Why AI coding agents waste 80% of the context they load — and how I fixed it

I've been tracking how Claude Code navigates a large codebase and there's a pattern that bugs me. Let's say you ask it to refactor a function in a FastAPI project with ~800 files. Before writing a sin