Letting an AI loose on your whole repo is fine — if you keep a few guardrails. Here’s the safe setup.
1. Work on a branch
Never let it run on main:
git checkout -b ai/feature-x
If anything goes sideways, you throw the branch away.
2. Plan before big changes
Use plan mode (Shift+Tab) so Claude shows its approach
before editing across files. Approve, then let it run.
3. Keep changes reviewable
- Ask for small, focused commits, not one giant change.
- Review diffs before approving edits — don’t blind-accept.
- Run your tests after each chunk:
npm test
4. Give it the map
A good CLAUDE.md (commands, conventions, “don’t touch” folders) keeps it from wandering into the wrong files.
5. Use permission modes deliberately
Default asks before each action. Auto-accept is faster but only use it on a throwaway branch where you’ll review the final diff anyway.
The pattern: branch → plan → small reviewed steps → tests. That’s how you move fast
without breaking main.