [asset-xxx] appearing independently in the action description
• The final output must include anti-collapse and high-image-quality fallback constraints
• Complex multi-person scenes must add strong positional constraints
• All dialogue and audio information provided by the user must be fully preserved and must not be discarded

Step 5: Output the result

Provide:

1. Problem analysis: Briefly analyze the problems, gaps, or points needing optimization in the original prompt
2. Optimized standard prompt: The complete optimized prompt content
3. Optimization notes: Explain specifically what was optimized and why it was optimized that way
4. [Based on reference documents] At the end of the reply, you must clearly indicate the document names and specific sections in the reference directory on which this optimization is based

Examples

Example 1: Original prompt optimization

Original input (user):

"A cat playing with a ball in a park, sunny weather"

---

Original prompt

公园里一只猫玩球，天气晴朗

---

Problem analysis

The original prompt is too brief and lacks key elements:

1. The subject's characteristics are unclear (the cat's breed, color, appearance)
2. The prop is unclear (the ball's type, color, size)
3. Environmental details are insufficient (the specific park environment, vegetation, time)
4. Missing storyboard and camera movement descriptions
5. Missing image quality and style requirements

---

Optimized standard prompt

一只毛茸茸的橘猫在郁郁葱葱的公园里欢快地扑击一颗红色橡皮小球。金色阳光穿过橡树，在草地上投下斑驳光影。

0-6 秒：橘猫俯身盯着小球，镜头特写，聚焦猫爪和球的互动；
6-12 秒：猫跳起来扑向小球，镜头中景跟拍，阳光在毛发上反光；

4K 高清，细节丰富，自然写实风格，暖色调，黄金时刻光影，整体氛围轻松愉悦；面部稳定不变形、五官清晰、人体结构正常、动作自然流畅、不僵硬、画面无卡顿、无闪烁。

---

Optimization notes

• Added texture details to the subject (fluffy orange cat)
• Clarified the type of ball (small red rubber ball) to improve clarity
• Added environmental details (lush park, oak tree, grass)
• Clarified the lighting quality (golden sunlight, dappled light and shadow)
• Split the storyboard by time, with camera movement specified for each slice
• Added quality keywords and standard fallback constraints

[Based on reference documents]

• references/prompt-guide.md - Standard three-part structure specification
• references/engineering-prompt-methodology.md - Image quality fallback constraint specification

---

Example 2: User provides only a Task ID

Original input (user):

"Help me optimize the prompt for the task cgt-20260511154725-k5hdz"

Handling method: Do not query the task ID, and do not attempt to obtain the original request body. Reply to the user:

The ability to query the original prompt or request body by task ID is not supported. Please directly paste the original prompt, or provide the complete request body content that can be publicly shared, and I will optimize it based on the content you provide.

---

Example 3: User provides the complete request body

Original input (user):

"Help me optimize the prompt in this request body: {...}"

Handling process:

1. Extract the text prompt from the request body
2. Delete internal fields or non-public fields, retaining only the general fields required by the public API
3. Process according to the prompt optimization workflow

Output result:(Same as the optimized prompt format)

---