Simple AI prompt produces playable FPS in browser

Developer Matt Shumer posted a demo showing Claude Opus 5 generate a playable first-person shooter two days after the model’s release. The model produced textures, meshes, animations and audio inside the browser without any external asset files, according to Shumer’s repository and logs.

The finished build runs on Three.js and WebGL2 and contains about 55,000 lines of code across 11 subsystems in Shumer’s repository. All visual and audio assets are created at load time inside the browser; the project uses no downloaded models, HDR images, image files or audio files.

Shumer published the three-paragraph prompt and a critic log. The prompt instructed Opus 5 to build a shooter comparable to recent Call of Duty titles, to split work among subagents with separate memory, and to iterate pieces against a critic until they could pass blind side-by-side comparisons with real game footage. Shumer wrote that the model “one-shotted this game” and described the method as a “Gauntlet Loop.” The critic log shows scores rising from about 3.59 toward just above 5 out of 10 across many rounds; the logged comparisons show the real game winning each recorded test.

The run used Claude Code features that spin up subagents in isolated contexts, an Ultracode setting to increase the model’s reasoning effort, and Anthropic’s /loop skill to repeat fix-test-adjust cycles. Shumer allowed repeated cycles until the critic stopped naming new gaps and he ended the session.

Other developers reproduced or adapted the approach. One used the identical three-paragraph prompt on Opus 5 and reported roughly ten hours of work and about 1.3 million tokens to produce a browser-playable build called Operation Blackout. Another applied the short prompt to an alternative large model, and a third developer used a longer, 20-section prompt with Cursor and Opus 5 on high effort, producing a different browser-playable shooter called Dust Corridor.

Researchers and developers note that agentic coding setups route tasks to subagents and critics that read files, run code and examine outputs. The field also recognizes data contamination as a possible factor: common shooter patterns and Three.js examples are widely available in public repositories and tutorials, which models may have seen during training. Shumer and the follow-up builders published prompts, code and critic logs but did not publish a formal check for training-data overlap.

Shumer’s materials include the full prompt, the game code and the critic log tracking iterative improvements. Other builders have made their versions available to play in the browser, and reported runtimes and token usage vary depending on model choices and orchestration settings.