Agent skills specification
An open specification for packaging instructions, scripts, and resources into a skill an AI agent loads on demand across different clients.
Category
Industry concept
See all terms
What is the Agent skills specification?
The Agent skills specification is an open standard for packaging a skill, a folder with a name, a description, an instructions file, and any supporting scripts or resources, so an AI agent can discover it and load it when the task calls for it. A skill built to the specification is a plain folder any compatible client can read, not code tied to one vendor.
Before this specification, a team's instructions for a repeated task lived wherever someone wrote them: a pinned message, a doc, a custom prompt copied between chats. Nothing pointed an agent to the right instructions at the right time, and every client needed its own format. The specification fixes both problems with one file format and a mechanism for an agent to find and load the matching skill without loading every skill's full instructions.
How does a skill differ from a system prompt?
A skill is different from a system prompt: a system prompt loads in full on every request, while a skill stays off to the side until its short description matches the task at hand, and only then does the agent read the full instructions. This is what lets a company hold hundreds of skills without bloating every request. A system prompt sets the agent's standing behavior, while a skill adds capability for one situation.
How does the Agent skills specification work?
A skill is a folder containing a SKILL.md file with two parts: a short YAML header giving its name and a description of when to use it, and a body with the instructions. An agent's runtime scans the headers of every available skill, matches the current task against their descriptions, and only then reads the full body of the one skill that fits. The folder can also hold scripts or reference files the instructions point to, loaded only if the task reaches that point.
Qontext stores a skill as a file in the context repository, so a skill written once by one person is available to the whole company. Anyone can load a listed skill into their own client through an install link and pull updates.