clearsmog

nano-banana

Generate images via Gemini (Google) and get Typst embedding code. Use when the user needs AI-generated illustrations, conceptual diagrams, or photorealistic images for documents.

clearsmog 2 Updated 3mo ago

Resources

1
GitHub

Install

npx skillscat add clearsmog/claude-skills/nano-banana

Install via the SkillsCat registry.

SKILL.md

Parse $ARGUMENTS into flags for the bundled script and run it in one Bash call. The script handles filename generation, API key checks, and Typst code output.

Flags: -p prompt, -d dir, --width, --caption "...", -i edit-image, -r 1K|2K|4K, -n count, -m model, -a aspect-ratio

Pass --typst when generating images for Typst documents (the typical case). If --output/-o is not given, omit it (script auto-generates from prompt + dir).

Run exactly one command:

GEMINI_API_KEY="$GEMINI_API_KEY" uv run --script {baseDir}/scripts/gemini_imagen.py -p "<prompt>" --typst [other flags...]

If GEMINI_API_KEY is empty in this shell, read it directly from fish universal variables (avoids shell greeting pollution): `printf '%b' "$(sed -n 's/.*GEMINI_API_KEY://p' ~/.config/fish/fish_variables)"`.

Print the script's stdout to the user. Done.

Prompt construction

When generating the -p prompt from the user's description, follow these rules:

  1. Describe scenes, not keywords — "A three-legged wooden stool on a white background, each leg labeled..." beats "stool, risk parity, three legs"
  2. Specify style explicitly — Include art style, lighting, and composition (e.g., "flat vector illustration", "isometric 3D render", "watercolor sketch")
  3. Use white backgrounds for documents — Add "on a clean white background" for images that will be embedded in Typst documents
  4. Describe spatial layout — "On the left... on the right..." helps Gemini compose multi-element scenes
  5. Don't rely on text labels — Gemini often misspells text in images; describe the concept visually instead of asking for labeled diagrams

Options

Aspect ratios (-a): 1:1 square, 16:9 landscape, 9:16 portrait, 3:4, 4:3, 21:9 banner. Omit for model default.

Models: default = Gemini 3 Pro (quality, 4K). -m gemini-2.5-flash-image = fast/cheap drafts.