Claude Fable 5 Turns a 2022 Raccoon Heist Concept Into a Playable Game

The prompt gave the agent two old concept images, asked for a browser-based 3D game, required mobile controls, and told it to work independently without requesting further design decisions. It also asked the agent to commit frequently and maintain a notes file.

The resulting repository contains a playable Three.js game, generated textures and title art, a development log, and an exported Claude Code transcript. The repository states that Fable 5 built the project from a single prompt. Willison's write-up says the agent began with a working index page, added game systems in later commits, and used Playwright checks across desktop and mobile layouts. Visual assets were generated through OpenAI's gpt-image-2 API and committed as static files, so the deployed game does not call the image API during play.

This is a documented individual experiment, not a controlled benchmark. It shows that a coding agent can turn a compact product brief and reference images into a deployed interactive prototype while managing files, commits, tests, and supporting assets. The public transcript and repository make the workflow unusually easy to inspect.

The evidence does not establish that one-shot generation reliably produces production-ready games. Willison describes the finished game as mediocre while calling the single-prompt starting point impressive. That distinction matters: the experiment demonstrates breadth of execution and rapid prototyping, but it does not measure maintainability, accessibility, security, or repeatability across projects.

For developers evaluating coding agents, the useful artifact is the workflow trail. The prompt defines deployment constraints, the commit history exposes intermediate choices, the notes file records changes, and the transcript shows how the agent used tools. Those materials support a more grounded review than judging the final visual demo alone.