Getting the thinking right up front is what makes AI build the right thing
Clear specifications were always worth having. Now that both people and AI tools generate from them, imprecision scales — and software that merely looks right is cheaper to produce than ever.
The arrival of capable code generation did not change what makes software succeed. It changed the cost of the failure mode. Vague intent has always produced software that merely looks right; what is new is how quickly, and how plausibly, it can now be produced.
A model given a loose description will return something confident and complete. It will not stop to ask which of the four possible readings you meant, because nothing in the request signals that there were four. The output compiles, it demos, and it encodes an interpretation nobody chose.
Precision is the input, not the output
This makes the up-front work more valuable rather than less. A specification precise enough to hand to an AI tool is the same artefact that was always worth writing: the behaviour stated plainly, the constraints named, the edge cases decided rather than discovered.
The difference is that the artefact is now load-bearing in a way it wasn’t. Imprecision used to be absorbed by a developer who had context, who would notice the gap and come and ask. Generated code absorbs nothing. It fills the gap silently, and it fills it with whatever was statistically likely.
Judgement is what doesn’t transfer
What remains irreducibly human is deciding what should be built, and knowing when the thing that came back is subtly the wrong shape. That judgement comes from understanding the business the software serves — which is the same place all the other answers come from.
The tools are good. They repay clear thinking and they punish hand-waving, both more sharply than people do.