Written but Unread: Do AI Agents Read CLAUDE.md?
In a Claude Code eval, AI agents told in CLAUDE.md to read a playbook opened it 0 times in 69 trials, while an Agent Skill loaded it every time. What went unread was the judgment: measuring how far apart two colors are, and sending brand changes to a person for sign-off. For human oversight of AI agents, the lesson is to check the transcripts, because a capable model's output looks finished either way.
