Hooks — Runtime layer

🌐 日本語

NOTE

Hooks are scripts that fire at defined lifecycle events (PreToolUse, PostToolUse, UserPromptSubmit, etc.). They run outside the LLM's context — deterministic checks and side effects that the model never sees.

How it works

• Defined in settings.json under hooks.<event>.
• The LLM cannot disable, bypass, or even read its own hooks — that's the point.
• Best used for guardrails (lint, secret scan, format) that you want to be unskippable.

Related chapters

• Part 7 — Runtime Layer (overview)
• settings.json
• Hooks Lifecycle
• Why Not in Context
• Configuration Reference

Structural problems it addresses

• Instruction Decay — enforcement happens regardless of memory drift.
• Sycophancy — the LLM can't talk its way past a hook.
• Prompt Sensitivity — phrasing doesn't change what the script does.

See also

• Topic: CLAUDE.md — for instructions the LLM is expected to follow voluntarily.
• Topic: Plugins — for distributing hooks alongside skills and MCPs.
• Feature Index