What is a Claude skill?

A Claude skill is a written recipe that Claude follows for one job, such as replying to a customer about a late delivery or putting together a quarterly review. It's a small folder of instructions, sometimes with files or short programs alongside, and Claude opens it only when a task calls for it. Anthropic's developer pages also call them Agent Skills.

The Claude details in this post come from Anthropic's own pages, as of October 2026.

Anthropic's help pages list skills on every Claude plan, Team and Enterprise included. They need a setting called code execution and file creation switched on, which lets Claude run small programs and make files. Claude keeps each skill's name and short description in view and reads the rest only when a request matches. So a company can add many skills, and Claude still opens only the one a job needs. If you want the technical detail, there's a longer piece on what a skill costs before it runs on amitkoth.com.

An example skill

Here's a short example we made up for a company that sells kitchen equipment, so it isn't a client's.

Name: late delivery replies

When to use it: a customer writes about an order that hasn't arrived.

• Find the order number in their message, and if it's missing, ask for it before anything else.
• Apologize once, in the first line, without blaming the delivery company.
• Give a new delivery date only if the order record shows one, and never guess a date.
• If the order is more than 10 days late, offer free delivery on the next order and copy in the customer service manager.
• Sign off with the writer's own name.

Claude always sees the first two lines. Anthropic calls the second one the description, and Claude matches each request against it to decide whether to open the rest. The steps are what the best person on that desk knows and a new hire doesn't. A longer skill can also hold files, such as a price list or your brand colors, which Claude reads only when a step points to them.

How is a skill different from a project or a connector?

Three other Claude features get mixed up with skills. Anthropic's help pages separate them this way:

• A project is a workspace whose files Claude reads at the start of every chat inside it. A skill works in any chat, and loads only when its job comes up.
• Custom instructions are standing preferences for all your chats, such as "keep answers short". A skill covers one job.
• A connector links Claude to another system, such as your email or your shared drive, so it can read or act there. A skill tells Claude what to do once it's in.

Anthropic's own explainer on skills gives a rule of thumb. If you keep typing the same instructions into chat after chat, or copying them from one project to the next, they belong in a skill.

Who should supply what goes into a skill?

The person who does the job best. We see a skill as closer to a job instruction sheet than to software, so the steps should come from whoever knows the work, and that's rarely someone in IT. Anthropic's tutorial on skills says "You also don't have to be technical to start creating skills." It suggests describing a task and letting Claude turn the description into a properly formatted skill file.

Claude can handle the format, but it can't know the steps your best person takes without thinking, like never guessing a delivery date. The same goes for any instruction you give Claude, and an amitkoth.com post explains why describing the work gets better results than telling Claude it's an expert.

We build a company's first skills this way during setup. The person who does each job best tells us how they do it, and we build the skill with them and test it before anyone else gets it. Some guides start with the file format and treat skills as code first. For a skill that runs programs, they have a point.

Can a skill run programs or reach the internet?

It can run programs, and in some places it can reach the internet. A skill can carry scripts, which are short programs Claude runs during the job, for example one that checks the totals in a spreadsheet. Anthropic's developer documentation says a skill's internet access depends on where it runs. On claude.ai it might have full, partial or no access, depending on your settings. In Claude Code, the version of Claude for writing software, a skill can reach the internet like any other program on that computer.

The same page says to treat skills "like installing software". The help pages say to install skills only from sources you trust, and to review one before switching it on, even when a colleague shared it. An Enterprise company can turn on a scan that checks uploaded skills for malicious content. A skill with no scripts, like the example above, still gives Claude instructions, so someone should read every skill before the whole company gets it. Your IT team or outside IT company looks after security, and we agree the settings with them.

Who can use a skill once it's made?

At first, only the person who made it. Anthropic's help page on using skills says custom skills "aren't visible to colleagues until you share them." On Team and Enterprise, a skill reaches other people in three ways:

• Its maker shares it with named colleagues. They can use it but not change it, and they get the maker's updates the next time they use it.
• Its maker publishes it to the company's library. If the company's publishing setting is Requires review, an Owner (one of the top admin roles in Claude) or someone with permission to review checks it first.
• An Owner adds it for everyone, and it appears in each person's list of skills.

Anthropic's admin guide says Team accounts start on Open, where a published skill goes into the company's library with no review, and Enterprise accounts start on Requires review. From November 4, 2026, Enterprise accounts with security scanning on move to Open unless someone has picked a setting. One of Anthropic's developer pages still says skills on claude.ai can't be shared with the whole company or managed by admins. The help pages for Team and Enterprise describe the three routes above, and we follow those.

The same admin guide says Claude keeps a version history for skills a company adds in its admin settings, with who added each version and when. For an uploaded skill, an Owner can put everyone back on an earlier version. A skill someone made only for themselves shows no version history there.

We also keep the original of every skill in a library your company owns, where each change is saved with who made it, when and why. Claude works from a copy, and if you stop working with us or change tools, the library stays yours. If your company hasn't set Claude up yet, how to set up Claude for your company covers where to start. Once skills are in use, people report where one went wrong, and your AI board, a small group of your own people, chooses which fixes go in each month. That's the monthly work we do with you.

If you'd like help building your first skills, get in touch.