I've been using frontend skills like Anthropic's frontend-design and ui-ux-pro-max in Claude Code and Cursor. On blank landing pages they help a lot. Things kept going sideways whenever I dropped them into an existing React or Next.js codebase just to fix a settings view or clean up a data table.
Half the time the agent put oversized editorial headings into an admin layout, or wrapped sections in 100vw and gave mobile a horizontal scrollbar. And once a long session hit /compact, whatever layout rules I'd typed at the start fell out of context anyway.
I wanted the agent to stop guessing purely from prompts and verify what Chrome actually renders, so I split my setup into an MIT skill called atelier-ui.
Splitting one giant prompt into a ~5 KB router
Keeping 1,500 lines of marketing typography rules loaded while you're editing an admin table doesn't end well.
Instead of dumping everything upfront, SKILL.md acts as a tiny router with three steps:
-
scripts/detect-env.mjsreadspackage.jsonand checks which local browsers are installed once, so the skill never assumes global CLIs are present. -
01-intake.mdsettles open questions in one turn and writes the brief to.atelier/BRIEF.md. Even if a long session compacts twice, the constraints stay on disk. - The agent loads only the build module for your task:
02-brand-site.mdfor marketing pages,03-app-ui.mdfor dashboards and existing codebases, or04-3d-visual.mdfor Three.js and WebGL scenes.
The line in 03-app-ui.md that mattered most in practice is disabling landing-page hero gating and dealt themes from library/themes.json completely. Existing Tailwind tokens and components/ui/ primitives stay untouched, numeric tables keep tabular-nums, and controls have to handle default, hover/focus-visible, disabled, loading, empty, and error states.
Checking mechanical CSS bugs at 0 LLM tokens
Asking an agent in prose to remember prefers-reduced-motion and avoid transition: all works for roughly 200 lines. Moving those checks into scripts/lint-ui.mjs takes them out of the prompt completely.
Here's one rule from lint-ui.mjs:
{
id: 'no-100vw',
severity: 'FAIL',
exts: ['.css', '.html', '.jsx', '.tsx', '.vue', '.svelte', '.astro'],
test: (line) => /\b100vw\b|(?:^|[^\w-])w-screen(?![\w-])/.test(line),
msg: 'Avoid `100vw` / `w-screen` (causes horizontal scrollbar overflow). Use `100%` / `w-full`.',
},
Two small fixes cut out most of the false positives when running this on real repos. stripCommentsPreserveLines() replaces <!-- -->, {/* */}, and /* */ comment blocks with spaces while keeping line breaks intact (and ignores https:// when stripping // comments) so reported line numbers still match your editor. The no-outline-none rule also explicitly allows Tailwind patterns like focus-visible:outline-none focus-visible:ring-2 instead of rejecting valid focus rings.
Checking real Chrome pixels without installing Playwright
Even clean CSS breaks on mobile once real webfonts and sticky headers render. At the same time, forcing everyone who installs a SKILL.md to pull a 200 MB Playwright and Chromium bundle slows down quick setups and clutters package.json.
05-qa-gate.md delegates first: if impeccable or agent-browser is already on your PATH, it uses that. When neither is installed (the usual case on a clean machine), scripts/check-ui.mjs runs with zero npm dependencies by driving whichever Chrome, Edge, Chromium, or Brave binary is already on disk over Node 22's built-in WebSocket client.
In one pass it starts a local static server (resolveSafeFile() blocks .. and sibling-prefix path escapes) or attaches to --url http://localhost:3000. It fails immediately on the layout regressions that hurt most in practice: 1440x900 or 390x844 horizontal overflow, and sticky headers overlapping <h1>. The same run also checks for unloaded webfonts, zero-size SVGs or canvases, oklch WCAG AA contrast errors, console.error output, and blank WebGL frames under --shots. Passing --full --sheet stitches desktop and mobile captures into one review-sheet.png, so a zero-context critic pass only needs a single image turn.
3 live builds and how to try it
I put together a three-mode showcase (yashh1323.github.io/atelier-ui-showcase) with Nordhavn Timber (brand site), DriftGate Console (dark ops UI), and Meridian 04 (WebGL2 astrolabe). Eleven headless Chrome smoke tests in tests/smoke.mjs run against them in GitHub Actions (docs/PROOFS.md).
Include the GitHub prefix if you test the installer, because bare atelier-ui on npm belongs to a different author:
npx github:yashh1323/atelier-ui install
npx github:yashh1323/atelier-ui doctor
# Or via vercel-labs/skills
npx skills add yashh1323/atelier-ui
Repo: github.com/yashh1323/atelier-ui (MIT)
It won't rescue an unclear brief or supply good design taste on its own, and check-ui.mjs needs Node 22+ plus a Chromium-family browser installed on the machine.

Top comments (0)