Enforced in Code, Not in Taste

Almost everything I publish is made with AI, and people tell me it does not look like it. That is a strange compliment to receive, because it implies a house style for machine-made work — and there is one. It is worth saying out loud what produces it, and then what actually prevents it.

Why the default is convergence

A generative model produces what is most probable given everything it has seen. Ask for a landing page and you get the average landing page: centred headline, gradient, three feature cards, a call to action in a rounded rectangle. Ask ten times and you get ten variations on the mean. This is not a flaw in the tool. It is the tool working exactly as described. Averaging is the mechanism.

Which means the sameness people complain about is not caused by laziness, or by AI lacking taste. It is caused by nothing in the process pushing away from the centre. The output converges because nothing makes it diverge.

What I do instead, stated plainly enough to check

The visual system for this studio is not a mood board or a set of preferences I try to remember. It is code, and it is small enough to describe in a paragraph.

There is a token file with ten accents and three themes, and nothing may introduce a colour that is not in it. There is a renderer that draws every carousel at 540 by 675 and exports at twice that, so every slide that has ever shipped is dimensionally identical. There is a frame component that supplies the watermark, the page counter and the footer, so those cannot drift between decks. And there is a written canon with eight checks that run before anything ships.

None of that is unusual. Design systems have existed for decades. What is different is where it sits: in the generation path rather than beside it. The constraint is not a document I consult after making something. It is the thing doing the making.

The part that matters: a rule that can fail

Two of those eight checks are about form rather than fact. One says every visual frame must contain an actual visual. The other says no more than two consecutive slides may share a layout.

They sound trivial. They are the two that do the most work, because they are the two that catch me.

This week I built a carousel arguing that evaluators should publish their thresholds before seeing the system. The facts were checked hard — every date, every figure, sourced and dated. Then I rendered it and three of the seven slides were paragraphs in boxes. Not diagrams. Text, wearing the shape of a diagram. The check caught it, and the three were rebuilt as a convergence diagram, a pair of threshold bars, and a timeline.

Two days later I did it again. A post about a regulatory deadline moving, whose single most visual fact was two dates jumping forward in time — and I rendered it as a table of rows. A timeline drawn as a list. Caught again, rebuilt as an actual time axis with the old dates ringed and arcs carrying them forward.

I am telling you about both failures because an essay claiming that rules produce quality has to show the rule catching something. A rule that never fires is not a constraint. It is a description of what you were going to do anyway, which is taste with better branding.

What a system does that a prompt cannot

A prompt produces one artifact. If it is good, that is a result. It is not a system, and you find out the difference at the hundredth artifact rather than the first.

The question a system answers is not "is this good" but "is this the same as the ninety-nine before it, in the ways that matter, and different in the ways that should differ". Colour, proportion, typography, the mark, the counter — identical, because they are supplied rather than chosen. Structure, argument, the shape of each figure — different every time, because the canon forbids repeating a layout and the argument decides the length.

You cannot prompt your way to that. Not because the model is not capable, but because consistency across a hundred artifacts is not a property of any one generation. It is a property of the constraint they all passed through.

The honest limit

This is expensive, and I would rather say so than sell it as a trick.

It took building a renderer, writing a token file, and — the slowest part by far — writing down rules specific enough to fail. "Make it distinctive" cannot fail. "Every visual frame must contain a visual" can, and did, twice this week. Getting from the first kind of sentence to the second is most of the work, and it is not work a model does for you, because it is the part where you decide what you actually mean.

Which produces an uncomfortable conclusion for anyone hoping this is a shortcut. The differentiator is not the AI. Everyone has the same models. The differentiator is the constraint you were willing to write down and then be caught by — and that remains as scarce as it ever was, because it is unpleasant and nobody can do it for you.

Constraint is the opposite of a prompt. A prompt asks for something. A constraint refuses things, including things you have already made and would rather keep.