Imported from cjnhust/chubby-skills (owned/codex-image-render/SKILL.md). Install upstream with npx skills add cjnhust/chubby-skills --skill codex-image-render. Copyright stays with the author.

Codex Image Render

Use this skill only as a leaf renderer when the current turn clearly exposes the built-in image_gen tool.

This skill does not own workspace creation, style normalization, prompt staging, review policy, or downstream routing. It only executes rendering through the built-in tool and saves outputs to explicit target paths.

Do not use this skill when the user explicitly asks for provider, model, API, CLI, batchfile, jobs, json, or the existing baoyu-image-gen script path. In those cases, return control to baoyu-image-gen so it can stay on its API/CLI path.

Inputs

• saved prompt files and/or raw prompt
• optional reference images or edit targets
• explicit output path per asset

Rules

• Require an explicit output path for every asset.
• Treat saved prompt files as the source of truth when they exist.
• For multiple assets, issue one built-in image_gen call per asset.
• For local reference images or edit targets that are not already visible in the conversation, load them with built-in view_image first so they are available to the built-in path.
• Support reference-image generation by capability equivalence. Do not describe it as CLI --ref.
• Do not promise provider, model, batchfile, jobs, json, or session semantics. This skill uses the built-in tool only.
• The built-in tool saves under $CODEX_HOME/generated_images/... by default. After generation, move or copy the selected result to the requested workspace path.
• Do not overwrite an existing output unless the caller or user explicitly asked for replacement. Otherwise create a sibling versioned filename.
• If the current turn does not actually expose built-in image_gen, stop and return control to the caller instead of guessing.

Workflow

1. Confirm the current turn clearly has built-in image_gen.
2. Normalize prompt inputs.
   • If prompt files were supplied, read them and assemble the final prompt.
   • If only a raw prompt was supplied, normalize it without changing the requested semantics.
3. Normalize image inputs.
   • Label each image as reference image or edit target.
   • Load local files with view_image first when needed.
4. Call built-in image_gen once per asset.
5. Inspect the result, choose the intended final, and copy or move it from the default generated-images directory to the explicit output path.
6. Report the saved output path and prompt used back to the caller.