GPT Image Prompt Generator

基于 OpenAI 官方《GPT Image Generation Models Prompting Guide》的专业提示词生成技能。帮助用户按照官方最佳实践创建高质量、可控的 GPT Image 2 提示词。

Purpose

This skill guides users through creating professional-grade prompts for OpenAI's GPT Image 2 model. It follows the official prompting guide's methodology — structured prompts, explicit constraints, and mode-specific techniques — to produce reliable, production-quality results.

When to Use This Skill

Use this skill when the user:

• Wants to generate AI images using GPT Image 2 / DALL-E
• Needs help writing or improving image prompts
• Asks about image prompting best practices
• Wants to create specific types of images (logos, ads, UI mockups, infographics, etc.)
• Needs to edit existing images (style transfer, object removal, background change, etc.)
• Wants to add text to images
• Asks about quality, input_fidelity, size, or other GPT Image API parameters
• Mentions keywords like "提示词", "prompt", "出图", "AI绘画", "图像生成"

Workflow

Step 1: Understand the User's Need

Ask the user (or infer from context) the following:

1. Mode: Generate (text → image) or Edit (text + image → image)?
2. Scene/Use Case: What type of image? (See Scene Catalog below)
3. Subject: What is the main subject/content?
4. Style: Any specific visual style? (photorealistic, illustration, flat design, etc.)
5. Constraints: Text to include? Things to avoid? Specific requirements?
6. Output Size: Portrait, landscape, square, or custom?

If the user's request is vague, use AskUserQuestion to clarify the mode and scene type before proceeding.

Step 2: Select the Appropriate Template

Based on the user's scene type, select the matching template from resources/prompt-templates.md. Each template includes:

• Structured prompt format (Scene / Style / Mood / Constraints)
• Key parameters recommendation
• Common pitfalls to avoid

Step 3: Generate the Prompt

Follow these principles when constructing the prompt:

For Generate Mode (text → image):

1. Structure the prompt in this order:
• Background/Environment → Subject → Key Details → Constraints
2. Be specific about materials, textures, lighting, and atmosphere
3. Specify composition: camera angle, lens, depth of field, lighting direction
4. Add explicit constraints: "No watermarks", "No extra text", "No logos"
5. For text in images: use quotes, specify "EXACT, verbatim", define font style

For Edit Mode (text + image → image):

1. Clearly separate what should CHANGE vs. what must STAY THE SAME
2. Lock identity features when editing people (face, body, pose, hairstyle)
3. Recommend input_fidelity="high" for precision edits
4. Use "ONLY" to restrict changes to specific elements
5. Reiterate constraints in every iteration to prevent drift

Step 4: Recommend API Parameters

Based on the scene type, recommend appropriate parameters:

Step 5: Output Format

Present the final result to the user in this format:

## 📋 提示词 / Prompt

\```english
[Complete English prompt]
\```

### 中文翻译

\```
[Chinese translation]
\```

## ⚙️ 推荐参数

| 参数 | 值 | 说明 |
|------|-----|------|
| model | gpt-image-2 | 最新模型 |
| quality | [value] | [reason] |
| input_fidelity | [value] | [reason] |
| size | [value] | [reason] |
| [other params] | [value] | [reason] |

## 💡 关键技巧

- [Tip 1]
- [Tip 2]
- [Tip 3]

## 🐍 Python 代码

\```python
from openai import OpenAI
import base64, os

client = OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))

result = client.images.[generate|edit](
    model="gpt-image-2",
    prompt="""[prompt]""",
    [image=[...],]  # edit mode only
    [input_fidelity="[value]",]  # edit mode only
    size="[size]",
    quality="[quality]",
    [n=[value],]
    [background="[value]",]
)

# Save image
image_bytes = base64.b64decode(result.data[0].b64_json)
with open("output.png", "wb") as f:
    f.write(image_bytes)
\```

Scene Catalog

Generate Mode (text → image)

Edit Mode (text + image → image)

Prompt Structure Patterns

Pattern A: Structured Prompt (for complex scenes)

[Scene description]

Subject:
[Detailed subject description]

Style:
[Visual style, medium, reference]

Mood:
[Atmosphere, emotion, tone]

Constraints:
- [Constraint 1]
- [Constraint 2]
- No watermarks, no logos

Pattern B: Creative Brief (for ads/marketing)

[Brand/Client]: [name]
[Target audience]: [description]
[Concept]: [idea]
[Tagline]: "[exact text]"
[Visual direction]: [style, composition, color]
[Constraints]: [what to avoid]

Pattern C: Specification (for UI/slides/charts)

Create a [deliverable type] for [product].

Include:
- [Element 1]: [details]
- [Element 2]: [details]
- [Element 3]: [details]

Design requirements:
- [Layout spec]
- [Color scheme]
- [Typography]
- [What to avoid]

Pattern D: Precision Edit (for image editing)

[Action verb] [target] [from/in] [location].

Preserve:
- [Element 1]
- [Element 2]
- [Element 3]

Do not change anything else.

Best Practices

1. Always write prompts in English — GPT Image 2 understands English best
2. Structure over length — A well-organized short prompt beats a rambling long one
3. Explicit constraints — Always state what NOT to do, not just what to do
4. Iterate with small changes — Don't rewrite the entire prompt; tweak one element at a time
5. Separate change from constancy — In edit mode, clearly distinguish what changes vs. what stays
6. Use "ONLY" for surgical edits — "replace ONLY X" is stronger than "replace X"
7. Quote exact text — For text-in-image, always use quotes and "EXACT, verbatim"
8. Specify font for text — "bold sans-serif, high contrast, centered, clean kerning"
9. Lock identity in edits — List every feature that must not change when editing people
10. Match lighting in composites — When combining images, explicitly request matched lighting/shadows

Limitations

• This skill generates prompts for GPT Image 2; other models may need adjustments
• Text rendering in images may not be perfect on first try — iterate
• Complex multi-image workflows may require multiple API calls
• The skill provides prompt guidance, not actual image generation (requires OpenAI API access)