Skip to main content

exit 0 Means Pass, exit 2 Means Block — Taming AI with Claude Code Hooks

Published 1 min readMarkdown

CLAUDE.md guidance is strong, but long sessions can dilute adherence. Hooks enforce behavior at execution time.

Semantics

  • exit 0: allow
  • exit 2: block
  • stderr: feedback to Claude
  • stdout: inject context

Common production patterns

  1. PreToolUse dangerous command block (rm -rf, destructive SQL)
  2. SessionStart re-injection after compaction
  3. PreCompact async backups
  4. StatusLine token visibility
  5. PostToolUse auto-format/lint

Use CLAUDE.md for “what to do.” Use hooks for “what must never happen.”

One exit 2 is often more reliable than 100 lines of “please don’t.”



Originally published at Jidong Lab

1 view

No comments yet. Be the first to comment.

More from this blog