AI is very good at code now. Most of what I build is UI, and I can describe a screen and have a working version in a few minutes. I still spend a lot of time on UI, sometimes more than I'd like to admit, because the hard parts of UI were never typing.
I wrote before that AI writes the code and the engineer still owns the system. This post is about the part I own on the interface side: taste.
Link to this sectionI give it everything
My repos have a rules file with the non-negotiables: colours only from design tokens, rounded corners, one icon library, no emojis in the UI, and where every kind of file goes. I run a memory plugin so fixes and decisions from past sessions carry over. The model can read every component I've already built. Skills written by developers who know what they're doing help too.
All of that solves the problem to an extent. I still get slop, and quite often: layouts that break at a real width or in dark mode, a rule it followed yesterday and skipped today, a new component built right next to one that already does the job, a mistake we fixed last week made again with the fix sitting in memory. Sometimes it does the easy 80% and calls it done. Sometimes it treats something I never said as a requirement and leaves out what I did say.
None of that fails a type check. The code compiles, the tests pass, and the page is wrong.
Link to this sectionYou can't one-shot UI
I've stopped trying. No matter how good the skills are or how detailed the prompt is, a first pass at UI has some fuckup in it. The only time it comes out right in one go is when I tell it to copy something that already exists.
AI also can't decide which kind of component a problem needs. Is this a dialog or an inline panel, a dropdown or a segmented control, a table or cards? It picks one and moves on. So now I ask for several variants, pick one and tweak it from there. The model is fast at making options. Choosing between them is my job.
Even with that, a screen can take me a long time to think through. Good UI follows patterns people already know, and knowing which pattern fits takes time no model saves me.
Link to this sectionIcons are a Mahabharat
Lucide is everywhere now, so seeing it sets off the slop alarm in my head. I use Phosphor or Tabler instead, which look better, but the library isn't the problem. The model picks bad icons: ones unrelated to the action, or the most literal option it can find. For a grid view, a matrix-style icon would work. The model never thinks of that.
It also uses one icon for an action on one page and a different icon for the same action on another. So I keep icons in one config and every page reads from it. Once the choice is made there, the model can't make it again somewhere else.
Link to this sectionUX copy turns into an essay
Left alone, the model fills the UI with text: long explanations, instructions on every screen, paragraphs nobody has time to read. It doesn't reach for a tooltip or an info popover to hide detail until someone asks. It rarely does inline errors. It shows one error after the form fails and that's it. And it puts the text in odd places, under the wrong field or far from the thing it explains.
Good UI copy is short, sits right next to what it describes, and appears when it's needed. That takes deciding what to leave out, and the model is bad at leaving things out.
Link to this sectionIt fixes one place
When I point out a problem, it fixes the spot I pointed at. If the same problem is on five other screens, I have to ask again, or it stays broken on four of them. It doesn't go looking for the rest of the pattern unless I tell it to.
Link to this sectionNothing replaces testing it by hand
There's no way around manual testing. I click through every screen, resize it, switch themes, open every dropdown and trigger every error. Most of the problems above only show up that way. A passing test suite tells me the code runs. It doesn't tell me the page is right.
Link to this sectionWhat I do about it
I ask for variants instead of one answer. I centralize anything that has to be consistent, like icons, labels and status styles, so there's one place to get it right. When I report a bug, I say "fix this everywhere" and then check that it did.
I remove code as often as I add it. The model only ever adds: a new component, a new helper, another style next to the old one. It almost never deletes the variant I didn't pick or the component nothing uses anymore. So deleting is a habit I keep on purpose. Less code means fewer places for slop to hide and less for the model to copy from next time.
The rules file keeps changing too. Every time I correct something, it goes into the rules or the docs, so I fix it once instead of every session. I also remove rules that stopped being true or that I no longer agree with. As the repo grows, my taste grows, and the rules are how the agent's taste grows with it.
So don't only build. Build taste while you build. AI might take over most of the code, but I don't think it takes over taste, at least not yet. That part you get by doing the work, testing it yourself and writing down what you learn.