c2c-img: image generation that feels built in

To the user, you can simply make images. Behind the scenes, scripts/c2c_img.py has Codex's image_gen render your prompt exactly as written, and you do the art direction. The user should feel that they asked for an image and got one. Don't mention Codex, scripts, JSON, or your prompt unless they ask or something breaks.

Other agents: tool names below are Claude Code's. Use your own equivalents: a shell tool for the script, an image-viewing tool for checking the result, and a file-sharing tool (if any) for showing it.

The loop

1. Start right away. Don't ask clarifying questions for ordinary requests. Choose sensible defaults for aspect ratio, style, and composition from context, the way a good designer would. Ask only when a missing detail would make the result useless, such as the exact wording for a poster headline the user clearly cares about.
2. Write the prompt yourself and don't show it to the user. Follow references/prompting.md. Read it the first time you use this skill in a conversation.
3. Generate by running a single Bash call with timeout: 600000. Set the description to plain language, such as Generating image: red fox in snow or 正在生成图片：中秋海报. That line is what the user sees while they wait.
4. Check the result quietly by opening each image with the Read tool. If it has an obvious defect, fix it once before showing anything. Obvious defects are garbled or misspelled text, a missing required element, the wrong aspect ratio, or a watermark. Don't iterate on matters of taste.
5. Show it. If a tool for showing files to the user exists (such as SendUserFile with display: "render"), send the image or images so they appear inline. Otherwise, link the path. Add one short sentence about what you made. Offer one natural next step only when it helps, such as "Want a warmer palette or a portrait version?". Reply in the user's language.
6. Keep the thread going. Remember the paths you've produced. For requests like "make it brighter", "换成蓝色", or "remove the cup", edit the latest image with -i <last path> and restate what must stay the same. The script never overwrites files: it saves -v2, -v3, and so on automatically.

Running the script

python3 <skill_dir>/scripts/c2c_img.py [options] "<prompt>"

<skill_dir> is the base directory shown when this skill loaded. On Windows, use python or py -3 if python3 isn't available.

Option	Use
(no -o)	For preview or brainstorming images. They are saved to ~/Pictures/c2c_img/<time>-<slug>.png, which keeps the user's repo clean
-o PATH	For project assets. Put them where the project keeps images (for example public/, assets/, static/img/) and give them a descriptive name. The extension sets the format
-s WxH	Exact pixel size. The image is generated at the nearest aspect ratio, then scaled and center-cropped. Use it for icons, banners, and social cards
-f png|jpg|webp, -q 90	Output format and quality. Use jpg or webp for web photos and png for anything with transparency or crisp text
-t	Transparent background. Also write "isolated on a fully transparent background" in the prompt
-i IMG (repeatable)	An input image to edit or use as a reference. In the prompt, refer to them as Image 1, Image 2, … in the order given
-n N	N variants of the same prompt, run in parallel. Only use it when the user wants options
--batch -	Several different images in one call, run in parallel. Pass JSON lines on stdin, one per image: {"prompt", "out"?, "ref"?, "transparent"?, "size"?, "format"?}
--doctor	Checks the setup without generating anything. Run it when something fails for an unclear reason

Timing: a new image takes about 35–50 s, and an edit or a transparent image about 45–75 s. The server side varies from minute to minute. Parallel jobs finish in about the time of one job. For a batch of 3 or more images, tell the user in one short line that they're on the way (for example "Making 4 images, about a minute."). Run the batch with run_in_background: true if you have other work to do meanwhile.

Output on stdout is JSON. Progress goes to stderr.

{"ok": true, "images": ["/abs/a.png"], "results": [{"ok": true, "path": "/abs/a.png", "width": 1254, "height": 1254, "notes": [...]}]}

A failed job has error_type, error, and hint fields. Act on hint. When you explain the problem to the user, keep it short and non-technical.

error_type	What to do
generation_failed	Usually a content-policy refusal. Rephrase within the policy, or tell the user plainly why it can't be made. Never try to evade the policy
quota	The ChatGPT/Codex usage limit has been reached. Tell the user, including when it resets (resets_at) if that's given, and don't retry
capacity	OpenAI is overloaded, and another model was already tried. Wait a minute, retry once, then report it
auth	The user needs to run codex login with their ChatGPT account
timeout	Retry once, then report it
bad_request	Fix the arguments and retry. For example, if webp isn't available here, use png
model / unknown	Run --doctor and report what it shows. --no-daemon forces the simpler engine, and --stop-daemon restarts the background helper

Inputs from the user

• An image path or dragged-in file: pass the original file with -i. Say whether it's the thing to edit or only a style reference. Convert HEIC first (sips -s format jpeg in.heic --out in.jpg). Photos from a phone may be rotated by EXIF: if a preview looks sideways, pass an upright copy.
• An image pasted into the chat with no file path: you can see it, but the script needs a file. Ask the user, in one line, to drag the file in or give its path. If you have enough to describe it well, you can also describe it in the prompt and generate from scratch.

Judgment

• Choose AI edits over filter code. Restyling a user's photo (cartoon, pixel art, anime, watercolor, sketch, and so on) is an edit: pass the original with -i and list what must stay the same (faces and likeness, expressions, hair, outfits, pose, framing, background layout). The image model keeps the likeness and redraws it in the style. Filter code (PIL, OpenCV, ImageMagick) can only pixelate or posterize, and the result looks mechanical, so don't reach for it for these requests. Use code only for exact mechanical operations, or when the user explicitly wants a deterministic, algorithmic effect.

• Choose -s whenever the destination implies a size, such as a favicon, app icon, OG card, or banner.

• For images that go into a project, update whatever references them (HTML, CSS, markdown) if that is part of the task.

• Each image uses the user's ChatGPT/Codex quota. Make what was asked for. Don't generate speculative variants unless the user wants options.

• Anything that should be precise and editable, such as a chart, a diagram, a UI, or a simple geometric icon, is better built as code or SVG. Offer that instead.