I Analyzed Dozens of AI Agent Rules Files. Most Are Making Your Agent Worse.

An ETH Zurich study found that LLM-generated rules files decrease AI coding agent success by 3% while increasing cost by 20%, and even human-written rules only marginally improve success (+4%) at similar cost overhead. The author analyzed thousands of GitHub rules files and categorized rules into four tiers—Essential, Helpful, Redundant, and Improve Codebase—concluding that only non-inferable and exploration-saving rules are worth keeping, while auto-detectable or enforceable-via-tooling rules waste context window tokens.