gpt-imagen

GPT Imagen

Use this skill when the user wants Claude to actually produce an image asset, not just brainstorm one.

Provider routing

The gpt-imagen binary handles provider detection automatically. Do not check for API keys or Codex login status yourself. Always run the binary and let it decide. It will:

1. Use Codex if the CLI is installed and logged in
2. Fall back to the OpenAI API if an API key is configured
3. Print a clear error if neither is available

Never refuse to run the binary because you think a provider is missing. Never ask the user for an API key before trying. Just run it.

Routing rules

1. If the request references the current project, read only the minimum repo context first.
2. Turn the request into a compact structured prompt.
3. Run the gpt-imagen binary to generate the image.
4. If the user names an output path, pass --out.
5. If the user provides reference or edit images, pass repeated --image flags.
6. Report the saved file path and the provider path used.

Prompt shaping

Use this structure when the user prompt is vague:

Use case: <category>
Asset type: <where it will be used>
Primary request: <main ask>
Scene/backdrop: <setting>
Subject: <visual subject>
Style/medium: <photo | illustration | 3D | pixel art | infographic>
Composition/framing: <viewpoint and layout>
Lighting/mood: <lighting and tone>
Color palette: <palette>
Constraints: <must include or preserve>
Avoid: <negative constraints>

Execution

cat >/tmp/gpt-imagen-prompt.txt <<'EOF'
<structured prompt>
EOF

gpt-imagen --out <optional-path> - < /tmp/gpt-imagen-prompt.txt

When images are present:

gpt-imagen --image /abs/path/ref1.png --image /abs/path/ref2.png --out <optional-path> - < /tmp/gpt-imagen-prompt.txt

Constraints

• Do not use this for deterministic SVG or exact publication charts.
• Do not claim exact typography or pixel-perfect interface fidelity.
• Prefer one strong image first, then iterate.