heydeer

Best Claude Code skills (2026): 12 worth installing

· · 13 min read

The best Claude Code skills come down to about a dozen, out of thousands published. This list says what each one does, when it earns its place, how to install it and what to watch.

The best Claude Code skills in 2026 are Anthropic’s skill-creator, frontend-design, webapp-testing and mcp-builder, Jesse Vincent’s Superpowers, Matt Pocock’s engineering skills, and Trail of Bits’ security plugins, plus task skills such as agent-browser and ci-audit. Install a few that match work you repeat, read each SKILL.md and its scripts first, and keep the list short so Claude picks the right one.

What Claude Code skills are and how Claude loads them

A skill is a folder with a SKILL.md file in it. The file opens with YAML frontmatter that names the skill and describes when to use it, and the Markdown below holds the instructions. Anything else in the folder, such as scripts, templates or long reference documents, is there for Claude to run or read when the instructions point to it. This is the example from Claude Code’s own documentation:

---
name: summarize-changes
description: Summarizes uncommitted changes and flags anything risky. Use when the user asks what changed, wants a commit message, or asks to review their diff.
---

## Current changes

!`git diff HEAD`

## Instructions

Summarize the changes above in two or three bullet points,
then list any risks you notice.

The loading model is what makes skills cheap to keep around. In a normal session, Claude sees only each skill’s name and description. The body loads when you type /summarize-changes or when Claude decides a request matches the description, and from then on it stays in the conversation. The line starting with ! runs before Claude reads the skill, so the current diff arrives already inlined.

Where the folder sits decides who gets the skill:

  • ~/.claude/skills/<name>/ is personal and loads in every project on your machine.
  • .claude/skills/<name>/ in a repository is a project skill. Commit it and your team gets it too. When a personal and a project skill share a name, the personal one wins.
  • A plugin carries skills under its own skills/ folder, and they run namespaced, as /plugin-name:skill-name.

Claude Code also ships bundled skills, including /code-review, /debug, /loop and /doctor, so a fresh install already reviews diffs. Our guide to the Claude Code review skill covers that one. Claude skills aren’t limited to the terminal either: the format is the open Agent Skills standard, which Claude.ai, the Claude API, Codex and Cursor read as well. If you also use those tools, see Codex skills and Cursor skills.

The best Claude Code skills in 2026

The picks below favor skills that change the result of everyday work, come from a source you can name, and have a repository you can read. They’re ordered by how many developers will get value from them, not by install counts. Install commands use either Claude Code’s /plugin command or the open-source skills CLI; the npx skills guide explains the CLI’s flags and where it puts files.

SkillFromBest forInstall with
skill-creatorAnthropicWriting and testing your own skillsPlugin or skills CLI
frontend-designAnthropicUI that doesn’t look templatedPlugin or skills CLI
Superpowersobra (Jesse Vincent)A full plan, TDD and review workflowPlugin
grill-me, tdd and friendsmattpocockSmall, composable engineering habitsPlugin or skills CLI
webapp-testingAnthropicChecking a local web app with PlaywrightPlugin or skills CLI
agent-browserVercel LabsDriving a real browser from the shellnpm + skills CLI
mcp-builderAnthropicBuilding MCP serversPlugin or skills CLI
docx, pdf, pptx, xlsxAnthropicReading and producing office filesPlugin
Security pluginsTrail of BitsSecurity review of diffs and codebasesPlugin
react-best-practices, web-design-guidelinesVercel LabsReact, Next.js and UI auditsSkills CLI
ci-auditHeyDeerSlow, expensive GitHub ActionsSkills CLI
find-skillsVercel LabsFinding the next skillSkills CLI

1. skill-creator (Anthropic)

The skill that makes the rest of this list less necessary. skill-creator interviews you about a procedure, drafts the SKILL.md, then tests it: it runs your prompts with and without the skill in clean subagents, grades the output against assertions you write, and tunes the description so the skill triggers on the right requests. The first skill most teams should write is one about their own codebase, and this is the tool for it.

/plugin install skill-creator@claude-plugins-official

Watch for: an eval loop spends tokens and time. Use it on skills you’ll share, not on a three-line personal shortcut.

2. frontend-design (Anthropic)

Guidance for visual design when Claude builds or reshapes a UI: choosing an aesthetic direction, typography and layout that don’t read as default templates. It’s the most installed Anthropic skill on the skills.sh leaderboard, and the difference shows on the first landing page or dashboard Claude produces with it.

/plugin install frontend-design@claude-plugins-official

Watch for: it pushes toward distinctive design. In a product with an established design system, your own components and tokens should win, so say so in the request or in CLAUDE.md.

3. Superpowers (obra)

Less a skill than a working method built from more than a dozen of them. Superpowers makes Claude ask what you’re actually building before it writes code, turn the answer into a spec and a step-by-step plan, then carry the plan out with subagents, red-green TDD and a review between tasks. Skills such as brainstorming, writing-plans, systematic-debugging and verification-before-completion trigger on their own.

/plugin install superpowers@claude-plugins-official

Watch for: it owns the process. For a one-line fix the brainstorming step is overhead, and it works best when you let it run its full loop on real features.

4. Matt Pocock’s skills: grill-me, tdd and more

The opposite philosophy to Superpowers: small skills you compose yourself. /grill-me has Claude interview you about a plan until every branch is decided, /grill-with-docs does the same while building a project glossary and decision records, tdd enforces a red-green-refactor loop, and diagnosing-bugs works a hard bug through reproduce, minimise, hypothesise and fix. The grilling skills are among the most installed on skills.sh.

claude plugin install mattpocock-skills@claude-plugins-official

Watch for: run /setup-matt-pocock-skills once per repository to set your issue tracker and labels. The collection includes its own code-review skill; installed into .claude/skills/ through the skills CLI, it replaces Claude Code’s bundled /code-review.

5. webapp-testing (Anthropic)

Lets Claude check a local web app instead of assuming its change works. Claude writes Python Playwright scripts, and a bundled with_server.py starts and stops your dev servers around them, so it can click through a flow, read browser logs and take screenshots. Useful any time “it compiles” isn’t the same as “it works”.

npx skills add anthropics/skills --skill webapp-testing -a claude-code

Watch for: Python and Playwright need to be installed. Claude Code’s bundled /run and /verify skills cover some of the same ground without setup, so try those first.

6. agent-browser (Vercel Labs)

A Rust CLI that gives Claude a browser it drives one shell command at a time: snapshot returns the page as text with a ref for each element, and Claude clicks or types by ref, then reads the page again. The skill itself is a short stub that tells Claude to load instructions from the installed CLI, so the two never drift apart.

npm install -g agent-browser && agent-browser install
npx skills add vercel-labs/agent-browser

Watch for: it downloads Chrome for Testing on first install. Pick it or webapp-testing, not both, so Claude has one way to open a browser.

7. mcp-builder (Anthropic)

A guide for building MCP servers in Python (FastMCP) or TypeScript (the MCP SDK), with reference material on tool design and scripts for evaluating the result. When you need Claude to reach an internal API, this gets a server with well-shaped tools instead of a thin wrapper around every endpoint.

npx skills add anthropics/skills --skill mcp-builder -a claude-code

Watch for: it’s only useful while you’re building a server. Install it in that project rather than globally.

8. docx, pdf, pptx and xlsx (Anthropic)

The document skills behind Claude’s file features in the Claude apps, published as open source. They read, edit and create Word, PDF, PowerPoint and Excel files with bundled scripts: filling PDF forms, merging and splitting PDFs, building spreadsheets with working formulas. In Claude Code they matter when a repository produces reports or ingests documents. Both of Anthropic’s plugin sets come from the same marketplace:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

Watch for: Anthropic labels the repository as demonstration and reference material, and the behavior in Claude.ai can differ from these copies. Test on your own files before relying on them.

9. Trail of Bits security plugins

A marketplace of security skills from a firm that does security audits for a living. Developers will get the most from differential-review, a security review of a diff that uses git history to spot regressions and estimates blast radius by counting callers, plus static-analysis for CodeQL and Semgrep, and supply-chain-risk-auditor for npm, PyPI and Go dependencies.

/plugin marketplace add trailofbits/skills
/plugin install differential-review@trailofbits

Watch for: install individual plugins, not the whole marketplace’s worth. Many target smart contracts, C and Rust, and the static analysis plugin needs CodeQL or Semgrep on your machine.

10. react-best-practices and web-design-guidelines (Vercel Labs)

Two rule sets for frontend work. react-best-practices holds Vercel Engineering’s React and Next.js performance rules, ordered by impact from request waterfalls and bundle size down to micro-optimizations. web-design-guidelines audits UI code against more than 100 rules for accessibility, focus states, forms, animation and typography.

npx skills add vercel-labs/agent-skills --skill react-best-practices --skill web-design-guidelines

Watch for: where a rule conflicts with your lint config or house style, the lint config should win, so note the exceptions in CLAUDE.md.

11. ci-audit (HeyDeer)

Coding agents push far more commits than people do, and every push reruns CI. ci-audit reads your last two weeks of GitHub Actions runs with a read-only script and opens with two numbers: what you could save each month keeping your runners, and what you could save if you switch them too. Behind those numbers it finds serial job chains, caches that silently miss, duplicate push and pull_request triggers, flaky jobs and macOS legs that never catch anything Linux missed. It changes nothing until you pick a fix, and each fix arrives as a small pull request.

npx skills@latest add heydeer/skills --skill ci-audit

Watch for: it’s built for GitHub Actions and needs Python 3.8 or later plus read access through the gh CLI or a GH_TOKEN. It runs only when you type /ci-audit. The ci-audit page lists everything it checks.

12. find-skills (Vercel Labs)

A skill for finding skills. When you ask “is there a skill for X”, it searches the open skills directory and suggests what to install. It tops the skills.sh leaderboard, and it’s a reasonable way to explore once you know what you need.

npx skills add vercel-labs/skills --skill find-skills

Watch for: a popular result isn’t a reviewed one. Treat each suggestion as a link to read, the same as any other skill on this page.

Claude skills examples: what the good ones share

Read a few of the SKILL.md files above and the same choices keep coming up. They’re worth copying when you write your own.

  • The description says when, not just what. skill-creator’s description lists the requests that should trigger it: create a skill, edit one, run evals, tune a description. Claude matches requests against this text, so vague descriptions mean a skill that never loads.
  • Side effects wait for you. ci-audit sets disable-model-invocation: true, so Claude can’t start a long audit because CI came up in passing. Deploys, commits and anything that posts somewhere deserve the same line.
  • Scripts run as black boxes. webapp-testing tells Claude to run its helper with --help instead of reading the source, which keeps a large script out of the context window.
  • Long reference material sits in separate files. The pdf skill keeps its form-filling guide in FORMS.md and loads it only for form work.

Skills are also the right home for procedures that have outgrown CLAUDE.md: the body loads only when used, while CLAUDE.md is read every session. The CLAUDE.md best practices guide covers what belongs on each side.

How to judge a skill before installing it

A skill is instructions plus code, from someone you probably haven’t met. Five minutes on its repository is cheap compared with what it can do once it’s loaded.

  1. Read the SKILL.md. It’s usually short. Check that it does what the README claims and nothing more, such as quietly sending data somewhere or editing files outside the task.
  2. Read every script it calls. Scripts run on your machine with your permissions. Look for network calls, credential reads and writes outside the project.
  3. Check allowed-tools and ! lines. allowed-tools pre-approves tools for the turn that invokes the skill, and !`command` lines run before Claude even reads the skill. Claude Code applies a project skill’s allowed-tools even in a -p run in a folder you’ve never trusted, so review skills checked into a repository before running Claude Code there.
  4. For plugins, look past the skills. A plugin can add hooks, background monitors and MCP servers that run as processes with your user privileges, outside Claude Code’s permission rules. Read hooks/hooks.json, .mcp.json and bin/, or run claude plugin details <name> after installing to list them.
  5. Check names. A personal or project skill with the same name as a bundled one replaces it. Two community skills called code-review can’t both be /code-review.

skills.sh shows security audit results from Gen, Socket and Snyk for public skills, and the skills CLI prints them before it asks you to confirm an install. They’re a useful first filter, not a substitute for reading the files. Vercel, which runs the site, says plainly that it can’t guarantee the security of every listed skill.

Skills also change after you install them. Plugins from a marketplace with auto-update turned on update in the background, and a repository you install from can change tomorrow. For skills your team depends on, commit a copy into .claude/skills/ so changes reach you as reviewable diffs.

The Claude Code skills marketplace and plugins

Claude Code has no separate store for skills. Skills are distributed in plugins, and plugins come from marketplaces: a marketplace is any repository with a .claude-plugin/marketplace.json catalog. Anthropic runs two general ones.

  • claude-plugins-official is added automatically the first time you start an interactive session. It holds Anthropic’s own plugins, such as skill-creator, frontend-design and code-review, plus partner plugins like Superpowers.
  • claude-community holds third-party plugins their authors submitted. Add it with /plugin marketplace add anthropics/claude-plugins-community.

Anthropic’s topic marketplaces, such as anthropics/skills, and vendor ones like trailofbits/skills are added the same way. Run /plugin and open the Discover tab to browse what your marketplaces offer, or search Claude Marketplace on the web, which shows install counts and marks some plugins as Anthropic verified.

Standalone skills, outside any plugin, are listed on skills.sh, which ranks them by installs through the skills CLI. The CLI installs the same SKILL.md folder into Claude Code, Codex, Cursor and dozens of other agents at once. Choose a plugin when you want automatic updates and a skill that stays namespaced; choose the CLI when you want plain, editable files in your repository. The npx skills guide walks through that route.

Keep your skill list short

Every installed skill costs something, even unused. Claude Code keeps every skill’s name and description in context so Claude knows what’s available, within a budget of about 1% of the context window. When the list overflows, Claude Code drops descriptions, starting with the skills you invoke least, and a skill without its description is one Claude can’t match to your request.

Run /skills and press t to sort by token count, or /skill-doctor to see how often each skill is actually used. Turn off what you haven’t touched in a month. Two skills that do the same job, such as two TDD workflows, are worse than one: Claude has to guess which you meant.

The best Claude Code skills are the few that match work you repeat. Start with skill-creator and one workflow collection, add domain skills such as security or frontend rules when a project needs them, read each SKILL.md and its scripts before installing, and prune with /skill-doctor so Claude keeps picking the right one.

The skills you write for your own codebase tend to beat any of the above, because they hold what no public skill knows: how your services authenticate, which directories are generated, what a safe migration looks like. They’re useful beyond your editor, too. HeyDeer reads Agent Skills from .agents/skills/ and follows the ones a repository check names during pull request review; a symlink makes an existing .claude/skills/ folder count.

HeyDeer is an AI code review agent for GitHub pull requests. It reviews every pull request against the whole repository, follows the procedures your skills describe for repository checks, and leaves approval of any pull request that changes skills, AGENTS.md or CLAUDE.md to a person, so a change to your agent’s instructions always gets human eyes.

Sources

  1. Claude Code Docs: Extend Claude with skills
  2. Claude Code Docs: Commands reference
  3. Claude Code Docs: Anthropic’s marketplaces
  4. Claude Code Docs: Plugin security and trust
  5. GitHub: anthropics/skills
  6. GitHub: anthropics/claude-plugins-official
  7. GitHub: obra/superpowers
  8. GitHub: mattpocock/skills
  9. GitHub: trailofbits/skills
  10. GitHub: vercel-labs/agent-skills
  11. GitHub: vercel-labs/agent-browser
  12. GitHub: vercel-labs/skills
  13. GitHub: heydeer/skills
  14. skills.sh: leaderboard and security audits
  15. Agent Skills specification

Frequently asked questions

What are the best Claude Code skills?

The best Claude Code skills for most developers are skill-creator and frontend-design from Anthropic, the Superpowers workflow, Matt Pocock’s grill-me and tdd skills, webapp-testing or agent-browser for checking UI changes in a browser, and mcp-builder for MCP servers. Add domain skills such as Trail of Bits’ security plugins or Vercel’s React rules when your work calls for them.

What is a Claude Code skill?

A Claude Code skill is a folder with a SKILL.md file: YAML frontmatter with a name and a description, then Markdown instructions, plus optional scripts and reference files. Claude sees each skill’s description, loads the full instructions when a task matches or when you type /skill-name, and runs the scripts the skill points to.

Where do Claude Code skills live?

Claude Code skills live in ~/.claude/skills/<name>/SKILL.md for personal skills that apply to every project, in .claude/skills/<name>/SKILL.md for project skills you commit to the repository, and inside plugins, where they run as /plugin-name:skill-name. Organizations can also deploy skills through managed settings.

Is there a Claude Code skills marketplace?

The Claude Code skills marketplace is the plugin system: Claude Code adds Anthropic’s official marketplace, claude-plugins-official, on first run, and you can add others such as anthropics/skills or trailofbits/skills with /plugin marketplace add. Claude Marketplace on the web lists plugins with install counts, and skills.sh ranks standalone skills by installs from the skills CLI.

Are Claude skills safe to install?

Claude skills are as safe as the files in them. A skill’s instructions steer what Claude does, its scripts run on your machine, and a plugin can add hooks and MCP servers that run with your user permissions. Read SKILL.md, the scripts it calls and any allowed-tools line before installing, and prefer sources you can name.

Do Claude skills work outside Claude Code?

Claude skills follow the open Agent Skills format, so the same SKILL.md folder works in Claude.ai, the Claude API and other coding agents such as Codex and Cursor. Claude Code adds its own frontmatter fields, such as disable-model-invocation and allowed-tools, which other tools may ignore.

How many Claude Code skills should I install?

Install as few Claude Code skills as cover the work you repeat. Claude Code lists every skill’s name and description in context, within a budget of about 1% of the context window, and drops descriptions of the least-used skills when the list overflows. Run /skills or /skill-doctor to see what each one costs and turn off the ones you never use.

Is there a Claude Code code review skill?

Claude Code ships a bundled /code-review skill, also available as /review, that reviews the current diff, a branch or a pull request for correctness bugs. Installing another skill named code-review in your personal or project folder replaces the bundled one, so check names before installing review skills from community collections.

Start with your first PR.

Free plan. No credit card required.

Get started free