Article
Tacit Taste Is Why AI UI Prompts Fail
When you cannot describe the website you want, that is not always a prompt failure. Some design judgment is tacit and only becomes visible through iteration.
A common AI design failure looks like this:
You ask for a website. The model makes one. It is technically fine. You hate it.
You ask for changes. It gets closer, then worse, then generic again.
After ten rounds, the most frustrating part is not that the model failed.
It is that you still cannot say exactly what you want.
I do not think that is only a prompt skill problem.
Some taste is tacit
There are preferences we can state clearly.
Use smaller type. Remove the gradient. Make the layout denser. Use a table instead of cards. The hero is too tall.
Those are easy.
The hard part is the judgment behind them.
Why does one dense interface feel professional while another feels cramped?
Why does one plain layout feel confident while another feels unfinished?
Why does one screenshot make the product feel real while another feels like stock decoration?
You often know it when you see it before you can explain it.
That is tacit taste.
A full spec is not the same as taste
The instinct is to solve this by writing a better prompt.
Sometimes that helps. A vague prompt produces vague work.
But if the thing you want is not fully explicit in your own mind, no prompt can completely encode it upfront.
The missing part only appears through interaction with the artifact.
You see the draft. Something feels wrong. You compare. You remove. You keep. You discover the rule after the mismatch.
That is not inefficiency. That is design thinking.
AI changes the speed, not the need for judgment
AI makes iteration cheap.
That is good. It lets you externalize a lot of possibilities quickly.
But cheap iteration can also hide weak judgment. If you keep asking for versions without naming what changed in your own preference, the process becomes roulette.
The better workflow is to turn each rejection into a rule.
Not just:
Make it cleaner.
But:
The layout feels like a marketing landing page. This product needs an operational tool surface: tighter spacing, visible state, less decorative copy, and no oversized hero card.
That kind of feedback turns tacit taste into usable guidance.
Slowly.
The real lesson
If you cannot describe what you want, it does not mean you are bad at prompting.
It may mean the preference is not verbal yet.
The job is not to pretend everything can be specified in advance. The job is to build a loop where the artifact helps you discover the rule.
AI is useful in that loop.
But it does not remove the loop.