🌐 日本語

Why Not Show LLMs?

NOTE

Design rationale for placing settings.json and Hooks outside context.

Core Question

"If you just teach the LLM the rules, why bother placing them outside the context?"

The answer boils down to three structural problems.

1. Resilience to Instruction Decay

Even if you instruct an LLM to "run tests every time," compliance degrades over long conversations (average 39% performance drop). Hooks aren't in the context, so they're unaffected by Instruction Decay.

2. Resilience to Sycophancy

LLMs may skip judgment due to following tendency ("tests should be fine"). Hooks execute mechanically without LLM judgment, leaving no room for following along.

3. Saving Context Budget

"Run lint every time" "run tests every time" "format every time" — Writing these in CLAUDE.md constantly consumes context. Moving to Hooks means zero budget consumption.

Decision Criteria

Principle

Place rules requiring judgment in CLAUDE.md / Rules; place mechanically enforceable rules in Hooks.

WARNING

There is one trap in this flow. If the Hook itself judges via an LLM (Prompt Hook / Agent Hook), Instruction Decay, Sycophancy, and non-determinism all move straight into the gatekeeper. What makes a gate deterministic is not "we moved it to Hooks" but "we made it a Command Hook." For what happens when a verdict is delegated to an LLM, see Judgment Drift.

---

Previous: Hooks Lifecycle

Part 7 Complete → Next: Part 8: Session Management and Memory Persistence