The failure modes of Claude Code in a guided 60h project

I guided Claude Code to build a fuzzer for Obsidian Sync without ever reading the code myself. I got a functional prototype in ~30h, saving an estimated ~10h.

Then I tried to finish it off, and the next ~30h turned into a bug treadmill, where at each turn Claude broke as much as it fixed.

This is the experience report: what went wrong, why research (and Anthropic's own docs) suggest this is to be expected, how an YC-alum AI-tooling founder's public talks tell a similar story, and maybe how to salvage the good bits.