A case has been revealed in which Anthropic’s Claude Opus 5 implemented a playable first-person shooter game with only a short, three-paragraph prompt, widening debate over how to design prompts for artificial intelligence coding.

On July 28 local time, blockchain outlet Decrypt reported that AI investor and former Hyperlight CEO Matt Shumer 공개released a video showing the model building the game on its own, without external assets, just 2 days after Claude Opus 5 was launched.

The demo Shumer released is a browser-based FPS. “Claude Opus 5 one-shotted this game,” he said, adding that everything shown in the demo is custom code.

The prompt posted on GitHub consists of three short paragraphs. It instructs the model to make a shooter game at a “Call of Duty” level, split the work across multiple sub-agents, and have a separate harsh critique agent blind-compare each output against actual gameplay footage. It also includes a condition that the final output must be perfect.

The key was to provide a clear comparison target and a repeated evaluation structure instead of listing detailed functions at length. Shumer called the approach a “gauntlet loop.” It gives agents verifiable criteria rather than vague instructions, breaks the work into small pieces, and prevents the creator of the result from evaluating it. He said he did not specify a renderer, and did not separately include a game system list or a detailed definition of AAA-level quality.

The process used Claude Code features. Sub-agents worked with separate contexts and tool access, and an “ultracode” setting let the model set its own execution plan and distribute work across up to 16 agents. Anthropic’s built-in “/loop” function made it repeat revisions, tests and adjustments. Shumer did not set a number of iterations, and continued improving the output each time the critique agent pointed out new issues.

The finished game runs on a Three.js and WebGL2 base. The code totals about 55,000 lines and is divided into 11 subsystems. Textures, meshes, animations and sound were all generated when the browser loaded. It did not use downloaded models, images or audio files. However, the critique log Shumer released shows the score rose from 3.59 out of 10 to just over the 5-point range, but “Call of Duty” remained ahead in every recorded round.

There were also cases later in which the same prompt was run again. Former hedge fund manager and podcaster James Altucher said he used the same prompt to make “Operation Blackout” with a little over 10 hours and about 1.3 million tokens. A Prompt Silo developer released results from applying the same request to OpenAI’s top model, “Sol 5.6 Ultra.” By contrast, developer Leon Lin did not use Shumer’s short prompt and instead created a detailed document of about 20 sections, including ragdoll physics and shadow maps, and tested it in Cursor. The result, “Dust Corrido,” also ran in a browser.

However, the follow-on outputs did not go through the blind comparison test Shumer ran. As a result, how consistently the same three-paragraph prompt works across other models or workflows has only been confirmed in a limited way so far.

Limitations have also been cited separately from the technical achievement. The FPS genre is an area where many examples have already accumulated in open code, tutorials and developer forums. Three.js itself has an example of pointer-lock camera control, and patterns for mouse-look, WASD movement and ray-based gunfire handling have long been widely shared. In such an environment, it is difficult to rule out the possibility that a code-generation model trained on public repositories recombined a similar structure.

Researchers call this problem training data contamination. It refers to cases where nearly identical examples are already in the training data and reconstruction, rather than new reasoning, determines performance. No inspection procedure to verify this was disclosed in this FPS case. As a result, the outcome needs to be interpreted not as evidence that AI designed a shooter game from scratch without any prior examples, but as a case showing how strongly agent-style coding tools can operate in a well-documented genre.

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