Generate Images

This skill generates images using the Nano Banana model via ~/.claude/scripts/generate_image.py.

Always use this skill when the user asks to:

• Generate, create, or make an image
• Create icons, logos, banners, or visual assets
• Edit, modify, or transform an existing image
• Generate mockups, diagrams, or illustrations
• Create any visual content

Prerequisites:

• GEMINI_API_KEY must be set in ~/.claude/settings.json under env
• Script: ~/.claude/scripts/generate_image.py (runs via uv run)

Reference images: Users can store named images in ~/.claude/images/ for use as editing sources. Filenames describe the content (e.g., myself.jpg, company-logo.png, office.jpg).

uv run ~/.claude/scripts/generate_image.py --check-key

If API_KEY_MISSING: inform the user they need to set GEMINI_API_KEY in ~/.claude/settings.json or get one at https://aistudio.google.com/apikey

List available reference images:

ls ~/.claude/images/ 2>/dev/null

Matching rules:

• Match user mentions to filenames (without extension). Examples:
  • "add a hat to myself" → look for myself.jpg, myself.png, etc.
  • "put my dog in a park" → look for my-dog.jpg, dog.jpg, etc.
  • "update the company logo" → look for company-logo.png, logo.png, etc.
• Match is case-insensitive, try with and without hyphens/underscores
• If a match is found, use --edit mode with the matched file as source
• If no match and the user clearly references a personal image, ask them to place it in ~/.claude/images/

Also check the current project for relevant images if the user references project assets:

• Look in assets/, images/, public/, static/, or project root

• If user specifies a path, use it
• If editing a reference image, save to the current project (not back to ~/.claude/images/)
• If inside a project, use a sensible location (e.g., assets/, images/, public/, or project root)
• Default filename: descriptive kebab-case with .png extension (e.g., hero-banner.png, app-icon.png)

• Subject: What to generate (e.g., "a minimalist logo of a rocket")
• Style: Visual style (e.g., "flat design", "photorealistic", "watercolor", "pixel art")
• Details: Specific attributes (colors, lighting, composition, mood)
• Quality: Resolution hints (e.g., "high detail", "4K quality", "professional")

Example prompt structure: [Subject], [style], [details], [quality]

Text-to-image (new image):

uv run ~/.claude/scripts/generate_image.py "prompt here" --output path/to/output.png

Image editing with reference image:

uv run ~/.claude/scripts/generate_image.py "editing instructions" --edit ~/.claude/images/myself.jpg --output path/to/output.png

Image editing with project image:

uv run ~/.claude/scripts/generate_image.py "editing instructions" --edit path/to/source.png --output path/to/output.png

Options:

• --output PATH - Output file path (default: generated_image.png)
• --edit IMAGE - Source image for editing mode
• --json - Output metadata as JSON

1. Read the output image using the Read tool to verify it was created and looks correct
2. Report the file path and size to the user
3. If the result doesn't match expectations, refine the prompt and regenerate

Icon/Logo: "A minimalist app icon for a task management tool, flat design, blue and white color scheme, clean geometric shapes, centered composition, professional quality"

Banner/Hero: "Wide panoramic banner for a tech blog, abstract gradient background in purple and teal, modern typography space on the left, subtle geometric patterns, professional web design"

Edit with reference image: User says: "add a clown hat to myself" → Match myself.jpg in ~/.claude/images/ → Run: uv run ~/.claude/scripts/generate_image.py "Add a colorful clown hat to the person in this photo" --edit ~/.claude/images/myself.jpg --output clown-hat-myself.png

Edit with reference image: User says: "put my dog on a beach" → Match dog.jpg or my-dog.jpg in ~/.claude/images/ → Run: uv run ~/.claude/scripts/generate_image.py "Place the dog on a tropical beach with waves and sunset" --edit ~/.claude/images/dog.jpg --output dog-on-beach.png

• API key check passes
• Reference images matched when applicable
• Image is saved to the specified output path
• Output image has been visually verified via Read tool
• User is informed of the file location