| name | nano-banana |
| description | Generates images from text prompts using Google's Nano Banana models (Gemini 3.1 Flash Image / Gemini 3 Pro Image) via OpenRouter. Use when the user asks to generate, create, or make an image/picture from a text description, or mentions Nano Banana. |
nano-banana
Generates or edits an image from a text prompt (and optional input images) and saves it. Streams reasoning and progress to stderr so the user gets feedback during the ~10-60s generation.
Usage
Invoke the wrapper script with the user's prompt:
./scripts/nano-banana.sh "<prompt>"
The script handles the API key (pass api/openrouter/image-editing), SSE streaming, and saving. It writes the PNG to the current working directory unless -o is given. After it returns, the last stdout line is Saved: <path> (model: <id>) — read the path with the Read tool to display the image to the user.
Model selection
- Default (
flash) — Gemini 3.1 Flash Image. Use for almost everything: cheap, fast, newest, "Pro-level quality at Flash speed."
-m pro — Gemini 3 Pro Image. Use only when the user explicitly asks for high quality, 4K, professional / commercial work, multi-subject identity preservation (3+ recognizable people), or precise in-image text rendering.
Output controls
-a, --aspect — 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9. Flash-only adds 1:4, 4:1, 1:8, 8:1.-s, --size — 1K (default), 2K, 4K. Flash-only adds 0.5K.-i, --image FILE — input image for editing/blending; repeatable. Pro accepts up to 5 (identity preservation across subjects).
The script validates aspect/size against the chosen model and rejects unsupported combos before hitting the API.
Common invocations
./scripts/nano-banana.sh "a red panda eating ramen"
./scripts/nano-banana.sh -a 16:9 -s 2K "wide cinematic mountain vista"
./scripts/nano-banana.sh -m pro -s 4K -o product.png "studio shot of a wireless mouse"
./scripts/nano-banana.sh -i photo.jpg "make it look like a watercolor painting"
./scripts/nano-banana.sh -m pro -i a.jpg -i b.jpg -i c.jpg "blend these three subjects into a group portrait"
./scripts/nano-banana.sh --no-stream "..."
Crafting prompts
For non-trivial generations (text in images, multi-image composition, photorealism, brand work), read references/prompting.md before writing the prompt.
Failure modes
pass api/openrouter/image-editing returned empty — entry missing in the user's password store.API error: ... — surfaced from OpenRouter (rate limit, model unavailable, content policy). Report it verbatim.No image returned — model produced text only (e.g. refused). Show the response if helpful.
Script Execution: Scripts should be executed from the skill directory. All scripts use Nix shebangs so no manual dependency installation is required.
Fixes
