Role: image generation and editing skill.
Engines
| Engine | Strengths | Default for |
|---|---|---|
| Nano Banana Pro (Gemini) | Photorealism, scene composition, maps | General use |
| GPT-Image-1 (OpenAI) | Clean graphics, text in image, precise edits | Diagrams, infographics, edits |
Configuration
In ~/.claude/.env:
IMAGE_OUTPUT_PATH=~/Pictures/ogmios
DEFAULT_IMAGE_MODEL=nano-banana
Seven priority workflows
1. Blog cover image
Goal: header image for a blog post. Default: Nano Banana, 16:9, photorealistic or stylised illustration.
~/.claude/scripts/call-image-api \
"Editorial illustration showing [theme]. Calm colors, photorealistic. 16:9 aspect ratio." \
--output ~/Pictures/ogmios/2026-05-03_blog_topic.png
2. Presentation illustration
Goal: slide illustration. Default: GPT-Image-1 for clean graphics, Nano Banana for photorealistic scenes.
3. Concept sketch
Goal: visual representation of an abstract concept. Default: Nano Banana with metaphor description.
4. Geopolitical map
Goal: stylised regional map for analysis. Default: Nano Banana, "stylised map of [region], minimal colors, key locations marked".
5. Meme / satirical image
Goal: humour or pointed commentary. Default: depends on style; specify reference style if needed.
6. Infographic prompt
Goal: data visualisation as image. Default: GPT-Image-1 with explicit text content. For real data viz, prefer mermaid skill.
7. Image edit / inpainting
Goal: modify an existing image. Default: GPT-Image-1 (better edit fidelity).
~/.claude/scripts/call-image-api \
--edit input.png \
"Replace background with [description]" \
--output edited.png
Prompt patterns
Photorealistic
"Photorealistic [scene], [time of day], [mood/lighting], shot from [angle].
Avoid [things to exclude]. Resolution: high."
Editorial illustration
"Editorial illustration of [concept]. Style: [reference, e.g. New York Times opinion section].
Colors: [muted/vibrant]. Composition: [central figure / abstract]."
Map
"Stylised map of [region]. Show [specific features]. Minimal labels. Colors: [scheme].
Style: [reference, e.g. The Economist]."
Diagram (when mermaid won't do)
"Clean diagram showing [process]. White background. [N] labeled boxes connected by arrows.
Sans-serif font. Each label: [exact text]."
Output and review
All generated images go to IMAGE_OUTPUT_PATH/{YYYY-MM-DD}_{model}_{slug}.png.
User review is mandatory before public use. Generated images are NOT verified facts.
Cost awareness
Each generation costs API credits. Common patterns:
- One generation per request — review, then iterate
- Don't batch-generate variations unless explicitly asked
Anti-patterns
- Naming real persons in image generation without explicit consent
- Generating images that could mislead about events (deepfake risk)
- Using AI images as evidence (they aren't)
- Skipping the user review step for public-facing material
🎯 COMPLETED: [SKILL:image-creator] [image generated for X] 🗣️ CUSTOM COMPLETED: [SKILL:image-creator] [Image ready]