Anthropic shipped a feature this week that's easy to miss in a product-update email but matters directly to anyone administering a Claude Team or Enterprise organization: members can now publish a skill or plugin they've built to a shared, org-wide library everyone on the team can use. The part worth an admin's immediate attention: for Team organizations, the default policy is Open — meaning any member can publish directly to that shared library with no review step at all — unless someone with Owner or Primary Owner access changes it first.
TL;DR
| Question | Answer |
|---|---|
| What shipped? | Org-wide publishing for member-built skills and plugins, into a shared org library |
| Who can publish? | Any member, by opening a skill/plugin and selecting "Publish" |
| What's the default policy for Team orgs? | Open — no review step before something reaches the org library |
| What are the other policy options? | "Requires review" (admin approves each submission) and "Off" (no org-wide publishing) |
| Where do admins change this? | Organization settings → Skills → Policy → Publishing |
| Who can change it? | Owners and Primary Owners only |
| Does review apply to updates too? | Yes — every future version of an approved skill goes through the same policy |
What actually changed
Before this update, a skill or plugin a member built was effectively personal or project-scoped — useful to them, not automatically visible to the rest of the organization. The new publishing flow closes that gap directly: open a skill or plugin, select Publish, and it's submitted for the whole org. What happens next depends entirely on the organization's publishing policy, which lives in Organization settings → Skills → Policy → Publishing and has three states:
- Requires review — a member's submission (the files, and what changed if it's updating something already published) goes to an admin, who approves it or sends it back with feedback before it reaches the shared library.
- Open — the submission publishes directly to the org library, visible to every member, with no review step.
- Off — org-wide publishing is disabled entirely; skills and plugins stay scoped to whoever built them.
Once something is approved and in the library, every future version follows the same policy — an update to an already-published skill isn't grandfathered in as automatically trusted just because an earlier version passed review.
The default worth checking today: Open, not Requires Review
The detail buried in Anthropic's own rollout notes is the one that matters most operationally: Team organizations default to Open. That means, unless an admin has already gone in and changed the setting, any member of a Claude Team org can publish a skill or plugin straight to the shared library right now, and it becomes available to every other member's Claude without anyone reviewing what's actually in it first.
That's a meaningfully different trust model than most organizations apply to other kinds of shared internal tooling — a shared code repository typically has some form of code review before a change merges; a shared org-wide AI capability, under the default setting here, doesn't. This is the same category of concern explainx.ai raised in its own verification look at agent skills as a potential security threat: a skill or plugin can carry instructions, and depending on what it wraps, real tool access, that Claude will treat as legitimate simply by virtue of being present in the org library — which makes "who can add something to that library, and does anyone check it first" a genuine governance question, not just a convenience setting.
What to actually do about it
If you administer a Claude Team or Enterprise organization, the concrete action here is small and specific: go to Organization settings → Skills → Policy → Publishing and confirm which of the three states your org is actually running under, rather than assuming. For most organizations with any sensitivity around what capabilities get shared broadly — anything touching customer data, internal systems, or credentials via MCP connectors — switching to Requires review costs little (a lightweight approval step for genuinely useful submissions) and closes the specific gap the Open default leaves. Organizations that would rather not have org-wide skill sharing at all can set it to Off without losing the ability for individuals to keep using skills scoped to themselves.
This pairs directly with the broader admin-controls gap explainx.ai flagged in its coverage of the Claude Chat/Cowork merge earlier this month, which noted Anthropic hadn't yet published detailed Team/Enterprise controls for the new Docs/Slides/Design surfaces shipped alongside that merge. This skill-publishing policy is a separate, specific answer for one part of the admin surface — worth checking on its own, not assuming it covers everything an admin might want to control across Claude's fast-shipping feature set this quarter.
How this compares to how most teams handle shared internal tooling today
It's worth putting this default in context against how the average engineering or ops team already governs shared internal tools, because the contrast is where the real risk sits. A shared code repository almost universally has some form of required review before a change merges to a branch other people build on — a pull request, a required approver, at minimum a CI check. A shared internal wiki or documentation space usually has an edit history and often some approval gate for anything linked from a top-level index. A shared package registry, even an internal one, typically has some publish gate beyond "any employee can push a new version that everyone else's build immediately picks up."
An org-wide Claude skill or plugin library defaulting to Open sits, by default, closer to the least-governed end of that spectrum than most teams would choose deliberately if asked — not because Anthropic made an obviously wrong call (Open lowers friction for genuinely useful sharing, which is a real and legitimate goal), but because a default that ships as "no review" is easy to never revisit once it's live, the way an unreviewed default that quietly becomes the actual practice does in a lot of tooling launches. The concrete action item here isn't "distrust the feature" — it's "make the review policy for org-wide AI capabilities a deliberate choice your organization made, not the setting nobody looked at."
What a "skill" actually is, for readers new to the term
For anyone administering a Claude Team org without deep hands-on Claude Code experience, it's worth being concrete about what's actually being published here, since "skill" undersells how much capability can live inside one. A Claude skill is a packaged set of instructions — and, depending on how it's built, references to tools, connectors, or external services — that Claude loads and follows when relevant, the same underlying mechanism explainx.ai covers in its complete guide to agent skills. A plugin extends that further, potentially bundling multiple skills, commands, and integrations together. Publishing either org-wide means every member's Claude can pick up and act on whatever instructions and access that package contains — which is exactly why the review-versus-no-review question matters more here than it would for, say, a shared document template.
Honest limitations
- The Open default applies specifically to Team organizations — Anthropic's own notes don't specify whether Enterprise organizations default differently, so Enterprise admins should verify their own setting directly rather than assuming parity with Team.
- Review, where enabled, checks the submission's files and diff — Anthropic's description doesn't detail whether review includes any automated security scanning, or is purely a human reading the submitted files.
- This is a governance/access-control feature, not a content-moderation one — "Requires review" gates whether something reaches the org library, not an ongoing audit of skills already published before this feature shipped.
- No mention of retroactive review — skills and plugins already in use individually before this update aren't described as being pulled into any review process; the policy applies going forward, to new publish actions.
What this means for builders
If your organization has active Claude Code or Claude Team users building their own skills already, this feature is likely to surface sharing that was previously informal (a skill passed around via Slack or a shared doc) into a structured, org-visible library — genuinely useful for avoiding duplicated effort across a team. The tradeoff is that "structured and visible" isn't the same as "reviewed," and the setting that determines which one you actually have is one admin click away from being checked, right now, in Organization settings.
Related on explainx.ai
- Claude Cowork and Chat merge into one Claude — plus Docs, Slides, Design
- What are agent skills? A complete guide
- Agent skills as a security threat: explainx.ai verification
- How to use Claude connectors and MCP servers
- Top 10 Claude Cowork use cases
- Claude platform: computer use, browser, skills, Files API GA
- Claude unified memory across Chat and Cowork
- Official docs: Anthropic — Organization settings and skills policy
This post is sourced to Anthropic's own product-update announcement to Claude Team/Enterprise administrators, dated September 2026. Policy defaults and settings paths reflect that announcement; verify current behavior directly in your organization's admin settings before relying on any default described here, as Anthropic may adjust rollout defaults after initial launch.