Skills

A skill is reusable know-how you make available to an AI agent: a written procedure, the reference files that go with it, templates, and — where needed — processing scripts.

Where a knowledge base gives the agent documentation, a skill gives it a method: what to do, in what order, with which tools.

The principle: the agent reads only what it needs

A skill is disclosed in three stages, which is what lets you attach several of them to one agent without flooding its context window:

1. At startup, the agent sees only the name and description of each available skill. That is very little text, and it is enough for it to decide which one is relevant.
2. If it judges the skill useful, it reads the full instructions (the SKILL.md file).
3. If the instructions point to a reference, a template or a script, it loads them only at that moment.

This is why the description matters as much as the instructions: it is the only thing the agent relies on to decide whether to activate the skill. Write it by answering "when should this be used?" rather than "what is this?".

tip

"Checks the liability and termination clauses of a commercial contract" is a good description. "Contracts skill" is not: the agent will not know when to reach for it.

Creating a skill

From the Manager → Skills tab:

1. Click New skill.
2. Give it an identifier in lowercase with hyphens (analyse-contrats). It also serves as the folder name on export.
3. Write the description.
4. Choose the storage space — required in order to attach files. If there is none, ask your administrator to create one (Admin → Skills).

Importing an existing skill

Rather than writing a skill from scratch, you can import one published on GitHub or agentskill.sh: in the New skill window, choose Import from URL, paste the link, select the storage space, then Import.

On GitHub (public repositories), three kinds of link are accepted:

Link	What is imported
A folder: https://github.com/anthropics/skills/tree/main/skills/pdf	That skill
A file: …/blob/main/skills/pdf/SKILL.md	The skill in that folder
A repository: https://github.com/anthropics/skills	Every skill it contains (25 at most)

Each folder holding a SKILL.md becomes a skill, named after that file's name field. The folder's files are stored under scripts/, references/ or assets/ according to where they sit and their extension. The agent still finds them under the path the original instructions use (reference.md, agents/grader.md…).

Some files are not imported: unsupported extension (.xsd, .js…), file too large, more than 200 files. They are listed after the import, skill by skill. A skill whose name already exists in your organization is skipped, and the reason is shown.

warning

Imported scripts are inert: the skill is created with script execution turned off, and each script must be approved (see two-party approval). Read them first: you are running code written by a third party.

Without authentication, GitHub allows 60 calls per hour per server. An import uses one or two. Beyond that, a message asks you to try again later.

The skill page opens on four tabs.

Instructions

This is the heart of the skill: the SKILL.md file. It starts with a metadata block delimited by ---, followed by the instructions in Markdown.

---
name: analyse-contrats
description: Checks the liability and termination clauses of a commercial contract.
---

# analyse-contrats

## Overview

This skill checks that the mandatory clauses of a commercial contract are
present and compliant.

## Instructions

1. Locate the "Liability" section. Check that a figure-capped limit is stated.
2. Locate the "Termination" section. Check that a notice period is mentioned.
3. Compare the thresholds found against those in `references/seuils.md`.

A status bar under the editor tells you live whether the format is valid and how many tokens the skill represents. Beyond 5000 tokens it starts to weigh heavily on the agent's context: better then to move the detail into a reference file, which the agent will load only when needed.

The Preview button switches between the editor and the rendered output.

The authoring assistant

If you would rather describe your need in plain language, click Authoring assistant, explain what the skill should do, and the AI proposes a compliant SKILL.md.

The proposal is inserted into the editor but never saved automatically: review it, fix it, then save. The cost of that generation appears in AI costs like any other model call.

Files

Three families, matching the three folders of a skill:

Folder	Contents	Formats
Scripts	Code executed in an isolated environment	.py, .sh
References	Documentation read on demand by the agent	.md, .txt
Assets	Templates and data used by the skill	documents, spreadsheets, images…

Files stay one level deep (references/seuils.md, not references/2026/seuils.md), so the agent reaches them in a single read.

Reference them explicitly in your instructions — the agent will open them only if you tell it when to.

Try it

This tab runs a script the way the agent would, with the arguments and standard input of your choice, and shows its exit code, standard output and errors. It is the same execution path as during a job: what you test here is exactly what the agent will get.

Settings

• Skill active — an inactive skill can still be connected in the editor but is never passed to the agent.
• Allow script execution — off by default. Without this option the agent can read your scripts but never execute them.

Scripts and their isolation

A skill's script does not run on the application server. It runs in a dedicated container, deliberately stripped of anything that could be useful to an attacker:

• no network access — neither the Internet, nor the database, nor storage;
• no credentials — the container knows no password and no API key;
• read-only filesystem, with a temporary workspace destroyed after every run;
• strict limits on memory, CPU, process count and duration: a script that loops is interrupted.

A script therefore has access only to what you hand it: its own code, its skill's assets, its arguments and its standard input.

Two-party approval

As a safeguard, a freshly uploaded script is inert. It must be approved by an administrator before it can run: whoever writes the code is not whoever authorises its execution. The "Awaiting approval" badge shows you where each script stands.

An administrator can turn this two-party check off (Admin → Skills), but it is not recommended: every uploaded script would then become immediately executable.

Connecting a skill to an agent

In the pipeline editor, through the Skill node:

1. Drag a Skill node from the library (AI Analysis category).
2. Select the skill in the right-hand panel.
3. Connect its output to the skills port of an Agent or Supervisor node — the pink port, below the node.

You can connect several skills to one agent, and the same skill to several agents. A skill connected to a supervisor is shared by all of its subagents.

The Skill node is not part of the document flow: it receives nothing and transforms nothing, it simply declares what the agent is allowed to draw on.

Exporting a skill

The Export (.zip) button produces the complete folder in the standard Agent Skills format. The archive is directly reusable in any other compatible environment, and makes a readable backup of your work.