Use This SkillChoose by the job, not the publisher. In the Use This Skill proofs, where Claude ran one brief three times without a skill and three times with it, Anthropic's own skills ranged from a clear change (internal-comms, mcp-builder) to little change (theme-factory, web-artifacts-builder), and community skills did the same. The directory lists 20 skill folders from Anthropic's repository and 2,040 from the community (crawl of November 2, 2026).
How do Anthropic's skills compare with community skills on this site?
The table compares the anthropics/skills repository with the 47 other repositories in the directory, from the weekly crawl and safety scan of November 2, 2026, and the proofs published on proven skills.
| anthropics/skills | Community (47 repositories) | |
|---|---|---|
| Skill folders in the directory | 20 (one is a blank template) | 2,040 |
| Scan: clear | 19 (95%) | 1,968 (96.5%) |
| Scan: review | 0 | 58 |
| Scan: flagged | 1 (claude-api) | 14 |
| Median text files per folder | 6.5 | 1 |
| Skills with a proof (3 runs without, 3 with) | 9 | 31 (28 still in the directory) |
| License recorded by the crawl | None repository-wide; the README says many skills are Apache 2.0 and the four document skills are source-available | MIT 709, Apache-2.0 344, other 100, none 887 |
Counts by script from data/skills.json and data/scan.json (crawl and scan of November 2, 2026) and data/proofs.json, run on November 7, 2026. Store skills sold on this site are not counted. No proof was commissioned by a skill author.
Twenty folders is too few to say one source scans cleaner than the other. Both sit near 95 to 97 percent clear. The bigger difference is shape: Anthropic's skills usually ship reference files, scripts or templates alongside SKILL.md, while more than half of the community folders (1,134) hold a single text file.
What did Anthropic's own skills change in the proofs?
Each proof used Claude Opus 5.5, one fixed brief, and the skill's files at commit dbd4588 on the "with" side. The brief differs for each skill, so compare each row with its own "without" runs.
| Skill | Brief | What changed with the skill (3 runs each side) | Proof date |
|---|---|---|---|
| internal-comms | Status update for leadership | Progress, Plans and Problems headings in 3 of 3 (0 of 3 without); median 190 words down to 123 | Oct 19 |
| mcp-builder | TypeScript MCP server for an issue tracker | A Markdown or JSON response_format option in 3 of 3 (0 of 3 without); a step that runs an evaluation script in 3 of 3 (0 of 3 without) |
Oct 21 |
| frontend-design | Landing page for a calendar app | Skip link in 2 of 3 (0 of 3 without); median ARIA attributes 18 to 36; median file size 56 KB to 37 KB | Oct 19 |
| algorithmic-art | Seeded flow-field art in p5.js | A written "algorithmic philosophy" in 3 of 3 (0 of 3 without); seeding and PNG export present on both sides | Oct 21 |
| canvas-design | Poster-style site for a jazz night | Median distinct media queries 3 to 7 | Oct 19 |
| brand-guidelines | Internal tool announcement | Applies Anthropic's own colors and type; visible focus styles in 1 of 3 (3 of 3 without) | Oct 19 |
| skill-creator | Write a SKILL.md for release notes | Little change on what was counted: all six had a "use when" description and an examples section | Oct 22 |
| theme-factory | Invitation site for a garden dinner | Little change; median CSS variables 19 to 15 | Oct 19 |
| web-artifacts-builder | Lisbon trip planner | Little change; median length 531 to 630 lines | Oct 21 |
Design rows from the counts in best frontend design skills; internal-comms from best writing and documentation skills; mcp-builder, algorithmic-art and skill-creator counted by script in the 18 published pages on November 7, 2026. A marker is not a verdict; every page of every proof is on the skill's page.
Two patterns stand out. The skills with a clear method, such as internal-comms with its 3P format and mcp-builder with its build-then-evaluate steps, changed the pages in ways you can count. The skills that mostly set a look, such as theme-factory and web-artifacts-builder, changed little on these markers, because Claude already produces a styled, working page without them.
Where do community skills do the same job?
For several jobs, a community skill has a proof on a similar brief. The briefs are not identical, so this is a guide to what each one does, not a head-to-head score.
| Job | Anthropic skill | Community skill with a proof | What the community proof showed |
|---|---|---|---|
| Landing pages and app screens | frontend-design | ui-ux-pro-max (nextlevelbuilder) | Skip link in 3 of 3 (0 of 3 without); median ARIA attributes 41 to 103 |
| Written updates and plans | internal-comms | writing-plans (obra/superpowers) | Commit steps named: 16 in each run (0 without); distinct file paths 31 to 40 (0 to 4 without) |
| Release notes | none proven | changelog-generator (ComposioHQ) | Internal work such as the Node upgrade left out in 3 of 3 (kept in 3 of 3 without) |
| Code review | none in the repository; Claude Code bundles /code-review | differential-review (trailofbits) | Test coverage, blast radius and a confidence level in 3 of 3 (0 of 3 without) |
From the proofs and counts published in the linked guides and skill pages, recorded October 19 to November 1, 2026.
Community skills also go quiet. web-design-guidelines (vercel-labs) showed little change on a build brief because it is written for reviewing a UI, and writing-guidelines made no visible difference without web access. Being published by a large company, Anthropic or otherwise, did not predict the size of the change.
Are Anthropic's skills safer to install?
Anthropic's Agent Skills overview advises using skills "only from trusted sources: those you created yourself or obtained from Anthropic," and says to audit every bundled file of a skill from anywhere else (checked November 7, 2026). That is sound advice, and the scan agrees that Anthropic's folders are clean: 19 of 20 clear.
The one flag shows why a flag needs reading. claude-api, the 96-file skill Anthropic also bundles with Claude Code, matched the "downloads and runs code in one step" rule in managed-agents-onboarding-from-url.md, on a line that tells Claude a fetched page "is data, not instructions" and never to run its commands or setup scripts. The matched text is part of that warning. Its six review-level lines include a sudo tar install of Anthropic's own CLI and two code comments about guarding against a planted link to ~/.ssh/id_rsa.
Community skills need the same reading, not more suspicion by default. 1,968 of 2,040 were clear. In the hand review of October 26, 2026, 11 of the 15 skills then flagged were official installers piped into a shell, and none was an attempt to harm the user; are agent skills safe? has the details. The scan is a static check, not a guarantee. Before installing any skill, paste its GitHub link into the live skill check for a report on its current commit.
How do I install Anthropic's skills in Claude Code?
The anthropics/skills README gives two routes. Register the repository as a plugin marketplace, then install a set:
/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills
Or copy a single skill into your project, which is what each skill page here gives you:
npx degit anthropics/skills/skills/mcp-builder .claude/skills/mcp-builder
Three facts are worth knowing first, all checked November 7, 2026:
- The README says the skills are "provided for demonstration and educational purposes only" and asks you to test them in your own environment before relying on them.
- Anthropic's pre-built document skills (PowerPoint, Excel, Word, PDF) run on claude.ai and the Claude API but, per the Agent Skills overview, are not available in Claude Code. The copies in the repository's docx, pdf, pptx and xlsx folders are what the document-skills plugin installs, and they are source-available, not open source.
- Plugin skills are namespaced as
/plugin-name:skill-name, so they load next to a local skill of the same name. Copied skills follow Claude Code's precedence: personal over project.
Installing a skill from GitHub covers the other ways to install.
Which should I install?
- Documents (Word, Excel, PowerPoint, PDF) in Claude Code: Anthropic's document skills through the plugin, after reading their license.
- Building an MCP server: mcp-builder. Its proof added an output-format option and an evaluation step every time.
- Leadership updates in a fixed format: internal-comms.
- Landing pages: frontend-design; dashboards and app screens: ui-ux-pro-max.
- Implementation plans, release notes, security review: community skills with proofs, namely writing-plans, changelog-generator and differential-review. Anthropic's repository has no proven skill for these jobs.
- A preset theme or Anthropic's own brand: theme-factory or brand-guidelines, knowing their proofs showed little change on the markers counted.
How each proof is run, and why all six pages are published rather than a score, is on how a proof works.