The layering is the part I'd keep: CLAUDE.md / AGENTS.md for what must be true every session, skills for long how-tos that only load on a trigger, and hooks for the few lines the model must not get a vote on.
I'd also keep "usually follows" honest. Style and conventions can live in the always-loaded file; "never push to main" and "never print the production URL" belong somewhere the harness enforces, not somewhere the model is politely asked.
One small addition to the probe-test habit: when a rule fails the probe, move it down a layer before you rewrite the wording. A longer CLAUDE.md line that still only "usually" works is still the wrong place.
Which layer do you find people overstuff first — the always-loaded file, or skills that never get a clear trigger?