How to share Claude skills across a team: six patterns

Six ways to share a Claude skill with colleagues, from copying the folder to the organization library with owner review, shown with one example skill.

By Nolli research, The Nolli team · Oct 20, 2026 · 9 min read

Short answer

There are six patterns, in order of control: copy the skill folder by hand, share a Git repo, publish a plugin in a marketplace, use the organization skills library with owner review, have an admin roll skills out, or upload API skills for a product you build. Copying costs nothing but drifts. The library adds owner review and automatic updates on Claude Team and Enterprise plans.

On this page

Patterns: Six, from hand copy to API skills
Read: Anthropic’s public docs, 2026-10-19
Example: One made-up skill walked through all six

Key takeaways

Start by copying the folder to one colleague, and move up a pattern only when you need review, versions or group access that the last one lacks.
On Claude Team and Enterprise plans, the organization library with “Requires review” is the one pattern whose docs describe owner approval and automatic updates together.
Anthropic’s pages describe how to deliver a skill, not how to check that it gives right answers once people use it.

A skill is a folder with a SKILL.md file in it. Sharing a skill means getting that folder in front of your colleagues’ copies of Claude, and there are six ways to do it. They differ in how much control you get (who reviews the skill, what happens when you fix it, who receives it) and in how much work you take on.

We read three of Anthropic’s public pages on 2026-10-19 and reuse facts from two more we read on 2026-10-16, listed in Sources. We tested nothing. We walk one small skill through all six patterns, then compare them. Where Anthropic’s pages say nothing, we say so. For the basics, see our guide on what AI skills are.

The example: one skill, one folder

Our example is called customer-call-recap. We made it up as an illustration. It turns messy call notes into a short recap. Think of a skill as a recipe card for the cook, which here is Claude. The card says how your team makes one dish, and the cook reads it only when someone orders that dish.

A minimal skill folder (an illustration)

The Agent Skills specification requires name and description in the header. The name uses lowercase letters, numbers and hyphens, up to 64 characters, and must match the folder name. The description can run to 1,024 characters. The second layout is the same skill inside a plugin, as Anthropic’s marketplace walkthrough lays it out.

customer-call-recap/
  SKILL.md

SKILL.md contains:

---
name: customer-call-recap
description: Turns raw call notes into a short recap with decisions, open questions and next steps. Use when someone pastes call notes or asks for a follow-up summary.
---

Write the recap in this order:
1. Decisions made on the call, one line each.
2. Open questions, with the person who owns each.
3. Next steps, with an owner and a date if the notes give one.

If the notes do not say who owns something, write "owner not stated". Do not guess.
Keep the recap under 150 words.

The same skill inside a plugin:

customer-call-recap-plugin/
  .claude-plugin/
    plugin.json
  skills/
    customer-call-recap/
      SKILL.md

Six patterns, five needs

A full circle means the pages we read describe the thing plainly. A half circle means they describe part of it, or a pattern you would have to build yourself. An empty circle means we found nothing, which does not mean it cannot be done.

What each pattern says it does, from the pages we read

	1 Copy by hand	2 Git repo	3 Plugin marketplace	4 Library with review	5 Admin rollout	6 API skills
Review: a second person approves	none	fully	none	fully	partly	partly
Versioning: history and a way back	partly	fully	fully	fully	partly	fully
Access control: who gets which skill	partly	partly	partly	fully	fully	none
Discovery: people can find it	none	partly	fully	fully	fully	none
Updates reach everyone	none	partly	partly	fully	partly	partly

none partly fully · Our reading of Anthropic’s public pages, read on 2026-10-19 and 2026-10-16 (see Sources). It is a judgment, not a count. Pattern 1 gets half circles for versions and access because Claude’s built-in peer sharing sends new versions and lets the owner revoke.

1. Copy the folder by hand

What it is. You send the folder to a colleague, in chat or as a zip file, and they put it where their Claude looks. In Claude Code, a personal skill sits in ~/.claude/skills/<name>/SKILL.md and loads in all that person’s projects on that machine, but not in Cowork or cloud sessions. In the Claude app, users can upload skill files to their own skills list, unless an admin turned that off.

Claude also has a built-in form of copying. With Skill sharing on, a user shares a skill with named colleagues, who see it under “Shared with you.” The copy is view-only and stays off until the recipient turns it on. A new version reaches recipients at their next use, the owner can revoke access, and no owner approval is needed. Skill sharing is on by default for Team plans.

Who it suits. A few people, and a skill you are still shaping. How to set it up: zip or copy customer-call-recap/, send it, and ask the person to test it on one real set of notes. What it cannot do: a file copy has no review and no single source of truth. A copy you sent in March does not hear about your fix in April.

2. Share a Git repo

What it is. You commit the folder into a repository. Claude Code’s docs say a project skill at .claude/skills/<name>/SKILL.md loads in sessions in that repository, and committing it gives your team the skill. Anthropic’s enterprise guide goes further: keep skills in Git as the single source of truth, review changes through pull requests, and keep rollback in the history. It also says skill authors should not review their own skills.

Who it suits. Teams that already work in a repo, which mostly means engineers. How to set it up:

Put the folder at .claude/skills/customer-call-recap/ in the repo.
Open a pull request, and have someone other than the author approve it.
Teammates pull the repo. In a session there, Claude loads the skill when a task matches, or they type /customer-call-recap.

What it cannot do. It reaches only people who work in that repo. The pages we read describe no automatic path from a repo to a person using the Claude app, other than a file upload. Git also tells you who changed a skill, not who used it.

3. Publish a plugin in a marketplace

What it is. A plugin is a bundle that can hold skills, agents, hooks and MCP servers. A marketplace is a catalog of plugins: a marketplace.json file in a Git repository. Anthropic says marketplaces give centralized discovery, version tracking and automatic updates, and can live on GitHub, GitLab or another Git host.

Who it suits. Claude Code teams that want colleagues to install a skill by name, ship several skills together, or share across repos. How to set it up: wrap the folder as in the second layout above, add a plugin.json with a name, description and version, and list the plugin in .claude-plugin/marketplace.json. Push it. Each colleague runs /plugin marketplace add with your repo, then /plugin install customer-call-recap-plugin@my-plugins. Plugin skills carry the plugin name, so the command becomes /customer-call-recap-plugin:customer-call-recap.

One detail matters for updates. If you set version, users only receive an update when that field changes, so bump it on every release. The overview promises automatic updates, yet it also says users refresh their copy with /plugin marketplace update. We could not tell from the page which one happens by default.

What it cannot do. In the part we read, the page describes no review step and no usage counts. An installed plugin is copied to a cache, so it cannot point at files outside its own folder. It is also a Claude Code feature, so it does not reach the Claude app.

4. Use the organization library with owner review

What it is. On Claude Team and Enterprise plans, users can submit a skill to the organization’s library. In Organization settings > Plugins & skills > Policy, you set Publishing to “Requires review.” An owner then approves each submission and every later version. They see every file, the scan result and, for an update, what changed. They get a “possible duplicate” hint when one fits. They cannot approve their own submission.

Two details make this the pattern we would pick. A reviewer sees a fixed version, so the author’s later edits reach nobody until a new version is approved. And once a version is approved, everyone who uses the item gets it automatically, while the rest stay on the last approved version. A published skill also stays in the library if its author leaves.

Who it suits. A company where most people use Claude and someone owns the standards for a skill. How to set it up: choose “Requires review,” name the reviewers (owners, or roles with Libraries set to “Can manage”), have the author submit customer-call-recap, and when approving, choose whether it goes to everyone or to specific groups. Groups need an Enterprise plan.

What it cannot do. A reviewer sees files and changes. The page describes nothing that shows whether a skill gives right answers on real work. It covers Claude, so a second assistant means a second library.

5. A managed rollout by an admin

What it is. The reverse direction. An organization owner uploads a skill (a .zip with a SKILL.md) or creates one in the settings, and it is provisioned to everyone straight away. It appears in each person’s Customize > Skills, on by default, though people can toggle it off. On Enterprise plans, bundle skills into a plugin and assign the plugin to a group, so marketing gets its ten skills and nobody else does. The Inventory tab can set default access to Available to install, Installed by default, Not available or Required.

Provisioned skills also load in Claude Code for people who sign in with their Claude account, and Claude Code has an enterprise location in its managed settings directory for machines your organization deploys. One prerequisite: skills need code execution, so it must be on in the Policy tab.

Who it suits. A skill everybody should have, such as a house style or a checklist, with one accountable owner. How to set it up: upload customer-call-recap as a zip in Organization settings > Plugins & skills, then check it appears for a test user. What it cannot do: the page describes an owner’s own upload as going in directly, with no second reviewer. It also does not say what happens for users when an owner uploads a changed version. We would pair this pattern with a pull request in Git.

6. API skills, for a product you build

This pattern is for software you write that calls Claude, not for colleagues using the Claude app. Our evidence is thin, so this section is short. We did not read the API walkthrough, and we give no steps. Start with Anthropic’s Agent Skills overview.

What Anthropic’s enterprise guide says, as we read it on 2026-10-16: custom skills uploaded through the API do not appear on claude.ai or in Claude Code, and the reverse. Pin production to a specific version, because if you leave the version out, a new version uploaded by anyone changes what your production agents run. Usage analytics were not available through the Skills API at that time, so you log which skills are in each request yourself. What it cannot do: share a skill with a colleague in the Claude app.

One skill’s trip from a laptop to the team

You do not need all six. Skills can move up as they earn trust. This is one route for customer-call-recap.

From one person’s folder to a reviewed, shared skill

01One laptopThe folder sits in a personal skills folder. It works for one person.
02Copy to a colleaguePattern 1. Now there are two copies, and nobody owns the fix.
03Commit to a repoPattern 2. One source of truth, with pull-request review and history.
04Wrap as a pluginPattern 3. Installed by name, and a version bump ships a change.
05Submit to the libraryPattern 4. An owner reviews a fixed version and chooses who gets it.
06Admin offers itPattern 5 settings: everyone, or a group, installed by default or required.

A fix starts again in the repo, then a new version is approved

An illustration with a made-up skill. The route is one reasonable order, not a rule from Anthropic. Pattern 6 is a separate branch, for skills inside your own product.

Effort against control

Effort to run, and control over what reaches people

Our reading of the pages in Sources, not measurements. Effort rises to the right: from sending a file to building something in your own product. Control rises upward, using the grid above. Pattern 4 sits high because it is the one whose pages describe review and automatic updates together.

Pick this, skip that

Our position, for a company on Claude Team or Enterprise: keep the source of truth in Git and publish through the library with “Requires review.” Use hand copies only in the first week, while one person still edits the skill daily.

Copy by hand if
two or three people are testing a draft. Stop when a third version appears.
Use Git if
your team already reviews code in pull requests and wants the same for skills.
Use a plugin marketplace if
your team works in Claude Code and wants install by name and several skills per bundle.
Use the library if
you want owner approval and automatic updates for people in the Claude app.
Use an admin rollout if
one accountable owner decides that everyone, or a whole group, gets the skill.
Use API skills if
the skill belongs inside software you build, not on colleagues’ desks.

What the docs do not say

None of the pages we read describes checking whether a shared skill gives right answers on real work after release. Review checks what is in the files. Counting shows use, not correctness.
The Help Center text we read stopped part way through its section on monitoring shared skills, and the marketplace text stopped before its hosting section. We did not read the missing parts.
We reused the plugin admin page and the enterprise guide from our 2026-10-16 reading and did not re-fetch them. Check settings names against your own admin console.
We could not find a statement of what a user sees when an owner changes an already uploaded skill in a managed rollout.

Where Nolli fits

Where Nolli fits, and where it does not

Nolli is not a seventh pattern. It does not replace your repo or Claude’s library, and this post does not describe Nolli storing, versioning or approving skills. From our own pages: Nolli learns how your company works, from work on company devices, the email and calendar your company connects and the logs of the AI agents people use. It shows each person’s skills, including know-how that only one person has, and which skills people use and on which work. For each skill it shows the pass rate against your own checks, who uses it and the time it saves, and flags the ones that need work. While people work, it points them to the skill that already does the job. Your company decides what Nolli learns from. There is no public price. Request a demo. See Nolli for AI enablement teams.

Request a demo

Questions people ask

How do I share a Claude skill with my team?

Send the skill folder to a colleague, commit it to a shared Git repo, publish it as a plugin in a marketplace, or, on Claude Team and Enterprise plans, submit it to the organization library or have an owner provision it. The more people and the more change, the more you want review and versions.

Can I share Claude skills without an admin?

Yes. You can copy the folder, use Claude’s Skill sharing to send it to named colleagues, commit it to a repo, or host a plugin marketplace in Git. Skill sharing is on by default for Team plans, but an admin can switch it off.

Do shared Claude skills update automatically?

It depends on the pattern. In the organization library, an approved new version reaches everyone who uses the item automatically. Peer-shared skills update at the recipient’s next use. A plugin with a set version updates only when you change that version, and a hand copy never updates.

Can I give a skill to one team only?

On Enterprise plans, yes: bundle the skill into a plugin and assign the plugin to a group. The Help Center says a skill uploaded in the settings directly goes to everyone in the organization.

Who should approve a skill before it is published?

Someone other than the author. Claude’s review queue does not let a reviewer approve their own submission, and Anthropic’s enterprise guide advises that authors do not review their own skills.

Sources

“Provision and manage skills for your organization.” Claude Help Center, Anthropic, accessed October 19, 2026.
“Extend Claude with skills.” Claude Code Docs, Anthropic, accessed October 19, 2026.
“Create and distribute a plugin marketplace.” Claude Code Docs, Anthropic, accessed October 19, 2026.
“Skills for enterprise.” Claude Docs, Anthropic, accessed October 16, 2026.
“Manage plugins for your organization.” Claude Help Center, Anthropic, accessed October 16, 2026.
“Agent Skills.” Claude Docs, Anthropic, accessed October 3, 2026.
“Specification.” Agent Skills (agentskills.io), accessed October 16, 2026.

Plain-text copy: /blog/how-to-share-claude-skills-across-a-team.md