Ask a model for a responsive page. You will get md:text-2xl lg:text-4xl.
That is not a technical opinion, it is a corpus effect. The breakpoint was the normal way of writing responsive for a decade, so it fills most of what the machine has read, and it gives back the average of what it has read. Ask it for fluid sizing, with the bounds and the curve, and it executes that very well. It will never propose it on its own.
The whole question of the craft sits in that gap. Not in « is it going to replace me », which is a bad question, but in « who decides what it executes ».
It goes faster than me, and I have no misplaced pride about denying it
Repetitive execution goes to the machine: build work, variants, refactoring, the same screen rolled out across ten formats. It does not tire and does not lose intensity at the twentieth iteration. I do.
This site is the proof. Its concept had been sitting in a drawer for two years, whole, for want of the production time it asked for. The idea has not changed by a single word. The cost of producing it, on the other hand, fell far enough to make feasible what was not.
An intent cannot be subcontracted
Deciding that a title weighs more than an image, that a margin breathes here and tightens there, that one direction is worth more than another: that assumes an intent. I have never managed to subcontract one, to a human or to a machine.
The machine, for its part, is quick to propose the consensual. That is often correct. It is never a stance.
A closed vocabulary lowers the error rate
On this site no colour is hard-coded and no size is set in pixels: named roles, a scale in clamp(), 99% fewer breakpoints with EVA CSS.
When the vocabulary is closed, the machine gets it wrong far less often than on free CSS. It has fewer choices, so fewer occasions to take the most probable instead of the most apt. Constraining the tool costs two days, once. Proofreading its output costs every day.
The heading held at that speed
Too little machine, and you pay execution time at human rates. Too much machine with no hand on it, and the project slides towards the smooth and the interchangeable.
An agency that works with me is not buying raw speed, it is buying a heading. Put like that, the distinction sounds thin.
It can be measured, though. Twenty sizes across four steps make 80 declarations to write, to test and to redo every time the mockup moves. The same twenty sizes in fluid make 20. That is the kind of gap a model will not go looking for on its own, and that I put back at the top of every build pass, in three lines of instruction.



