| name | image-studio |
| description | Generate Cowrite article illustrations, Xiaohongshu images, presentation visuals, covers, and structured diagrams with Codex's built-in image_gen tool. Use for 配图, 整篇配图, 小红书图片, PPT 图片, 封面, 逻辑图, 流程图, 架构图, or any Cowrite task that needs raster image generation. |
Image Studio
Create polished raster images with Codex's built-in image_gen tool. Never call LabNana, Gemini, OpenAI API/CLI, a local generation script, or any other external image provider. Do not require or request an API key.
If the runtime does not expose the built-in image_gen tool, stop and explain that image generation requires Codex with built-in image generation. Do not silently downgrade or insert a placeholder.
Image types
| Type | Default ratio | Reference |
|---|
| Xiaohongshu cover or series | 3:4 | references/types/xhs-cover.md |
| Presentation visual | 16:9 | references/types/ppt-illustration.md |
| Structured chart or diagram | adaptive | references/types/chart.md |
| Article logic illustration | 16:9 | references/types/article-logic.md |
Read the matching type reference before prompting. For style, select one file from references/styles/ and keep its palette, typography, spacing, and visual language intact.
Confirmation rules
Confirm these four decisions before generation:
- image type and aspect ratio;
- complete content and required exact text;
- one visual style;
- automatic generation or prompt review first.
Treat a Cowrite command that explicitly lists these confirmed values as confirmation already completed. Do not ask again.
Prompt assembly
Read references/prompt-templates/base.md. Build a concise production prompt containing:
- use case and intended placement;
- exact aspect ratio and composition;
- selected style with complete palette and visual identity;
- one core idea and no more than 3-6 essential concepts;
- exact visible text, kept short and primarily Simplified Chinese;
- no logo, watermark, signature, page number, fake UI text, or decorative filler.
For a multi-image set, repeat the complete shared visual identity in every prompt. Do not assume the image model remembers an earlier call.
Built-in generation workflow
- Call Codex's built-in
image_gen tool directly. For a brand-new image, do not pass reference-image parameters.
- Generate one distinct asset per tool call. A batch request means repeated built-in calls, not an API or CLI batch script.
- For later images that must match an earlier generated image, use the first accepted image as a style reference only when the built-in tool and current context support it. Still repeat the full visual identity in the prompt.
- Inspect every result for composition, text accuracy, palette, watermark, cropping, and consistency. Retry once with one targeted correction when necessary.
- For a project-bound image, use the tool's saved output path or output hint, then copy or move the accepted image from Codex's generated-images area into a writable project directory. Never leave a referenced project asset only under the default Codex output directory.
- Never overwrite an existing asset unless explicitly requested. Use stable numbered filenames or a version suffix.
The built-in tool controls the actual raster dimensions. Express 16:9, 3:4, 2K-quality intent, and other layout constraints in the prompt; do not invent unsupported tool arguments.
Cowrite insertion workflow
For a Cowrite task:
- Read the latest page and revision before planning.
- Generate and inspect the requested image or image set with built-in
image_gen.
- Copy accepted results into a local writable project directory.
- Upload only accepted final images with
cowrite_upload_asset.
- Re-read the page and confirm every exact anchor.
- Insert with
cowrite_insert_after and the latest revision. For multiple anchors, insert from bottom to top while preserving article-order numbering.
- Never rewrite unrelated text, upload prompts, or insert failed images.
Design principles
- One image, one idea.
- Structure before decoration.
- Generous whitespace.
- Correct, minimal Chinese text.
- A single coherent style per series.
- No AI clichés, unnecessary icons, fake product logos, or unexplained visual noise.
