Let's get this out of the way: I don't believe AI can replace a designer's taste. Taste is a very specific mix of temperament, upbringing, everything you've looked at and the personal traumas that shaped you — and how all of that comes through in the work.

But I do believe taste can be turned into a system of checks.

When an agent builds a screen, a piece of copy or a component, it often builds it neatly:

Then you look at it — dead.

Just dead. No soul, no character — the thing I look for in a good designer's work. The thing that grabs you in the first seconds when you look through festival and award entries.

If I tell an agent "this isn't it" a hundred times, that's no longer a mood — it's a repeating signal.

And repeating signals can be collected into a system.

We started collecting these signals into taste — a skill named exactly what it is: taste.

How it appeared: I was building myself a web dashboard for working with agents and their output. I edited, the agent rebuilt, I edited again...

— you're learning to see things the way I do, write that down

That's how the skill was born.

Right now it holds 352 signals and 64 calibrated principles (home lab, measured July 2026).

It's an automated reviewer — a system that looks at the agents' work before my final review.

Below are the 4 checks it all started with.

1. Navigation doesn't lie

A "back" button often looks like you're taking care of the user.

You know — someone landed in the wrong place, let them go back.

But in interfaces it's often not care. It's a symptom.

If a user constantly needs an emergency exit, the system did a poor job explaining where they are and how to leave.

Good navigation doesn't make people escape.

It holds the space so you never have to hunt for a way back.

So the taste check asks:

Is the "back" button really needed — or is it masking a murky information architecture?

2. Menus don't multiply

Another common pattern: the sidebar, the burger and the bottom nav all lead to the same places.

Sometimes it looks like convenience.

More ways to get there — better, right?

On the screen it's noise.

When one section is reachable from many touchpoints, no habit forms: noisy navigation models spread across the system, adding cognitive load and eating up space

An agent loves adding one more path "for convenience".

And suddenly you don't have a product — you have a train station, a crossroads of every route.

During review the skill asks:

Is this navigation element here because the navigation model defined it — or did we just shove it in?

3. States have an address

Anchors drive me mad, and agents adore them: any model building a simple page will hang links in the top nav, like landing pages from 2k17. Except they're not links! They're anchors that scroll the page to the right section. I just hate them.

So my agent has a hard ban on anchors. Always a link. Only a link.

A link has a URL. You can open it, send it, save it, come back to it via browser back or a swipe on your phone.

4. Neat doesn't mean alive

The most dangerous thing about AI design: it's very good at being fine.

Clean.

Smooth.

Safe.

Looking like everything at once.

And that's usually the exact moment character dies.

Take Comic Sans.

I adore it. Comic Sans is a very particular taste. Meta-irony and a quick test.

If an idea falls apart without proper packaging, maybe there was nothing inside.

An agent will almost always try to "improve" a spot like that.

Swap it for something respectable.

Smooth it over.

Align it.

Make it professional.

And kill the test.

Same with visual language.

Dark background, neon, mesh gradients, monospaced captions — and the screen starts smelling like crypto.

Everything can be neat.

But there's less trust.

Did we actually improve this spot — or did we just scrub the life out of it?

A little crooked means a little alive, says my art director.

What came out of it

At first it was just voice notes and edits.

"Too much here."

"Out of place."

"Leave it crooked, it's more alive."

"Smells like crypto."

"An em dash is not a hyphen."

Then it became signals.jsonl.

Then principles started emerging from the signals.

Today taste holds 352 signals and 64 calibrated principles.

That's 352 small collisions with real work.

And my reviewer agent runs on them.

Not the final judge.

My first filter.

It looks at an agent's work before I do and catches the typical mistakes that make me fume a little.

It saves my time.

And stops me spending it on the same things over and over.

In the production pipeline the same approach lives as the Taste layer — I showed how it's wired into AI Review in "AI-native in 100 Days".

Finale

You can argue about whether taste is subjective.

Of course it is.

But subjectivity doesn't stop you from structuring it.

If an expert says "this isn't it" a hundred times, that's not noise anymore — that's training material.

The only question is whether we collect it — or redo it every time, trusting the almighty prompt "don't make mistakes".

I chose to collect. From every conversation and every iteration the agent gathers signals, then brings me selections from my trend radars (radar.looi.ru) or fresh community-awarded apps and sites for review. It forms the signals itself; I calibrate the patterns: they carry different weights, and some are hard stops — they restart the generation right away.

Agents with taste are my way of putting one more layer between the model's output and my final review.

An agent that doesn't bring you the same tidy corpse every generation.

Want to give your AI agent taste — write me: art@looi.ru · t.me/artloooi.

Carousel for this article8 cards, PDF for LinkedIn