This skill should be used when the user asks to "generate an image", "create a cover image", "make an illustration", "generate artwork", "create a hero image", "generate a logo", "make a banner", "create an image for", "generate a thumbnail", "make an icon", "create visual content", "edit this image", "restyle this image", "compose these images", or needs AI-generated or AI-edited images for their project. Provides a CLI with JSON output optimized for LLM consumption.
This skill should be used when the user asks to "generate an image", "create a cover image", "make an illustration", "generate artwork", "create a hero image", "generate a logo", "make a banner", "create an image for", "generate a thumbnail", "make an icon", "create visual content", "edit this image", "restyle this image", "compose these images", or needs AI-generated or AI-edited images for their project. Provides a CLI with JSON output optimized for LLM consumption.
imagegen
Text-to-image and image-editing CLI backed by Google Gemini's Nano Banana 2 family. JSON output.
CLI
Path: ./imagegen-cli/imagegen (relative to this SKILL.md).
Requires uv (install). First call provisions the Python environment automatically.
Subcommands: generate (alias gen). Pass --image <path> (repeatable) to edit, restyle, or compose existing images instead of generating from scratch.
Always run imagegen generate --help to discover current flags and defaults. Do not memorize flag values — they evolve with the API.
API key
The CLI looks for the key in this order:
GEMINI_API_KEY environment variable.
.gemini-key file (raw key, no quotes) found by walking up from the current working directory — project-local override. Add .gemini-key to .gitignore.
~/.gemini-key in the user's home directory — machine-wide default.
Use the default for almost everything. Switch to gemini-3-pro-image-preview for high-stakes assets (legible long text, complex multi-element compositions). Switch to gemini-2.5-flash-image only for cost reasons on simple prompts.
Prompting (these are thinking, creative models)
These models reason about composition and fill creative gaps on their own. Provide intent and context — not exhaustive specs.
Give what only the user knows: brand, audience, where the asset goes, mood, must-include elements, must-avoid elements.
Let the model choose lens, lighting, palette, and layout unless the user specified them.
One narrative sentence beats a tag list. Describe the scene, not keywords.
For exclusions, prefer rewriting the prompt positively (e.g. "an empty street" instead of --negative-prompt "people, cars"). Use --negative-prompt only as a last resort.
Iterate by saving the output and re-invoking with --image <previous-output> plus a small directive (e.g. "warmer lighting, keep everything else").
Output rules
Always pass --output <path> and place the file in a project-appropriate assets directory (e.g. src/assets/, public/images/). Use a descriptive filename matching the content.
After the call returns ok:true, use the Read tool on the produced file path to confirm the image landed and looks right.
Never generate copyrighted characters, real people's likenesses, or harmful content.
Editing / composition
Pass --image <path> one or more times in addition to the prompt. Same generate command. Examples of when to use it:
Add or remove elements from an existing image.
Apply a style transfer.
Compose a scene from multiple references (e.g. a model + a garment).
