There is a specific look people get when their AI stops working for them. I’ve seen it enough times now that I can spot it before they finish the sentence. They lean in, lower their voice a little, and say something like “I don’t know, man, the model just feels dumber lately…“.

Let me tell you what actually happened.

The honeymoon

For the first stretch it works. You open the tool in the morning, sketch out what you want, and the code lands before lunch. The roadmap holds. You demo it to someone who matters and they get that look on their face. You start telling people this is the future and you half believe you’ve been let in on a secret early.

I felt this myself. I’ve watched entire teams feel it. It is real and it is intoxicating and it is not a lie. The tools genuinely are that good in that window.

The problem is the window closes and nobody warns you.

The plateau

Then one week the same prompts stop landing. The thing you asked for comes back subtly wrong. You correct it, it drifts somewhere else, and a task that used to take twenty minutes eats your whole afternoon. You don’t say anything out loud, because saying it out loud makes it real…

So you do the thing everyone does. You blame the tool. Maybe the newer model will fix it. Maybe the other agent. Maybe the framework everyone is posting about this month. And here is the cruel part: each new thing works. For about eleven days. Long enough to convince you the last one really was the problem, short enough that you are back here again by the end of the month, standing in the same spot with a different logo open on your screen.

I’ve done this. I’ve watched founders burn a quarter doing this. The tool graveyard gets deep.

The part nobody wants to hear

It was never the tools.

The model didn’t get dumber. You started reaching for it differently. In the honeymoon you were working on a clean, small, well-understood problem, and you were giving it clear direction because you still had the whole thing in your head. As the codebase grew and the context got messier, you kept feeding it the same short prompts and expecting the same magic, except now the magic has to guess at everything you left out.

The tool didn’t change. The gap between what you’re asking and what you’re actually specifying widened, and the tool filled that gap with confident nonsense. That is not a model problem. That is a thinking problem.

The people who get past the plateau are not the ones who found the right tool. They are the ones who realised the tool was only ever as good as the thinking they handed it, and they started spending their effort on the thinking instead of the shopping.

What to do instead

Stop swapping. The next tool will also work for eleven days.

When something stops landing, resist the reflex to blame the model and ask the more uncomfortable question: what do I actually know about this problem that I never wrote down? Nine times out of ten the answer is most of it. The magic in the early days was not the tool being brilliant. It was you being clear without noticing.

Get clear on purpose and the honeymoon comes back. Not because the model improved. Because you did.

Idea