Guides · Updated November 21, 2026

Best agent skills for React and Next.js, tested

By Tineessa Nelson, Use This Skill

Install the Vercel skill that matches your job. In proofs run three times without and three times with on Claude Opus 5.5, vercel-react-best-practices kept all 500 list items and used content-visibility, vercel-composition-patterns gave the component a state, actions and meta contract, vercel-react-view-transitions used React's <ViewTransition> and vercel-react-native-skills chose LegendList and expo-image, each in 3 of 3 runs and 0 of 3 without.

Which React and Next.js skills have been run side by side?

The Use This Skill directory lists 33 skills that name React, React Native or Next.js in their name or description (crawl of November 16, 2026), and all 33 were clear in the scan that day. Four of them have published proofs: Claude Opus 5.5 did one fixed brief three times without the skill and three times with it, and all six pages of each proof are on the skill's page.

Skill Job Brief Proof date Scan (Nov 16)
vercel-react-best-practices Performance Rewrite a slow Next.js product list (sequential client fetches, 500 cards, full lodash, sort on every render) Oct 21, 2026 Clear, 76 files
vercel-composition-patterns Component APIs Refactor a Modal with fourteen props used by a dozen screens Oct 21, 2026 Clear, 14 files
vercel-react-view-transitions Animation A React photo gallery where photos expand into a detail view and the grid animates when filtered Oct 21, 2026 Clear, 9 files
vercel-react-native-skills Mobile Review and fix a React Native feed screen Nov 1, 2026 Clear, 42 files

All four come from vercel-labs/agent-skills under the MIT license. None of the proofs was commissioned by Vercel or anyone else.

What did vercel-react-best-practices change?

All six runs moved the fetching to the server, streamed the slow part behind Suspense and removed the full lodash import. The difference was in how they handled the 500 cards and what they cited.

What was measured (3 runs each) Without With
Kept all 500 cards and added content-visibility: auto 0 of 3 3 of 3
Cut the page to 24 products with server pagination 2 of 3 0 of 3
Added a virtualized list (@tanstack/react-virtual) 1 of 3 0 of 3
Added tag-based cache invalidation (revalidateTag) 3 of 3 0 of 3
Named the skill's rules by ID (such as async-parallel) 0 of 3 3 of 3 (14 to 15 rules each)

Without the skill, Claude changed what the page shows (fewer products per page) or added a library. With it, Claude kept the brief's 500 cards and used one CSS property, which is the skill's rendering-content-visibility rule; one page added that pagination or virtualization is for catalogs "well beyond 500 items." Every change on the "with" pages carries the rule ID it came from, so you can look each one up in the skill. Neither side is wrong on caching: the runs without the skill added a caching layer the brief did not ask for.

What did vercel-composition-patterns change?

All six runs replaced the fourteen props with small composed parts (Modal.Header, Modal.Footer and similar) and rewrote the three call sites. The skill changed where the state lives.

What was measured (3 runs each) Without With
Context defined as a contract of state, actions and meta 0 of 3 3 of 3
Providers that can be swapped while the parts stay the same 0 of 3 3 of 3
Written for React 19 (use() for context, ref as a plain prop) 0 of 3 3 of 3

Without the skill, a root component owned the state and the parts read it with useContext. With it, every page described the context as an interface any provider can implement, which is the skill's state-context-interface rule, and every page noted the React 19 changes from its react19-no-forwardref rule.

What did vercel-react-view-transitions change?

What was measured (3 runs each) Without With
Animated with React's <ViewTransition> and addTransitionType 0 of 3 3 of 3
Animated by measuring elements by hand (FLIP with getBoundingClientRect) 3 of 3 0 of 3
React version loaded from the CDN 18.3.1 canary
Respected prefers-reduced-motion 3 of 3 3 of 3

Without the skill, Claude wrote its own animation code on React 18. With it, Claude used React's built-in API and CSS ::view-transition rules. The "with" pages loaded React canary because the skill said so at the commit that was proven. On October 27, 2026 the skill's text changed: it now says ViewTransition and addTransitionType "became stable in React 19.3" and recommends stable React 19.3 or newer outside Next.js. React's reference page for <ViewTransition>, read on November 21, 2026, is published under react@19.3. If you install the skill today, expect the stable version.

What did vercel-react-native-skills change?

All six runs found the bug that matters most: the effect has no dependency array, so the screen fetches forever.

What was measured (3 runs each) Without With
List rebuilt with LegendList (@legendapp/list) 0 of 3 3 of 3
List rebuilt with FlatList 3 of 3 0 of 3
Images moved to expo-image 0 of 3 (1 suggests it as a next step) 3 of 3
contentInsetAdjustmentBehavior for the header and notch 0 of 3 3 of 3
Named the skill's rules by ID 0 of 3 2 of 3
Median output tokens 19,436 25,811
Median time per run 159 s 217 s
Input tokens per run 429 48,973

The skill's own rules say to use "a list virtualizer like LegendList or FlashList instead of ScrollView" and "expo-image instead of React Native's Image," and the "with" pages did both. This proof is the only one of the four with token and time data, and it is also the most expensive skill to load here: about 49,000 input tokens per run.

Counts from a script run on November 21, 2026 over the 24 published proof pages, checked by reading the matching passages. Token and time figures are from data/proofs.json; they are recorded only for proofs run from October 23, 2026. Rule names were matched against the skills' rule files at the proven commits.

Which React skill should I pick?

For visual design rather than code structure, read best frontend design skills. The Frontend category list has all 51 Frontend skills, proven ones first.

How were the proofs run and measured?

Each run sent the brief to Claude Opus 5.5 once, with the skill's SKILL.md and reference files loaded for the "with" side and nothing for the "without" side, and asked for a single HTML page. The runs had no tools, so the code was written, not executed. Every proof records the skill's exact commit; three of the four skills have not changed since they were proven, and the fourth changed in the one line described above (repeated in two files) (how to see what changed in a skill). The rules are on how a proof works, and every proof is on proven skills. Why three runs a side matters is explained in how to test whether a skill helps.

The script counts what each page's code and text contain: imports, APIs, CSS properties and rule names. It does not judge whether the code is correct or faster.

Are these React skills safe to install?

All four were Clear in the November 16, 2026 scan, which looks in every file of a skill's folder for patterns such as downloading and running code in one step, sending data out, reading credential files, or text that tries to override your instructions. In the Frontend category, 48 of 51 skills were clear, 2 need a manual look and 1 was flagged. The scan is a static check, not a guarantee; are agent skills safe? explains what it can and cannot catch. Before installing any skill, paste its GitHub link into the live skill check.

How do I install one in Claude Code?

Each skill page has a one-line install. For vercel-react-best-practices:

npx degit vercel-labs/agent-skills/skills/react-best-practices .claude/skills/vercel-react-best-practices

That puts the skill at .claude/skills/vercel-react-best-practices/SKILL.md, which Claude Code loads for that repository; ~/.claude/skills/ makes it available in all your projects. Claude uses a skill when your request matches its description, or you can invoke it directly with /vercel-react-best-practices.

Questions

Do I need all four Vercel React skills?

No. Each covers a different job: performance (vercel-react-best-practices), component APIs (vercel-composition-patterns), animation (vercel-react-view-transitions) and mobile (vercel-react-native-skills). Install the ones that match your work. Claude Code keeps each installed skill's description in context and loads the full text only when the skill is used, so an unused skill costs little, but a skill you never need is still one more description for Claude to weigh.

Does a React skill make Claude slower or use more tokens?

It can. In the React Native proof on November 1, 2026, each run with the skill read 48,973 input tokens instead of 429, the median output rose from 19,436 to 25,811 tokens and the median time from 159 to 217 seconds. The other three proofs ran before tokens and time were recorded. That is one brief on one model, not a rule.

Is there a skill for writing tests in a Next.js app?

Yes. nextjs-test-writer plans and writes tests for an App Router app, and its proof is compared with two test-first skills in the guide to the best skills for writing tests. It is sold by this site; its proof follows the same rules as every other proof.

Are these the most popular React skills?

They are among the most listed. This page ranks nothing by installs or stars; it reports what changed when each skill was loaded on a fixed brief. Popularity tells you many people installed a skill, not what it does to Claude's output.