Claude Code Permission Modes: When to Run Hands-Off with Auto Mode — Transcript

Full transcript

1. 0:02Permission mode lets you decide once what's safe to run without you.
2. 0:06You already cycle through several of them with shift-tab, manual, accept edits, and plan.
3. 0:10The rest are where hands-off Claude code lives and the one to lead with is auto.
4. 0:15Claude runs without prompts while a separate classifier reviews each action before it executes, stepping in on ones it identifies as dangerous.
5. 0:23So here are the permission modes.
6. 0:25First, we have manual, which reads only without prompting.
7. 0:28Everything else asks.
8. 0:29Accept edits, reads, file edits, and common file system bash commands.
9. 0:33This is for iterating on code that you review after the fact.
10. 0:36Plan reads only.
11. 0:37It researches and proposes without editing.
12. 0:39Auto mode accepts everything with a classifier model reviewing each action before it runs.
13. 0:44Don't ask.
14. 0:45Only pre-approved tools allowed.
15. 0:47Everything else auto-denied with no prompt.
16. 0:49This is great for your CI pipelines.
17. 0:51And then bypass permissions, skips all check.
18. 0:54This is equivalent to the dangerously skipped permissions.
19. 0:57Only run it inside of an isolated container or virtual machine.
20. 1:00Press shift tab to cycle through manual, accept edits, plan, and auto.
21. 1:04The status bar shows which one that you're in.
22. 1:07Auto mode is where Claude runs, but before each action executes, a separate classifier model reviews it.
23. 1:12The classifier guards intent.
24. 1:14It blocks the moves that escalate beyond your request, production deploys, and migrations, force pushing or piping download code into a shell, sending sensitive data to external endpoints, and destroying files that exist for the session.
25. 1:27It allows the everyday work, local edits in your project, installing dependencies from your log file, read-only requests, and pushing to your own branch.
26. 1:35Now, the classifier won't catch whether the code works or not.
27. 1:38So if you ask Claude to refactor authentication and it writes broken authentication, the classifier waves it through because broken isn't dangerous.
28. 1:46Use auto mode together with a stop hook that runs your tests.
29. 1:49Auto mode watches what Claude is trying to do while the hook confirms that the code actually runs.
30. 1:53Auto mode guardrails are still evolving, so check the docs for the current block and allow lists.
31. 1:58Don't ask is the right move whenever no human is there to approve prompts.
32. 2:02CI, scheduled jobs, overnight batches.
33. 2:05Anything off that list gets auto-denied.
34. 2:07So your pipeline keeps moving instead of hanging on an approval no one will give.
35. 2:15There are several permission modes and you can reach the everyday ones by cycling shift tab.
36. 2:19Auto is the hands-off default.
37. 2:21The classifier checks intent before each action runs and a stop hook checks correctness after.
38. 2:27Don't ask covers unattended pipelines and bypass permission belong only inside isolated containers and VMs.
39. 2:33Match the mode to the job.