Hooks: Automating Reactions To Tool Calls

A hook is a small script Claude Code runs at a defined point in its lifecycle — before a tool call, after a tool call, when a session starts, when it stops. The hook can read what the agent is about to do, transform it, log it, or block it. Hooks are how you turn project rules from optional suggestions into enforced behavior.

The lifecycle events

Three hooks worth building

1. 1PreToolUse blocking destructive bash: refuse `rm -rf`, `git push --force`, `db:reset` unless explicit confirmation
2. 2PostToolUse format-on-edit: run Prettier or Biome after every Edit so files always land formatted
3. 3UserPromptSubmit context injection: pre-pend the result of `git status` so every prompt has fresh repo state

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": { "tool": "Bash" },
        "command": "./scripts/block-destructive.sh"
      }
    ],
    "PostToolUse": [
      {
        "matcher": { "tool": "Edit" },
        "command": "pnpm prettier --write $CLAUDE_TOUCHED_FILE"
      }
    ]
  }
}

Where hooks beat CLAUDE.md

CLAUDE.md says 'don't push to main directly.' A hook BLOCKS pushes to main, period. Anywhere the cost of a mistake is high enough that 'usually respects' isn't good enough, a hook is the right answer. The model is no longer the only line of defense.

The big idea: prompts persuade, hooks enforce. Build hooks for the rules you actually need to hold.