Claude Fable 5 autonomously builds a browser-based 3D raccoon heist game from a 2022 prompt
An agentic coding session shows Fable 5 creating a playable Three.js game, generating textures via OpenAI’s image API, and shipping 7 commits with automated testing—all from a single prompt and two reference images.
1 source · cross-referenced
- Claude Fable 5 completed a 3D browser game from a four-year-old prompt and two reference images without further human input.
- The agent used Three.js, OpenAI’s gpt-image-2 for textures, and Playwright for cross-device testing, committing 7 times with automated playthroughs.
- The resulting game includes guard flashlights, a scent-tracking dog, procedural audio, and mobile touch controls, but the author judges it ‘mediocre’ as a finished game.
Simon Willison used a 2022 GPT‑3 prompt and two DALL‑E images to test whether Claude Fable 5 could autonomously build a playable browser game. He provided the original materials and a concise prompt instructing Fable 5 to create a mobile‑friendly 3D browser game, generate textures via OpenAI’s image API, and commit changes frequently for preview.
Fable 5 chose Three.js, vendored the library, and wrote a Python script to generate seven ground/wall textures using OpenAI’s gpt‑image‑2. It then implemented the game loop, guard flashlights, a patrolling dog with scent tracking, and a procedural WebAudio soundtrack without any audio files. The agent also added mobile touch controls and automated testing via Playwright across desktop and phone viewports.
The session produced seven commits on a GitHub branch, each verified by automated Playwright playthroughs. Fable 5’s notes.md documents design decisions such as the dog’s 12‑unit detection radius and 17‑unit escape gap, plus fixes for rendering bugs discovered during testing.
Willison characterizes the result as an impressive technical achievement but a mediocre game: the mechanics are simple, the difficulty low, and the entertainment value limited. He emphasizes that while agents can rapidly assemble functional software, designing genuinely fun experiences remains a human skill.
- Aug 6, 2026 · Simon Willison — everything
Meta unveils Muse Spark 1.2 with Muse Code co-training and long-horizon coding benchmarks
Trust79 - Aug 5, 2026 · Simon Willison — everything
llm-anthropic 0.26 adds support for Claude 5 models and server-side tools
Trust84 - Aug 4, 2026 · Google AI — Blog
Google releases three new Gemini models and Gemini Robotics ER 2 for agentic workflows
Trust71