Claude Code: Gauntlet Loop tutorial and game review

The speaker explains Matt Schumer's Gauntlet Loop prompting method, then reviews a Burnout Paradise-inspired browser game built with Claude Code. The method gives an AI agent a clear goal and concrete references, splits the work into smaller tasks, and assigns separate agents to build and judge each result. The speaker recommends keeping the loop running until the user stops it.

For this experiment, a meta prompt produces a game-building prompt. Firecrawl searches for reference screenshots, audio resources and research. The speaker reports about 34 hours of work, 251 subagents and 412 million tokens, mostly cache reads. The demo includes drifting, boost, crash animations and controls for time of day, road dampness and car color. Oncoming traffic has no collision handling.

A progress page compares generated scenes with the original game's screenshots. The agents use grids to inspect individual areas and refine details such as reflections and lighting. The speaker says the first run performed poorly and required a second loop focused on performance.

The result remains a demo: it lacks races, time attacks, collectibles and a gameplay loop. The speaker also describes features breaking when agents add new ones. For cost control, he favors lower effort settings and suggests using Opus 5 to lead cheaper models, though this video does not present those experiments.