Packages of instructions an AI agent opens only when a task calls for them.
Agent Skills are packages of files that teach an AI agent how to do a particular job. At the centre is SKILL.md, a file holding a name, a short description and instructions. A skill can also bundle scripts or templates.
At first the agent reads only each skill's name and description, opening the full instructions only when a task matches, so many skills take up little of the context window. Anthropic launched the format in October 2025 and made it an open standard that December.
But installing a skill means letting someone else's instructions and code into your agent. In January 2026 Cisco analysed 31,000 agent skills and found at least one vulnerability in 26% of them.
Think of it like this
Imagine a rack of rule cards in the sports hall, one for each game. The teacher usually just skims the titles, and on dodgeball day pulls out only the dodgeball card to read closely. An agent keeps its skills' titles in mind the same way and opens the instructions when they're needed.
For example
Ask an agent to "make our class presentation in the school template", and if a school-template skill is installed, it pulls out those instructions and the template file.
Common mix-ups
Mix-upInstalling a skill retrains the AI model and makes it smarter.
ActuallyA skill doesn't change the numbers inside the model. The agent just reads the instructions into its context while working, so if you remove the skill, the know-how goes with it. That's also why a skill can move unchanged to another agent that supports the same format.
Mix-upIf a skill has been published somewhere public, it's safe to install.
ActuallyAnthropic advises installing skills only from trusted sources, and otherwise checking the bundled code and any instructions to connect to outside networks first. SKILL.md is plain text, so you can open it and read it yourself before installing.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- How should a skill's one-line description be written so an agent knows when to use it?
- Activity
- In groups, pick a job such as "write a nutrition report from the lunch menu" and make a skill card: a name and one-line description at the top, step-by-step instructions below. Another group guesses from the description alone when the card should be used. Then read the launch record (E0723) for the read-only-when-needed approach and the Cisco analysis (E0838), and check each other's cards for risky lines such as "send this somewhere".
Records about this
Source-checked moments from AI history.
The first large-scale measurement to treat skill marketplaces as an unvetted supply chain
Released a specification and a reference SDK; the second standardization attempt after MCP
Packages procedural knowledge as folders and loads it only when needed, in three stages: metadata, body, bundled files