Skip to content
Ben Serkan Zararsız

Teaching an AI your taste

AI can build a screen in minutes. It cannot tell whether the screen feels right. Here is how I turn taste into rules a machine can follow, and what stays human.

On this page
  1. Intro
  2. Adjectives are not instructions
  3. Put taste in tokens
  4. Rules, not vibes
  5. Review like a creative director
  6. The brief behind every project

Today I work as a design engineer. With AI-assisted engineering I take projects from brand and interface all the way to production code. The tools build astonishingly well and judge surprisingly badly. Ask for a “polished, modern” dialog and you get something that is, technically, a dialog.

The gap is taste, and taste does not travel through adjectives. It travels through rules. The better I can say why a decision is right, the more of my judgment an agent can apply without me in the room. This is what it looks like in practice.

Adjectives are not instructions

“Make it feel premium” means nothing to a machine and, honestly, not much to a junior designer either. What works is the reason behind the feeling. A button feels responsive because it scales to 96 percent on press, within 140 milliseconds. A dialog feels calm because it has one primary action and the rest stay quiet. Once the reason is explicit, it can be written down. Once it is written down, it can be followed.

So the first step is uncomfortable: explain your own decisions. Why ease-out and not linear, why 8 pixels and not 10, why two type sizes and not five? If the answer is “because it looks better”, keep digging. The real answer is usually logical, and it usually fits in one line.

Buttons: 40px tall, 8px radius, 600 weight. One primary per screen (indigo). Secondary: white, 1px ring. Tertiary: text only. Press: scale .96, 140ms ease-out.

Numbers, roles, one accent

Make the buttons feel premium and modern.

Three guesses at “premium”

Put taste in tokens

The most reliable way to carry taste into code is tokens. Colour, spacing, radius, type, motion: when every value has a name and a place, an agent cannot invent a new grey. Since late 2025 there is a stable, shared format for exactly this, so tokens move between design tools and code without translation. Brand and product as one system, now readable by machines too.

Tokens also make the second test cheap: switch the theme. A component built from tokens survives dark mode without a single change. A component with hard-coded values breaks within the first second. That is a taste test an agent can run on its own, before I ever see the screen.

Invoice #2041

Due in 14 days · €1,840.00

Sent

Built from tokens

Invoice #2041

Due in 14 days · €1,840.00

Sent

Hard-coded white, inherited text

Rules, not vibes

My rule files read like a strict brief: durations by element type, one easing per situation, type rules with numbers, spacing on an 8-pixel grid, one saturated colour per screen, and a short list of things that are never allowed. Tables and short lists work better than paragraphs, because an agent applies them mechanically, and mechanical is exactly what you want for the parts of taste that can be stated.

The point is to stop spending judgment twice. Every rule I write once is a decision I never have to make again, and one the agent cannot get wrong. Agentic tools are already in most product teams this year. The teams that benefit and the ones that drown in output often use the same model. The difference is how clear their standards are.

Delete “Website relaunch”?

All 24 files will be deleted. This cannot be undone.

After the rules: one danger action, short copy

Delete project?

This will permanently delete “Website relaunch” and all of its files. This action cannot be undone. Please confirm that you really want to proceed.

Before: three equal buttons, four sentences

Review like a creative director

Rules get you to good. The last ten percent, from good to right, is still a human looking at the result and saying no. I review AI output the way an art director reviews a layout: does the hierarchy hold at a glance, does the motion start where the user acted, is there anything on the screen that has not earned its place.

This is also where the responsibility stays. As generating interfaces becomes trivial, judgment, context and taste become the actual work. The machine builds. The human decides what is right. That division is the job description, and I like it that way.

The brief behind every project

In my projects, every agent gets the same brief:

  1. Tokens for colour, spacing, type, radius and motion, in one file.
  2. Duration and easing rules by element type.
  3. Typography rules with numbers: sizes, line length, tabular figures.
  4. A list of forbidden things: no ease-in on entrances, no scale from zero, no third primary button.
  5. Reference screens that already feel right.
  6. A review pass, by me, before anything ships.

Craft until it ships still applies, only the hands have changed.

Sources

  1. W3C Design Tokens Community Group: Design Tokens Format Module 2025.10, first stable version (October 2025)
  2. Nielsen Norman Group: AI in UX: Work Is Faster, But Practitioners Have Mixed Feelings (2026)
  3. Nielsen Norman Group: State of UX 2026: Design Deeper to Differentiate (2026)