I keep reading this, but with Claude Code in particular, I consistently find it gets smarter the longer my conversations go on, peaking right at the point where it auto-compacts and everything goes to crap.

This isn't always true--some conversations go poorly and it's better to reset and start over--but it usually is.

I've found there usually is some key context that is missing. Maybe it is project structure or a sampling of some key patterns from different parts of the codebase, or key data models. Getting those into CLAUDE.md reduces the need to keep building up (as large) context.

As an example for one project, I realized things were getting better after it started writing integration tests. I wasn't sure if that was the act of writing the test forced it to reason about the they black box way the system would be used, or if there was another factor. Turns out it was just example usage. Extracting out the usage patterns into both the README and CLAUDE.md was itself a simple request, then I got similar performance on new tasks.

This is my exact experience as well. I wonder if I should switch to using Sonnet so that I can have more time before auto-compact gets forced on me.