AI makes it tempting to begin with a detailed specification: describe the feature, hand it to the model, and wait for the code.
It feels efficient. It is where learning stops.
A specification captures what we think we know. Early in a product, that is not enough. We have not seen how people use it. We do not know the awkward edge cases, the missing context, or where the technology pushes back.
So start with a prototype. Make the smallest working version that lets someone use the idea. Put it in front of real users. Watch what happens. Change it. Repeat.
This is not an argument against specifications. It is an argument about order. The best specification is the one you write after the prototype, not before it.
Prototype-driven development puts the person using the product first. The functionality follows from what you learn by watching them.
AI makes these cycles faster. That is the opportunity: not to lock in our assumptions faster, but to test them faster.