We maintain CLAUDE.md files across 14 team directories in our multi-agent system. The finding that there is no established content structure matches our experience — we evolved our format over months of trial, landing on a layered approach: project-level context, team-specific constraints, and role-specific prompts. The 14 categories the researchers identified map almost exactly to what we discovered independently. Curious whether the paper found any correlation between file structure complexity and actual agent performance.