Inheritable Taste Through Skill Files
Emil Kowalski's thesis that design taste - the accumulated judgment about what looks good, what feels right, what to avoid - can be encoded into skill files and passed between agents. The theoretical foundation for Kevin's entire taste-* skill taxonomy.
The Thesis
Design taste has historically been considered non-transferable: you either have it or you don't, and even designers who have it struggle to articulate why something works. Emil's insight is that skill files change this equation. A skill file can encode not just procedures ("how to center a div") but judgment ("when to use 12px vs 16px spacing, and why the answer depends on information density").
The mechanism is inheritance: Agent A develops taste through exposure to good design and encodes it in a SKILL.md. Agent B reads that skill file and inherits the taste without needing the original exposure. The skill file is the gene - it carries design DNA between agent instances.
This is distinct from style guides (which encode rules) and design systems (which encode tokens). Taste encodes the decision-making process - when to break the rules, when to deviate from the system, when something that follows every rule still looks wrong. Source: X/@emilkowalski, 2026-05-07
The Motion Taste Payload
The bookmark's four image cards are the important artifact. They are not vague praise for "taste"; they are a compact decision system that an agent can apply while reviewing UI motion.
| Decision surface | Encoded rule |
|---|---|
| Easing | Entering/exiting viewport uses ease-out; moving or morphing on-screen uses ease-in-out; hover/color changes use ease; constant motion uses linear; default back to ease-out. |
| Duration | Micro-interactions should land around 100-150ms; standard UI such as tooltips and dropdowns around 150-250ms; modals and drawers around 200-300ms. UI animation should stay under 300ms unless the surface is rare, large, or explanatory. |
| Asymmetry | Exits can be roughly 20% faster than entrances. Larger elements and longer travel can justify longer durations; frequent UI should get shorter and quieter. |
| Practical polish | Pressable controls need subtle scale(0.97) feedback; entrances should start from about scale(0.95) plus opacity, not scale(0); popovers should scale from their trigger; sequential tooltips should skip delay/animation after the first; hover effects need @media (hover: hover) and (pointer: fine). |
| Review failures | Flag transition: all, ease-in on UI elements, keyboard-triggered animation, UI durations above 300ms, keyframes on rapidly-triggered elements, same-speed enter/exit, Framer Motion x/y props under load, and all elements appearing at once. |
These rules now live in design-engineering-polish so they are executable, not just inspirational. Source: local artifacts wiki/assets/x-bookmarks/2052400581528916288/image-01.jpg through image-04.jpg, reviewed 2026-07-01
Kevin's Implementation
Kevin's wiki has 13+ taste-* skills, routed through taste-skills-playbook:
| Skill | Encodes |
|---|---|
taste-core |
Metric-based rules, component architecture, LLM bias overrides |
taste-soft |
$150K agency feel - fonts, spacing, shadows, depth |
taste-gpt |
GSAP ScrollTrigger, editorial typography, AIDA structure |
taste-minimalist |
Warm monochrome, flat bento, editorial restraint |
taste-brutalist |
Swiss typography + military terminal aesthetics |
taste-brandkit |
Premium brand identity generation |
taste-imagegen-web |
Per-section website design references |
taste-imagegen-mobile |
App-native screen concepts |
Each skill is a taste gene - a distilled judgment about what "good" looks like in a specific context. The playbook routes incoming tasks to the right gene based on context signals. This is the resolver pattern applied to aesthetic judgment.
Why It Matters
The compounding implication: taste improves with use. Each time a taste skill fires and Kevin approves or corrects the output, the skill can be refined. Over time, the skill files encode not just Emil's taste or Kevin's taste but a hybrid that reflects actual project decisions. The skills become a persistent aesthetic memory that survives across agent sessions, model versions, and tool migrations.
This also explains why Kevin's skill taxonomy exists at all - 13 taste variants isn't redundant, it's genetic diversity. Different projects need different taste genes. The brutalist gene would be wrong for a luxury SaaS landing page, and the soft gene would be wrong for a developer tool dashboard.
Timeline
- 2026-07-01 | Deep-reviewed Emil's four skill-file taste cards and resolved the follow-up link to
emilkowal.ski/skill. Added the concrete easing, duration, practical tips, and review-failure rules to this page and the executabledesign-engineering-polishreference. Source: X/@emilkowalski, 2026-05-07; https://emilkowal.ski/skill - 2026-05-07 | Emil Kowalski shared the "inheritable taste through skill files" thesis on X, with four motion-rule cards and a follow-up link to his skill page. Source: X/@emilkowalski
- 2026-05-17 | Compiled into dedicated wiki concept page. Kevin's taste-* skill taxonomy identified as direct implementation of this thesis. Source: wiki compilation, 2026-05-17