ワンクリックでManusで任意のスキルを実行

codex-imagegen

スター10

フォーク2

更新日2026年6月22日 14:23

Generate or redesign raster images (PNG mockups, illustrations, photos, UI design polish) from inside Claude Code by driving Codex CLI's built-in $imagegen, billed to ChatGPT-included usage instead of per-image OpenAI API charges. Use whenever the user wants to create/generate an image, make a mockup or illustration, produce a photo, polish or redesign a UI from a screenshot, or get a PNG out of a session — even if they don't mention Codex or "imagegen". This is the preferred image-generation path because Anthropic models cannot generate raster images and this route avoids API billing.

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

Manusで実行

ソース

Takazudo

Takazudo/claude-resources

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Manusで実行

ファイルエクスプローラー

2 ファイル

SKILL.md

readonly

name

codex-imagegen

description

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.

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: ~/.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.

このリポジトリの他の Skills

同じリポジトリ

issue-sweep

Takazudo/claude-resources

Sweep open GitHub issues — optionally narrowed by label — and drive each to completion via `/big-plan -m -a`. Collects candidates, triages out work that needs careful human judgment, confirms once, then handles the rest autonomously one issue at a time. Use when: (1) User says '/issue-sweep', 'sweep issues', 'sweep open issues', 'handle the open issues', 'clear the issue backlog', or 'do the issues', (2) The user wants to batch-process a label's worth of issues (e.g. `agent-found`, `mac`), (3) After a work round that left a pile of follow-up issues. Options: "`-f`/`--filter LABEL` keeps only issues carrying that label; `-ex`/`--exclude LABEL`" drops issues carrying it; with no options it sweeps ALL open issues. Locally-checkable "verify X" tasks are verified directly and closed (not run through `/big-plan`). Skips issues that need careful human judgment (huge multi-major version bumps, super-big epics, design calls); verification tasks that can't be auto-verified locally (auth/deploy/external/subjective) are

2026-06-2510

design-polish-loop

Takazudo/claude-resources

Iterative visual design-polish loop for an existing or WIP web UI. Each round: capture the live page as screenshots, judge it from the rendered pixels, and apply CSS polish toward a better-looking target using /css-wisdom's tight-token strategy — then re-capture and compare. The target direction is set by codex-imagegen 'north-star' mockups generated from the current screenshot, so the loop pursues a genuinely better look instead of nudging the existing one in place. Use when: (1) user says 'design-polish-loop', 'polish loop', 'polish this page/design/UI', (2) user wants to level up / upgrade the look of a current design or WIP project, (3) user wants AI-proposed redesign directions applied to a real rendering page, (4) user calls a design plain, dated, flat, or unpolished and wants it improved. Main use case: improving the current design of a WIP project. It polishes what already renders — not brand-new-from-scratch pages.

2026-06-2210

dev-bump-zudo-deps

Takazudo/claude-resources

Bump every `@takazudo/*` registry dependency in the current project to its newest version, resolving each by the channel it currently tracks. Use when: (1) User says '/dev-bump-zudo-deps', 'bump zudo deps', 'bump the @takazudo packages', 'update zfb', 'update zudo-doc', or 'update the takazudo toolchain to latest', (2) A new `@takazudo/*` prerelease or release has shipped and this project should adopt it, (3) Routine first-party dependency-update rounds. Resolution rule: a dep currently on the `next` prerelease line (`0.1.0-next.N`) → newest `next` release; a dep on the literal `latest` tag → `dist-tags.latest`; a plain stable semver → newest stable. Skips `workspace:`/`file:`/`link:` specs and leaves operator style (exact / `^` / `~`) and prerelease channel intact. A bundled script does the discovery + semver resolution; this skill drives install + verify + report.

2026-06-2210

big-plan

Takazudo/claude-resources

Plan implementation by breaking work into one epic GitHub issue + child sub-issues. Use when: (1) User says '/big-plan', (2) User wants to plan an implementation before coding, (3) User wants to split a feature into small issues for parallel agent team work, (4) User references existing issues (e.g. 'implement issue #123', 'plan all open issues', 'plan recent 3 open issues'). Auto-reads project-scope l-lessons-* skills (from /retro-notes) before planning. Codex (-co) is the DEFAULT Step 5 plan reviewer — it runs even when no reviewer flag is passed; -op/-gco add more second opinions (runnable in parallel). Use -a/--auto to run the chain autonomously: skip the Step 6 confirmation wait (pausing only when something needs careful consideration), auto-create the issues, then auto-invoke the implementation skill in the same session — /x-wt-teams for a multi-sub-issue plan, /x-as-pr for a single-sub-issue plan — forwarding -a so multi-wave plans auto-chain wave after wave. Use -m/--merge to also merge the final PR i

2026-06-1810

dev-purge-gh-cache

Takazudo/claude-resources

Purge GitHub Actions caches across every non-archived repo for one or more GitHub owners (users or orgs). Use when (1) the user says "purge gh cache", "clear actions cache", "delete CI cache", or invokes /dev-purge-gh-cache, (2) the user wants to free up GitHub Actions cache storage for an account/org, (3) the user names one or more GitHub owners (e.g. "Takazudo", "zudolab") and asks to wipe their Actions caches. Handles fan-out across all repos, parallel survey, bulk delete, and real-time verification.

2026-06-1810

dev-upstream-report

Takazudo/claude-resources

Report bugs and improvement ideas found in Takazudo's own upstream packages as GitHub issues on the upstream repo. The current project depends on packages built by the user under github.com/takazudo and github.com/zudolab (e.g. zudolab/zudo-doc, Takazudo/zudo-front-builder, the @takazudo/* npm scope). Use when: (1) User invokes /dev-upstream-report — with a description to report something now, or with no args to enable upstream-watch mode for the rest of the session, (2) During development a bug, limitation, or missing feature is traced to one of these upstream packages rather than the current project, (3) User says 'report upstream', 'upstream issue', 'file it on the package repo', or blames one of these packages for a problem.

2026-06-1610

name

codex-imagegen

description

codex-imagegen

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

Cost — this burns ChatGPT plan usage fast

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: ~/.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.