This AI Agent Reads Your Article and Illustrates It. Then It Critiques Itself.
There's a specific pain that every content creator knows: you've written something good, and now you need images to go with it. You can either spend an hour in a design tool, or you can ship it with no visuals and feel vaguely bad about that.
A developer called ε°δΊ (Xiaohu) just open-sourced something that might change that calculation. It's called IP Studio β an AI agent that reads your article, picks an illustration type, generates the image, and then reviews its own output. It runs inside Claude Code. It's free. And it ships with 31 original character designs.
How It Works
IP Studio isn't a wrapper around an image API. It's a pipeline with actual decision-making at each step.
Here's what happens when you run it:
- The agent reads your article to understand the content β not just keywords, the actual meaning and tone.
- It plans the illustration type based on what it reads. It picks from three formats: emotion graphics (for mood and human moments), explanatory diagrams (for processes and systems), and four-panel comics (for narrative or humor).
- It generates the image through your connected API key.
- It reviews its own output β checking if the illustration actually matches the content β and revises if something's off.
That last step is the part worth paying attention to. Most image-generation workflows are one-shot: you prompt, you get an image, you decide if it's good enough. IP Studio treats generation as a draft and review as part of the process. It's the same self-correction loop that good human designers use β just running automatically.
31 Characters, 5 Visual Styles
One of the more thoughtful parts of the project: the developer included 31 original character designs to give the illustrations a consistent visual identity. Fifteen are hand-drawn line art characters; sixteen are meme-style figures built on wordplay.
The default rendering is hand-drawn line art with light color washes β the kind of illustration style that looks clean alongside text without fighting for attention.
If that doesn't fit your aesthetic, you can switch between five visual modes: 3D blind box, black-and-white line art, and three more. The same characters render in each style, so your illustrations stay coherent across a project even when you change the look.
Setup Is Minimal
You need Python 3 and an OpenAI-compatible image API key (it defaults to GPT-image-2, but any compatible endpoint works). Clone the repo, configure your key, and it's ready to run. The developer specifically called out support for Claude Code and Codex as agent environments.
If you'd rather not use an image API at all, the agent can output prompts instead of images β you copy them into whatever image tool you prefer and generate there.
The whole thing is designed to be lightweight and composable, not a locked-in product.
Why the Self-Review Loop Matters
The image-generation space is full of tools that produce output. What's harder to find is tools that know when their output is wrong.
IP Studio's self-check step isn't just quality control β it's what makes this an agent rather than a generator. An agent has a goal (illustrate this article well) and takes actions to reach it (generate, evaluate, revise). A generator produces output and stops. That distinction sounds academic until you're shipping content at scale and you don't have time to manually review every image.
The agent model β plan, act, evaluate, revise β is the pattern that makes AI useful for real workflows rather than one-off tasks.
What This Means If You Use OpenClaw
The best part about IP Studio isn't the 31 characters or the five visual styles. It's that it demonstrates something ClawWorld users already know: the most powerful AI tools aren't chat interfaces. They're skills β composable agents you build once and slot into your workflows.
IP Studio is exactly that. It's a skill for Claude Code that you install, connect to your content pipeline, and run as part of publishing. It doesn't replace your judgment β it handles the parts of the job that don't require it.
On OpenClaw, that's the whole idea. You define what you want done. The agent handles the execution, checks its own work, and hands you something ready to use β not something that still needs half an hour of cleanup.
Creative work is the last frontier where AI agents have felt genuinely clunky. Projects like this one are evidence that's changing.