Claude Opus 5 “one-shotted” a first-person shooter game with a prompt so simple, and results so impressive, that some readers struggled to believe it.

Others reran the prompt and said the results checked out.

Matt Shumer, the maker of the Claude-assisted game, calls the method a Gauntlet Loop: give an agent a real bar to beat, split the work among fresh critics, and never let the builder grade its own homework.

Just two days after Claude Opus 5 shipped, AI investor and former HyperWrite CEO Matt Shumer posted a video of a fully playable first-person shooter the model had built entirely on its own.

“Claude Opus 5 one-shotted this game,” he wrote, adding that not a single external asset made it into the build.

Now, you may think such a high-quality output required a long, detailed, and careful prompt to guide the AI model through the complexities of building a polished first-person shooter.

Think again.

Claude Opus 5 one-shotted this game. EVERYTHING you see in this demo is custom code... not a single external asset was used. AI games are going to be amazing. (sound on) pic.twitter.com/zc7C61kgv1 — Matt Shumer (@mattshumer_) July 25, 2026

Claude Opus 5 one-shotted this game.

EVERYTHING you see in this demo is custom code... not a single external asset was used.

AI games are going to be amazing. (sound on) pic.twitter.com/zc7C61kgv1

— Matt Shumer (@mattshumer_) July 25, 2026

The prompt behind the project ran three short paragraphs and was published in full on GitHub. It told Opus 5 to build a shooter at the level of the most recent Call of Duty games, to fan out subagents — workers that each get their own separate memory and a narrow job — to tackle pieces individually, and to keep looping on every piece with a separate, harsh critic until it held up against real Call of Duty footage in a blind side-by-side. The result, according to the prompt, should be “utterly perfect.”

That approach reverses much of the advice prompt engineers have given during the vibe-coding boom. The common guidance was to specify criteria instead of adjectives: say what “good” means rather than simply asking for it.

Shumer’s version does almost the opposite, asking its subagents to be “utterly wowed” and leaving the actual definition to a critic Opus 5 built for itself.

That was close to the entire brief. Shumer later wrote that he never specified the renderer, listed the game’s systems, or defined what “AAA quality” needed to include. He has started calling the approach a Gauntlet Loop: hand an agent a real, inspectable bar instead of a vague instruction, let it split the job into small pieces, and route each piece through a critic that never sees the builder’s own reasoning for its choices.

Two Claude Code features support that loop. Subagents spin up in isolated context windows with their own instructions and tool access, so a critic grading the weapon model does not inherit the builder’s excuses for why it looks the way it does. Ultracode is a Claude Code setting that pushes the model to its 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, designed for repeated fix-test-adjust cycles, is what kept the run from stopping the moment the game looked decent. Shumer did not specify a number of rounds. He let the critic keep naming new gaps and kept the builder chasing them for hours before he closed the session himself.

The finished build runs on Three.js and plain WebGL2, with roughly 55,000 lines of code spread across 11 subsystems. Every texture, mesh, animation, and sound is generated inside the browser at load time — no downloaded models, HDRIs, image files, or audio files. Shumer’s published critic log shows the score climbing from 3.59 out of 10 to just above 5, still trailing the real game in every round.

That level of transparency also made the result easier for others to test, which matters in a coding world where polished demos can hide a lot of hand-tuning. Skeptics assumed hours of hidden manual coding, so Shumer published the entire prompt and codebase. That is when copycats started.

Same trick, three different builders

James Altucher, the former hedge fund manager and podcaster, ran the identical prompt and said he spent “a little over ten hours” and roughly 1.3 million tokens on Opus 5 to get there. His build, Operation Blackout, plays free in the browser.

The developer of Prompt Silo pointed the same request at OpenAI’s rival flagship instead, posting “Sol 5.6 Ultra with same prompt” — Sol being the top tier of the three-model GPT-5.6 family OpenAI made generally available on July 9 alongside the cheaper Terra and Luna versions.

Sol 5.6 Ultra with same prompt. pic.twitter.com/1xwBQw5nRh — Rich · Atom Tan Studio (@atomtanstudio) July 26, 2026

Sol 5.6 Ultra with same prompt. pic.twitter.com/1xwBQw5nRh

— Rich · Atom Tan Studio (@atomtanstudio) July 26, 2026

Developer Leon Lin took the opposite approach and used the usual detailed prompt. Rather than copy Shumer’s short version, he set out to “reverse engineer a prompt for this game,” producing a document that runs some 20 sections deep and spells out everything from ragdoll physics to cascaded shadow maps. He fed that into Cursor using plain Opus 5 on high effort, with no subagents and no ultracode, and the result — a market-street shooter called Dust Corridor — also plays in-browser and looks strong.

None of the follow-up builds has faced the blind test Shumer ran on his own project. His critic log still shows real Call of Duty winning every round he logged — the bar Altucher and Atom Tan Studio are chasing with his exact three-paragraph prompt, and the one Leon Lin is chasing with roughly 20 sections of his own.

How much of this is actually new?

Agentic coding tools like Claude Code write software the way a supervised junior engineer might: they read files, run code, look at the screenshots they generate, and hand pieces of the job to subagents and critics that check the result against a stated goal. That loop is real, and Shumer’s Gauntlet Loop is a genuine way to structure it.

None of that, on its own, proves the model designed a game from imagination rather than recombining code patterns it had already absorbed.

Three.js ships its own pointer-lock camera controls as an official example, and that base pattern — mouse-look, WASD movement, and 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 almost certainly seen hundreds, if not thousands, of near-identical shooters before it ever read Shumer’s prompt.

That does not make Claude of Duty fake, but it does make “built from scratch” a harder claim to fully credit, so take those results with a grain of salt.

Researchers who study code-generating models have a name for the broader issue: data contamination, when a model does 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 that kind of contamination. Shumer’s own repo does contain Claude’s own creativity, if it is fair to call it that way. That is a reason to read “one-shotted a AAA game” as a capable agent working inside one of the most heavily documented genres in programming, not as proof an AI designed a shooter with no prior art to lean on.