Add one evidence-backed line to AGENTS.md ("When you change code, delete any comment that only records its history.") and the evals/ A/B harness that justified it: an undirected agent runs in a sandbox with/without a candidate sentence, scored by an LLM judge. On the reproduced case, that line takes Opus 4.8 from 0% to ~94%, while vague "write good comments" phrasings stay at baseline.