simonwillison.net web signal

Willison builds 3D raccoon game one-shot with Claude Fable 5

TL;DR

  • Simon Willison used Claude Fable 5 inside Claude Code for web to build a playable 3D raccoon game in about 34 minutes and 7 commits.
  • The agent picked Three.js, generated all textures with OpenAI's gpt-image-2, wrote a procedural WebAudio soundtrack, and shipped to GitHub Pages with Playwright tests.
  • Willison called the finished game mediocre but the starting point impressive, concluding that designing games that are fun remains a uniquely human trait.

A single prompt, about 34 minutes of wall-clock time, seven commits, and you have a playable 3D browser game where a raccoon steals items from a suburban yard while flashlight-wielding guards and a scent-tracking dog try to catch it. That is what Simon Willison walked through on his blog after handing Claude Fable 5, running inside Claude Code for web, a design brief and telling it to work independently rather than ask any further design questions.

The build is a decent snapshot of where end-to-end agent tooling has landed. Claude picked Three.js for rendering, generated all textures and title art through OpenAI's gpt-image-2 image model, wrote a procedural WebAudio soundtrack with no audio files on disk, and pushed the whole thing to GitHub Pages with Playwright tests along the way. It found and fixed two of its own bugs: a raccoon that was invisible on mobile because the canvas was rendering at 2x size, and a win screen button that was unresponsive because of inherited CSS.

The interesting part for anyone thinking about agents in production is how modest Willison is about the result. His own summary is that as a finished game project it is mediocre, but as a starting point from a single prompt it is very impressive. The gameplay itself, he says, is thin and too easy, and his closing observation is that designing games that are fun remains a uniquely human trait.

The honest caveat is that a demo like this is a best-case artifact, not a benchmark. What the write-up does not give you is a token bill, a bug count beyond the two Willison caught, or a sense of how much of the seven-commit trajectory depended on the specific brief he wrote up front. Take it as a data point on how far an unattended agent can now carry a self-contained web project, not as evidence that any brief will land this cleanly.

For teams that live on prototypes, whether hackathons, pitch decks, learning materials, or throwaway internal tools, the floor of runnable-in-under-an-hour has visibly moved. What agents still cannot supply is the taste that makes any of it worth playing.

Shared on Bluesky by 2 AI experts