Topics›AI Drift Prevention and Review Culture›Agent Configuration Has a Half-Life: Evidence for Periodically Auditing CLAUDE.md/Skills
HighAct: HighFull content
Agent Configuration Has a Half-Life: Evidence for Periodically Auditing CLAUDE.md/Skills
AI Maker·August 27, 2026
New
Summary
Addy Osmani argues coding-agent configuration needs periodic auditing, citing Anthropic's finding that removing 80%+ of Claude Code's system prompt caused no measurable eval regression, a study finding 62% of repos had lint-related config leakage, and a paper showing generic (multi-developer) skills often outperform hyper-personalized ones. Recommends encoding must-hold rules as tests/hooks/permissions rather than prose, and running /doctor periodically.
Why this signal
Directly actionable, evidence-backed methodology for Neo's #5 focus area — pairs with the AI Maker 'audit Claude Code setup' piece the same week into a two-part playbook Thomas Engineering could run this quarter.
Original email
AI Maker · 8/27/2026
Audit your Agent files