Knowing when to stop

I was finishing a slide deck, and I went back to the AI for another pass. Then another. The deck already worked. I wasn’t fixing anything, I just wanted to see if one more try would surface a sharper version. None of that is new. What caught me was that I didn’t stop. I used to stop around here, because another pass still took real effort. This time the effort was gone, but the cost of not stopping wasn’t. It had just moved somewhere I couldn’t see.

I iterate like this constantly, and rarely to fix a mistake. I’m looking for the better idea I haven’t found yet, and often another pass turns up something I wouldn’t have reached on my own. It can look like trying too hard. Maybe it is. But it’s how I’ve worked for a long time, well before AI, and more often than not it gets me somewhere better.

Each pass feels like progress, and the next one might be better. The refining is satisfying on its own, so the pull doesn’t fade when the gains do. I’ll keep going after the better idea has stopped showing up, fishing in a spot that’s run out of fish.

The brake used to be the cost

Iterating used to cost something. Another pass at a paragraph meant writing it. Reworking a design meant another hour. If I was leaning on someone else’s time, there was a limit to how many times I could say “one more tweak” before I felt it.

That cost was a brake. I’d reach a point where going again wasn’t worth the effort, and I’d stop. The cost made that call for me. I never had to decide the work was finished, so I never learned how. The friction did it.

Now there’s no brake

Making another pass isn’t my work anymore, I just ask for one.

Not for everything, but on the harder calls I’ll run two of them, Claude Code and Codex. I treat them as two separate reviewers, each digging into the question on its own. I’m not looking for them to agree, the useful part is where they don’t: one flags a claim the other smoothed over, something I’d have read straight past on my own. They argue it out through me, and I get a better answer than either gave me alone. It doubles the loops. I even tell them to push back if they think it’s already fine, and sometimes they do. I can still ask again, and I usually do.

And I don’t sit and wait. The tools are still slow, so a pass takes a few minutes, but while one’s running I’m in another tab doing something real, three or four of these going at once.

So the brake never comes back. The cost used to end the session for me. Now nothing does, unless I decide to.

Where the cost actually went

For a while I told myself this was fine. I’m not wasting time, I’m parallelizing. And that’s true, as far as the clock goes.

It’s not true about attention. The time I’d have spent waiting, I spend on something else. But the attention each loop pulls from me doesn’t parallelize. It comes off the top of one finite pile, and that’s the thing I can’t get back. Every loop I spend tuning a deck that already works is attention I don’t have for the harder calls around it: deciding what really matters, noticing what still feels thin, or walking away before my judgment gets worse.

That’s the part that gets lost when people say iteration is cheap now. Making it cheap didn’t make it costless. It just moved the cost from a place I notice, the clock, to a place I don’t, my attention and whatever it was supposed to be on.

Does this thing deserve another pass? When iterating was expensive, I didn’t have to ask, the cost answered for me. Now that it takes no work from me, I have to ask on purpose. And I’m not in the habit, because I never needed to be.

Some things have a wall

The way I’m learning to tell the difference is to ask what I’m pushing against, and whether it pushes back.

Some things have a wall. A script either works or it doesn’t. A deadline arrives. Someone signs off, or sends it back. The effort has somewhere to go, and it comes back as a better result. You can tell when you’re done, because the thing tells you.

Other things have no wall. A single word with three options that all work. Spacing nobody will notice. A slide that already says what it needs to. You push, and nothing pushes back. The effort is real, there’s just nowhere for it to land, so it cycles into more passes and more heat, never a clearer answer.

AI can’t tell whether another pass has somewhere to land. It will refine the thing that’s already fine just as readily as the thing that still needs work, for as long as I keep asking. That difference is still mine to notice: whether I’m pushing against a real wall, or just pushing into open air.

What I’m actually trying to change

I’m not going to stop chasing the better version. It isn’t a flaw I’m planning to fix. It’s curiosity, and it’s part of how I do work I care about. I’d rather aim it than dull it.

The habit I’m after is smaller and harder than “polish less.” It’s pausing before the next loop to ask whether the thing in front of me has a wall, and whether the attention it’s about to eat is attention I owe something else.

So I stopped. The deck is fine, and there’s a talk to rehearse.