Skip to content

Best Frontend Skills: Design, React, Next.js and Mobile

The frontend agent skills worth installing in September 2026 fall into three jobs: design skills (frontend-design, ui-ux-pro-max, design-taste-frontend, impeccable) that fix the direction before UI code exists, stack skills (vercel-react-best-practices, the Next.js skills in vercel/next.js, Expo, Remotion) that keep code current, and web-design-guidelines, which reviews the finished UI. One skill per job is enough.

Your agent built the pricing page in four minutes: a purple gradient hero, three identical cards, Inter everywhere, and a signup form with no autocomplete attributes and no visible focus state. The next page looks the same, because nothing told the agent what your product looks like or what “done” means for a UI. This page is for frontend developers who want the agent to produce UI they can ship, and for tech leads who need that result checked by gates rather than by someone squinting at every diff.

What you get from this guide to frontend skills

Section titled “What you get from this guide to frontend skills”
  • A short list of frontend skills by job, with a dated popularity line and the right repository for each one
  • Install commands for Claude Code, Codex and Cursor, including the plugin routes
  • A worked example: a landing page built with ui-ux-pro-max, then reviewed with web-design-guidelines
  • A four-stage loop (design system, build, accessibility and performance review, Chrome DevTools trace) with the gate that closes each stage
  • The install traps that make a skill silently do nothing, and how to recover

Which frontend skills are worth installing in 2026?

Section titled “Which frontend skills are worth installing in 2026?”

Popularity, as of 2026-09-26. “sk” is all-time skills.sh installs from the third-party LinklyAI/best-skills scrape dated 2026-09-26 (secondary; re-read the skill’s skills.sh page before you quote a number). “Plugin” is installs on Anthropic’s plugin directory (claude.com/plugins). Stars are GitHub stars. The three metrics are not comparable with each other.

JobSkill (repository)What it makes the agent doPopularity (2026-09-26)
Design directionfrontend-design (anthropics/skills)Commits to a purpose, audience and aesthetic before writing UI, away from templated defaults923,303 sk; plugin 1,134,112
Design systemui-ux-pro-max (nextlevelbuilder/ui-ux-pro-max-skill)Generates palettes, type pairings and UX rules from bundled data before code371,037 sk; 130.7k stars
Anti-generic tastedesign-taste-frontend (Leonxlnx/taste-skill)Applies layout, type and spacing constraints and a ban list of “AI look” patterns520,147 sk; 90.2k stars
Design vocabulary and lintimpeccable (pbakaus/impeccable)24 /impeccable commands plus a detector with 61 deterministic rules293,766 sk; 71.3k stars
React performancevercel-react-best-practices (vercel-labs/agent-skills)70 rules in 8 categories; request waterfalls and bundle size come first744,240 sk
UI reviewweb-design-guidelines (vercel-labs/agent-skills)Reviews UI code against 100+ interface and accessibility rules, reports file:line667,369 sk
Next.jsNext.js skills (vercel/next.js)Cache Components and partial prefetching adoption, the next dev looprepo 142.4k stars (whole project)
MobileExpo skills (expo/skills)Expo Router, native UI, SDK upgrades, EAS builds and updates2.6k stars
VideoRemotion skills (remotion-dev/skills)Scaffolds and renders React video compositions4.7k stars

Two things the table implies. First, the design skills overlap: all four push the agent away from generic UI, so install one of them, not four (see how to choose one design skill). Second, only web-design-guidelines and impeccable check work after it is written. The others shape generation, which is exactly the output you still have to verify.

How do you install the frontend skills in Claude Code, Codex and Cursor?

Section titled “How do you install the frontend skills in Claude Code, Codex and Cursor?”

The skills CLI (npm skills 1.7.0) installs the same SKILL.md folders into every agent; the flags below come from its --help. It matches --skill against the skill’s frontmatter name, not its folder name, and it needs a space, not =, after --skill. Run these in a terminal at the repository root so the skills land at project scope and go into version control.

Terminal window
# Anthropic's plugin directory (if the marketplace is missing: claude plugin marketplace add anthropics/claude-plugins-official)
claude plugin install frontend-design@claude-plugins-official
# Vercel's React and review skills, into .claude/skills
npx skills add vercel-labs/agent-skills --skill vercel-react-best-practices web-design-guidelines -a claude-code
# ui-ux-pro-max as a plugin, from its own marketplace
claude plugin marketplace add nextlevelbuilder/ui-ux-pro-max-skill
claude plugin install ui-ux-pro-max@ui-ux-pro-max-skill

Plugin skills are namespaced: the plugin’s skill runs as /frontend-design:frontend-design, and it also fires on its own when you ask for UI. Claude Code also bundles a /design skill (research preview); try it before you add a third-party one for the same job.

The other items in the table install the same way in every agent. Add -a claude-code, -a codex or -a cursor to choose the target:

Terminal window
npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"
npx impeccable install # then run /impeccable init inside the agent
npx skills add vercel/next.js
npx skills add expo/skills --skill '*'
npx skills add remotion-dev/skills

In Claude Code, Expo is also a plugin: claude plugin install expo@claude-plugins-official. In Codex, impeccable installs a project hook that you approve under /hooks after install, and again after any update that changes .codex/hooks.json.

Context cost. A skill costs only its name and description until it fires. After you install a plugin, claude plugin details frontend-design@claude-plugins-official prints its always-on token cost (about 78 tokens for this plugin, measured with Claude Code 2.1.283 on 2026-09-26); check /context in a session as well. MCP servers such as Chrome DevTools add tools to every session. Claude Code and Codex defer the tool schemas with tool search, which is on by default in both (Claude Code 2.1.283, Codex 0.157.1), so the cost is lower than the full schema, but it is not zero; add these servers at project scope only in frontend repositories.

How do you build a landing page with ui-ux-pro-max and review it with web-design-guidelines?

Section titled “How do you build a landing page with ui-ux-pro-max and review it with web-design-guidelines?”

This is the example the rest of the page generalises. The product is a hypothetical developer-tools SaaS with a Next.js App Router front end.

  1. Generate and save the design system before any page code. ui-ux-pro-max can persist it to design-system/<project>/MASTER.md, which becomes the file the agent reads on every later page.

  2. Review the design system, not the code. Check the contrast ratios, the type scale and the anti-patterns list. This is the one artifact a human approves line by line, because every later page inherits it.

  3. Build the page against the approved file.

  4. Review the result with web-design-guidelines. The skill fetches the latest guidelines before each review and reports terse file:line findings.

  5. Expect findings such as SignupForm.tsx:42 input missing autocomplete="email" or a missing focus style on the pricing toggle. The review is done when the second run reports zero accessibility findings and any remaining design notes are written down as accepted.

What does a frontend loop from design system to performance trace look like?

Section titled “What does a frontend loop from design system to performance trace look like?”

The example above covers design and review. A shippable frontend change needs two more checks the skills cannot do on their own: an automated accessibility and design gate, and a measurement of the running page. The loop has four stages, and each ends in an artifact you can check without reading the component code.

StageSkill or toolGate that closes the stageWho signs off
1. Design systemui-ux-pro-max or frontend-design (or /impeccable init, which writes PRODUCT.md)An approved MASTER.md or direction statement with contrast ratiosDesigner or tech lead
2. Buildvercel-react-best-practices, the vercel/next.js skills, Expo skills for mobileType check, lint and unit tests green; no raw colour values outside the tokensCI
3. Accessibility and design reviewweb-design-guidelines, /impeccable audit, npx impeccable detectZero open accessibility findings; detector output attached to the pull requestCI plus reviewer of the findings list
4. Runtime traceChrome DevTools MCP (performance_start_trace, performance_analyze_insight, lighthouse_audit)LCP, CLS and the Lighthouse accessibility score within the budget you setReviewer of the trace summary

Stage 3 is where a deterministic check replaces eyeballing. impeccable’s detector runs its 61 rules with no model and no API key, so the same input gives the same findings in CI:

Terminal window
npx impeccable detect --json src/ # CI-friendly JSON output
npx impeccable detect http://localhost:3000 # the rendered page, using an installed Chrome

The detector exits 2 when it finds primary findings and 1 when a target cannot be scanned, so a CI step fails on either; exit 0 means a clean scan (impeccable 4.1.0).

Stage 4 needs the Chrome DevTools MCP server (npm chrome-devtools-mcp, 1.10.1 on 2026-09-23; 52.6k GitHub stars on 2026-09-26). By default it sends usage statistics to Google, and its performance tools can send trace URLs to the Chrome UX Report (CrUX) API. The commands below turn both off with --no-usage-statistics and --no-performance-crux:

Terminal window
claude mcp add chrome-devtools --scope project -- npx chrome-devtools-mcp@latest --no-usage-statistics --no-performance-crux

Or install the plugin, which adds seven skills such as debug-optimize-lcp and a11y-debugging but pins an older server (1.9.0): claude plugin install chrome-devtools-mcp@claude-plugins-official. The plugin starts the server with its default telemetry and gives you no place to add the opt-out flags, so the flags above cannot be passed. Usage statistics can still be turned off by exporting CHROME_DEVTOOLS_MCP_NO_USAGE_STATISTICS=1 before you start Claude Code (they are also off when CI is set). The CrUX lookup has no environment switch, so use the claude mcp add command above if that matters (checked on chrome-devtools-mcp 1.10.1).

A lab trace on one machine is a smoke test, not a performance budget. Put the numbers in the pull request, and keep the regression check in CI with the approach in performance testing. For keyboard and screen-reader coverage that a static review cannot see, add the Playwright and axe checks from accessibility testing.

All four design skills attack the same problem, generic “AI look” UI, from different angles. Installing several gives the agent competing instructions.

Your situationChooseWhy
New UI, no design system, you want the agent to commit to a directionfrontend-designSmallest footprint; first-party plugin in Claude Code
Greenfield product that needs tokens and UX rules written down firstui-ux-pro-maxProduces a persisted design-system file the agent reads on every page; needs Python 3 for its search script
Marketing surfaces: landing pages, portfolios, redesignsdesign-taste-frontendOpinionated ban list and variants; see Taste Skill
A team that wants shared design verbs and a lint-like checkimpeccable/impeccable audit, polish and critique plus the deterministic detector; see Impeccable
An existing design system in Figma or shadcn/uinone of the above firstUse the Figma MCP server or the shadcn/ui MCP server so the agent reuses your components

web-design-guidelines is not in this table because it reviews rather than designs. It pairs with any of them.

Which stack skills keep React, Next.js and mobile code current?

Section titled “Which stack skills keep React, Next.js and mobile code current?”

React. vercel-react-best-practices fires on data fetching and performance work and applies its request-waterfall rules first. The same repository has vercel-composition-patterns (compound components, React 19 API changes) and vercel-react-view-transitions.

Next.js. The Next.js skills moved into the framework repository: npx skills add vercel/next.js installs next-cache-components-adoption, next-cache-components-optimizer, next-dev-loop and the partial prefetching skills. The retired vercel-labs/next-skills repository and its next-best-practices and next-upgrade skills are gone; that knowledge now ships as bundled docs and the AGENTS.md or CLAUDE.md that next dev generates in Next.js 16.3 and later. Next.js 16.2 ships the same version-matched docs in node_modules/next/dist/docs/ but does not write the instructions, so add a line to your AGENTS.md that tells the agent to read that directory. On Next.js 16.1 or earlier, run npx @next/codemod@canary agents-md to download version-matched docs to .next-docs/ and index them in AGENTS.md. Working with the running dev server is covered in dev-server MCP servers.

Mobile. expo/skills has 26 skills (24 in the Expo plugin, plus expo-skill-eval and expo-migrate-module), including an expo-overview router, expo-router, expo-native-ui, expo-upgrade and EAS skills such as eas-update and eas-workflows. expo-skill-feedback controls its telemetry. For React Native without Expo, add vercel-react-native-skills from vercel-labs/agent-skills, or Callstack’s react-native-best-practices (callstackincubator/agent-skills), which is a consultancy’s community set, not the vendor’s.

Video. remotion-dev/skills holds 12 skills. /remotion-create Make a promo video for a record store scaffolds a composition, and /remotion-render renders it.

Vue, Angular and Svelte. We checked these collections with npx skills add <repo> --list on skills CLI 1.7.0:

StackInstallSkills found
Vue, Nuxt, Vite, Vitestnpx skills add antfu/skillsvue, nuxt, pinia, vite, vitest, unocss, nitro and 4 more (11)
Angularnpx skills add angular/skillsangular-developer, angular-new-app (2)
Sveltenpx skills add sveltejs/ai-toolssvelte-code-writer, svelte-core-bestpractices

For official vendor collections outside the frontend (Stripe, Supabase, Cloudflare and others), see official vendor skills.

When frontend skills break, and how do you recover?

Section titled “When frontend skills break, and how do you recover?”

The skill never fires. You ask for a page and get the old generic output. Usually the install matched nothing (see the traps above) or the skill went to a directory your agent does not read. Recover: run npx skills list, then check /skills in Codex or the skill list in Claude Code. Reinstall with the frontmatter name and an explicit -a target. For a plugin skill, invoke it by its namespaced name once to confirm it loads.

Two design skills fight. The agent mixes a taste-skill ban list with a ui-ux-pro-max palette and produces neither. Recover: remove all but one design skill (npx skills remove <name>) and put the chosen design system file path in CLAUDE.md or AGENTS.md so every session starts from it.

The skill teaches the wrong framework version. Next.js guidance from an old skill copy contradicts your installed version. Recover: npx skills update for skill copies; for Next.js, prefer the AGENTS.md that next dev generates, because it is matched to your installed version.

The review passes but users still hit problems. web-design-guidelines reads code; it cannot see focus order at runtime, colour contrast over images, or a slow third-party script. Recover: add the Chrome DevTools trace and a Playwright plus axe test for the page, and treat the static review as one gate of three.

The agent “fixes” findings by deleting features. Asked to clear the detector, it removes the animation or the form. Recover: say in the prompt that the visual design and behaviour must not change, require before/after counts, and review the diff summary for removed components rather than every line.

Mobile and web advice collide. A monorepo with a Next.js app and an Expo app gets React Native advice in web files. Recover: install web skills in the web package and mobile skills in the mobile package (project scope per directory), not at the monorepo root.

For vetting a skill’s scripts and hooks before a team rollout, see skill supply-chain security.