I shipped three things last week. A working prototype on Monday, a marketing dashboard on Wednesday, and a small internal tool on Friday that I honestly can’t believe I lived without. A few years ago, any one of those would have eaten a month. Now it’s an afternoon and more coffee than I’d like to admit.
If you build with AI, you know the feeling. It is genuinely remarkable, and I don’t mean that as a figure of speech. I mean I sometimes sit back and laugh at what just happened.
Here’s the part nobody puts on the box. The faster I can generate something, the faster it drifts out of sync with everything else I’ve already generated.
The reason is simple, and it doesn’t go away with a better model. AI doesn’t know what my product is supposed to feel like. It knows what I asked for five minutes ago. Those are very different things, and that difference turns out to be the whole game. Coherence isn’t something that lives inside any single screen or paragraph. It lives across all of them, and across time. No single prompt can hold that, so no single generation can protect it.
So I end up with more screens, more copy, more components, all produced quickly, and all just slightly off from one another. And here’s the trap: each piece, on its own, looks fine. Nothing throws an error. Nothing is broken. It’s just that the homepage sounds like one company and the onboarding sounds like its slightly more anxious cousin. Every one of those was a small, reasonable decision, made five minutes apart, none of them wrong on its own. That’s exactly what makes it sneaky. You don’t notice it in any one place. You feel it everywhere.
This is where you start to hear the phrase I hear constantly, and have said to myself: “I got it about 85% of the way, and then things started going in circles.” The same randomness that makes these tools feel so creative early on is what makes them so hard to pin down later. And if you’re building something you actually intend to sell, that last stretch, the precision and the finishing, is the part that matters most. It’s also the part AI is worst at, because finishing is mostly about coherence, and coherence is exactly the thing it can’t see.
Now here’s the uncomfortable bit, and the reason I think this gets worse before it gets better. I don’t believe the next model fixes this. For a while, I think it makes it worse.
A better model is better at giving me what I asked for. It is not better at knowing what I should have asked for. It generates more, more convincingly, and faster, which means the gap between what I can produce and what the thing is actually supposed to be gets wider, not narrower. We are becoming world-class at manufacturing parts. We have not gotten any better at making them fit.
I could be wrong. Maybe the tools eventually learn to hold the whole picture in their head the way a good designer or a founder does. But even then, somebody still has to decide what the whole picture is. That decision has never been a problem with generation. It’s a judgment problem, and judgment is the one thing we keep assuming will sort itself out.
Here’s what I’ve actually learned from hitting that 85% wall more times than I’d care to admit. The mess at the end is never one big mistake. It’s a thousand small ones. A button here, a word choice there, a data model that made perfect sense for the screen you were on and no others. AI will happily make every one of those decisions with you, five minutes at a time, and not one of them will look wrong in the moment. They just quietly stop adding up to the same thing.
What experience actually buys you isn’t the ability to clean that up later. It’s the ability to see those small decisions coming and make them in service of the whole, from the first prompt instead of the last. Once you’ve watched enough projects start going in circles, you recognize the choices that are going to calcify into a real problem three weeks out. So you make them differently, on purpose, while the cost of changing your mind is still basically zero.
That’s the part I care about, and it’s why I don’t think the answer is to use AI less. It’s to put that judgment in the room from the very beginning, pointed at the exact same tools everyone else is using, and get more coherent results out of them because someone is holding the whole the entire way through. Same tools, same speed, a very different thing at the end of it. (Could you wait until it’s a mess and bring someone in to untangle it? Sure. People do, and we’ll help when they ask. But that’s the expensive way to learn a lesson you can get for free up front.)
AI made it easy to build the parts. It didn’t make it any easier to care what you’re building. That part is still a person’s job. And honestly, I think that’s good news.