| name | generate-image-openai |
| model | haiku |
| effort | medium |
| description | Generates images using OpenAI/DALL-E/GPT image models with an iterative two-phase workflow: fast drafts for iteration, high-quality final render on approval. Triggers on 'generate an image', 'create a picture', 'make me an illustration', 'DALL-E', or any image creation task. Do NOT use for diagrams or flowcharts (use /diagram or /excalidraw-diagram). |
OpenAI Image Generation — Iterative Workflow
Two-phase workflow: fast drafts for iteration, high-quality final render on approval.
Cost per successful session: ~$0.21 (3 draft iterations + 1 final render)
Phase 1: Clarify Before Generating
Before touching the script, ask targeted questions to nail the prompt. Typical clarifiers:
- Subject/scene: What's in the image? Who or what is the focus?
- Style: Photorealistic, illustration, watercolor, vector, cinematic, cartoon?
- Composition: Wide shot, close-up, overhead, centered?
- Mood/lighting: Warm/cool, dramatic/bright, moody/clean?
- Background: Specific setting or transparent/solid color?
- Text content: Any words/labels to include?
- Aspect ratio: Square (1:1), landscape (3:2), portrait (2:3)?
Don't over-ask — 2-3 targeted questions if the prompt is ambiguous. Clear prompts skip directly to generation.
Phase 2: Draft Iterations
Model: gpt-image-1-mini at high quality — fast (~10s), good quality, cheap ($0.052/image)
.claude/scripts/generate-image.sh "<prompt>" [output_dir] [filename]
After each generation:
- Read the saved PNG with the Read tool to view it
- Assess against the user's intent — composition, style, accuracy
- Note what's working and what needs refinement
- Adjust the prompt and regenerate (target 2-3 iterations max)
If the draft is clearly far off (wrong style, wrong subject), surface that before burning more iterations — a prompt restart is cheaper than 3 bad drafts.
Phase 3: Present for Approval
After 2-3 draft iterations, present to [Your Name]:
- Display the best draft
- Note what changed across iterations
- Identify any remaining gaps
Two outcomes:
- Approved → proceed to Phase 4 (final render)
- Major changes needed → restart from Phase 1 with new direction
Phase 4: Final Render
On approval, generate the high-quality final:
.claude/scripts/generate-image.sh "<refined-prompt>" [output_dir] [filename] gpt-image-1 high
Model options for final:
| Model | Cost | Best for |
|---|
gpt-image-1 high | ~$0.167 | Highest OpenAI quality, transparent PNG support |
gpt-image-1-mini high | $0.052 | When draft quality is already sufficient |
Note: Google imagen-4.0-ultra ($0.060) delivers better quality per dollar for final renders and supports 4K output — but requires a separate script (not yet implemented). OpenAI gpt-image-1 is the current final-render default.
Script Reference
.claude/scripts/generate-image.sh "<prompt>" [output_dir] [filename] [model] [quality] [size]
| Parameter | Default | Options |
|---|
prompt | required | Text description |
output_dir | current dir | Any path |
filename | auto from prompt | Custom (no extension) |
model | gpt-image-1-mini | gpt-image-1-mini, gpt-image-1 |
quality | high | low, medium, high, auto |
size | auto | 1024x1024, 1536x1024, 1024x1536, auto |
Prompt Tips
- Style: "watercolor", "photorealistic", "vector logo", "cyberpunk", "flat illustration"
- Composition: "overhead shot", "close-up portrait", "centered on white background"
- Lighting: "warm amber lighting", "moody cinematic", "soft diffused light"
- Text: explicitly state the text content and typography style
- Negative space: "minimal", "clean background", "product on white"
