The wins came from three places: 1. Pre-computing upfront in a clever way that doesn't block the main thread
2. Avoiding NAPI overhead for small result sets
3. Iteratively asking Claude to find perf improvements for faster file search

→ View original post on X — @bcherny,