Simon Willison took a four-year-old tweet and turned it into a working 3D game in a single Claude Fable 5 session. The prompt had two images and a few lines of instructions. No further design decisions. The model built the whole thing.
The original tweet from August 2022 had a GPT-3 text completion describing "Raccoon Heist", a game where a team of raccoons pull off daring heists, and a DALL-E generated screenshot of what that game might look like. Four years later, Willison fed both into Claude Fable 5 running in Claude Code for web. The model wrote the game, generated textures through OpenAI's gpt-image-2 API, and committed it to a GitHub Pages repo. The result is a playable 3D browser game at simonw.github.io/raccoon-heist.
How did Claude Fable 5 build the game with just a tweet?
Willison wrote his prompt in the notes app on his phone. No tech stack specified. He told Fable to build for the browser, make it mobile friendly, use OpenAI for textures, work independently, and commit often so he could preview the results. The model chose Three.js on its own, generated raccoon-themed image textures, and built an isometric 3D street with a bank to rob.
The most instructive part of Willison's write-up is the workflow trick. He set up GitHub Pages to deploy from the Claude Code branch, which meant each push gave him a live URL within 30 seconds. That loop, write and preview instantly, turned a speculative experiment into something that felt like pairing with a game designer.
What makes this more than a novelty?
A model that can go from "here's a screenshot and a concept prompt" to "here's a playable game, no more questions needed" has moved past the novelty stage. Fable 5 is a coding model that makes design decisions, picks its own stack, handles image generation as a subroutine, and pushes to production without asking permission. The raccoon game is a toy, sure. But the pattern is real: give the model a goal, assets, and a deployment target, and it ships.
I wrote about how far coding models have come in a year yesterday, and Fable 5's one-shot game build is a better demo than any benchmark number. 95% on SWE-bench is abstract. Watching it build a raccoon heist game from a four-year-old tweet and an SVG prompt is concrete.
Frequently Asked Questions
Did Claude Fable 5 build the game from scratch?
Yes. Willison gave it a prompt and two images. No starter code, no framework hints. Fable 5 chose Three.js, generated textures via OpenAI, and built the entire game in one session without asking for clarification on any design decision.
Can I play the Raccoon Heist game?
Yes. The game is live at simonw.github.io/raccoon-heist. The source code is on GitHub at github.com/simonw/raccoon-heist.
What was the original tweet from 2022?
The tweet combined a GPT-3 text completion describing a raccoon-themed heist game with a DALL-E generated screenshot. Willison used it as the creative brief for Fable 5, providing both the text concept and the visual reference.
How did Willison test the game while Fable built it?
He configured GitHub Pages to deploy from the Claude Code branch. Every time Fable committed and pushed, the game was live at a public URL within 30 seconds. This let him preview incremental progress without interrupting the model's workflow.
Could this approach work for building real applications?
The same pattern, prompt plus assets plus deployment target, is already being used for prototypes and small applications. The limitation is scope: complex multi-service apps still need architecture decisions a model can't make alone. But for single-page applications, tools, and games, the barrier just dropped significantly.
Top comments (0)