LLMs can’t tell you if your UI feels good

Despite all the advancements in LLMs, there's one area they can only attempt to emulate: tactility. They can verify logic, catch bugs, even generate entire components. But whether that button press feels satisfying? Whether the animation has the right weight? Completely blind to it.

The feeling of a button responding to user’s press, or a menu smoothly flying out of a button with certain energy and feel is hard to describe to a machine. What’s more, 250ms transition applied to a popover enter transition likely won’t be adequate for a modal. Why? It’s too abstract. What feels satisfying to touch or eye in one case because of weight and timing likely won’t work in another.

Here’s what’s backwards: people are trying to close this gap by feeding LLMs static screenshots and trying to teach them to see what they can’t evaluate anyway. Connect it to browser dev tools, burn tokens on image analysis, get a “yes” or “needs more work” answer. Feedback loop closed, right?

Wrong. You’re still left with the problem of what “feels good”, and no amount of screenshot analysis will fix that.

Instead of trying to solve the unsolvable, focus on what LLMs are great at: predictability and pattern recognition. Let AI write correct, stable frontend code that gets you most of the way there. Then fine-tune using your perception and intuition—the stuff you’ve built over years of practice.

After all, you’re the only one who can judge what pressing a button feels like.