Summary
Nate Herk runs an unusual experiment: he gives Claude Code a deliberately ugly coffee shop website and tells it to “make it better” 100 times in a row, once with Sonnet 5.5 and once with Opus 5.5, both at medium effort.
The prompt required the model to inspect the site, think outside the box, and ship one bold change per round, since earlier attempts stalled after 20 to 25 rounds when the model decided nothing more could be done. Each run started in a clean room with no skills, CLAUDE.md, plugins, or memory. Herk walks through milestone rounds, such as the first generated image, first animation, and first AI-generated video, and tracks how the code grew in size.
The cost difference is significant: the Sonnet run came to about $29.93 in API billing with 11 images and two videos, while the Opus run cost around $137 with 54 images and three videos. Herk observes that both models eventually start overengineering, removing features and changing fonts just to satisfy the ship-every-round rule, and argues that human taste is still needed to push results from good to great.
📺 Source: Nate Herk | AI Automation · Published October 30, 2026
🏷️ Format: Comparison
