Shumer's Three-Paragraph Prayer Convinces Opus 5 to One-Shot a Call-of-Duty Clone on Vibes Alone 🤖

Two days after Anthropic shipped Claude Opus 5, AI investor and former Hyperwrite CEO Matt Shumer posted a video on July 25, 2026 showing a fully playable first-person shooter the model had built entirely from scratch. "Claude Opus 5 one-shotted this game," he wrote on X, adding that every asset in the demo was custom-generated and "not a single external asset was used." The build runs on Three.js and plain WebGL2 across roughly 55,000 lines of code organized into 11 subsystems, with every texture, mesh, animation, and sound generated by the model.

The prompt behind the build, which Shumer published in full on GitHub, ran only three short paragraphs. It instructed Opus 5 to build a shooter at the level of the most recent Call of Duty games, to fan out subagents—isolated workers with their own context windows and narrow tasks—to handle pieces individually, and to route each piece through a separate harsh critic until it held up against real Call of Duty footage in a blind side-by-side. The target output, per the prompt, should be "utterly perfect," and subagents were asked to be "utterly wowed," with the actual definition left to a critic Opus 5 constructed for itself.

Shumer later clarified that he never specified the renderer, listed the game's systems, or defined what "AAA quality" needed to include, and has begun calling the approach a "Gauntlet Loop": hand an agent a real inspectable bar rather than a vague instruction, split the job into small pieces, and route each piece through a critic that never sees the builder's own reasoning for its choices. The technique leans on two Claude Code features—subagents, which spin up in isolated context windows with their own instructions and tool access, and Ultracode, a setting that pushes the model to its top reasoning effort and lets it write its own orchestration plan across as many as 16 agents at once, capped at 1,000 per run—alongside Anthropic's built-in /loop skill for repeated fix-test-adjust cycles. Shumer set no round limit, letting the critic keep naming new gaps until he closed the session himself.