generate images from Claude Code
on your ChatGPT subscription

1Why this exists

You already pay for ChatGPT. You already have Codex CLI installed. And Codex CLI happens to ship with a built-in $imagegen tool. So why pay again for OpenAI Images API credits when you want Claude Code to generate an image?

This skill bridges Claude Code and Codex CLI: Claude phrases the prompt, the skill shells out to codex exec '$imagegen ...', and the PNG lands wherever you asked. ~40-70 seconds per image, zero marginal cost on top of your subscription.

2What it can produce

A tiny Shiba Inu wearing a red bow tie, watercolor illustration — first sanity test.

Tech-blog cover: a shield protecting a server rack, pastel 3D illustration.

Tech-blog cover: a friendly bot bridging Telegram and a code editor.

These were generated by the same script in this repo, with a shared pastel-illustration style prompt appended to each per-image concept. It scaled cleanly to a 26-image batch.

3Just want raw Codex (no skill)?

If you don't need Claude Code in the loop, call $imagegen straight from Codex:

# interactive
codex
> $imagegen a tiny shiba inu wearing a red bow tie, watercolor, no text

# non-interactive (one-shot)
codex exec -C "$(pwd)" -s workspace-write --skip-git-repo-check \
  '$imagegen a tiny shiba inu wearing a red bow tie, watercolor, no text'

Either way, the PNG lands at ~/.codex/generated_images/<session-id>/ig_*.png. This skill exists because the non-interactive form is awkward to script — you have to parse Codex's stdout for the session id, find the PNG, copy it out, and dodge Codex's sandbox eating your cp. The wrapper handles all of that.

Availability: $imagegen shipped as a built-in skill in Codex CLI v0.117.0 on 2026-03-26; upgraded to gpt-image-2 in v0.123.0 on 2026-04-23. Make sure your Codex is up to date with npm i -g @openai/codex.

4Install

Prerequisites: Codex CLI installed and codex login done.

# clone the repo straight into your Claude Code skills directory
git clone https://github.com/yazelin/codex-imagegen-skill ~/.claude/skills/codex-imagegen
chmod +x ~/.claude/skills/codex-imagegen/codex-imagegen.sh

Restart Claude Code and the codex-imagegen skill is available.

5Verify

Inside Claude Code, just ask:

Generate a cute Shiba Inu with a red bow tie in watercolor style,
save to /tmp/test-shiba.png

Or run the bundled script directly (text-to-image, 2 args):

~/.claude/skills/codex-imagegen/codex-imagegen.sh \
  "a tiny shiba inu with a red bow tie, watercolor, no text" \
  /tmp/test-shiba.png

6Multi-image edit · new

Pass 1–4 reference images as trailing positional args to do composition / outfit-swap / scene-merge / style-transfer / text-localization. The script builds the canonical Use case: image-edit / Input images: Image 1: … / Image 2: … scaffolding that gpt-image's edit mode keys off, and emits one --image flag per reference into codex exec.

# composition: subject from image 1 placed into scene from image 2
~/.claude/skills/codex-imagegen/codex-imagegen.sh \
  "place the subject from image 1 into the scene from image 2; match lighting and perspective" \
  /tmp/composite.png \
  ~/Pictures/person.png \
  ~/Pictures/kitchen.png

# outfit-swap with 2 garment references
~/.claude/skills/codex-imagegen/codex-imagegen.sh \
  "replace the clothing on image 1 with the garment from image 2 and the shoes from image 3; preserve face, body, and pose" \
  /tmp/outfit.png \
  ~/Pictures/model.png \
  ~/Pictures/jacket.png \
  ~/Pictures/shoes.png

For multi-image prompts, refer to inputs by their position in the arg list: image 1, image 2, etc. Typical use cases:

• composition — "place subject from image 1 into the scene from image 2"
• outfit-swap — "replace clothing on image 1 with garments from image 2 and image 3"
• style-transfer — "apply the visual style of image 1 to the subject in image 2"
• text-localization — "translate every text element in image 1 to Japanese; preserve typography"
• sketch-to-render — "turn the line drawing in image 1 into a photorealistic render"

The 4-image cap is a practical default — gpt-image accepts more, but in practice 2–3 references give the cleanest composition. Bump the cap in the script if you have a real need.

7What's in the box