A skill worth stealing

Stop your AI from generating boring, generic slop.

Ana — who runs Folderly with me — asked me to explain how to design with AI. So here we go: the skill, the before/after, and what I built with it, all in one place.

A model with no art direction returns the internet's average — the purple glow, the floating blobs, the fake dashboard. taste-skill is the written-down taste that makes it execute your decision instead.

Full disclosure, because it's the whole point: I didn't make this. It's Leon Lin's open-source project — 45,000+ stars, MIT-licensed, works in Codex, Cursor, and Claude. I share a fork of it, the way you'd hand a friend a great tool. This page is what stealing a skill well looks like: here's the skill, here's what I built with the discipline, take it — and credit Leon.

Install in one line
npx skills add https://github.com/Belkins/taste-skill
That's my shared fork. The original, always current, is tasteskill.dev, by Leon Lin.

Slop, then taste.

Same prompt. The only variable is the art direction.
A generic-AI slop landing page: purple glow, floating blobs, a fake dashboard with invented stats
Before · no art direction "Moon-base landing page for an email product." With no intent supplied, the model returns the average: a purple-blue glow, floating orbs, gradient headline, and a fake dashboard claiming 99% · 10× · INFINITE. Every tell on the skill's own ban-list.
A directed moon-base hero for Reach: flat illustrated world, one muted blue, real headline
After · the discipline Same brief, one locked intent: a moon base where the inbox is the territory and emails are capsules landing on a "PRIMARY INBOX" pad. One world, one accent, real type. An art-direction study for Reach by Folderly, made with an image model (ChatGPT, not Claude).
flicked.email's three landing pages — chaos, hype, calm — side by side
One product, three intents flicked.email's three live pages — chaos, hype, calm — in three heading typefaces, one constant brand mint. Intent is the input; the model is the executor.
The Playbook's own design system style guide
The proof you're standing in This book runs on the same discipline: a small set of written-down tokens the next screen has to obey. The page you're reading is the receipt. See the style guide →

It's a collection.

Install all of them at once, or one by name. The repo lists the exact install names.
imagegen-frontend-web flagship

Art-directed website images — one image per section, the anti-slop ban-list, a consistent palette across the whole set.

imagegen-frontend-mobile

Premium app-screen concepts and flows — readable hierarchy, multi-screen consistency, clean phone-mockup framing.

design-taste-frontend

The anti-slop frontend skill: reads the brief, infers the direction, ships landing pages and portfolios that escape the template look.

high-end-visual-design

Design like a high-end agency — the exact fonts, spacing, shadows and structures that make a site feel expensive, and the defaults that make it look cheap.

redesign-existing-projects

Audit an existing site, find the generic AI patterns, and upgrade it to premium without breaking anything.

brandkit

Premium brand-guideline boards, logo systems and identity decks — lock a brand world so everything downstream stays on it.

image-to-code

Generate the design image first, analyze it, then build the site to match — section-specific, no cards-in-cards.

minimalist-ui · industrial-brutalist-ui · gpt-taste

Style skills — commit to one art direction (clean editorial, raw brutalist, or GSAP-motion editorial) instead of averaging them into mush.

Sixty seconds.

  1. Install: npx skills add https://github.com/Belkins/taste-skill (my fork) — or one skill: --skill "imagegen-frontend-web". Prefer the source? Swap in Leonxlnx/taste-skill.
  2. Write the intent first: what is this for — loud or calm, the emotion, the positioning. One sentence, before you prompt.
  3. Ask for the design. The skill fires, kills the slop defaults, and holds one world across every frame.
  4. Cull the misses. Image models are stochastic — the skill biases toward your taste; you still pick.

Honest note: the image-generation skills hand the prompt to an external image model (ChatGPT, Midjourney, etc.) — Claude doesn't generate images. The skill is the art direction; the image model is the brush. That's the which-model-for-which-job rule made literal.

What's real, what's not
  • The skill is Leon Lin's, not mine — the original is tasteskill.dev / Leonxlnx/taste-skill (45k+ stars, MIT, with its own research corpus on model "laziness"). I share a fork at github.com/Belkins/taste-skill so it sits with everything else I share — credited to its author, not claimed.
  • The receipts are art-direction studies. flicked.email is a live product; the Reach moon-base set is a pre-build concept (the name isn't locked) — both made with an image model and the discipline, not shipped-and-scaled claims.
  • No invented numbers. The slop image's "99% · 10× · INFINITE" is the parody; nothing here is a metric I don't have.
Where this sits in the book
Stay close

The next edition lands when this list says it does.

No course. No paywall. Operator playbooks weekly. 10K+ subscribers.