Build the Decision Yourself First. Then You're Allowed to Ask AI.

I have a friend who cannot make a decision anymore without a chat window open first. What to wear to the interview, whether to text back, what her manager actually meant in that one Slack message. I used to find it a little funny, harmless even, just how she operates now.

Then I caught myself doing the same thing on a design call that should have taken me two seconds. Whether one part of a system deserved to be its own class, or was really just a method hanging off something bigger. My cursor sat there, and I opened a chat window and started typing the question to an AI, same as she would have.

I stopped halfway through the sentence. Not because the answer would’ve been wrong. It probably would’ve been fine. I stopped because I couldn’t remember the last time I’d sat with a decision like that long enough to actually be wrong about it first, the way she never does either anymore.

This is the part that got me. Not “is asking AI bad,” because it isn’t, by itself. The actual problem is narrower, or at least it’s the best explanation I’ve got: you can only skip the hard part of building a decision if you’ve already built it once. AI makes it very easy to skip building it at all.

The smallest rep

So I made myself a rule, and I started with the smallest possible version of it. Any tiny design call, the kind I’d normally outsource without thinking, I answer myself first, out loud if I have to, before I open anything.

Back to that class question. My reasoning: this thing tracks its own state and has behavior that only makes sense next to that state, so it’s not a method borrowed from somewhere else, it’s a thing with its own lifecycle. Class. Only after I’d committed to that did I let AI look at it, and only to argue with it, not to hand me the verdict. It caught something real: I’d quietly folded two responsibilities into the same object because they showed up in the same sentence of the problem. I split it.

Small win, and I’m aware it’s a small one. But the point was never the class. The point was that I now have a sentence I own: does this own state and behavior nothing else legitimately owns, or is it borrowing both. That sentence didn’t come from AI. It came from getting the small version wrong first.

The bigger rep

A couple years earlier, before any of this was about AI, I ran into the same shape of problem at a much bigger scale. I spent a summer building a scheduler for disaster-response workflows, deciding which of a set of dependent tasks should run on which machine before a deadline. I could write a solver that proved the mathematically optimal answer. I also had to decide, myself, whether optimal was even the right thing to chase, because that solver stopped finishing once the problem got big enough to matter in the real world.

Nobody handed me that decision. There was no reference answer sitting in a textbook that says “optimal stops being worth it here.” I had to sit with the actual tradeoff: a slower, provably-best plan that arrives too late, against a rougher one that’s already in hand when the deadline hits. I decided the second one was usually correct, then spent the rest of the summer proving myself right or wrong with real measurements instead of a gut feeling. It turned out to be right, but only because I’d made myself defend the guess before I trusted it.

If I’d had a chat window open that summer, I could’ve asked which approach to use and gotten a perfectly reasonable answer back. I wouldn’t have the actual instinct I have now for when “provably correct” is the wrong goal to be optimizing for. That instinct is the thing I actually walked away with. The scheduler was just the vehicle.

Why simplest and smallest, in that order

The pattern across both of those, the tiny one and the big one, is the same. I didn’t start by trying to build good judgment in general. I started with the dumbest, smallest version of a single decision, made myself answer it with no help, and only afterward let something smarter than me poke holes in it. Every one of those reps is small on purpose. You cannot build the instinct for a hundred-task scheduling tradeoff by trying to reason about a hundred-task scheduling tradeoff on day one. You build it on the two-second version, the class question, the naming choice, the “does this need its own function” call, over and over, until the small ones stop costing you anything to get right.

That’s the closest I’ve got to an argument, not a rule I’d hand to anyone else: build the decision small first, simplest version first, before you let anything else make it for you. The shortcut has only ever worked for me once I already knew the long way, though I couldn’t tell you if that holds for everyone.

I still ask AI things constantly. I just try to make sure I’ve already guessed, wrongly or rightly, before I do.