I maintain CLAUDE.md files across projects. Most of them made my agents worse before I figured out the pattern.
File structure, import patterns, framework conventions? This is all noise. The model already knows this from reading the code. Especially if an LLM wrote your context file.
Knowledge that the model can't infer? That's THE thing, like that one business rules that contradict defaults, or the decision that looks wrong but exists for a reason.
Researchers at ETH Zurich tested this across repositories. Context files decreased agent success rates and increased costs by 20%. Theo from t3.gg in his video put it best: it's like making a developer sit through every meeting before writing code.
The context files that work answer one question: what would surprise a senior engineer reading this codebase for the first time?
Everything else is noise you pay 20% more for.
Open your context file right now. If you can't point to one line the model couldn't figure out alone, delete the rest.