AI Pulse · Under the Radar

A developer named Matt Shumer fed Claude Opus 5 a single prompt and got back a working first-person shooter. The game, called Claude-of-Duty, runs in a browser, includes 55,000 lines of code, and generates every texture, mesh, and animation on the fly. No art assets. No follow-up prompts. One instruction, one output.

The repo picked up 2,381 stars in five days, one of the fastest-growing projects we tracked this week. Within 48 hours, a second builder posted a similar one-shot WebGPU demo. The pattern is repeating, which means we’re watching a new category form in real time.

Why this matters if you run a company

This isn’t about games. It’s about what happens when an AI can hold 55,000 lines of logic in working memory, coordinate procedural generation across multiple subsystems, and ship something that actually runs. The same capability applies to internal tools, data pipelines, customer-facing dashboards, anything with enough structure to describe in a prompt.

The catch is knowing what to ask for. The prompt file is public, and it’s detailed. Shumer didn’t just say “build me a shooter.” He specified render loops, collision logic, weapon behaviour, UI layout. The model did the work, but the instruction set was tight. That’s the skill gap right now. Most teams don’t know how to write that kind of spec, and most vendors won’t teach you because they’re selling managed services, not capability transfer.

If you’re building something like an AI command centre that routes tasks to the right model and logs what works, you start collecting those specs. You learn what Claude handles in one shot, what needs iteration, and where humans still close the loop. That’s the infrastructure play here, not the game itself.