AI Pulse · Under the Radar

Matt Shumer built a working first-person shooter that looks surprisingly close to a AAA title. He used Claude Opus 5, zero external assets, and a technique he calls “fan out sub-agents.” The model spun up parallel tasks, looped visual QA against reference footage from the real game, and iterated until the output matched. The whole thing runs in Three.js.

When people accused him of faking it, he open-sourced the exact prompt and repo. It hit 572 stars in 24 hours, the fastest first-day climb we tracked this harvest outside the mega-repos. Within a day, two other builders copied the pattern. One cloned a Trevor Noah site, another replicated the prompt structure for a different project. The technique is spreading.

An independent audit ran the game past 11 critics who scored it between 3.5 and 5 out of 10 against the original. That tempers the hype honestly. It is not production-ready, but it is far closer than anyone expected from a single prompt session. The gap between “proof of concept” and “looks like the real thing” just collapsed by months.

This matters because the loop is tightening. A year ago, you needed a team to prototype something visual and interactive. Six months ago, you needed a developer who could wrangle APIs and patch together libraries. Now you need a good prompt, a reference, and patience to let the model iterate. If your business touches prototyping, design validation, or customer demos, this is the kind of workflow we build into an AI command centre so you can test ideas at this speed without hiring a studio.

The technique is out in the open. The stars and forks tell you people are already running with it. If you have been waiting for AI to handle more than text, this week moved the line.