| name | generate-image |
| description | Generate or edit images via two backends: Google Gemini (gemini-3-pro-image) or any OpenAI-compatible endpoint (gpt-image-2, dall-e-3). Use when the user wants to create, draw, illustrate, or edit an image, icon, logo, poster, or concept art. |
Generate Image (gemini / openai backends)
Turn a text prompt โ optionally with reference images โ into one or more images
via one of two explicit backends. --backend has no default โ always pick
one (gemini or openai), or set IMAGE_BACKEND. The script does the API
call, file saving, and configuration; your job is to pick the backend, craft a
strong prompt, and wire up the flags.
gemini โ Google's native Gemini API via google-genai. Full feature
set: --aspect-ratio, --size tiers (1K/2K/4K), and multi-image
edit/compose (-i a.png -i b.png). Use GEMINI_API_KEY / GEMINI_IMAGE_MODEL.openai โ any OpenAI-compatible image endpoint (OpenAI official,
DashScope, new-api gateways, ...). Use OPENAI_API_KEY / OPENAI_BASE_URL /
OPENAI_IMAGE_MODEL. Supports gpt-image-2, dall-e-3, etc.
Prerequisites
-
uv available (the script is a self-contained uv run script; deps install on first run).
-
The API key for the chosen backend. The script resolves it progressively, so any one works:
export GEMINI_API_KEY=... (gemini backend) or export OPENAI_API_KEY=... (openai backend), or- a
.env file (checked in order: $PWD/.env, then ${CLAUDE_PLUGIN_ROOT}/.env), or
--api-key ... on the command line.
CRITICAL -- Never paste the API key into chat or commit a .env. If the key is missing,
the script prints exactly how to set it โ relay that to the user rather than guessing.
Workflow
1. Clarify intent (only if genuinely ambiguous)
A one-line request like "draw a fox in a spacesuit" needs no questions โ just generate.
Ask (via AskUserQuestion) only when a choice would materially change the result and you
cannot reasonably default it, e.g. aspect ratio for a "banner vs. avatar", or whether an
attached image should be edited vs. used as style reference.
2. Craft the prompt
Read references/prompting.md before writing a non-trivial prompt. In short: describe
subject, composition, lighting, style, and mood in concrete terms; put any literal text to
render in quotes; and for edits, state what to change AND what to keep. A vivid one-paragraph
prompt beats a terse phrase.
3. Run the script
Invoke it directly (the shebang runs it through uv):
${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate_image.py "PROMPT" --backend BACKEND -o OUT.png [flags]
Flags:
| Flag | Purpose | Default |
|---|
--backend | gemini (native Google API) or openai (OpenAI-compatible endpoint). No default โ also via IMAGE_BACKEND. | required |
-o, --output | Output path (.png/.jpeg) | generated.png |
-i, --input | Reference/input image (repeatable). gemini: edit/compose multi-image. openai: edit single image (first wins). | none |
--aspect-ratio | 1:1 2:3 3:2 3:4 4:3 4:5 5:4 9:16 16:9 21:9. gemini: via image_config. openai: via extra_body (Gemini-compat endpoints only). | model decides |
--size | gemini: 1K/2K/4K. openai: free string (1024x1024, auto, ...). | model decides |
--count | Number of images. gemini: N separate calls. openai: one call, n=N. | 1 |
--model | Model id or alias (else GEMINI_IMAGE_MODEL/OPENAI_IMAGE_MODEL). gemini: pro/flash/raw id. openai: raw id (e.g. gpt-image-2). | backend default |
--quality | low/medium/high/auto โ openai backend only. | model decides |
--response-format | b64_json/url/none โ openai backend only. url is downloaded to disk; none omits the param. Some gateways require url or reject the param (use none). | b64_json |
|
Models (pass the alias to --model or set the backend's env var):
| Backend | Alias / id | Notes |
|---|
| gemini | pro (default) โ gemini-3-pro-image | highest quality |
| gemini | flash โ gemini-2.5-flash-image | faster / cheaper |
| openai | gpt-image-2, dall-e-3, gpt-image-1, ... | any id the endpoint serves |
Examples:
generate_image.py "podcast cover art" --backend gemini --aspect-ratio 16:9 --size 2K -o cover.png
generate_image.py "put the watch from image 1 on the wrist in image 2" --backend gemini -i watch.png -i wrist.png -o composite.png
generate_image.py "a red bicycle" --backend openai --base-url https://api.tu-zi.com/v1 \
--model gpt-image-2 --size 1024x1024 --response-format url -o bike.png
generate_image.py "a red bicycle" --backend openai --base-url https://api.openai.com/v1 --model gpt-image-2 -o bike.png
When the user wants choices to pick from, request --count 2 (or more) and show all outputs.
With -i, the prompt becomes an edit/compose instruction over the supplied image(s).
4. Report
Tell the user the saved path(s). If nothing was returned, it is usually a safety-filtered
prompt โ say so and offer a reworded prompt. Output files are images: reference them by path;
do not try to inline their bytes.
Configuration is progressive (the key best practice)
Backend, key, base URL, model, and quality are each resolved by
lib/progressive_env.py in this order, stopping at the first hit: CLI flag โ
process env โ .env chain โ built-in default. This is why the same command
works in a project with a local .env, in a shell with exports, or with
everything overridden inline โ and why a newer model or a different endpoint can
be selected with export GEMINI_IMAGE_MODEL=... / export OPENAI_BASE_URL=...
without touching code. See references/prompting.md for the parameter reference.
Files
scripts/generate_image.py โ the generator (google-genai for gemini, openai for openai).references/prompting.md โ prompt-writing guide and full parameter reference.${CLAUDE_PLUGIN_ROOT}/lib/progressive_env.py โ shared progressive config resolver.
