Guide

Share AI skills across a company: write it once, everyone has it

To share an AI skill across a company, take the prompt that works out of the one chat it lives in and write it down once as a skill: a SKILL.md file that says, in plain language, when to use it, what to give it, the steps and what good looks like. Have a colleague who did not write it test it on a real job, then publish it. In Claude, Team and Enterprise plans can share a skill with named people, publish it with Publish to org or have an owner provision it for everyone. In a ChatGPT workspace you share from the Skills tab or publish to the workspace library. Give each shared skill one owner.

Shared AI skill. A shared AI skill is a skill, a folder of plain-language instructions led by a SKILL.md file, that its author shares, publishes or has provisioned to colleagues in Claude or ChatGPT, so everyone’s assistant does the job the same way from one copy.

Written by Tayyeb Mahmud, founder and Principal Architect. Claude Certified Architect – Professional · LinkedIn profile of Tayyeb Mahmud

1AYM is an OpenAI Select Partner.

Checked 11 October 2026. Anthropic’s help centre, Claude and Claude Code documentation, developer platform guidance, engineering blog and Claude Academy pages on skills, sharing, projects and organisation settings, and OpenAI’s Help Center, ChatGPT documentation, developer documentation and OpenAI Academy pages on skills, plugins, projects and workspace administration, were read on this date. Menus, plan availability and defaults change often at both vendors.

Stuck in one person’s chat · 28 seconds, with captions

The method: five steps from one person’s prompt to everyone’s skill

Find the prompt people already ask for
Pick the prompt colleagues already ask its author for, or paste into their own chats, because that demand is the evidence it works.
Write it down once
In the author’s work account, ask the assistant to turn the prompt into a skill in plain English: when to use it, what to give it, the steps, what good looks like and the mistakes to avoid, with names and figures taken out.
Have someone else check it
A colleague who did not write it reads it cold and lists what is unclear, then runs it on a real job whose right answer they already know.
Publish it with an owner
Share or publish it in Claude or ChatGPT to the people who do the job, with one named owner, the audience and a review date written down.
Improve the one copy
When someone finds a better way, the owner changes the master once, checks it again and lets the change reach everyone, rather than each person fixing a private copy.

Five steps for moving a working prompt out of one person’s chat and into every colleague’s assistant, and what it looks like when each step is skipped.
Dimension	What it looks like when the step is skipped
Find the prompt people already ask for	A library of skills written from a blank page that nobody asked for.
Write it down once	Copies pasted into a dozen chats, each edited a little, and nobody sure which one is right.
Have someone else check it	A skill that works for its author, because the author fills every gap without noticing.
Publish it with an owner	A shared skill with nobody’s name on it, which nobody updates and nobody retires.
Improve the one copy	Private fixes in private copies, none of them reaching anyone else.

Why the best prompt in the company sits in one person’s chat

Somewhere in your company there is a prompt that does a job properly. It writes the bid summary in the house format, or the month-end commentary finance signs off without edits, or the customer reply that gets the refund policy right. One person built it over weeks of corrections, and it lives in their chat history.

Everyone else gets it the way office knowledge has always moved. They ask for it, someone pastes it into a message, and the copies drift apart, each edited a little, until nobody can say which one is right. When its author is on holiday, the good version is on holiday too. When they leave, it usually leaves with them, because nobody else knew where it was kept or which copy was the one that worked.

Neither product moves it for you. Even once it is saved as a skill, Claude keeps a custom skill out of colleagues’ sight until its author shares it, and an admin can see its name and sharing status but not its files [1]. A prompt that only exists in someone’s chats is harder still to find. That is a key-person risk sitting on the work the assistant does best, and it is one of the cheapest risks in an AI rollout to remove.

Both vendors set a low bar for when to act. OpenAI’s rule of thumb is that a prompt you keep reusing, or a workflow you keep correcting, should probably become a skill [2]. Anthropic’s Claude Code guidance makes a playbook pasted for the third time the trigger [3]. A prompt colleagues ask for by name has cleared both.

The fix: write it down once, check it, publish it

The film’s fix has four moves. The author writes the prompt down once, in plain English. Someone checks it. It is published, so every colleague’s assistant has it. And because there is now one copy, improving it once improves it for everyone. Each move maps onto something Claude and ChatGPT already do.

The container is a skill. In both products a skill is a folder whose one required file, SKILL.md, carries a name, a description and plain-language instructions, and the assistant reads only each skill’s name and description until a request matches one, then loads the rest [4, 5, 6]. So a company can carry dozens of shared skills without every chat carrying all of them. The description decides whether a skill is ever used, because it is the only part Claude reads before choosing one [5].

Sharing a skill also ends the copy problem. In Claude, a skill shared with colleagues is view-only for them, stays off until each person turns it on, and recipients get the author’s updates automatically [1]. There is one master, and everyone else uses it.

The prompt: write it down once

The author types this in their work account, either in the chat where the prompt already does its job or in a new chat with the prompt pasted above it. Replace [the job] with the task in the words colleagues use for it. The film shows the move rather than the words, so this prompt is ours, built from what both vendors say a skill should contain.

In Claude, Code execution and file creation has to be on in Settings > Capabilities, and skills live in Customize > Skills [1]. In ChatGPT, skills are on eligible plans, under Plugins in the sidebar, on the Skills tab [7].

Type this where the prompt already works

Turn the prompt I use for [the job] into a skill a colleague could use without asking me. Plain English. Say when to use it, what to give it, the steps, what a good result looks like and the mistakes to avoid. Take out names, figures and anything confidential, and list what you took out.

What comes back is a draft. Read the list of what it took out first, then the rest. Save it to your own account and use it for a few days before anyone else sees it: OpenAI’s guidance is to keep a skill’s first version local while you iterate, and to share it broadly only once it is ready [2].

Why each part of the prompt is there

Each clause asks for something a colleague needs and the author never had to write down, because the author carries it in their head.

“A colleague could use without asking me”: This is the test the whole skill has to pass. Anthropic’s version is the colleague test: show a prompt to someone with little context, and if they would be confused, Claude will be too [8]. Naming that reader makes the assistant write down the context the author supplies without noticing.
“Plain English”: A skill is plain-language instructions in a Markdown file [4], so the people who know the job can write it, check it and keep it current. Nobody has to translate it for the colleague who reviews it.
“When to use it”: This becomes the description, the only part the assistant reads before deciding to use the skill [5]. Anthropic’s advice is to say what the skill does and when to use it, in the words a request for that task would use [4]. A vague description is usually why a published skill sits unused.
“What to give it, the steps, what a good result looks like”: These are the parts OpenAI Academy lists for a typical SKILL.md: what the skill does, the inputs it needs, the steps, the output format and the final checks [9]. Anthropic’s starting point is the same: the task in one sentence, and what a good result looks like [4].
“The mistakes to avoid”: The author’s corrections are the valuable part of the prompt, and they are the part a copy-paste loses. Anthropic’s engineering team advises capturing what worked and the common mistakes when a skill is written, so it holds the context Claude actually needed [10].
“Take out names, figures and anything confidential”: A shared skill travels whole. Anthropic says everyone you share a skill with receives its files, which is why passwords and keys never go in one; a connector lets each person sign in as themselves [4]. The list of what was taken out gives the author a quick thing to check.

When the prompt already exists in several versions

Sometimes the prompt is not locked in one chat at all. It has already been passed around, and each team has its own edit. Do not pick a favourite. Collect every version, paste them into one chat in the work account and ask for a single skill that settles the differences in the open.

Variation: merge copies that have drifted

Here are the versions of the prompt our team uses for [the job]. Merge them into one skill a colleague could use without asking anyone. Where the versions disagree, list each difference and ask me which is right before you write the skill.

Each disagreement is a decision two people have been making differently, often without knowing it. Settle each one with whoever owns the job, so the published skill carries an answer rather than a compromise. Then run the colleague’s check below as normal.

Have someone who did not write it check it

The author is the wrong person to test a skill, because they fill every gap without noticing. Anthropic’s Cowork course says to test-drive a skill before sharing it, because a teammate will phrase things differently and hit edge cases you work around without noticing [11]. So the check belongs to someone else, ideally a person who does the job but did not write the prompt.

Two passes cover it. First, the colleague reads the skill cold, in a new chat, with this prompt.

The colleague’s check, in a new chat

Read this skill as someone who does this job but has never seen it before. List every step that is unclear, every term it assumes I know and every question it would make me ask its author. Quote the line each time. Do not rewrite it.

Second, they run the skill on one real job whose right answer they already know, asked the way they would normally ask, and compare. Anthropic’s Claude 101 course gives the same test for any recurring task: try it on past work where you know the answer before you trust it [12]. OpenAI’s plugin guide adds a run on an input with missing or unclear details, to see whether it flags the gaps [13].

Fix what the check finds one change at a time, re-running after each, so you know which change made the difference [11]. And do not take a skill firing as proof that it helped. Anthropic’s Claude Code documentation says to compare the same prompts with and without the skill, in a fresh session [14].

Where sharing lives in Claude and ChatGPT

The method is the same in both products. Who can share, how far a skill reaches and who decides are not, and both vendors move these menus often.

Where to make, share, publish and control a skill in Claude and ChatGPT, checked 11 October 2026
Dimension	Claude	ChatGPT
Make it	Turn on Code execution and file creation in Settings > Capabilities; skills are in Customize > Skills. Ask Claude to write the skill from the chat, as above	On eligible plans, open Plugins in the sidebar, then the Skills tab, and choose Create with chat, Create with editor or Upload from your computer, or ask ChatGPT in a chat and install the draft
Share with named people	Team and Enterprise: Share, to specific people where the organisation allows it. Sharing with a group is Enterprise only, and the Share button only appears when an owner has sharing switched on	From the skill’s more options menu choose Share, then search for people or groups or copy a sharing link
What colleagues get	A view-only skill that stays off until they turn it on, with the author’s updates arriving automatically	Access to the skill. With the admin permission Enable skills installing, members can install skills for other members
Publish to the company	Team and Enterprise: Publish to org puts it in the organisation’s library. Where review is required, a reviewer approves it and decides how it is offered and to whom	Publish it to the workspace library
On for everyone	Owners provision it from Organization settings > Plugins & skills. It is on for everyone by default and each person can switch it off. Some groups only is Enterprise, through a plugin assigned to a group	Plugins, which can bundle skills, start private. Sharing one and publishing it to the workspace each need their own permission, and recipients install it unless an admin already has
Checks before it spreads	Team plans start with publishing Open, with no review, so switch review on if you want one	Uploaded skills are scanned, and some are marked Needs Review or Blocked. OpenAI says the scan does not replace your own review
Admin control	Owners can turn off User-created skills, so members use only provisioned and built-in ones. Admins see a custom skill’s name and sharing status, not its files	Admins decide who may create, upload, share, publish and install skills. Where the admin page is available, they see each skill’s owner, access, users and 30-day invocations, and can change its owner

The Claude column is from Anthropic’s documentation and help centre on skills, on using them and on managing them for an organisation [1, 5, 10, 15], and the ChatGPT column from OpenAI’s Help Center article on skills and its guide to building plugins [7, 13]. OpenAI’s Help Center names eligible Business, Enterprise, Healthcare and Edu workspaces for skills [7], and sharing a skill with colleagues in Claude needs a Team or Enterprise plan [1].

Which assistant the company standardises on is a separate decision, and our comparison of Claude Enterprise and ChatGPT Enterprise sets out what each plan publishes on price, admin controls and data location.

One point matters for the film’s promise that improving a skill once improves it for everyone. Anthropic states that people a skill is shared with get the author’s updates automatically [1]. The OpenAI pages read for this guide do not say either way for a shared ChatGPT skill, so check it in your workspace before you rely on it. OpenAI’s guidance for shared plugins is to have a teammate try a representative task before the wider team relies on a change [13], which is a sound rule for any shared skill in either product.

If skills are switched off at work

Some companies switch personal skills off. Respect that decision rather than routing around it with a personal account. In Claude, owners can turn off user-created skills [15], and in a ChatGPT workspace admins decide who may create skills [7]. Send the written skill to whoever runs the workspace and ask for it to be published or provisioned, which is the outcome you wanted anyway.

While that is decided, a shared project is the nearest thing. Paste the written method into the project’s instructions, which in Claude apply only to chats inside that project [16] and in ChatGPT apply across the project’s chats [17]. Claude Team and Enterprise projects can be shared with colleagues with Can view or Can edit access [18], and ChatGPT project invitees get chat or edit access [19]. A project reaches only the people invited to it, so treat it as a stopgap.

One owner, one copy, one review date

A published skill with no owner decays quietly. The process changes, the template moves, and the skill keeps producing last quarter’s version, now for everyone at once, which is worse than one person’s stale prompt. Both vendors ask for the same few things before a skill goes beyond its author.

OpenAI’s admin checklist: Before a plugin or skill goes company-wide, confirm its source, accountable owner, intended audience and review date, review everything it bundles, test it with non-sensitive data and the least access it needs, and record who owns re-review and retirement [20].
Anthropic’s Cowork course: For a shared plugin, which bundles skills: one named owner, a test run before every publish, specific names, and a regular review that retires what nobody runs [21].
OpenAI’s plugin guide: Share with a sample prompt beside the link and say who maintains it, and have a teammate try a representative task before the wider team relies on a change [13].

When someone finds a better way, the owner types

Update the [skill name] skill with this fix: [what went wrong and what should happen instead]. Change only what the fix needs. Add one dated line at the end saying what changed and why, then show me every line you changed.

In practice, put a short block at the top of every shared skill’s instructions with its owner, audience and review date, and let the prompt above add a dated line for each change at the end. Keep the master copy somewhere the company owns, such as a shared drive or a repository, so the skill does not depend on one person’s account. In ChatGPT, admins can change a skill’s owner from the admin page where it is available [7].

Either way the owner changes the one copy, and nobody keeps a private edit of their own. In Claude the owner can also highlight text in the skill file beside the conversation and choose Edit with Claude [22], and OpenAI’s guide says to ask for a correction to be added to the skill [23]. Run the colleague’s check again before the change reaches everyone.

What goes wrong, and the honest limits

Sharing a skill spreads a method, and it spreads the method’s mistakes just as fast. These are the ways it fails, and each one can be planned for.

It guides, it does not enforce: A skill is instructions the assistant reads and usually follows. Anthropic is plain that a “never do this” line in an instruction file is the wrong tool, and that real guardrails are hooks, permissions and managed settings [24]. OpenAI’s builder documentation tells skill authors to make clear that a user’s explicit instructions take priority over the skill [25]. A rule that must hold every time belongs in a control, not in a shared skill.
Someone still reads the output: A shared skill that has worked for months can still get a figure wrong or follow an example that no longer fits. Every draft it produces is read by the person who sends it, as it was before the skill existed.
A mistake reaches everyone at once: One copy means one fix, and it also means one error copied into every colleague’s work. That is why the check and the review date matter more for a published skill than for a private one, and why a change is tried by a teammate before the wider team relies on it [13].
Data travels with it: Everyone a skill is shared with receives its files [4]. Build shared skills in the company’s work account, keep confidential material out of consumer accounts and never put a password or key in one. Read any skill a colleague shares before you turn it on, which Anthropic’s help centre also advises [1]. OpenAI scans uploaded skills and says the scan does not replace your own review [7].
Review is not on everywhere by default: Claude Team plans start with publishing set to Open, with no review [1, 15]. If you want a second pair of eyes before a skill reaches the whole company, switch review on deliberately rather than assuming it is there.
Moving a skill loses its settings: OpenAI documents three separate routes for a skill, a ChatGPT workspace skill, a folder on a machine or in a repository, and a plugin, and moving one between them does not carry over ownership, sharing or roles [26]. Decide the route before you publish, and expect to set sharing up again after a move.
Too many skills: Every useful prompt can become a skill, and a library of near-duplicates is hard to trust. The regular review in Anthropic’s course exists to retire what nobody runs [21], and OpenAI’s admin page shows each skill’s 30-day invocations where it is available [7], which is the number to look at before a skill’s review date.

Where 1AYM fits

One person can move one skill out of their chat with the steps above. A company doing it for every team needs decisions somebody owns: which switches are on, who may publish, who reviews, which skills every team gets on the first day and who owns each one when its author moves on. Setting up and rolling out Claude and ChatGPT for companies is core 1AYM work, including the instructions, skills and permissions around them, and we set those publishing rights, reviews and owners up as controls in the admin console under our AI governance implementation work.

The nearest published case is engagement file D-01, at a large international marketing agency, where we contribute engineering to an internal Claude Code skill platform that now carries around 600 production skills. The skills are written across the business by the people who do the work; our part is the plumbing and reliability beneath them. It can start as a small fixed-scope statement of work, which can begin within a day of the scope being signed. If that would help, book a 30-minute call from the end of this page or email enquiries@1aym.com.

For engineers: where shared skills live in Claude Code and Codex, distribution from version control and a test set

The detail behind the steps above, from Anthropic’s and OpenAI’s documentation as read on 11 October 2026. Paths and limits move between releases, so check each one against the version you run.

Claude Code locations: A skill’s location sets its reach: ~/.claude/skills/ is personal, .claude/skills/ in a repository is shared with everyone working there, and plugins or managed settings reach a team or an organisation. On a name clash, an enterprise skill wins over a personal one and a personal one over a project one [14].
Codex locations: Codex scans .agents/skills in every directory from the working directory up to the repository root, then $HOME/.agents/skills for the user and /etc/codex/skills for an administrator [6]. Skills stored in the repository are available in Codex cloud tasks, while personal skills on a local computer are not synced to cloud environments [27], so a team’s skill belongs in the repository rather than in someone’s home folder.
Account sync: Skills enabled in a Claude account load in Claude Code when you sign in with the same account. The sync is one way, from the account down, and checks about every 10 minutes while a session is active and about every 40 minutes when idle. Cowork loads the skills enabled for the account and does not read ~/.claude/skills/ [14].
Distribution from version control: In Claude, giving a skill to only some groups is Enterprise only, by bundling it into a plugin assigned to a group [15]. In ChatGPT, admins can import a plugin marketplace, a JSON catalogue in a public or private GitHub repository, from Admin, Plugins, Add, Import marketplace; new marketplaces sync daily and Sync now requests an update [28]. Either way the skill’s source can live in a repository with a history.
Review like code: Anthropic’s advice for a shared instruction file is to give it an owner and review changes to it like code [24]. Shared skills in a repository deserve the same: changes arrive as pull requests and the skill’s owner approves them.
Descriptions at scale: Automatic use depends on the description, so OpenAI advises keeping it concise, with clear scope, and putting the key use case and trigger words at the front in case descriptions get shortened [6]. Distinct names and descriptions matter more as a shared library grows, because two skills that describe the same job compete for the same requests.
A test set: OpenAI’s five kinds of request make a reusable test: direct requests that should trigger the skill, indirect ones with the same goal, incomplete inputs that should prompt a question, requests that should not trigger it, and edge cases where it must not invent information. If it fires at the wrong time, fix the description; if it picks the right job but results vary, fix the instructions [25]. In Claude Code and Cowork, skill-creator runs the same prompts with and without a skill and reports pass rates, timing and token counts [4].
Enforcement: A skill is context the model reads. Anthropic’s triggers put something that must happen every time into a hook, and a setup a second repository needs into a plugin [3]. Keep hard rules in hooks, permissions or CI, and keep the skill for the method.

Sources

Questions people ask about sharing skills

How do I share a skill with my team in Claude?

On Team and Enterprise plans, use Share on the skill in Customize > Skills to give it to specific people, where your organisation allows sharing; sharing with a group is Enterprise only. Shared skills are view-only, stay off until each person turns them on, and recipients get your updates automatically. To reach the whole organisation, use Publish to org, or ask an owner to provision it for everyone.

How do I share a skill in ChatGPT?

In a workspace on an eligible plan, open Plugins in the sidebar, then the Skills tab. From the skill’s more options menu choose Share, then search for people or groups or copy a sharing link, or publish it to the workspace library. Admins decide who may create, upload, share, publish and install skills, so the options you see depend on your role.

What is the difference between sharing and publishing a skill?

Sharing gives a skill to named people. Publishing puts it in the organisation’s or workspace’s library, where everyone who can see the library can find it. In Claude, publishing can require a reviewer, who approves the skill and decides how it is offered and to whom, and owners can also provision a skill so it is on for everyone by default.

If I update a shared skill, do colleagues get the change?

In Claude, yes: Anthropic says people a skill is shared with get the author’s updates automatically. OpenAI’s pages read for this guide do not say either way for ChatGPT. Its guidance for shared plugins is to have a teammate try a representative task before the wider team relies on a change. In both, check a change before it goes out, because a mistake in a shared skill reaches everyone too.

Can admins stop people creating their own skills?

Yes. In Claude, owners can turn off User-created skills so members use only provisioned and built-in skills. In a ChatGPT workspace, admins decide who may create, upload, share, publish and install skills.

Can my admin see the skills I create?

In Claude, a custom skill is not visible to colleagues until you share it, and admins can see its name and sharing status but not its files. In ChatGPT, where the admin page is available, admins can see each workspace skill’s owner, access, users and 30-day invocations.

Should a shared prompt be a skill or project instructions?

A skill, if it is a method people use in many different chats, because it loads wherever a request matches its description. Project instructions, if it belongs to one body of work, because they apply only inside that project. If skills are switched off at work, a shared project is the nearest substitute, but it reaches only the people invited to it.

Is it safe to share a skill made from a real chat?

Only once it has been read. Everyone you share a skill with receives its files, so take out client names, live figures, passwords and keys first, and build shared skills in your work account rather than a personal one. Read any skill a colleague shares before you turn it on.

What happens to a shared skill when its author leaves?

Plan for it before they go. Give every shared skill a named owner and keep its master copy somewhere the company owns, such as a shared drive or a repository. In ChatGPT, admins can change a skill’s owner from the admin page where it is available. Anthropic’s pages read for this guide do not say what happens to a departed member’s shared skills in Claude, so do not count on a skill outliving its author’s account there.