Directing is rejecting

The models will give you something that compiles and looks finished. That is not the hard part anymore.

Directing is what you do next. You reject the plausible thing until the right one is left.

People talk about directing as if it were prompting: a better brief, more iterations, a longer system prompt. That is logistics. Prompting is how you start a run. Directing is how you decide the run is over, and which of the things it produced is allowed to stay.

Generation used to be the scarce part. You paid for it with hours, or with a person. Now a finished-looking interface is cheap, and the expensive part moved. It is the decision to keep something. Every extra variant, every motion that exists to prove the tool works, stays in the product. The rejected ones only cost the afternoon.

That is why the job did not become “write better prompts.” What changed is how easy it got to see who is designing and who is accepting. The models will keep offering the impressive take. Someone has to still be there after the generation finishes and say no.

At Smallstep I needed a badge for status, the small label that has to hold next to a title or inside a table cell. We already had the quiet one, a soft tint with dark type on a wash of color. The model proposed a solid fill instead, the loud pill with light type on a color field. It looked like a product. It failed contrast in three of its four colors, and we did not need a second family of badges. I rejected it.

Notice what the no was about. Not taste. Contrast, and the cost of a second family of badges. Neither shows in a screenshot, and that is where most of my rejections live. This month it was a blur effect Chrome would not render the way the mockup promised, a glass tint that dropped body text to 2.9:1 on a bright wallpaper, a hover state that would have fired on every phone tap. The models get the visible part right on the first try. The part a screenshot cannot show, they get right by accident, if at all, and they will never tell you which.

That is the whole shape of the work. Not a tour of projects. One no, defended.

A model cannot do that for you. It has no cost for being wrong in a way that still looks right. It will not sit with the thing in six months. It will not hear a teammate ask why the other version is gone. You will.