Skills

A Skill is a folder of instructions that teaches the agent how to handle one kind of task, such as the way your team checks a monthly report. The agent knows the name and description of every Skill it has, and loads the full instructions only when your request matches one. You can also name a Skill in your message, for example Use the Create a Skill Skill.

Standard Skills

The platform comes with Standard Skills. Every user can use them, and nobody in a Tenant can change them or switch them off. They are updated with each release of the platform, so you get the current version without doing anything. Today there is one: Create a Skill.

Find and read Skills in the Skill Library

Choose Skills in the sidebar, under New Conversation, to open the Skill Library. It lists every Skill you can use with its name, description, scope (Standard, Tenant or Personal), version and last change. A Standard Skill has no date for its last change, because it changes with a release of the platform.

Use the search field to find a Skill by a word in its name or description, and the buttons All, Standard, Tenant and Personal to show only one scope. Both narrow the list as you type or click.

Open a Skill to read it. The Skill page shows the Skill’s instructions (its SKILL.md), its version, the list of its files and, for a Standard Skill, its changelog. Choose a file in the list to read it. A file that isn’t text is listed but can’t be shown. A Standard Skill is read-only: it has no edit, export or rollback. A Personal Skill of yours has an Edit action.

While you only have Standard Skills, the Library explains what Skills are and how to create your own: start a Conversation and ask the agent, for example Make a Skill from how I check a monthly report. See Create a Skill for what makes a good Skill.

Tenant Skills

A Tenant Skill is one your Tenant shares. You find it in the Skill Library once an admin gives you, or a Group you are in, the Right to use it, and the agent then uses it for you like any other Skill, with its files under /skills/tenant/. Its page says what you may do with it: use it, and perhaps propose changes to it or accept them. Choose Export as ZIP to download it. Admins create Tenant Skills and decide who may use them: see Manage Tenant Skills.

Propose a change to a Tenant Skill

A Tenant Skill changes only through an accepted Skill Change Request. If you hold the Right to propose changes to it, its page has Propose a change, which opens the editor on its current version, and Propose a change from a ZIP, which takes a ZIP in the same format as an import. Describe the change in the field the editor asks for: reviewers read it next to the changes. Nothing about the Skill changes yet. You may have 20 open requests at a time.

The request has its own page with the changes against the version it is based on, the description, who proposed it and the version it is based on. Someone who holds the Right to accept changes sees it under To review in the Skill Library, with a count next to Skills in the sidebar, and accepts it or rejects it with an optional comment. Accepting makes the next version of the Skill. Reviewers may accept their own requests; the Skill’s history shows who proposed and who accepted each version. Your own requests are under My requests, with their state: open, accepted, rejected, withdrawn or Stale. You can withdraw one while it is open.

When one request is accepted, every other open request on the Skill is Stale: it was based on a version that is no longer the current one, so it can’t be accepted. Reject or withdraw it and propose the change again on the current version. A holder of the Right to accept can also roll a Tenant Skill back from its history; that is recorded as an accepted request.

Share a Personal Skill with the Tenant

You can ask to share one of your Personal Skills with your Tenant. Open its page and choose Share with the Tenant, then say what the Skill is for: an admin who holds the Right “Manage Skill access” reads your Skill and decides. The page tells you while the request waits, and the Skill stays yours until then. If a Tenant Skill has your Skill’s name already, the share is refused with that reason; rename your Skill and share it again. If you change the Skill after asking, the request is Stale: withdraw it or share again.

The admin finds the request under To review in the Skill Library and reads the whole Skill on the request’s page. Accepting asks who gets the Rights to use it, propose changes to it and accept them. Your Personal Skill then becomes the Tenant Skill, with the same versions, not a copy, and you may propose changes to it. Rejecting leaves your Personal Skill as it is, with the admin’s comment. The request is in your My requests.

Write your own Personal Skills

A Personal Skill is yours alone: nobody else in your Tenant sees it, uses it or finds it in their Skill Library, not even an admin. Write one in the editor:

1. In the Skill Library, choose New Skill.
2. Enter a Name of 1 to 64 lowercase letters, digits and single hyphens, such as board-pack-check. A Standard Skill’s name can’t be used, and you can’t have two Personal Skills with the same name. The name can’t be changed later.
3. Write the Description, at most 1,024 characters: what the Skill does and when to use it.
4. Write the Instructions, the body of the Skill’s SKILL.md, in Markdown.
5. Choose Add a file for scripts and resources next to SKILL.md. Give each file a path with / between folders, such as scripts/check.py. Rename a file by changing its path, and choose Remove to delete it.
6. Choose Create Skill.

The editor checks your Skill as you type and says beside each field what is wrong, for example a name with capital letters, a path that starts with / or goes up with .., or a description that is too long. Create Skill can be used once nothing is wrong. A Skill has at most 100 files, a version is at most 5 MB in all, and SKILL.md is at most 100 KB. You can have up to 100 Personal Skills.

Saving makes a new version of the Skill. An earlier version is never changed. To change a Skill, open it and choose Edit, change what you want and choose Save new version. If the Skill got a newer version since you opened the editor, the editor tells you and doesn’t save: open the Skill again and repeat your change. The editor edits text files only, so a Skill with a file that isn’t text can’t be changed in it.

Let the agent create a Skill

You can also ask the agent to write a Personal Skill for you, for example Turn what we just did into a Skill or Add a chart to my monthly-report Skill. It follows the Create a Skill Standard Skill to write it. After a task that clearly recurs, or one that needed several corrections, the agent may offer once in the Conversation to make a Skill of it.

The agent saves nothing on its own. Its proposal shows as a card that asks for your Approval, with:

• what it would save: a New Personal Skill with its name, or a New version of one of your Personal Skills;
• the Description;
• the Files, compared with the Skill’s current version: each file that would be added, removed or changed, with the lines that change. For a new Skill every file is new.

Choose Accept to save it, Edit to change the proposal first, or Reject to save nothing. Edit shows the proposal as JSON: change the description or a file’s text, then choose Accept with changes, and the agent saves what you changed. Once it is saved, the card says which version it is and links to the Skill, which opens in a new tab. The Skill’s history shows the version as made by the agent.

The agent never changes a Standard Skill: it offers a Personal Skill based on it instead. If the Skill got a newer version while the card was open, or the name is already taken, nothing is saved and the agent tells you why.

Let the agent propose a change to a Tenant Skill

Ask the agent to improve a Tenant Skill you may use, for example Make the board-pack-check Skill also compare the totals of every page. The agent never changes a Tenant Skill itself: if you hold the Right to propose changes to it, it proposes the change, and nothing happens until you approve.

The card that asks for your Approval says Change request for the Tenant Skill and that accepting opens a request. It shows:

• Why, the agent’s explanation for reviewers;
• the Description;
• the Files, compared with the version the agent read when it made the change.

Choose Accept to open the request, Edit to change the proposal first, or Reject to open nothing. Once the request is open, the card says so and links to its page, which opens in a new tab. The Skill stays as it is until someone who holds the Right to accept changes accepts the request: see Propose a change to a Tenant Skill. The request shows in the Skill’s history as made by the agent.

• Without the Right to propose, nothing opens, and the agent offers a Personal Skill based on the Tenant Skill, which you can change freely.
• If someone’s change was accepted after the agent read the Skill, the request is Stale: it can’t be accepted and appears under My requests as Stale, the card says so, and the agent offers to move the change onto the current version. If you agree, you approve a new card, and a new request opens.
• If you hold the Right “Manage Skill access”, you can also ask the agent to create a new Tenant Skill. The card then says New Tenant Skill, and accepting saves it right away, with no request.

How the agent picks and loads Skills

At the start of every Turn, the agent gets the list of the Skills you can use: every Standard Skill and every Personal Skill of yours, each with its name, such as personal/weekly-report, and its description. It never sees another user’s Personal Skills, and a Skill you couldn’t open yourself is never offered to it.

The agent chooses by the descriptions. When your request matches one, or you name a Skill in your message, it loads that Skill, says which one it uses and follows its instructions. A clear description of what the Skill does and when to use it therefore decides whether the agent picks it.

If you can use more than 40 Skills, the agent’s list holds only the Standard Skills and says how many more there are. The agent then searches them by the words in their names and descriptions, gets the best 20 matches, and loads the one that fits. It searches only the Skills you can use, so a large Skill Library works the same way, and your descriptions matter even more: the search finds a Skill by the words you used in its name and description.

A Turn uses the version of each Skill that was current when the Turn started, and keeps it to the end. If you save a new version while a Turn runs, the next Turn uses it. Two Conversations running at the same time each use their own versions.

Choose the Skills of a Conversation

Below the message field on Start, next to Integrations, the Skills button lists every Skill you can use. Standard Skills are marked Always on and have no checkbox. Every other Skill is ticked. Untick the ones the agent shouldn’t use in this Conversation, then send your first message.

Your first message fixes the choice for the whole Conversation, and later messages can’t change it. The agent offers a Turn only the Standard Skills and the Skills you left ticked, and only while you may still use them, each at its current version. If the list can’t be loaded, the agent uses all your Skills.

Inside a Conversation, the same Skills button lists its fixed Skills. To start a Conversation with just one Skill, open the Skill’s page and choose Start a Conversation with this Skill: Start opens with only that Skill ticked, and the Standard Skills stay on. A Conversation from before this choice existed keeps using all your Skills.

Pick a Skill with /

To make sure the agent uses a Skill, type / in the message field, on Start or in a Conversation. A / at the start of the message or after a space or line break opens a list with the category Skills. A / inside a word, as in data/q3.csv or und/oder, opens nothing.

The list holds the Skills the agent will have in this Conversation: on Start the Standard Skills and the ones you left ticked, in a Conversation its fixed Skills. Each entry shows the Skill’s name, its scope (Standard, Tenant or Personal) and its description. A name that exists in two scopes appears twice. While the Skills load, the list shows placeholders.

Type after the / to filter by name or description. Use the arrow keys to move, Enter or Tab to pick, and Escape to close the list and keep what you typed. If nothing matches, the / stays plain text and you can send the message as it is.

Picking an entry puts a chip into your message, showing the Skill’s name, in line with your text. The chip behaves like one character: the arrow keys cross it in one step, Shift with an arrow key or a click selects all of it, and Backspace right after it or Delete right before it removes all of it. Copying, cutting, pasting and undo treat it as one unit. You can pick several Skills in one message, in any order between your text. Behind the chip, your message still holds a short piece of text such as :skill[create-a-skill]{name=standard/create-a-skill}: that is what is sent, stored and copied. You can write that text yourself, and it becomes a chip once it is complete; incomplete text stays plain text. The message you sent, and the same Conversation when you open it again, shows the same chips. Before it answers, the agent loads every Skill your message names, once each, in the order you named them, and follows them. The Used Skill line shows each one among the steps. If a named Skill isn’t available in the Conversation, the line says Not loaded and the agent tells you. A new message loads the Skill again.

Pick a tool of an Integration with /

The same list also offers the tools of the Integrations of the Conversation. After Skills it has one category for each Integration, labelled with its name and a badge, Tenant or Personal. Open a category to see its tools: the tool’s title (or its name, if the server gives none), the Integration it belongs to and its description. A tool that changes something shows Asks first: the agent still asks for your approval before every call of it. While you type after the /, the list searches the Skills and the tools of every Integration at once.

The tools are asked from the Integrations the first time you type a /, with your own sign-in, and kept for the Conversation. Until they arrive, the Integration’s row shows placeholders. An Integration you haven’t signed in to, or whose sign-in ended, stays in the list with a hint, and its one entry, Sign in, takes you to the sign-in and back to the Conversation. An Integration that can’t be reached stays in the list too, with a hint. Nothing else can be picked from these Integrations. The titles and descriptions come from the Integration’s server and are shown as plain text.

Picking a tool puts text like :tool[Search pages]{name=wiki_search_pages} into your message. It tells the agent that you want it to use that tool in this answer; the agent works out the tool’s arguments from your message, and nothing is called for it. If the agent doesn’t have the tool in this answer, because its Integration is gone, can’t be reached or isn’t part of the Conversation, it tells you and carries on.

Pick an Integration Prompt with /

Some Integrations offer prompts: ready-made requests the Integration’s server provides. In an Integration’s category they come first, before its tools, marked Prompt, and show their title (or name) and description. While you type after the /, the list searches them as well.

Picking one puts text like :prompt[Weekly summary]{name=sales_weekly_summary} into your message. Only you can start a prompt: the agent never uses one on its own.

• A prompt without required arguments is loaded before the agent answers. A card named Used prompt shows which prompt was used and, expanded, its text. The agent follows it together with the rest of your message.
• A prompt with required arguments is left to the agent, which fills the arguments from the free text of your message and then loads it, with the same card. If a required argument is missing, it asks you.

The text of a prompt comes from the Integration’s server and is shown as plain text. Images and other content that isn’t text are named as not supported. If the agent doesn’t have the prompt in this answer, because its Integration is gone, can’t be reached or doesn’t offer prompts, the card says nothing was loaded, the agent tells you and carries on. Loading a prompt only reads, so it never asks for your approval.

See the history of a Personal Skill and go back

The Skill page of a Personal Skill has a History with every version, newest first. Each version shows who made it, when, and how: in the editor, by the agent after you approved a change, by importing a ZIP file, or by going back to an earlier version.

To see what changed, choose two versions under Compare two versions. The comparison lists each file that was added, removed or changed. For a text file it shows the lines that were removed and added, with a few unchanged lines around them. For a file that isn’t text, only its size is compared. At first, the page compares the current version with the one before it.

To go back, choose Roll back at an earlier version and confirm. This makes a new version with the content of the earlier one. No version is changed or deleted, so you can always go forward again from the history. If the Skill got a newer version since you opened the page, nothing is rolled back: look at the history again and repeat it.

A Standard Skill has no rollback. Its history is its changelog on the Skill page.

Delete a Skill

You can delete a Personal Skill you own, and, if you hold “Manage Skill access”, any Tenant Skill. Open the Skill and choose Delete. A confirmation asks first. For a Tenant Skill it says how many users may use it, and that open change requests on it are withdrawn.

Deleting is for good: the Skill, all its versions and its files are gone, and its name is free again. The Skill disappears from every list at once. In a Conversation that is already open, the agent no longer has it from the next message on. If you ask for it, the agent tells you it is gone. A Standard Skill can’t be deleted.

When a user leaves

When a user is removed from the Tenant, or deletes their account, their Personal Skills are deleted with their versions and files, and their open change requests are withdrawn. Tenant Skills they created stay as Tenant Skills.

Import and export Personal Skills as a ZIP

You can bring a Skill from elsewhere and take your own out, as a ZIP in the Anthropic format: one folder named like the Skill, holding SKILL.md and the files it needs, such as scripts and templates.

To import, press Import from ZIP in the Skill Library, choose the ZIP and press Import. It becomes a new Personal Skill with the name of the ZIP’s folder. To add a new version to a Personal Skill instead, open it and press Import a new version: the ZIP’s folder and the name in its SKILL.md must be the Skill’s name, and the earlier versions stay as they are. The dialog says when the ZIP is uploaded and when it is checked, and then which version it made.

A ZIP is refused, with the reason for each problem found, when:

• it isn’t a ZIP, is damaged, or is protected by a password;
• it holds more than one folder, or files outside the folder;
• it has no SKILL.md, or the frontmatter, the name or the description of SKILL.md is wrong;
• it holds a link, or a path with .., an absolute path or a backslash;
• it is larger than 6 MB, holds more than 100 files, or is larger than 5 MB once unpacked (the size is checked while it is unpacked, so a ZIP that expands without end stops early);
• the name is the name of a Standard Skill, or of another Personal Skill of yours.

A Skill from a ZIP may hold files that aren’t text, such as images. The Skill page lists them and says they can’t be shown. The editor would lose them, so such a Skill has no Edit: change it by importing a new version.

To export, open a Personal Skill and press Export as ZIP. Your browser downloads the current version as <name>.zip, which you can import again unchanged. A Standard Skill can’t be exported. Each export of a Personal Skill is recorded in the Audit Log with the Skill’s ID and version, nothing else.

See which Skill the agent used

When the agent loads a Skill, it says so in its answer and shows a line among the steps of its answer: Used Skill, followed by the Skill’s name, its scope and its version, for example Standard · version 1. Open the line to read the instructions the agent was given and the files the Skill holds. The line keeps the version the agent used, even after the Skill has changed.

If the agent asks for a Skill it can’t use in the Conversation, the line says Not loaded and the agent carries on without it.

Read a Skill yourself

The agent finds a Skill’s files in its files, at /skills/standard/<name>/ for a Standard Skill and at /skills/personal/<name>/ for one of your Personal Skills. You can ask it to open a file there, for example Show me the files of the Create a Skill Skill. These files are read-only: a Skill changes only when you save a new version in the editor, or accept one the agent proposes.

Scripts in Skills

A Skill can hold scripts, such as a Python file that adds up a table. When a task needs one, the agent copies the Skill’s files into the Conversation’s Sandbox and runs the script there, under the Skill’s name: skills/personal/<name>/ for one of your Personal Skills, skills/standard/<name>/ for a Standard Skill. The version the Turn uses is the one copied, and it is copied once per version in a Sandbox Session. The Sandbox gets only these copies of the files: no access to the Skill Library itself and no credentials.

Every file of the Skill is copied, also one that isn’t text, such as an image or a template spreadsheet, so a script can open it from the Skill’s folder. A Skill’s files may add up to 5 MB at most, the size of a version. Keep scripts as text, and put data on your Personal Drive.

What the Sandbox offers a Skill

A script uses only the packages the Sandbox has installed. The Create a Skill Skill lists them with their versions, the time limit per run (about five minutes) and the disk space, in the file sandbox-manifest.md, so a Skill written with it relies on the right things. Installing more with pip install works only if your Tenant admin granted the Right “Sandbox has internet access”, which is a separate Right: say in a Skill’s instructions when it needs the internet. See Let the agent run code for what the Sandbox can reach.