imagegen

imagegen — Claude Code plugin

Generate real images from inside Claude Code, with zero API costs. The plugin delegates to OpenAI's Codex CLI and its built-in gpt-image tool — billed against your ChatGPT subscription, not the OpenAI API.

🎯 The point: once installed, Claude generates images automatically whenever your task involves one — websites, READMEs, blog posts, slide decks, game assets, app icons, marketing pages. No / commands needed. No "may I generate this?" prompts. The plugin's broad-trigger SKILL watches every conversation and fires when an actual rendered image (not a placeholder) would land better than a description.

Slash commands (/imagegen "<prompt>", /imagegen:edit) are still there for explicit control — but the typical day-to-day flow is just "build me a roofing company landing page" and the images appear, wired into your HTML.

Quick start

In any Claude Code workspace:

/plugin marketplace add github.com/colin-automates/Codex-ImageGen--Claude-Code
/plugin install imagegen@imagegen-marketplace
/imagegen:setup

/imagegen:setup auto-installs Codex CLI if it's missing, auto-launches a browser to sign into ChatGPT if you're not logged in, and runs a test image to confirm the pipeline. When it reports green, you're done — Claude will generate images on its own from there.

---

How auto-activation actually feels

The headline feature. Once the plugin is installed, you don't think about it. Examples that fire it without any slash command:

You say	What happens silently
"build me a hiking-blog landing page"	Claude scaffolds the HTML, then generates a hero photo, 3 service-section illustrations, and an OG image — all wired into the right <img> slots.
"add a banner to the README"	A 1536×1024 hero image lands at ./docs/images/banner.png and the README's empty image syntax is filled in.
"I need a logo for a coffee app, flat vector, two-tone"	One image generated; the prompt is interpreted style-aware.
"generate enemy sprites: red, blue, green"	Auto-routes to set mode — three coherent sprites in one Codex session, sharing palette/style.
"redesign this slide deck"	Empty hero slots in .md/marp/reveal.js decks get filled per-slide.
"fix the broken <img src=""> in landing.html"	Single image, saved next to the page, src="" rewritten with a real alt.

Trigger framing: the skill is a pre-flight check across any project type, not a keyword matcher. Before Claude ships any build — websites, apps, games, CLIs, libraries, docs, decks, brand work — it asks "does any surface here need a generated image?" and fires the tool if yes. So build-mode requests trigger it even when you never say "image": "build me a landing page for X", "scaffold a SaaS app", "build a platformer game" (sprites + tiles + character art), "set up a CLI tool" (README hero + OG card), "design a logo for Y", "write a tutorial on Z" (header + section illustrations). Same for explicit verbs ("generate / create / make / build / draw / render / design / edit / redo") applied to images, websites, apps, games, libraries, decks, or any visual surface.

Non-triggers (intentionally narrow): SVG icons that match an existing icon set already wired into the project (lucide, heroicons, etc.), charts/graphs from real data, screenshots of running code, or anything you explicitly told Claude not to generate ("no images", "SVG only").

Multi-image intent auto-detects too: if your prompt has an explicit count ("three icons"), a comma/and list, pipe-separated parts, or set-keywords (set / series / pack / roster), Claude calls generate_image_set once instead of generate_image N times — keeps style consistent and uses fewer ChatGPT plan turns.

Slash commands (for explicit control)

Command	What it does
/imagegen "<prompt>"	Generate one image, or a coherent set if the prompt describes multiple.
/imagegen setup	One-time setup probe — installs Codex CLI if missing, auto-launches browser sign-in if needed. Same as /imagegen:setup.
/imagegen:edit <path> <prompt>	Edit an existing image with a natural-language instruction.
/imagegen:img <prompt>	Alias of /imagegen.