Anthropic Skills Repo Review 2026: 19 Skill Folders

On this page: My take At a glance Who it is for What it is Install and run Architecture Maintenance and activity Security and licence Cost Trade-offs Score Alternatives FAQ Evidence Keep exploring

Quick answer. anthropics/skills is Anthropic's public repository of Agent Skills for Claude: 19 skill folders, a template, the spec, and a plugin marketplace file for Claude Code. It is a good reference, but it has no releases or tags and its licences differ by folder, so pin a commit and read the scripts before you install.

My take

anthropics/skills is the place to read how Anthropic writes skills, and a reasonable source of ready-made ones. It is a reference collection, not a hardened product: the README itself says “These skills are provided for demonstration and educational purposes only.” It has no releases and no tags, so you cannot pin a version the usual way, and the licence differs by folder.

JAVIS view: install the two document and example bundles only if you will read what they run. JAVIS did not install or run anything; this is an editorial review from official sources. The same repository already has separate JavisReview reviews of two single skills, skill-creator and webapp-testing, so this page stays at repository level.

At a glance

Live numbers were read from the GitHub API on 19 Oct 2026 at 10:39 UTC and go stale within about 48 hours. Source: api.github.com/repos/anthropics/skills.

Who it is for

• Fits: Claude Code users who want to learn how skills are structured, or who want the document skills (Word, PDF, PowerPoint, Excel) that Anthropic published.
• Fits: people writing their own skills who want a template and worked examples.
• Does not fit: anyone who needs a version-pinned, audited package for a production or shared machine. Nothing here is signed or tagged.
• Check first: if you plan to redistribute or build on docx, pdf, pptx or xlsx, read their licence (see Security and licence).

What it is

The README says: “Skills are folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks.” It also says: “Each skill is self-contained in its own folder with a `SKILL.md` file containing the instructions and metadata that Claude uses.”

Top level of the repository, from the GitHub contents API: .claude-plugin, .gitignore, README.md, THIRD_PARTY_NOTICES.md, skills, spec, template. The README describes spec as “The Agent Skills specification” and template as “Skill template”.

The skills folder holds 19 sub-folders (counted by Python from the contents API and cross-checked against the full file tree; each has a SKILL.md). Name, plugin bundle (from marketplace.json), licence situation and the number of script files (.py, .sh, .js, .mjs) in each:

Totals: 77 script files out of 445 files in the whole repository. 8 folders have none: academy-guide, brand-guidelines, canvas-design, discernment-nudge, doc-coauthoring, frontend-design, internal-comms, theme-factory.

JAVIS did not review the individual skills one by one. Two of them have their own JavisReview pages (skill-creator and webapp-testing); the rest are only listed here.

Install and run

Not tested by JAVIS. Commands in Option B are quoted from the README; commands we assembled ourselves are labelled composed by us, not run. The README shows no expected output and JAVIS saw none, so none is claimed.

The main way here is to copy one skill folder: you install only what you chose and can read it before it loads. The marketplace route (Option B) installs whole bundles.

Option A (main way): copy one skill folder

The README does not describe copying. The Claude Code documentation does: a personal skill lives at “`~/.claude/skills/<skill-name>/SKILL.md`” and a project skill at “`.claude/skills/<skill-name>/SKILL.md`”. Copying one folder keeps the install to what you chose and lets you read it before it is ever loaded:

git clone https://github.com/anthropics/skills.git
cd skills
cp -r skills/frontend-design ~/.claude/skills/

Composed by us, not run. This sequence is JAVIS's composition from the two documented facts above; JAVIS did not run it. Replace frontend-design with the folder you want. Copying one folder leaves out the plugin wrapper, so the skill is not under a plugin name (inferred from the docs table, not tested).

Which skill is worth a look, and for whom (from each skill's own description in its SKILL.md; JAVIS did not test any of them, so no claim is made about how well they work):

• Office documents: docx, pdf, pptx and xlsx describe work on Word files, PDFs, slide decks and spreadsheets, so they suit people who handle those files daily. Their licence is proprietary (see Security and licence), so read it before reuse.
• Front-end builders: frontend-design describes guidance on visual design for new or existing UI, and webapp-testing describes a Playwright toolkit for testing local web apps.
• People writing their own skills or MCP servers: skill-creator describes creating and improving skills, and mcp-builder describes a guide for building MCP servers.

Option B: plugin marketplace (the README method, installs whole bundles)

The README says: “You can register this repository as a Claude Code Plugin marketplace by running the following command in Claude Code:”

/plugin marketplace add anthropics/skills

Then either use the menu (the README lists “Select `Browse and install plugins`”, “Select `anthropic-agent-skills`”, then “Select `document-skills` or `example-skills`”) or install directly:

/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

What each bundle holds comes from .claude-plugin/marketplace.json: document-skills has 4 skills (xlsx, docx, pptx, pdf); example-skills has 12. The same file also lists three single-skill plugins that the README does not mention: claude-api, academy-guide and discernment-nudge. The marketplace name is "name": "anthropic-agent-skills".

After install, the README says: “you can use the skill by just mentioning it.” Its example: “"Use the PDF skill to extract the form fields from `path/to/some-file.pdf`"”

Pin the version

The repository has no tags and no releases (both API lists are empty; the page shows "0 Tags"), so the documented tag pin does not exist here. The Claude Code docs say: “Add `#ref` to pin a branch or tag.” A branch such as main keeps moving. The one stable handle is the commit SHA read on 19 Oct 2026. This is the secondary route; the pinning commands below are composed by us, not run:

dbd4588f9e1033efb41dad4bef2f7947c8993d44

Short form dbd4588, "Update claude-api skill: dynamic workflows and the four workflow quickstarts (#1995)", committed 9 Oct 2026, 18:45 UTC. To use exactly that commit as a Claude Code marketplace, clone, run git checkout dbd4588f9e1033efb41dad4bef2f7947c8993d44 (standard git, composed by us, not run), then, from the folder where you ran git clone, add the clone as a local marketplace. The docs say a relative local path must start with “`./`”:

/plugin marketplace add ./skills

Not verified: whether Claude Code accepts a bare SHA after # (the docs text above names only a branch or tag), and how a local-path marketplace behaves with the "source": "./" entries in marketplace.json. JAVIS ran none of these.

Check that a skill loaded

• Run /skills. The Claude Code docs describe “The `/skills` menu” and say a plugin skill is available as “`/plugin-name:skill-name`”.
• Run /plugin and open the Installed tab; the docs say “Open the **Installed** tab in `/plugin`, which lists the plugin with its scope.”
• If the install summary says to reload, the docs say: “Run /reload-plugins to apply.”
• Then ask for a task that matches the skill and see whether Claude names it. JAVIS has not seen this output.

Architecture

JAVIS read the README, the marketplace file, every SKILL.md front matter, every LICENSE.txt and five scripts. It did not read the other 72 script files in full.

• Skill unit: a folder with SKILL.md. The README says the front matter requires only “`name`” and “`description`”. Some skills add a license field.
• Marketplace file: .claude-plugin/marketplace.json, version "version": "1.0.0", five plugins, every one with "source": "./" and a list of skill folders.
• Helper code: 77 script files, mostly under skills/*/scripts. Biggest groups: docx 15, pptx 15, xlsx 12, skill-creator 10, pdf 8. The same Office helper folder (scripts/office/) appears in docx, pptx and xlsx.
• Standard: the README points to “[agentskills.io](http://agentskills.io)” for the Agent Skills standard and says this repo is “Anthropic's implementation of skills for Claude”.

Maintenance and activity

The last commit is dated 9 Oct 2026, ten days before the read date, and changed the claude-api skill. Skills are updated unevenly: the screenshot shows some folders untouched for months. Because there are no tags, you cannot tell from the repository which state a teammate installed; compare commit SHAs instead.

Security and licence

Licence by folder

The GitHub API reports no licence for the repository as a whole and the repo root has no LICENSE file. Each skill folder carries its own LICENSE.txt, except one. JAVIS read all 19 folders:

The README matches this split: “Many skills in this repo are open source (Apache 2.0).” and “These are source-available, not open source” for the four document skills. Quotes from the proprietary LICENSE.txt (identical in docx, pdf, pptx and xlsx):

• “Use of these materials” ... “including all code, prompts, assets, files,” ... “is governed by your agreement with” “Anthropic regarding use of Anthropic's services.”
• Users may not: “- Extract these materials from the Services or retain copies of these” “materials outside the Services”; “- Create derivative works based on these materials”; “- Distribute, sublicense, or transfer these materials to any third party”.

JAVIS is not a lawyer. Plain reading: you may use the Apache-licensed skills, copy them and adapt them under Apache 2.0 terms (keep the notices). For docx, pdf, pptx and xlsx the file restricts copying and derivative works, so treat them as read-and-use-inside-the-Services reference, and get your own advice before shipping a modified copy. Two details: skill-creator has an Apache 2.0 LICENSE.txt but no license line in its SKILL.md; and frontend-design's LICENSE.txt has the Apache terms but lacks the line “Copyright 2026 Anthropic, PBC.” that the other Apache files end with.

Third-party notices

THIRD_PARTY_NOTICES.md opens with “THE FOLLOWING SETS FORTH ATTRIBUTION NOTICES FOR THIRD PARTY SOFTWARE THAT MAY BE CONTAINED IN PORTIONS OF THIS PRODUCT.” and lists BSD 2-Clause (“**imageio 2.37.0**”), GPL v3 (“**FFmpeg 7.0.2**”), MIT-CMU/HPND (“**Pillow 11.3.0**”) and SIL Open Font License fonts. JAVIS did not map which skill ships which component.

Read scripts before you run them

Skills are instructions plus code that Claude may run on your machine with your permissions. 77 script files sit in this repository. JAVIS read five and found what a reader should know before running them (these are observations, not a security audit; whether each is a problem was not assessed):

A tension worth knowing: the webapp-testing SKILL.md tells Claude “**Always run scripts with `--help` first**” and “DO NOT read the source until you try running the script first”. That is a token-saving rule for Claude, not a rule for you. As the person installing, read the source first.

Skills that change Claude's behaviour broadly

Some descriptions are written to fire often. The academy-guide description begins “Stop and check this skill before finishing any reply to a question about how” and the discernment-nudge skill tells Claude to append questions after substantive answers. The claude-api SKILL.md is 105,119 characters. The Claude Code docs warn: “Every skill in the [skill listing](#skill-descriptions-are-cut-short) adds to your context on every turn, whether or not Claude ever uses it.” and offer “Run `/skill-doctor` to see what each of your skills costs”. JAVIS did not measure any cost.

Checklist before you install

1. Pick the smallest unit you need: one folder (Option A) before a whole bundle.
2. List the scripts first: git ls-files "*.py" "*.sh" "*.js" "*.mjs" in the clone (77 files at the commit above).
3. Search them for shell, network and delete calls: grep -rnE "shell=True|subprocess|curl|wget|npm|pnpm|pip install|rm -rf" skills/<name>.
4. Read SKILL.md and the scripts it tells Claude to run, in full.
5. Check the folder's LICENSE.txt before you copy or modify.
6. Try it in a throwaway project with no secrets in the environment.
7. Record the commit SHA you used; update by checking out a new SHA on purpose.

Cost

The repository is free to read and clone. Using skills in Claude Code runs on your own Claude plan; the README says of Claude.ai only that the example skills are already available to paid plans. JAVIS has no affiliate link here and earns nothing from this page.

Trade-offs

Score

JavisReview editorial score: 66/100. A composite JAVIS set from official sources read on 19 Oct 2026; it is not a benchmark and not a hands-on test.

Alternatives

Figures are from the GitHub API on 19 Oct 2026. JAVIS did not compare their contents.

The Claude Code docs name claude-plugins-official as the official marketplace (“`claude plugin marketplace add anthropics/claude-plugins-official`”). Or write one skill yourself from the template folder: slower, but you know every line.

FAQ

Is it official from Anthropic?

Yes as to ownership: the GitHub API lists the owner as the anthropics organization. The README calls the skills demonstrations, not supported products.

Is it free?

Reading and cloning are free. Licence terms differ by folder; see Security and licence.

Can I use the docx, pdf, pptx and xlsx skills commercially?

Their LICENSE.txt ties use to your agreement with Anthropic and forbids copying and derivative works outside the Services. Read the file and get advice; JAVIS cannot say.

How do I get an exact version?

There are no tags or releases. Clone and check out a commit SHA, for example the one in this review.

Does the install add a script that runs by itself?

Not verified. The README does not say. Read the files before you enable a bundle.

Did JAVIS run it?

No. This is an editorial review from official sources. Not a hands-on test.

Evidence

All sources read on 19 Oct 2026 (GitHub data fetched about 10:39 UTC) and saved as local evidence copies; counts recomputed with a script of ours.

Not read: the contents of most script files, the Agent Skills spec, third-party licence texts in full. Disclosure: no affiliate relationship; JAVIS did not install or run anything; figures change.

Keep exploring

• agentskills.io, the standard the README points to (not read by JAVIS).
• Claude Code docs: skills.
• JavisReview reviews of the single skills skill-creator and webapp-testing, and of davila7/claude-code-templates.