Best Claude Skills, Ranked by Verdict

Claude skills are folders of instructions, scripts and resources that Claude loads when a task matches the skill's description. Among 10 Claude skills we ranked by verdict, Anthropic's Frontend Design leads with 76, Recommended. Here, Claude skills means Agent Skills, the SKILL.md packages Claude loads, not Claude's general abilities. Each row gives the safety grade and the Claude surface where public evidence shows the skill running. Our explainer covers how Claude picks a skill.

Two more Claude skills sit just below the top 10. Both are Conditional: Web Design Guidelines (66) and Beautiful Prose (62). Our library lists every skill we have reviewed.

What do Claude skills actually do?

Claude skills are folders of instructions, scripts and resources that Claude loads when a task matches. They teach Claude a repeatable procedure. Only a short description stays in context until the skill is needed.

An Agent Skill is a folder with a SKILL.md file at its root. The file opens with YAML frontmatter that holds a name and a description. Scripts and reference files are optional extras. Claude loads skills in the Claude app, the Claude API and Claude Code. The app covers claude.ai, Desktop and Cowork, and Claude Code is Anthropic's coding agent. Anthropic's Agent Skills overview describes three loading levels:

Claude's Help Center names four types of skills:

• Anthropic skills, such as the PowerPoint, Excel, Word and PDF skills.
• Custom skills, which you or your team write.
• Organization-provisioned skills, which admins provision for members.
• Partner skills, which partners such as Notion, Figma and Atlassian publish.

The format is an open standard documented at agentskills.io, and Claude Code "extends the Agent Skills open standard." Our guide has Claude skills explained in depth. Other agents read the same format, as our primer on Agent Skills across agents explains.

What is a Claude skill example?

A Claude skill example is Anthropic's Frontend Design, one 71-line SKILL.md that steers Claude away from stock page designs. Three ranked skills show the range, with token counts estimated from file size:

• Frontend Design holds one 2,348-token SKILL.md plus a licence file, with no scripts.
• PPTX pairs a 5,199-token SKILL.md with 15 Python scripts that validate and render decks.
• Humanizer holds one 8,087-token SKILL.md listing 26 AI-writing patterns with before-and-after examples.

What are the official Claude skills?

The official Claude skills are Anthropic's four pre-built document skills: PowerPoint, Excel, Word and PDF. Anthropic publishes more open-source skills, including the Claude API skill bundled with Claude Code. Our verdicts score them 66 to 72.

Anthropic's platform docs list the pre-built skills as pptx, xlsx, docx and pdf. They run on claude.ai, the Claude API, Claude Platform on AWS and Microsoft Foundry. On claude.ai, "These Skills are active when you create documents. Claude uses them with no setup required."

Per the platform overview, pre-built document skills "are not available in Claude Code." The Claude API skill "comes bundled with it." The Claude Code skills docs list /claude-api among bundled skills. Those docs differ on document skills: signed-in sessions load built-in skills "such as pdf and xlsx."

Our document verdicts reviewed the open-source folders in anthropics/skills at commit 8a1541c. Anthropic does not document whether the hosted pre-built skills match those files.

Can you use Claude skills for free?

Yes, you can use Claude skills for free: Claude's Help Center lists them on the Free plan. Anthropic's platform docs list custom skill uploads on paid plans only. Both require code execution.

The two Anthropic sources conflict on custom uploads for Free users:

The Help Center states the requirement plainly: "This feature requires code execution to be enabled." All 10 ranked skills are free to copy from public GitHub repositories. Humanizer, Find Skills, Claude SEO and Supabase Postgres use the MIT licence. The anthropics/skills repository shows no repository-level licence. Our guide shows how to turn on skills on each Claude plan.

Which Claude skills should you add to Claude?

Add Frontend Design when you build new pages, Humanizer for AI-sounding drafts and PPTX for decks. Add Claude API for Anthropic SDK code. Each Claude skill here has a verified Claude run and a published verdict.

1. Frontend Design, for a new landing page or prototype, verified in Claude Code. First test: run three no-direction briefs with and without it, then count pill buttons.
2. Humanizer, for an 800-word blog draft, verified in Claude Code, Cowork on the web and Claude Desktop. First test: feed it three fact-dense drafts, then diff every number and name.
3. PPTX, for a deck from notes or a company template, verified in claude.ai and Claude Code. First test: fill a 12-slide template, then check thumbnail labels against the real slide order.
4. XLSX, for a formula-driven budget model, pre-built in claude.ai and verified in Claude Code. First test: edit a chart-bearing template, run recalc.py, then open it in desktop Excel.
5. Claude API, for Anthropic SDK code, bundled with Claude Code and verified there. First test: request a Python streaming call with one tool, then run it with your own key.

How were these Claude skills ranked?

These Claude skills are ranked by verdict score, weighted toward each skill's worst failure.

The score combines 7 universal measures with 4 to 8 job-specific ones. Verified means dated public evidence shows a real run on that surface. Claimed means only the publisher says so. Our methodology page shows how we score and grade skills.

What changed since our review?

Three ranked Claude skills changed upstream after our 2026-09-30 review, per GitHub data checked on 2026-10-21. Verdicts stay as reviewed on 2026-09-30.

• Claude API: two commits to skills/claude-api landed on 2026-10-09. Commit 9d63080 makes Claude Haiku 5.5 the current Haiku and covers Managed Agents. Commit dbd4588 adds dynamic workflows and four workflow quickstarts (#1995). We reviewed commit 8a1541c.
• Claude SEO: release v2.4.2 shipped on 2026-10-04, and that day's commits edit only the README. We reviewed v2.4.1 (ff87fce).
• Find Skills: vercel-labs/skills released CLI v1.7.4 on 2026-10-19. The skills/find-skills folder last changed on 2026-07-10, so the reviewed SKILL.md (v1.7.0, 3694740) is unchanged.

A re-review that moves a score appears among the verdict changes we log.

How does each top Claude skill hold up?

Each of the 10 top Claude skills gets a short card below. Every card states the verdict, safety grade and verified Claude surface. It names the main limitation found at the reviewed commit.

1. Frontend Design: 76, Recommended

Frontend Design gives new pages a deliberate visual direction from a loose brief. Tessl, a third-party skill registry, measured a 1.27x lift on Sonnet 4.6 across 3 scenarios. It carries safety grade A and is verified only in Claude Code. Element defaults slip through: pill buttons appeared on 10 of 21 Opus 5.5 pages (#1808). Alternatives rank among Claude skills for UI and frontend design.

2. Claude API: 72, Recommended with caveats

Claude API keeps Anthropic SDK code on current model IDs, prices and API shapes. All 13 current model IDs and the 12-row price table match Anthropic's docs. It ships bundled with Claude Code, where it is verified, with safety grade B. Four retired models still read as usable in models.md (#1603). The frontmatter breaks strict loaders (#1622). Check our Claude API verdict before trusting its legacy IDs.

3. Humanizer: 72, Recommended with caveats

Humanizer rewrites AI-drafted posts, READMEs and PR text so they read as human. Blind judges preferred its rewrite 16 of 16 times (#229). A GPT-5.4 replay on 18 drafts cut a slop score from 76.3 to 35.4 (#250). It is verified in Claude Code, Cowork on the web and Claude Desktop, with grade B. Three worked examples add or change facts, and both fix requests (#233, #306) closed unchanged. Compare other writing skills ranked by verdict.

4. PPTX: 71, Recommended with caveats

PPTX builds and edits PowerPoint decks, including company templates. It is verified in claude.ai and Claude Code. Tessl measured a 1.53x uplift with Sonnet 4.6 across 3 scenarios. Safety grade B reflects its 15 local Python scripts. Thumbnail grids mislabel slides on decks of 10 or more (#1844). Its validator can pass files python-pptx cannot open (#1862). The full defect list sits in the PPTX verdict.

5. Find Skills: 70, Recommended with caveats

Find Skills has the agent search skills.sh, a public skill catalogue, and suggest installs. It is one code-free file of 1,368 tokens, with safety grade A. Tessl measured a 2.15x lift on Claude Sonnet 4.6 across 2 scenarios. It is verified in Claude Code and Codex. The skills.sh search index misses indexed skills (#1688). Read the Find Skills verdict before letting it install anything.

6. XLSX: 70, Recommended with caveats

XLSX builds formula-driven workbooks, then recalculates every formula through LibreOffice. It is pre-built in claude.ai and verified in Claude Code, with safety grade B. Tessl measured a 1.44x uplift with Sonnet 4.6 across 2 scenarios. Its recalc.py script crashes on Windows (#1120). Claude Code argument substitution corrupts its $B$6-style examples (#1779). The repair cases are in the XLSX verdict.

7. Claude SEO: 70, Recommended with caveats

Claude SEO is a broad, scripted SEO audit plugin for Claude Code. Its maintainer fixes most reported bugs within days, in tagged releases. Safety grade C reflects 169 script files and a PostToolUse hook. That hook fires on every edit in every project (#321). It claims claude.ai hosted plugins but is verified only in Claude Code. Learn what a C safety grade means before installing.

8. PDF: 68, Conditional

PDF gives copy-paste recipes for everyday PDF reading, extraction, merging and generation. It is pre-built in claude.ai and verified in Claude Code, with safety grade B. Its form helpers mishandle string choice options and rotated pages. One validator prints FAILURE yet exits 0 (#1770). Test one real form first, as the PDF verdict advises.

9. Doc Co-Authoring: 67, Conditional

Doc Co-Authoring runs an interview-driven workflow for specs, RFCs and decision docs. It is one instruction file with safety grade A, verified in Claude Code. Unattended runs produce nothing: in 30 Haiku 4.5 runs, the 15 that engaged wrote no file (#1718). Fix PR #1759 is unmerged. It belongs among Claude skills for drafting documents when a person answers its questions.

10. Supabase Postgres Best Practices: 66, Conditional

Supabase Postgres Best Practices loads compact wrong-vs-right SQL patterns on demand. Its 808-token entry file links 31 focused rule files, with safety grade A. It is verified on Claude Sonnet 4.6 in a Tessl eval and in Claude Code user reports. Its SECURITY DEFINER helper example breaks RLS policy evaluation (#390, #499). Review every policy it writes, as the Supabase Postgres verdict explains.

Which Claude skills work in the Claude app, the API and Claude Code?

Only Humanizer and PPTX have a verified run in the Claude app. Every ranked Claude skill has a verified Claude Code run. Anthropic's docs say custom skills do not sync across surfaces.

Anthropic's platform docs set different runtime limits per surface:

• On claude.ai, "Depending on user/admin settings, Skills may have full, partial, or no network access."
• On the Claude API, skills get "No network access" and "No runtime package installation."
• In Claude Code, skills get "Full network access."

The sync rules conflict. The platform overview says "Custom Skills do not sync across surfaces." The Claude Code sync docs disagree for signed-in sessions: "Claude Code loads the skills enabled for your claude.ai account." Terminal sync needs "Claude Code v2.1.273 or later," and the Help Center repeats that version. Skills verified only in Claude Code appear among Claude Code skills, ranked separately.

Which Claude skills should you avoid?

Avoid Landing Page Design (50, Not Recommended), which tells Claude to invent statistics. Use Skill Creator (57) with care before you run it. Its evaluation scripts report false results and run Claude sessions with your permissions.

• Landing Page Design (50, Not Recommended, grade A) claims "Claude.ai Projects" support. Its rules ask for organic-looking figures such as 47.2% and invented brands. PR #1 adds a sourcing step but sits unmerged and unanswered, per our verdict on Landing Page Design.
• Skill Creator (57, Conditional, grade C) has a description optimiser that reports false negatives (#1721, #1352, #1412). It counts crashed runs as passes (#1478). Its scripts spawn claude -p sessions with your permissions. Read our verdict on Skill Creator, or build and test a skill yourself.
• Web Artifacts Builder installs dependencies at latest with no lockfile, so pnpm 10.1+ aborts bundling (#1362).
• MCP Builder has an evaluation harness that fails at import on mcp 2.x (#1668).

Anthropic's platform docs advise: "Use Skills only from trusted sources: those you created yourself or obtained from Anthropic." Our checklist shows how to vet a Claude skill before you upload it.

What are the best Claude skills on GitHub?

The best Claude skills on GitHub sit mostly in anthropics/skills, with about 179,000 stars. It holds six of our 10 ranked Claude skills. Stars do not predict verdicts: its eight verdicts here range from 57 to 76.

ComposioHQ/awesome-claude-skills, a curated list with 76,196 stars at review, holds one skill we reviewed. Our Content Research Writer verdict gives it 58, Conditional. Claude Code reads custom skills from ~/.claude/skills/ for personal use or .claude/skills/ in a project. Our guide covers how to upload a skill from GitHub to Claude.

Where can you find more Claude skills?

Find more Claude skills in four places: Claude's Customize menu, Anthropic's GitHub repository, Claude Code plugin marketplaces and directories. None of these listings judges quality. Our library lists only skills with a verdict.

• Customize > Skills in Claude is where you enable example skills and upload your own, per the Help Center.
• GitHub's anthropics/skills repository holds Anthropic's open-source skills.
• Claude Code plugin marketplaces start with Anthropic's official anthropics/claude-plugins-official, installed with /plugin install <name>@<marketplace-name>.
• Directories such as claudeskills.info list skills without running them: "Directory listing is not evidence we have run the skill."

Listings sort by installs or stars, while our library sorts by verdict. You can browse every Claude skill we have judged.

Which Claude skills suit developers and coding?

For developers, the Claude skills that suit coding run in Claude Code. Claude API (72) keeps Anthropic SDK code current, and Supabase Postgres (66) adds safe SQL patterns. Coding-only skills rank on our Claude Code hub.

Claude API covers Claude API and Anthropic SDK code in eight languages. Supabase Postgres Best Practices guides schema, index and RLS work, with one broken helper example to fix. Find Skills (70) searches skills.sh from the terminal. Designer Skills Collection (78) and Brainstorming (76) lead the hub for coding skills verified in Claude Code. The cross-agent ranking covers the best agent skills for coding.

What does the community report about Claude skills?

The community reports on Claude skills mostly through Hacker News threads and GitHub issues. Our verdicts cite them beside each score. Humanizer draws both praise and criticism, while Frontend Design draws two critical comments.

Common questions

Do Claude skills sync between claude.ai, the API and Claude Code?

No, per Anthropic's platform docs: "Custom Skills do not sync across surfaces." Claude Code docs differ: signed-in sessions on v2.1.273 or later load your claude.ai skills.

What is a skill in Claude Code?

A Claude Code skill is a SKILL.md folder that Claude Code loads from ~/.claude/skills/ or a project's .claude/skills/. Our guide shows how to set up skills in Claude Code.

How do you create a Claude skill?

Write a SKILL.md file with a YAML name and description, then add instructions, per the SKILL.md format. Anthropic's Skill Creator (57, Conditional) drafts skills for you, though its trigger evals report false results.

Are Claude skills safe?

Not all of them: 1 of our 10 ranked Claude skills, Claude SEO, holds safety grade C. Run every skill through our vetting checklist before any upload.

Do Claude skills work with connectors and MCP?

Yes: Claude's Help Center says MCP connects Claude to external services, while skills teach Claude how to use them. Our guide explains when to use a skill or an MCP server.

Is there a Claude skills course?

Yes: Anthropic's academy.claude.com covers skills in its Claude Code 101 course. Anthropic publishes a complete guide to building skills for Claude as a PDF.

How often is this list updated?

This list changes when a re-review moves a verdict, and our corrections log records each change. Every verdict here dates from a 2026-09-30 review at a pinned commit.