AI investor and former HyperWrite CEO Matt Shumer posted a video of a fully playable first-person shooter game built entirely by Claude Opus 5 two days after the model shipped, using a three-paragraph prompt published on GitHub. The game, which Shumer said 'one-shotted' without external assets, generated immediate skepticism due to its quality and the prompt's simplicity. Shumer published the complete prompt and codebase after critics assumed hidden manual coding, prompting multiple developers to replicate the results with the same method he calls a Gauntlet Loop.
Claude Opus 5 Builds First-Person Shooter With Three-Paragraph Prompt
Matt Shumer posted on July 25 that Claude Opus 5 built a first-person shooter game with a prompt running three short paragraphs. The prompt, published in full on GitHub, instructed Opus 5 to build a shooter at the level of recent Call of Duty games, to deploy subagents for individual tasks, and to loop each piece through a separate critic until it matched real Call of Duty footage in blind side-by-side comparisons. The finished build runs on Three.js and WebGL2, with roughly 55,000 lines of code across 11 subsystems. Every texture, mesh, animation, and sound generates inside the browser at load time with no downloaded models or external files.
Shumer never specified the renderer, listed game systems, or defined AAA quality requirements. The prompt told subagents to be "utterly wowed" and left actual definitions to a critic Opus 5 built for itself. Shumer's published critic log shows scores climbing from 3.59 out of 10 toward just above 5, with real Call of Duty winning every logged round.
Developers Replicate Results With Identical and Modified Prompts
James Altucher, former hedge fund manager and podcaster, ran the identical prompt and reported spending over ten hours and roughly 1.3 million tokens on Opus 5. His build, Operation Blackout, plays free in the browser. The developer of Prompt Silo pointed the same request at OpenAI's Sol 5.6 Ultra—the top tier of the three-model GPT-5.6 family OpenAI made generally available July 9. Developer Leon Lin reverse-engineered a detailed prompt running 20 sections deep, specifying elements from ragdoll physics to cascaded shadow maps, and fed it into Cursor using plain Opus 5 on high effort with no subagents and no ultracode. His result, Dust Corridor, also plays in-browser.
None of the follow-up builds faced the blind test Shumer ran on his project. Shumer's critic log shows real Call of Duty winning every round he logged.
Gauntlet Loop Method Reverses Traditional Prompt Engineering Approach
Shumer calls his approach a Gauntlet Loop: hand an agent a real, inspectable bar instead of vague instructions, let it split the job into small pieces, and route each piece through a critic that never sees the builder's reasoning. Two Claude Code features carry the loop. Subagents spin up in isolated context windows with their own instructions and tool access, so a critic grading a weapon model doesn't inherit the builder's excuses. Ultracode is a Claude Code setting that pushes the model to top reasoning effort and lets it write its own orchestration plan, fanning work across as many as 16 agents at once, capped at 1,000 per run.
Anthropic's built-in /loop skill, built for repeated fix-test-adjust cycles, kept the run from stopping when the game looked decent. Shumer never specified a number of rounds and let the critic keep naming new gaps for hours before he closed the session.
Researchers Raise Data Contamination Concerns
Three.js ships its own pointer-lock camera controls as an official example, and the base pattern—mouse-look, WASD movement, raycasting for gunfire—has been forked and tutorialized across GitHub, gists, and dev forums for more than a decade. A coding model trained on public repositories has likely seen hundreds if not thousands of near-identical shooters before reading Shumer's prompt. Researchers who study code-generating models call the issue data contamination, when a model performs well on a task mainly because near-identical examples already sat in its training data, not because it reasoned out something new.
None of the first-person shooter builds published a check for data contamination. Shumer's repo does contain code the model generated, but that makes "one-shotted a AAA game" a capable agent working inside one of the most heavily documented genres in programming rather than proof an AI designed a shooter with no prior art to lean on.
FAQ
What did Matt Shumer's Claude Opus 5 prompt actually contain?
The prompt ran three short paragraphs published on GitHub. It told Opus 5 to build a shooter at the level of recent Call of Duty games, to deploy subagents for individual tasks, and to loop each piece through a separate critic until it matched real Call of Duty footage in blind side-by-side comparisons. Shumer never specified the renderer, listed game systems, or defined AAA quality requirements.
How did other developers replicate the Claude Opus 5 game-building results?
James Altucher ran the identical prompt and spent over ten hours and roughly 1.3 million tokens on Opus 5 to build Operation Blackout. The Prompt Silo developer used OpenAI's Sol 5.6 Ultra with the same prompt. Developer Leon Lin reverse-engineered a detailed 20-section prompt and fed it into Cursor using plain Opus 5 with no subagents and no ultracode, producing Dust Corridor.
