Claude Fable 5 One-Shots Playable "Raccoon Heist" Game from Single Prompt
Veteran developer Simon Willison has demonstrated the impressive capabilities of Anthropic's Claude Fable 5, using the LLM to generate a complete and playable "Raccoon Heist" game from a single text prompt. The project, initiated on the fourth anniversary of an earlier GPT-3 game concept tweet, showcases how far AI-assisted game development has progressed. Claude Fable 5, operating within Claude Code for web, autonomously created the top-down stealth mission where players control raccoons to collect loot while avoiding guards. The AI not only handled game logic, collision detection, and enemy AI, but also generated visuals using OpenAI's gpt-image-2 for textures. The resulting game, hosted on GitHub Pages, features 3D graphics, procedural audio, and mobile touch controls. Willison's experiment highlights the model's ability to interpret high-level instructions and produce complex, multi-file programming tasks in a "one-shot" manner without manual edits. While Willison notes the gameplay is "shallow and easy," it serves as a powerful proof point for AI's potential in accelerating game prototyping. The source code for the game is publicly available on GitHub.
Sources
Reporting credited to:
- Primary source — simonw.github.io
Byte Marlowe is the pen name of the Game Pixel AI desk. Reported facts are credited to the outlets above; analysis is our own. How we work.