AI rules files cut from 16,032 bytes to 7,796B

The article's starting point was accumulation: the author kept adding observations to the file the AI reads at startup until Claude's auto-loaded documents, together with shared rules, reached 16KB. Codex was even instructed to read two files first, so it read the same content twice.

The counterweight is research. A February 2026 ETH Zürich study (arXiv:2602.11988) measured rules files directly, finding that LLM-generated ones lowered success rates by about 2–3% and raised inference cost by over 20%, while human-written ones raised success slightly but still added up to 19% in cost. Overviews and directory listings were found useless, and the study noted that written instructions are followed so faithfully that unnecessary requests also get obeyed — producing detours rather than speed. OpenAI's official guidance, cited in the article, likewise favors a short, accurate AGENTS.md over a long, vague one.

The fix moved reasoning out of the always-read file: conclusions stayed, the 'why' went to a separate reference file opened only when the rules are being changed, and project-specific facts are now emitted at session start by a hook (SessionStart) showing the current project, first file to read, check commands, and prohibitions. Read-only-when-needed documents are listed with the moment to open them, such as before driving a browser or writing a new script. The author also suggests a small script that reports per-file byte counts and rough token estimates so a rules file that starts growing again is noticed quickly.