Human-Curated CLAUDE.md Files Outperform LLM-Generated Ones by 4.5-6.5 Percentage Points — LLM-Generated Files Increase Costs 20%+ While Reducing Success Rate

ArXiv·high signal

The ArXiv study 'Dive into Claude Code: The Design Space of AI Agent Systems' (2604.14228) found that human-curated CLAUDE.md context files yield roughly a 4-percentage-point improvement in task success rates, while LLM-generated files actually reduced success rates by 0.5-2% and increased inference costs by over 20%. The key insight: rules should respond to observed failure, not be generated speculatively. Write CLAUDE.md entries only when you encounter a repeated agent mistake. The study also confirmed that the four extension mechanisms — MCP servers, plugins, skills, and hooks — operate at different points of the agent loop with different context costs.