- If AI can act for you, build the stop button before anything else. Then test it until it's boring.
- Make your yes a separate step. Nothing goes out until you approve that exact thing.
- Let approval expire on its own. A yes from three hours ago shouldn't still be live.
- For anything you can't undo, make a copy first. No copy, no delete.
I built something that can text people as me.
The clever part — getting it to sound right, know my week, answer a question at 2am — took an afternoon.
The rest of the day went to the brakes. That's the part I'd actually tell someone else to copy.
Why I started with the off switch
The first thing I built wasn't a reply. It was a way to stop it.
One word, typed from my phone, and everything it's holding dies. Not paused. Dead.
Then I sat there and typed that word over and over until it was boring. Different times of day. Mid-job. While it was in the middle of writing something.
If the stop button only works when nothing is happening, it isn't a stop button.
If you're going to let AI talk for you, build the brakes before the engine.
Then I made my yes a separate thing
Here's where I think most people get it wrong. They hand the tool permission once, in general, and then hope.
I did it the other way around. It can't send an email, post anything, spend a dollar or publish a page. It can only ask.
It puts the actual thing in front of me — the real email, the real image, the real dollar amount — and waits. Nothing happens until I say yes to that exact thing.
And if the thing changes after I said yes, the yes stops working. That one took a while to get right. It's also the one that matters.
Three rules that came out of it
- The yes expires. Thirty minutes and it's gone on its own. I don't want something I approved this morning going out tonight.
- The yes works once. Approved, used, done. It can't be spent twice.
- Anything I can't undo gets copied first. If it's going to delete a post, a copy of that post and its image lands in a folder before I'm even shown the question. No copy, no delete.
Try this today
You don't need any of what I built to use the idea.
Next time you have AI write an email, a post or a message, don't let it send. Have it show you the finished thing and stop there.
Then read it out loud before you send it. Out loud, not in your head. You'll catch things your eyes skip right over.
That's the whole thing. The gap between “it wrote it” and “it sent it” is where you still live.
What I'd tell somebody starting this
The fun part of AI is what it can do. The part that lets you sleep is what it can't.
Build the second one first.
Questions people ask
Is it safe to let AI send emails or post for me?
It depends entirely on what sits between the AI and the send button. On its own, no. With a step where you see the finished thing and approve that exact version, it's fine, and you still get the speed. Start by letting it draft only. Add sending later, if ever.
What should I build first when setting up an AI assistant?
The stop button. Before anything clever, make sure one command halts everything, and test it over and over while the assistant is mid-task. If it only works when nothing is running, it doesn't work.
How do I stop AI from doing something I didn't ask for?
Approve specific things, not general permissions. A yes should cover one exact message at one moment, expire on its own, and only work once. And anything you can't undo should make a backup copy before it happens.
