On August 5, 2026, Simon Willison turned a four-year-old joke into a shipped, playable game. In 2022, he’d fed GPT-3 the prompt “write a detailed product description of a computer game where a team of raccoons go on heists” and paired it with a DALL-E image of a raccoon burglar crew. It was a bit, not a spec. This month he handed that old bit to Claude Fable 5 running in Claude Code for web, and it built an actual 3D browser game: seven commits, real bugs, a real fix log.
From GPT-3 gag to Three.js game
The original material was pure novelty: a GPT-3-generated product description (“In ‘Raccoon Heist’, you and your team of thieving raccoons are tasked with pulling off a series of daring heists…”) and a DALL-E screenshot mockup. Nobody built anything at the time. There was no reason to; it was a tweet, not a project.
Four years later, Willison created a new GitHub repository, opened a Claude Code session, and handed the model the old prompt and images with one addition: build it as a browser-based 3D game with mobile compatibility. The model worked largely on its own from there. It vendored the Three.js library, wrote Python scripts to generate textures and title art, built the core raccoon-heist gameplay loop, and used OpenAI’s gpt-image-2 to generate the title screen key art and seven ground/wall textures, committed as static assets alongside the code.
The bugs were real, and so was the fix log
This wasn’t a flawless one-shot. Three distinct bugs showed up during development, and the value of the exercise is in how they got caught and fixed, not in pretending they didn’t happen:
- Invisible characters on mobile. The raccoon, the dumpster, and the crew members simply didn’t render at mobile viewport widths — fixed by correcting visibility settings that only worked at desktop sizes.
- Canvas rendering at 2x on real phones. The root cause, once found: a
cssTextassignment was silently wiping out Three.js’s own inline sizing. This is the kind of bug that’s easy to miss without testing on an actual device, not a simulator. - The win screen wasn’t clickable. The star-rating div on the win screen inherited full-screen CSS meant for the title screen, which ate every tap before it could register as a button press. Fixed with a CSS specificity correction.
All three were caught because the build included Playwright-driven automated testing across desktop, portrait-phone, and landscape-phone viewports — checking pickup, banking, chase sequences, day/night progression, and win/retry flows. That’s the detail worth taking seriously: the bugs weren’t found by eyeballing a demo GIF, they were found by actually testing the built thing on the surfaces where it would break.
What actually shipped
Seven commits landed on the branch claude/3d-raccoon-heist-game-50n293, deployed via GitHub Pages with branch-based static file serving. The stack: Three.js for the game engine, WebAudio for a procedural soundtrack, gpt-image-2 for all visual assets, Playwright for cross-viewport testing. A guard-dog mechanic got added on the third night of iteration, giving the core loop actual escalating difficulty instead of a flat challenge curve.
Willison’s own verdict, worth quoting directly rather than summarizing: “As a finished game project, it’s mediocre. As a starting point from a single prompt I think it’s very impressive.” He’s also on record that the gameplay lacks depth and gets boring as it progresses, even with the difficulty mechanics in place. That’s not a caveat tacked onto marketing copy — it’s the actual assessment from the person who built it, and it’s a more useful data point than a polished highlight reel would be.
What this means if you have your own dead tweet lying around
The interesting part isn’t “AI made a game.” It’s the shape of the workflow: an old, low-effort concept (a tweet, a meme, a mockup image) fed into a coding agent with a single directive, producing a real repository with real commits, real automated tests, and real bugs that got triaged and fixed without the human touching a line of code. The finished game being mediocre by Willison’s own account is arguably the more honest signal — this isn’t a cherry-picked demo, it’s what a single prompt actually produces against a real testing bar.
If you’re a solo builder sitting on a joke concept, a scrapped pitch, or an old mockup that never got past the idea stage, this is a concrete, testable pattern: point a coding agent at it, ask for something deployable with real testing, and see what comes back. The value isn’t in perfection. It’s in going from “tweet from 2022” to “seven commits and a working build” without writing the code yourself.