takazudo/codex-imagegen

codex-imagegen

Anthropic models are vision-in / text-out — they cannot emit raster pixels. This skill produces real PNGs by delegating to Codex CLI's $imagegen (the gpt-image-2 model behind ChatGPT Images), and keeps it on ChatGPT-included usage by stripping OPENAI_API_KEY from the call so it never silently bills the OpenAI image API.

Use the script

scripts/codex-imagegen.sh runs the whole flow (ChatGPT-auth preflight → headless codex exec → save to disk → optional exact resize) and prints the saved path:

"$HOME/.claude/skills/codex-imagegen/scripts/codex-imagegen.sh" \
  --prompt "<what to make>" [--out <name.png>] [--in <image>]... [--size WxH]

• --prompt — what to generate, or how to redesign the --in image.
• --out — optional. A bare name/relative path lands in the repo-scoped cclogs dir; an absolute path is used verbatim; omitted auto-names from the prompt. See Output location.
• --in — attach an image as vision/edit source; repeatable (e.g. a screenshot to redesign + a brand/style reference).
• --size — exact output pixels, e.g. 600x600. $imagegen only emits preset sizes, so the script hard-resizes to your dims — sips on macOS, ImageMagick or Pillow on Linux/WSL (warns and keeps native size if none is present).

The script prints the absolute saved path. After it succeeds, Read that PNG to view/verify it and show the user.

Generate from scratch

scripts/codex-imagegen.sh --size 600x600 \
  --prompt "a realistic photo of a grey heron at the water's edge"

Redesign / polish from a screenshot

Pass the screenshot with --in. This is the design-polish path: the model works from the picture, not the app's source code.

scripts/codex-imagegen.sh --out redesign.png --in baseline.png \
  --prompt "redesign this pricing table to look premium — keep every number, column, and row identical; improve only spacing, typography, color, and dividers"

Output location

The script saves into the repo-scoped, Dropbox-synced cclogs dir — <cclogs-base>/<repo>/imagegen/ — resolved automatically by $HOME/.claude/scripts/get-logdir.js (the same {logdir} rule the other skills use; worktree-aware, _misc/ when outside a git repo). Don't hand-build the path: pass just a filename to --out, or omit it. An absolute --out overrides this when a specific destination is needed.

get-logdir detects the repo with git rev-parse from the current directory, so invoking the script from a non-repo cwd (a /tmp scratchpad, a Dropbox prototype dir) silently routes output to <cclogs>/_misc/imagegen/ instead of the repo. Pass --project-dir <repo-dir> to anchor resolution to the repo regardless of where the command runs — e.g. a design-polish loop that screenshots into /tmp should pass --project-dir "$REPO_ROOT" so mockups still land in <cclogs>/<repo>/imagegen/.

Cost — this burns ChatGPT plan usage fast

Each image costs roughly 90k–100k tokens of ChatGPT/Codex usage (image turns run ~3–5× heavier than text). Fine for a handful of design explorations; don't loop it dozens of times without telling the user. The script reports the tokens used per call.

Requirements & caveats

• Works on macOS and Linux/WSL. Pure-portable except --size, which prefers sips (macOS) then ImageMagick (magick/convert) then Pillow (python3 -m PIL) — generation itself needs none of these. get-logdir.js already resolves the WSL Dropbox cclogs path. The real per-machine requirement is Codex auth: $HOME/.codex/auth.json is not shared across machines, so run codex login (ChatGPT) once on each (Mac and WSL).
• Codex must be logged in with ChatGPT (codex login status → "Logged in using ChatGPT"). The script warns if not. If only API-key auth is available, this skill's billing premise doesn't hold — say so rather than silently spending API credits.
• The env var OPENAI_API_KEY would flip Codex to API billing; the script unsets it per-call to stay on ChatGPT usage. That's the whole point — don't remove that.
• Never trust exact text/numbers in the output. Image models regenerate (not pixel-copy) and drift on dense small text — price lists, tables, microcopy are the worst case. Treat a redesign PNG as visual direction only; re-key real data from the source and verify before any of it informs production code.