Archify
tt-a1i/archifyA skill that turns a description or a codebase into a validated, interactive standalone HTML diagram or one-page explainer.
Recommended
Check first Depth 5/5
What are the best Claude skills?
The best Claude skills are the ones that fit a job you do often, ship real expertise (reference files, scripts, tests) and ask for no more access than they need. A skill with 300,000 stars that wants broad shell permissions is a worse pick than a 5,000-star skill with a clean security status. Below are our picks by job, all taken from the skills we reviewed and all with a Recommended verdict.
How we pick the best Claude skills
We do not rank by stars or install counts. Both are easy to inflate and both reward being early, not being good. Three things decide a place on this list:
- Depth, 1 to 5. How much expertise ships in the repository. A one-page prompt is a 1; references, scripts and evals you can leave running unattended make a 5. See methodology.
- Verdict. Only skills we mark Recommended are eligible. Of 195 repositories we reviewed, 84 qualify and 109 are listed With caveats.
- Security status. No issues found, Review before installing or Broad access, computed from facts we record file by file. A Broad access skill can still be the best tool for the job, but it does not get a pick without a warning attached. The security page lists every flag.
Stars still appear below as context, as recorded on 2026-10-29. For the live ranking of the deepest Recommended skills, use the table at the top of this page; the lists here are the shorter, opinionated version.
Best Claude Code skills by job
Planning, TDD and debugging
Engineering process is where most people start, and where overlap bites hardest. Pick one primary workflow and add single skills around it.
- Superpowers: 15 skills that make the agent clarify the goal, write a plan and review its own work. Depth 3, security status Review before installing. Expect more questions up front and fewer half-finished results.
- addyosmani/agent-skills: 25 skills and 9 slash commands covering spec, plan, build, test, review and ship. Depth 3. It overlaps with Superpowers, so choose one.
- Matt Pocock’s skills: small, composable and easy to read in full, though depth per skill is modest. The tdd and diagnosing-bugs skills both have a clean security status and a depth of 3.
- planning-with-files: a planning skill for multi-hour tasks that survive context resets. Depth 4, but expect hooks that run on every prompt and tool call, so it carries a Broad access label.
More in planning, TDD and debugging and code review.
Documents and building your own skills
- anthropics/skills: Anthropic’s reference collection, depth 4. The pptx and docx skills are battle-tested, but their license is proprietary, so use them as a user, not a redistributor.
- skill-creator: the most complete official tool for writing and testing new skills (depth 5). It runs local scripts and starts claude subprocesses, so run it in a project you trust.
See documents and skill and agent tooling.
Frontend and design
Design is the most crowded category, with more than 50 entries, so depth matters most here.
- ui-ux-pro-max: searchable style, palette and font catalogs with a design-system generator. Depth 4. Only the optional image-generation scripts need third-party API keys.
- Taste Skill: an art-direction skill with open-licensed photos, working scripts and evals. Depth 4; budget real time and tokens per page.
- Hallmark: a depth-5 design skill with a real linter for AI-looking pages. The opinionated rules are the thing to check against your own taste.
- frontend-design: a one-page design brief from Anthropic. Depth 2, clean security status, and cheap on context.
- vercel-react-best-practices: 70 React and Next.js performance rules with before and after examples. Clean, no scripts, nothing to configure.
More in design and frontend.
Security audits and skill scanning
- Cloudflare’s security-audit skill: a multi-agent source audit in six phases with machine-readable findings. Depth 5; full runs cost a lot of tokens.
- Trail of Bits skills: expert-written audit skills, 97 of them. Install the plugins for your stack, not all of them. Depth 4 with a Broad access label.
- NVIDIA SkillSpector and Cisco skill-scanner: scanners for checking other people’s skills before you install them. Treat the score as a lead and read the findings.
More in security.
Marketing and SEO
- marketingskills: 52 skills for CRO, copywriting, SEO, analytics and paid ads, depth 4. Read the partner disclosure and keep ad-account access read-only. The seo-audit skill inside it has a clean security status.
- claude-seo: a 35-skill SEO suite that works without API keys. Review the installer first.
Browsers, research, codebases and cloud
- agent-browser: drives a real browser from the shell. The main risk is that it can hold your logins and cookies.
- last30days-skill: depth-5 research across Reddit, X, YouTube and Hacker News. Decide which sources and keys you are comfortable enabling.
- understand-anything: maps a codebase into an interactive knowledge graph. Plan for the token cost of the first full run.
- supabase-postgres-best-practices: a compact Postgres rulebook with a clean security status.
- Microsoft’s Azure skills: depth 4 and well maintained, but the agent acts on live resources, so use a least-privilege Azure identity.
Awesome Claude skills lists: how to use them
“Awesome” lists are discovery tools, not endorsements. ComposioHQ/awesome-claude-skills has 78,228 stars, and we rate it With caveats at depth 1: the list is useful, but the hundreds of bundled skills are templated wrappers that route through Composio’s hosted service. sickn33/agentic-awesome-skills catalogs roughly 2,900 community playbooks of uneven quality. Use lists to find names, then look the name up here before installing. Our GitHub repositories guide covers more of them.
Large skill packs and when to skip them
Some of the biggest repositories work better as a source of single skills than as a full install. Everything Claude Code has 292,012 stars, and its README counts 311 skills and 71 agents. We rate it With caveats because it is a sprawling harness with hooks and overlapping skills, so a minimal profile is the sensible route. The same logic applies to any pack: install the three skills you will use, not the three hundred you might.
How to try a skill safely
- Open the skill’s page here and read the verdict line and security reasons.
- Read its
SKILL.mdand any scripts once. It takes minutes. - Install it in one project first, not globally.
- Run the same task with and without it. The Claude Code docs suggest this kind of baseline comparison for judging whether a skill helps.
Install steps are in how to install Claude skills, and the full list is at /skills/.