I guided Claude Code to build a fuzzer to find Obsidian Sync bugs. It runs Obsidian instances in containers and makes them edit notes while it cuts their network connection.
And it works: it finds amusingly short sequences of operations that trigger data loss, and that you can verify yourself by running them manually on your own phone and computer.
I got a functional prototype in 30h, but the next 30h turned into a painful bug treadmill, where at every turn Claude broke as much as it fixed.
So I wrote an experience report: what went wrong, why research (and Anthropic's own docs) suggests that this was to be expected, and how to salvage the good bits.
I posted the full experience report in Substack; Blogger is not keeping up with the world.