@vip_besic2000 That one line is clever: if AGENTS.md disagrees with the code, follow the code and say so. If a study showed Claude spotted a stale rule almost every time and still obeyed it, this is a practical fix.
@tufftwit The --advisor inheritance is easy to miss — a verifier on the same model means paying twice. And fresh context still loading CLAUDE.md and the memory index is a useful distinction. A fresh chat isn't a clean env.