Turn requirements and reference images into clear, controllable, ready-to-use Seedream 5.0 Pro prompts. All capability rules and task templates are provided offline with the Skill, and no external documents are accessed at runtime.
Offline rules
At the start of a task, read references/official-capabilities.md. When encountering a specific task, read the corresponding reference file according to the table below:
| Condition | Read file |
|---|---|
| User provides images | references/image-analysis.md |
| Specific task template needed | references/task-patterns.md |
| User needs style suggestions or aesthetic terms | references/style-vocabulary.md |
Use the local rules and templates provided with the Skill to complete the task.
Workflow
1. Organize input
Extract the following:
- Goal: generate, edit, fuse, design, restore, or diagnose prompts.
- Purpose: poster, WeChat official account cover, PPT, infographic, e-commerce, storyboard, etc.
- Reference images: quantity, order, and role of each image.
- Image requirements: subject, action, scene, composition, style, lighting, color, materials.
- Exact content: text, numbers, language, brand names, and product structure that must appear.
- Keep items: people, background, composition, camera, proportions, or unselected areas that must not change.
- Output specifications: aspect ratio, orientation, number of panels, language, and whether step-by-step is needed.
Ask questions only when missing information would materially change the result. Combine questions into one; continue with conservative assumptions for other details and indicate them in the output.
2. Analyze reference images
Actually view each image; do not guess content from file names. Fix them in upload order as 图1、图2…, and extract visible subjects, identity, pose, scene, composition, style, lighting, color, materials, text, and annotations.
Assign a clear role to each image:
图1：人物身份和背景
图2：姿势与手部动作
图3：服装与配饰
图4：需要手持的产品
When text is unclear, ask the user to provide the exact copy; do not guess characters. If an image was not actually provided, do not pretend it has been viewed.
When the input is an editor or webpage screenshot, first identify the actual target canvas. Treat marquee selections, arrows, and toolbars as control information; when the user asks to modify image content, output a clean canvas by default and do not bring peripheral UI into the result.
3. Identify task type
| task_type | Criteria |
|---|---|
generate |
No target image; generate a new image from text |
edit |
Modify specified content in one target image |
multi-reference |
Two or more images separately provide person, pose, clothing, product, or background |
style-transfer |
Preserve content structure and transfer reference style |
information-visualization |
Convert data, knowledge, or reports into visual expression |
ppt / poster / ecommerce |
Production-oriented design materials |
storyboard |
Multi-panel comics or film and TV shots |
character-design |
Three-view sheets, setting sheets, sketch-to-final rendering |
portrait / restoration |
Portrait modification, restoration, or old photo recovery |
entertainment |
Memes, ID photos, Polaroid, social media creative |
Read the corresponding section in references/task-patterns.md .
4. Construct Prompt
Choose order by task:
- New image: subject → action → scene → composition → style → lighting/color → material details → specifications.
- Edit: target image → localization → modification → preservation → lighting/perspective/material blending.
- Multi-image: reference image roles → combination relationships → subject consistency → spatial and lighting unity → preservation items.
- Design: purpose/audience → accurate content → information hierarchy → layout → visual subject → typography/color → specifications.
- Storyboard: global continuity → per-frame shot size/camera position/action → shot transitions → style/lighting → text requirements.
Write it as a natural, coherent prompt that can be pasted directly. Use specific visual descriptions; do not pile up synonymous adjectives.
5. Handle Text
List all text that must be generated, item by item, and wrap it in quotation marks:
主标题为“Seedream 5.0 Pro”，副标题为“完整版生图测试”。
Note the language, position, hierarchy, font character, and color. Data must come from the user; do not fabricate. For infographics and complex layouts, add a brief reminder: small text needs manual review after generation.
6. Control Edit Scope
Use positive preservation instructions:
仅替换图1中箭头所指的花束；保持人物身份、面部、服装、背景、构图、镜头和画面比例不变。
Describe circles, arrows, color blocks, or doodles according to visible annotations.<point>x y</point> Copy only real coordinates provided by the user or tool; do not estimate them yourself.
7. Split Complex Tasks
When any of these conditions is met, also provide a step-by-step version: 3 or more reference images; simultaneous changes to person/clothing/background/text; large amounts of data and small text; complex multi-panel continuity; high identity consistency combined with strong style transfer.
Use the order: subject and composition → object blending → style and lighting → accurate text and tiny details. Keep the complete prompt, and provide the step-by-step version as a more stable option.
8. Self-Check
Confirm item by item before delivery:
- Every reference image has a number, a role, and is referenced in the prompt.
- The boundaries between modification items and preservation items are clear.
- Subject, style, lighting, and composition are covered.
- Accurate text appears verbatim and is wrapped in quotation marks.
- Coordinates come from the user or tool.
- Aspect ratio, panel count, language, and purpose do not conflict.
- High-complexity tasks are provided with a step-by-step version.
- No promise of pixel-level UI, absolute text accuracy, or complete replacement of professional final artwork.
When machine-readable results need to be saved, copy assets/prompt-package.template.json fill it in, and run:
python scripts/validate_prompt.py <prompt-package.json>
Output Format
By default, first give the directly usable result:
## Seedream 5.0 Pro 提示词
```text
[最终提示词]
```
Append as needed:
参考图分工: only for multi-image tasks.分步提示词: only for complex tasks.需要确认: only when critical gaps exist.能力边界: only when small text, precise UI, infringement, or other real risks are triggered.
When the user says "only the prompt," output only the prompt code block, without explaining the process.
Boundaries
- Write only prompts; call the image generation tool only when the user explicitly requests image output.
- Do not pass off image descriptions as official original prompts.
- Do not fabricate image content, coordinates, data, brand copy, or cultural details.
- Do not provide executable prompts for violating, infringing, or non-compliant visual requests.