Is Taste Skill an official Anthropic skill?

No. It is an MIT-licensed personal project by developer Leonxlnx, documented at tasteskill.dev. It works in Claude Code because Claude Code reads the open SKILL.md format, and the same files load in Cursor and Codex.

Yes — MIT license, no paid dependencies. The project is sponsor-funded, which is why the README carries a Kimi affiliate banner and sponsor logos; the skill files themselves contain no promotion.

What changed between v1 and v2?

v2, now the default under the same install name, is a substantial rewrite: brief inference with a stated Design Read, three 1-10 dials for variance, motion, and density, a redesign-audit protocol, and a strict pre-flight check. It is labeled experimental, and v1 remains installable as design-taste-frontend-v1.

Will it improve my dashboard or app UI?

Probably not — the skill's own header scopes it to landing pages, portfolios, and redesigns, and explicitly excludes dashboards, data tables, and multi-step product UI. For product interfaces, an alternative like pbakaus/impeccable casts a wider net.

Skill review · by Leonxlnx (Leon Lin; @lexnlin on X)

Taste Claude Skill

The taste skill Claude users search for is Taste Skill, an open-source library of SKILL.md files by solo developer Leonxlnx that steers AI coding agents away from generic "AI slop" frontends. The repo, Leonxlnx/taste-skill, sits at about 83,200 GitHub stars, is MIT-licensed, and was last pushed August 24, 2026 — popular and maintained, though it is a personal project, not an Anthropic product. It contains 13 skills: ten implementation skills, led by the flagship design-taste-frontend, alongside brutalist, minimalist, soft, redesign, and output-enforcement variants, plus three image-generation skills that produce reference boards rather than code. The one thing to know before installing: the default skill is a v2 experimental rewrite whose rules are still changing, and its own header scopes it to landing pages, portfolios, and redesigns — not dashboards, data tables, or multi-step product UI. A frozen v1 remains installable for anyone who depends on stable behavior.

View the sourceMIT licensed · 83,168★ · facts checked September 1, 2026

By Robert MacKelfresh

Founder, Automation Squad · September 1, 2026

Install

1Fastest path: run npx skills add https://github.com/Leonxlnx/taste-skill (the Vercel agent-skills CLI scans the repo's skills/ folder); add --skill "design-taste-frontend" to install only the default skill into your current project.
2Claude Code plugin route: the repo ships a .claude-plugin/marketplace.json, so run /plugin marketplace add Leonxlnx/taste-skill inside Claude Code, then /plugin install taste-skill@taste-skill to get all variants at once.
3Manual install for Claude Code: copy a skill folder (e.g. skills/taste-skill/) into ~/.claude/skills/ for all projects, or .claude/skills/ inside one repo — Claude Code discovers the SKILL.md automatically.
4Claude.ai on the web: zip a single skill folder with its SKILL.md at the root and upload it under Settings → Capabilities → Skills (skill upload requires a paid plan).
5To pin stable behavior instead of the experimental v2 default, install the frozen original: npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend-v1".

What it does

Brief inference: Before writing code, the agent reads the brief and states a one-line "Design Read" — page kind, audience, vibe, and target design system — and is explicitly banned from LLM defaults like AI-purple gradients, three equal feature cards, and Inter + slate-900.
Three configuration dials: DESIGN_VARIANCE, MOTION_INTENSITY, and VISUAL_DENSITY (1-10 scales, baseline 8/6/4) gate every layout, motion, and density decision; a signal table maps vibe words like "Linear-style" or "Awwwards" to dial values, and overrides happen conversationally.
Style variants: Standalone skills for chosen directions: brutalist (Swiss type, sharp contrast), minimalist (Notion/Linear-style editorial UI), soft (calm premium UI with spring motion), stitch (Google Stitch-compatible rules), and a stricter gpt-taste variant for GPT/Codex.
Redesign audit: redesign-skill audits an existing UI first, then fixes layout, spacing, hierarchy, and styling; the v2 default also carries a redesign-audit protocol and a strict pre-flight check before shipping.
Image-generation skills: imagegen-frontend-web, imagegen-frontend-mobile, and brandkit produce reference comps, mobile flows, and brand-kit boards (images only, no code) to feed back into a coding agent — image-to-code-skill chains generate, analyze, and implement into one workflow.

Common questions

Is Taste Skill an official Anthropic skill?: No. It is an MIT-licensed personal project by developer Leonxlnx, documented at tasteskill.dev. It works in Claude Code because Claude Code reads the open SKILL.md format, and the same files load in Cursor and Codex.
Is Taste Skill free?: Yes — MIT license, no paid dependencies. The project is sponsor-funded, which is why the README carries a Kimi affiliate banner and sponsor logos; the skill files themselves contain no promotion.
What changed between v1 and v2?: v2, now the default under the same install name, is a substantial rewrite: brief inference with a stated Design Read, three 1-10 dials for variance, motion, and density, a redesign-audit protocol, and a strict pre-flight check. It is labeled experimental, and v1 remains installable as design-taste-frontend-v1.
Will it improve my dashboard or app UI?: Probably not — the skill's own header scopes it to landing pages, portfolios, and redesigns, and explicitly excludes dashboards, data tables, and multi-step product UI. For product interfaces, an alternative like pbakaus/impeccable casts a wider net.

Before you install

Not an Anthropic product — a personal project by one developer (Leonxlnx), MIT-licensed, with support via GitHub issues, X DMs, or email only.
The default skill is a v2 experimental rewrite that is "actively iterating"; the docs warn section numbering and rule wording may change before a v2.0.0 stable cut, so a reinstall can silently change behavior.
Scope is narrower than the name suggests: the skill's own header says landing pages, portfolios, and redesigns — "not dashboards, not data tables, not multi-step product UI."
Opinionated defaults: baseline DESIGN_VARIANCE of 8 pushes asymmetric, experimental layouts with GSAP-flavored motion; conservative brands must dial it down conversationally or the output can overshoot.
There are zero GitHub releases — no version tags to pin against; the only stable pin is the separately preserved v1 skill.
The README is heavily sponsor-monetized (a Kimi affiliate banner offering bonus API credits, plus seven sponsor logos); the skill files themselves are clean, but the repo doubles as ad space.
The repo carries a disclaimer that crypto tokens using the project's name exist and are unaffiliated — a sign of active impersonation scams around the brand.
Several bundled skills are not aimed at Claude at all (gpt-tasteskill targets GPT/Codex; the three imagegen skills target image generators), so installing everything into Claude Code adds dead weight.

Facts come from the skill's own repository and docs on the date above, via the GitHub API — not from memory, and never from a sponsor.

Worth comparing

Impeccable (pbakaus/impeccable) — 64.6k stars; a broader design-language skill covering product UI, accessibility, theming, and design systems — a better fit for the dashboards Taste Skill excludes.
Emil Kowalski's skills (emilkowalski/skills) — 34.4k stars; taste and animation skills from the design engineer behind animations.dev — the "Agents with Taste" lineage, and its author sponsors Taste Skill.
Tastemaker (codeswithroh/tastemaker) — About 290 stars; grounds AI-generated UI in real reference images and a persistent per-developer taste profile rather than written rules — a different mechanism, far smaller community.

← All MCP servers & skills