Design Is a Short-Horizon Task

Everyone wants AI to design the whole thing. Give it the brief, get back the screen. The best models still can't do this, and design in particular is the wrong place to ask for it.

Why design resists the long horizon

A long-horizon task is one you can hand off and check only at the end. Book the trip, refactor the module, draft the email. The intermediate steps don't need your eyes; only the result does. Design is the opposite. It's judged at every step, because the quality lives in the details, and the details only show themselves when you look.

The distance between a plausible design and a good one is made entirely of those small decisions. The exact spacing. Where the weight sits. Whether this hierarchy actually reads or just looks like hierarchy. None of them are big enough to notice on their own, and all of them together are the whole difference. When a model runs the full horizon and hands you a finished product, every one of those decisions arrives at once, unjudged. Now you have to verify all of it in a single pass: catch every small miss, in every corner, with no sense of which step it went wrong on. That is harder than designing it was. It is much easier to feel each change as you make it, keep what's right, and build from there.

You can encode a lot, but not the eye

The strong objection is that you can encode your way past this. A design system. A set of skills. A CLAUDE.md full of your rules and principles. And you should. That is real leverage: it removes whole classes of decisions, holds consistency, and carries your taste forward so you aren't re-deciding the basics every time.

But encoded rules give you correctness, not rightness. They can enforce what you have already decided. They cannot perceive the experience the design actually creates, here, in this context. Someone still has to look at the result and feel whether it lands. A bigger model doesn't change this either. More capability gives you a better starting point for a step, not permission to skip judging it. The rules and the model get you most of the way; the eye is what finishes it, and the eye is yours.

Ask it to move, not to design

So I stopped asking AI to design and started asking it to move.

The unit that works is one focused decision. Explore three directions for this element. Tighten this spacing. Try a heavier weight here, then undo it. Push this section wider. Each is a short horizon: small enough that I can look at the result, react, and know immediately whether to keep it or throw it out.

Small units are also what make exploration cheap. When the horizon is one element, generating five versions of it costs nothing, and I can react to all five in seconds. That is the real gift. Not one answer, but fast, disposable breadth on a decision narrow enough to actually judge. Diverge wide, look, converge on the one, move to the next detail.

Perfecting the details is how you build for scale

Short horizons sound slow. They are the opposite. Because you are working one decision at a time, you can take the time to get each piece genuinely right, and then put it where it can be reused: a component, a token, a pattern the rest of the work inherits. Every detail you perfect this way stops being a one-off and becomes foundation. You aren't just polishing a screen, you are compounding a system. Moving carefully through small horizons is how the fast, scalable version gets built, not a detour around it.

Right is not a fixed thing

This is also why design never fully automates. Engineering has a correct answer you can specify and test against. Design is an experience. It has to look and feel right, and right is not always something you can encode, because it changes with context: the content, the moment, who is looking and what they came to do. You can hand a model every rule you have and it will still, sometimes, produce something that satisfies all of them and feels wrong. Only a person, looking, catches that.

The models will keep getting better at long-horizon work. Design won't be one of those horizons, because the thing that makes design good is a person perceiving it, closely, one step at a time.