Lesson 3 · 8 min

Hooks

Shell commands that fire on Claude Code events.

Hooks fire on events: PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, Notification, Stop. A hook is just a shell command; its stdout becomes context for the model and its exit code can block tool execution.

Production scenario

Real-world example: Auto-logging file edits to Slack

A startup wants on-call to see every file Claude Code touches in their critical-path repos. They add a PostToolUse hook that fires when the tool is Edit or Write:

#!/usr/bin/env bash
# .claude/hooks/log-edit.sh
TOOL=$(jq -r '.tool_name' <<<"$INPUT")
FILE=$(jq -r '.tool_input.file_path' <<<"$INPUT")
if [[ "$TOOL" == "Edit" || "$TOOL" == "Write" ]]; then
  curl -s -X POST "$SLACK_WEBHOOK" \
    -d "{\"text\": \"Claude edited \`$FILE\` in $(basename $PWD)\"}"
fi

The hook exits zero (doesn't block) and silently posts to #claude-activity. On-call has full visibility without slowing engineers down.

Why this matters: hooks are the right place for org-wide observability, policy, and notifications. They run for every developer automatically, with no per-user setup.

Knowledge points in this lesson

Hooks fire on harness events
PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, Stop
Hook = shell command; non-zero exit blocks tool
Hook stdout becomes model context
Use hooks for policy and telemetry
Don't bypass blocking hooks; fix root cause

Quick check

Claude CodeSelect one

Which file format defines a Claude Code slash command?