Claude Fable 5.2 guidance flags rule bloat; tester added 6 rules in 2 days

Anthropic's guidance is aimed at a common habit: after a model upgrade, old prompts and harnesses designed for earlier models can linger and quietly add cost or side effects. The new model came out on November 9, 2026, and the writer had been building instructions for Claude Code for four months — a CLAUDE.md, per-task procedure documents, and a memory the AI writes itself. So when the official page listed seven things to review, he applied each one to his own files.

What stood out was the timing. He had already added 6 promises and more than 3 watch mechanisms in those two days, often to patch a failure that had just occurred. A separate audit by the new model suggested the environment relied too much on words — rules, promises, memory — which are read long before the moment they are meant to matter, rather than on procedures or mechanisms that act at the point of use. The planned decision on 11月15日 is about what to cut and what to move into procedures or mechanisms.