Descriptions:
Nate Herk pushes Claude Code’s browser automation capabilities into unexpected territory: teaching it to play custom-built browser games through a self-correcting feedback loop. Rather than using a traditional automation script, Claude Code takes a screenshot of the game state, reasons about what went wrong, rewrites its own bot script, and retries — improving performance with each iteration.
The video walks through three progressively harder games. A simple block-runner is solved on the sixth attempt after the bot converges on a pixel-distance-based jump timing strategy. For Tetris, Claude Code first scores over 16,000 points in a single session, then is challenged to run two simultaneous instances and push past 25,000, then 80,000 — eventually reaching scores that eclipse the noted real-world Tetris world record of 40 million points by running in an uncapped hard-drop loop.
Beyond the entertainment value, the video illustrates a practical pattern: Claude Code’s browser automation isn’t limited to form-filling or scraping Best Buy results. The same agentic loop — read codebase, write automation script, observe result, self-debug — applies to any browser-based task without a formal API. Viewers interested in Claude Code’s computer-use capabilities will find this a concrete, reproducible demonstration of its iterative reasoning in action.
📺 Source: Nate Herk | AI Automation · Published March 11, 2026
🏷️ Format: Showcase
