No, Claude Didn't "Kill" Web Designers — But This 3-Skill Setup Is Worth Stealing

Why this works at all: skills are the new plugins

The plumbing behind this setup is Agent Skills, a format Anthropic launched in October 2025 and published as an open standard that December. A skill is just a folder with a SKILL.md file — written rules the AI loads when relevant. Cursor, OpenAI's Codex, Gemini CLI, and 40+ other tools have adopted the format, and Vercel's npx skills CLI installs any skill from GitHub into whichever agent you use with one command.

That's the real story here: expert designers are now packaging their judgment as installable text files, and the results are dramatic enough to fuel the hype.

The three skills, verified

Emil Kowalski's skills (~22k GitHub stars, free, MIT). Kowalski is a design engineer at Linear, previously Vercel, and the creator of the Sonner and Vaul React libraries. His repo is actually eight skills — the flagship emil-design-eng encodes shockingly specific rules: button feedback animations at 100–160ms, custom easing curves instead of CSS defaults, never animate from scale(0), always respect prefers-reduced-motion. His essay Agents with Taste explains the philosophy: agents "don't know what great feels like," but taste is transferable if you write the reasons down.

Impeccable (~52k stars, free, Apache 2.0) by Paul Bakaus — the guy who created jQuery UI and spent a decade at Google. Its pitch is a vocabulary problem: "You can't ask for 'more vertical rhythm' if you've never used those words." It ships 23 commands (/critique, /polish, /bolder, /typeset…), an anti-pattern detector, and a curated blocklist of AI clichés — Inter font everywhere, purple gradients, nested cards. It's now an a16z-backed company, and GitHub plans to pre-bundle it with Copilot. It's evolving fast — installs now go through npx impeccable install.

Taste Skill (~68k stars, free, MIT) by Leon Lin brands itself "the anti-slop frontend framework." It bans "AI tells" (em-dashes in copy, three equal feature cards, fake-precision stats), adds tunable dials for visual density and motion, and enforces a 50+ item checklist including WCAG contrast.

One common misconception worth correcting: the Taste skill is often described as letting Claude "pull inspiration from real websites instead of making generic designs." It doesn't browse or scrape anything — it's a written rulebook that references design languages like Linear's and Apple's as directions. The distinction matters: your output is still bounded by what rules someone wrote down, not by the live web.

The eyes and hands: Figma MCP + Playwright

The last step is real and underrated. Figma's MCP server (still beta) lets Claude read design files as structured data — tokens, components, layout — and, newly, write designs back to the canvas. Microsoft's Playwright MCP (~35k stars) gives the agent a browser: it builds your page, screenshots it, compares against the target, and fixes its own mistakes. Anthropic's own guidance notes that 2–3 iterations of this loop typically converge on a good match. This screenshot-iterate loop, more than any single skill, is why AI-built pages stopped looking broken.

Now the fact-check: did anyone actually get killed?

The capability claim is directionally true — for simple marketing sites, a skilled prompter genuinely can get a professional-looking page in minutes. The employment claim is not:

• The U.S. Bureau of Labor Statistics projects 7% growth for web developers and digital designers through 2034 — "much faster than average," with roughly 14,500 openings a year — and BLS explicitly factored AI into those projections.
• Upwork reports demand for AI-skilled freelancers rose 27%. The market is shifting toward designers who wield these tools, not away from designers.
• Quality gaps persist. One academic review found 308 distinct accessibility errors across just six AI-generated websites. AI output also converges on sameness — which is precisely why all three skills in this setup exist. They are patches for a documented weakness, authored by human designers.
• Even Figma's CEO, whose company arguably has the most to lose, says they're hiring in response to AI, not cutting.

The honest version of the claim: AI is compressing the market for low-stakes, templated web work, while raising the premium on strategy, brand, accessibility, and — ironically — taste. The very skills in this setup are human designers' judgment, exported as Markdown. That's not the death of web designers. That's web designers becoming infrastructure.

The takeaway

Steal the setup — it's free and it works:

npx skills add emilkowalski/skills
npx impeccable install
npx skills add Leonxlnx/taste-skill

Then add Playwright MCP so Claude can check its own work, and Figma MCP if you live in Figma:

claude mcp add playwright npx @playwright/mcp@latest

You'll get noticeably better output than raw prompting. Just skip the doom framing. The designers "killed" by this setup wrote the rules that make it work.

Frequently asked questions