by Chris Hunter
Share

TL;DR
- My agents can’t send a message, move money or permanently delete anything without a one-time pass from me.
- If the check itself breaks, it blocks. It shipped with 34 tests.
- Friday (my chief of staff) moved this site to a new host in one evening and repaired the 561 links the move broke.
This week was about brakes. My agents got more room to work, and I got a hard stop on the few things I never want them doing alone.
A hard stop before my AI sends, spends or deletes
My agents used to stop and ask permission constantly. Run this command? Yes. Edit that file? Yes. When everything needs a yes, you stop reading what you’re saying yes to, and a prompt you click without reading protects nobody.
So I turned the constant prompts off and put one hard stop in their place.
A small check runs before every command an agent tries to run. It blocks three kinds of action outright:
- Sending a message to a person.
- Moving money.
- Permanently deleting something.
Those are the ones you can’t walk back. An email that went to the wrong person stays sent. A deleted file with no copy is gone.
If the check itself breaks, it blocks. It fails closed. A guard that waves everything through when it errors is worse than having no guard, because you think you’re covered and you’re not.
When it blocks something, I give it a one-time pass. Not a standing permission. One pass covers one action, and the next one gets stopped again.
Testing caught a nasty one before launch. The setting that was supposed to make the agent stop and ask me first did nothing in the mode I run my agents in. No error, no warning. The rule looked like it was on, and it was quietly ignored. So this hook doesn’t ask. It blocks, and I decide.
It shipped with 34 tests.
It’s narrow on purpose. It doesn’t judge whether a command is a good idea. It only knows whether the command is one of those three, and if it is, it waits for me.
Narrow has a cost. The hook can’t tell a harmless delete from a bad one. Clearing out a scratch file gets stopped exactly like wiping something that matters, so some of my passes go to things that were never risky. I’ll take that trade. A few extra taps from me beats one message I can’t unsend.
If you’re handing an agent access to your inbox or your books, these are the three doors I’d lock first.
An “are you sure?” prompt only works if you read it, and nobody reads the fortieth one of the day. A block makes me stop and think, every time, about the only actions that can’t be walked back.
I moved this website in one evening
You’re reading this on a new host. This week I moved my personal site to a new home, along with the tool I use to manage it and the other sites inside that tool.
The move itself is the easy part. The cleanup is what everybody dreads: broken images, links pointing at the old address, pages that load fine for you and 404 for everybody else. You usually find them one at a time, when somebody emails you about it.
Friday (my chief of staff) handled the cleanup. It went through the links that broke in the move, flipped DNS and kept a one-line rollback ready the whole time in case anything went sideways.
The numbers on the cleanup:
- 1,219 references to the temporary address rewritten, plus 172 that pointed at the bare domain.
- 312 upload links and 249 page links that the move had broken, repaired.
If you’ve ever paid somebody to move your site and then spent the next month finding broken pictures, you know the move was never the expensive part. The cleanup is.
One evening, start to finish. A job I’d been putting off got crossed off in a night.
Need help building something like this?
Everything I write about here runs in my own business first. If the AI job landed on your plate, start with an assessment: where AI will make or save you the most money, and what it'd take to run it right.
Oct 1–8 build log: a local AI video editor that finished three podcast episodes, overnight builds from my AI agent, and two quiet sites.
Build log, Sept 21–27: an AI that checked which leads were real (495 of 495 tag changes, 0 errors) and an early-warning brief that went dark for a week.
Claude Code now reads AGENTS.md, the cross-tool standard behind Codex and Cursor. Why one portable instruction file matters, and what I'm doing about it.
Sept 14–20 build log: my AI agent got permission to approve blog posts but not publish them, and a holiday broke my Monday brief.
