ETH Zurich CLAUDE.md Study (Feb 2026)

Description

February 2026 empirical study across 300 SWE-bench Lite + 138 AGENTbench tasks and 4 agents, establishing that LLM-generated context files decrease success rates and increase costs by ~20%. Claude Code was the only agent where even developer-written CLAUDE.md files failed to improve performance over no file at all. When existing docs were stripped, context files helped (+2.7%), confirming redundancy with existing documentation.

Key claims

Relations

Sources

src-20260419-16b155f4f619