Claude Clue #19: Stop Babysitting Claude with AutoMode
Let Claude decide what's safe and what isn't...!
If you've been using Claude Code, you know the drill. Click yes, click no, make a comment, click yes again. You're basically a human permission button. The different modes (Default, Plan, AcceptEdits) each have their place, but they all require varying degrees of hand-holding. Now there's AutoMode, which lets Claude decide whether it actually needs to ask you for permission. It's not quite "do whatever you want" territory, but it's a nice middle ground where Claude handles the routine stuff and only interrupts you when something genuinely warrants your attention. Under the hood, it uses a separate classifier model to make these decisions, which you could totally simulate yourself for certain workflows using a Haiku model, just like we do when building an AI Judge.
Last week, I was refactoring a module and switched to AutoMode. Instead of approving every file read and minor edit, Claude just did the work. It still paused when it wanted to run a more complex command combination, which felt right. Your mileage will vary based on your security posture, and to be clear, you should never run Claude Code directly on production. Keep it in dev, then test and deploy through your normal process. Anthropic has a solid writeup on all the permission modes if you want the full picture.
Co-authored with Claude, naturally.
Concept from Tim "Roscoe" Rawson