I rejected 9 AI-generated UI concepts. Then I wrote rules.
Nine concepts for my design tool, and every one was the same busy screen. What I asked for instead, and how each spacing fault I caught became a check a machine runs.

In this story3 sections
It was a Tuesday in early November, and I was looking at 9 concepts for my design tool. Three directions, an editor of its own, five brand looks. Different names, different moods. I clicked through them and felt nothing. Then I felt tired.
They were the same screen. Layers on the left, canvas in the middle, a chat on the right. Every one was covered in little proofs of life: role badges, coloured agent tags, status pills, a timestamp on everything. Busy the way a stock photo of an office is busy. Nothing was louder than anything else, so nothing mattered. The names were mood words doing the job an idea should have done. Paper. Graphite. Machine.
I called it generic slop and asked never to see that kind of thing again.
Nine ways of not deciding
The annoying part is that I had asked for it. I wanted range, so agents worked in parallel and each one turned out a lot of screens, quickly and shallowly. A pile of options felt like progress. It was really nine ways of not deciding.
So I changed the brief, not the prompt. One concept at a time, designed in one place instead of farmed out. Remove before adding: the work comes first, and the agents stay quiet until they actually need you. Every piece checked on a full-size crop, pixel by pixel, before I see the whole screen. Show me a small piece early, not a finished deck late.
Later that same day I picked one. I called it Conversation: a chat that works like the assistants people already use, with the reasoning folded away and decisions sitting as cards in the thread. The next pass went dark grey, panels split by surface instead of outlines. My note on it was "cozy for designers, not a dev tool."
A home for the design
Then I built a home for it. A design hub, a small local site outside the product's code, where every round sits on one board under the one before it, made only from the approved kit, with every control working. No mockups that only look right in a screenshot.
That fixed the big ideas. It didn't fix the pixels.
Teaching a checker my eye
My eye kept catching the same faults. A padding a little off. A list whose left edge wandered. A switch sitting low because it lived inside a line of text instead of a row. I'd point at it, it got fixed, and next round something else drifted. Two days in, I typed "u cant learn it?" Fair question, honestly.
The answer was to stop relying on my eye for things a machine can measure. Spacing went onto a strict 4 px unit, with two tiny steps allowed inside controls and nothing else. When an in-between value crept in, I said no. And every fault I caught became a check. One rule mattered more than the checks themselves: before a fix was kept, its check had to fail on that exact bug. A check that would have passed anyway guards nothing. It just makes you feel safe.
- Concepts rejected on 3 Nov
- 9
- Checks behind one command, 4 Nov
- 14
- Checks green by the night of 5 Nov
- 41
- The one spacing unit
- 4 px
Now the hub runs those checks before I see anything, and my reviews are about the things nobody has measured yet. That is where taste is supposed to go.
If you catch the same spacing slip twice, in your work or someone else's, don't fix it a third time. Write it down as something a machine can measure, show the check failing on the bug in front of you, and only then fix it.



