| name | codex-imagegen |
| description | Generate or AI-edit raster images with Codex for banners, heroes, demo assets, photo edits, and cutouts. Not for compression or resizing. 日本語の依頼例:「画像を生成」「写真を加工」「バナー画像」「ヒーロー画像」。 |
Generate images directly in Codex
Use Codex's image generation capability directly. Do not start a nested codex exec process and do not register Codex itself as an MCP server.
Workflow
- Load the built-in
imagegen skill when it is available and follow its image inclusion rules.
- For a new image, describe the subject, composition, visual style, lighting, palette, aspect ratio, and intended placement.
- For an edit, inspect the source image first and include every target image in the edit request. Preserve identity, labels, and composition unless the user asks to change them.
- Generate to the requested workspace path or move the resulting asset there without altering unrelated files.
- Inspect the result visually. Retake only the failed image, not the whole batch.
Composition and crop safety
- Prefer native landscape, portrait, or square generation closest to the target ratio, then crop and resize with the project's normal media tooling.
- When cropping is expected, state the exact surviving region and treat the outside area as expendable background bleed.
- Keep text, faces, logos, and critical objects inside the safe central region.
- For Japanese text, inspect every character. Prefer adding production text in HTML/CSS or design tooling when exact typography matters.
Quality gate
- Verify subject fidelity, spelling, crop safety, contrast, and absence of unintended artifacts.
- For image-to-image work, compare before and after to confirm required details survived.
- Budget two or three focused attempts for important assets; trial one image before generating a batch.
- Apply
images-media after generation when the asset also needs format conversion, responsive variants, compression, or LCP/CLS optimization.
