The AI aesthetic

Idioms of a zeitgeist#

Some patterns belong to the technology, like streaming text, made for chat. Others spread: shimmering text, which in AI land implies thinking, is being repurposed for any asynchronous task. Tiny, thin icons in desktop Electron AI apps clash with the system grain of native macOS apps beside them.

Then the list that names the post: the aesthetics his brain associates with AI are beige and cream colours, orange accents, serif typefaces, and “whack-a-mole UI controls,” where clicking a toggle repaints the interface so you have to chase it.

Why sameness happens#

A year earlier, in Some Miscellaneous Thoughts on Visual Design Prodded By The Sameness of AI Company Logos, prompted by Radek Sienkiewicz on the sameness of AI company logos, Jim gave a second cause. Code either works or it does not, so however bad it is you can say “Yeah, but it’s working.” Visual design is subjective forever, so decisions go to whoever has the best language or the most senior title, and the bad logo emerges from everyone’s feedback without anyone suggesting it.

Individuals and small teams “have a sense of taste that doesn’t dilute through the structure and processes of a larger organization.” In The Tumultuous Evolution of the Design Profession he adds the economics: bad design does not visibly hurt margins, so designers rise into “performing taste.” Less craft, more optics.

Maggie Appleton describes the production side. In Vibe Code is Legacy Code she agrees with Steve Krouse that vibe code and legacy code are both code nobody understands, and admits the pull to let the agent go. Her note on Gas Town (January 2026), Steve Yegge’s vibecoded agent orchestration system, reads it as speculative design fiction: once agents churn through code, design becomes the limiting factor.

Her own blockers are never build time but “Is that transition subtle enough? How composable should this be? Is this the right metaphor?”, decisions agents cannot make without your context and taste. Gas Town’s biggest flaw is that it was poorly designed, vibe designed as well as vibe coded. Language is a poor medium for easing curves, and she always needs to touch the CSS.

Write the rules down#

The response is that taste can be articulated and therefore transferred. In Agents with Taste Emil Kowalski says agents do not know what great feels like, so “if you know what great feels like, describe the rules, then give them to your agents.” His unit is a skill file per aspect of the interface. The argument is the balloon again: “almost every taste decision has a logical reason if you look close enough,” so it can be written down.

## Practical Tips| Scenario                        | Solution                                   || ------------------------------- | ------------------------------------------ || Make buttons feel responsive    | Add `transform: scale(0.97)` on `:active`  || Element appears from nowhere    | Start from `scale(0.95)`, not `scale(0)`   || Shaky/jittery animations        | Add `will-change: transform`               || Hover causes flicker            | Animate child element, not parent          || Popover scales from wrong point | Set `transform-origin` to trigger location || Sequential tooltips feel slow   | Skip delay/animation after first tooltip   || Small buttons hard to tap       | Use 44px minimum hit area (pseudo-element) || Something still feels off       | Add subtle blur (under 20px) to mask it    |

The Easing Decision Flowchart is deliberately strict so the agent does not invent rules; a typography excerpt caps body text at about 65ch and puts tabular numerals on price columns. Any taste decision can be packaged this way, layout, icons, colour, and he writes them with Anthropic’s skill-creator.

## Easing Decision FlowchartIs the element entering or exiting the viewport?├── Yes → ease-out└── No    ├── Is it moving/morphing on screen?    │   └── Yes → ease-in-out    └── Is it a hover change?        ├── Yes → ease        └── Is it constant motion?            ├── Yes → linear            └── Default → ease-out## Duration Guidelines| Element Type                      | Duration  || --------------------------------- | --------- || Micro-interactions                | 100-150ms || Standard UI (tooltips, dropdowns) | 150-250ms || Modals, drawers                   | 200-300ms |**Rules:**- UI animations should stay under 300ms- Larger elements animate slower than smaller ones- Exit animations can be ~20% faster than entrance- Match duration to distance - longer travel = longer duration

His demo has Claude Code audit a dialog animation against the rules and return a before and after table. The whole blog became one design engineering skill at emilkowal.ski/skill, and the repository README states the limit: the skills are a side effect of domain expertise, which AI amplifies rather than replaces, “a shortcut to stand out in a sea of slop.”

Review is part of the loop#

Kathryn Gonzalez said something compatible at Config 2023, in Design Systems in Transition: AI can bridge the gap between our taste, intentions, and ability to execute, and a design system becomes the language by which AI tools understand your design, so systems matter more, not less. She defines taste as critical judgement paired with deep domain expertise, used to set an organisation’s standards, and predicts more people will design, code, and build.

Maggie’s bet is the most specific: code-must-be-close for serious work today, drifting toward code at a distance as harnesses mature, made safe by validation loops, tests, and specialised reviewing agents, because “the more agents can validate their own work, the better the results.” Emil’s review skill is one such loop.

Jim has a two-frame test. His Job Screening Blog Post is two GIFs of a button being clicked; in one the heading shifts up a pixel, in the other it does not. “Can you tell the difference between them? Do you care? If not, we might not be a good fit.” Taste is noticing the pixel, caring, and being able to say why; generated interfaces only raise the price of the third.

So, what to practise. Do Emil’s exercises and write the reason before reading his. When a rule you repeat in review is stable, write it down with its why, as an agent could follow it; you will learn which preferences have reasons and which are habits. And when the screen is beige, orange, serif, shimmering, and shifting by a pixel, notice it, care, and say what you would change.