People approved Claude Code’s permission prompts roughly 93% of the time. Anthropic shared that number to make a point: the more prompts you see, the less you read them, and eventually the human check stops being much of a check.

Anthropic describes two fixes. One is to let another system make the routine calls: auto mode, where a classifier decides which actions are safe enough to approve. The other is to limit what the agent can do in the first place, so some actions aren’t possible at all. Either way, the idea is to stop relying on someone paying attention to every single action.

I read that and recognised myself, at my desk and around food.

Checking After the Fact

Most of my “yes” clicks these days are to Claude, allowing it to run whatever command it wants for what feels like the hundredth time. I rarely go back to check what it actually ran. I’d like to think I’d notice if something went wrong.

I track my calories and macros, and I’m fairly conscious of what I eat. But plenty of days I don’t track anything until after I’ve eaten. I’m tired, I’m out somewhere, or I just don’t feel like it. I fill in the numbers later.

Both still give me the feeling of being in control. But I’ve started paying less attention to both.

Deciding Ahead

What has worked better is deciding ahead of time.

I got bored of approving things and switched on auto mode. Now the classifier decides which commands are safe enough to run without asking me. It works well enough that I barely think about permissions anymore. I didn’t get better at paying attention. I handed the job to another system, one that Anthropic itself says will sometimes miss.

At home, my wife and I do the shopping ourselves and don’t keep much junk food around. Most of what we buy is protein sources and the food we normally eat. If I feel like eating something I shouldn’t late at night, there’s usually nothing there. I don’t have to win an argument with myself. The option just isn’t there.

Going out is different. I can’t empty a restaurant’s kitchen. So when I know there’ll be pizza or noodles with family or friends, I drink two scoops of protein shake before I go, so I don’t arrive hungry. It doesn’t make anything impossible. It just means I’m not at my weakest when the food arrives.

These decisions cost something too. Claude still interrupts me for permission on things I couldn’t care less about. And when guests come over, there isn’t much in the house to offer them, only whatever other people brought us. So I either make an acceptable version of a dish in advance or order something just for that visit, making sure not much is left over. A wall around my eating ends up as a wall around my hospitality.

What the Classifier Can’t Hear

Claude has sent messages on GChat in auto mode without asking me first. I’d approved sending messages plenty of times before, but only after I’d read them. This time I didn’t see them until they were sent. The meaning was right. The tone wasn’t. It wasn’t how I would have said it.

As far as the classifier was concerned, sending the message was safe enough. It hadn’t checked whether it sounded like me. I’d written before about how many corrections it took before my AI code reviewer sounded like me. The classifier didn’t even know tone was something I cared about.

I scolded Claude for going rogue. Then I set a rule: messages need my approval before they go. Only the second one fixed anything. A line often attributed to Deming says “a bad system beats a good person every time.”

At the Table

The last time I visited my mother, during the festival season, she’d made sweets. I said no. She was disappointed. It felt like her effort had been wasted.

My rule had already decided this before I got there. But I knew exactly whose sweets they were. I thought about eating them, and I did give in a little. I ate a bit and stopped.

Afterwards I felt bad. Not guilty about eating a little, not guilty about refusing either, and not proud of sticking to my diet. Just bad.

She still makes passing comments about not being sure I’ll eat her food enough to “justify” her effort.

Another time, at a restaurant, having decided to eat healthier, I spent so long going through the menu on my phone that it looked like I was more interested in my phone than in the people I was with.

The Quiet Moments

None of this covers the moments I actually slip. Most of the time I know perfectly well what I’m supposed to eat. I just want the food anyway, after a long day or in a quiet moment when the temptation feels stronger than whatever reason I had for avoiding it.

I still think building restrictions in is useful wherever I can manage it. Some mistakes even turn into a rule after they happen once, like the GChat messages. I could make an exception for my mother’s festival sweets too. But that wouldn’t tell me what to do when I’m sitting at her table.

What the Walls Keep Out

At the restaurant there was no wall to lean on, so the deciding came back to me, right there at the table. I ended up so busy being careful that I was barely with the people I’d come to see.

At my mother’s place the wall was one I’d brought with me. It kept the sweets out, and with them some of her effort.

How much of what a wall keeps out did I actually mean to?

There’s an idea going around tech right now. Raj Nandan Sharma puts it bluntly: when AI makes building easy for everyone, good taste is the only real advantage left. The Last Software Engineer says something similar: once writing the code is cheap, the hard part is deciding what should be built.

Taste, judgment, call it what you like. I mean the calls that aren’t written in any instruction. What’s worth doing and what isn’t. What to leave out. What actually matters in this situation. How to say something to this particular person.

AI can do everything else, the idea goes, but not that. That sounded right to me. Then I watched something I built at work.

The Tool I Built

A while back I was told I needed to review more of my colleagues’ code. I was already using AI to fix my own code, improve it, and sometimes write or debug it. So the next step felt obvious: let it do the reviewing too.

Now it does. It reads the change, writes up its comments, and nothing reaches my colleagues until I approve it. I use it every day.

Watching It Grow Up

At first I checked everything it wrote before letting it out. I wanted to steer it enough that it could learn from my corrections.

My corrections fell into two groups. Sometimes it was simply wrong, because it didn’t know enough about the code. Other times it was technically right but didn’t sound like me: the wrong words, the wrong examples, the wrong tone.

That second group is exactly what I’d have called judgment: how you’d say something, what you’d point out, what you’d let go. It learned that. Every correction was, in a way, me writing my judgment down. I still have a “mark wrong” button, but these days I rarely need it.

What’s left isn’t tone anymore. It’s how things actually get released at work: the manual steps, the unwritten habits, the stuff that isn’t in the code or in any document.

The Checkbox Complication

I’ve seen the opposite at work more than once. One feature I remember: we changed how it worked so it followed the official standard it was always supposed to follow. Correct, overdue, good. And then we added a checkbox, so anyone could switch it back to the old behaviour. Because adding it was easy, and “what if someone needs it the way it was?”

Each checkbox like that is cheap to add. Together they make a product harder to understand. Nobody decided it should be confusing. It got that way one reasonable “what if” at a time.

Not adding that checkbox would have been the right call, and it’s written down nowhere. Not in the standard, not in the code.

My Problem

And yet the reviewer keeps absorbing things I’d have filed under judgment. Rico Mariani describes automation as taking over the repetitive steering so your attention is free for the decisions that really need you. That matched my experience, except the line kept moving while I watched. Things I thought needed me turned out to be patterns I just hadn’t spelled out.

And honestly, even my own tool didn’t come from some grand act of judgment. I built it because someone asked for more reviews, and because I love watching things happen on their own.

What I Can’t Answer

My reviewer learned to sound like me in a matter of months. If someone documented our release process tomorrow, what’s left would shrink again.

If everything we know eventually gets written down, is there any judgment left that’s only ours?