Standardize how you teach AI agents your team's workflows
This approach bundles instructions and resources into a single folder so that AI agents can follow specific business procedures or specialized knowledge. Agents load this content only when needed, allowing them to work efficiently without keeping vast amounts of information in memory at all times. It serves as an open specification for safely conveying specific company or team contexts to AI.
Specialized knowledge in a single folder
An agent skill consists of a folder structure containing a SKILL.md file. This file includes metadata such as the skill's name and description, along with instructions on how the agent should perform specific tasks. You can include various resources in the same folder as needed, such as executable scripts, reference documents, and templates. This packaged knowledge is version-controllable and portable, meaning it can be used as-is in other agent products.
Loading information in stages
Agents do not read the content of all skills at once; instead, they process information in three stages. At startup, they load only the names and descriptions of available skills to understand what situations they can be used for. The full instructions are loaded into the context only when the actual task matches the skill's description. Subsequently, the agent executes code or loads reference files according to the instructions to perform the task. Because detailed content is loaded only when necessary, context usage does not increase significantly even if you have many skills.
Origins of the standard and how to participate
This format was initially developed by Anthropic and released as an open standard. It is currently adopted by several agent products, and contributions from across the ecosystem are welcome. You can learn how to participate in the standard's development by referring to the CONTRIBUTING.md file. The code is distributed under the Apache 2.0 license, and the documentation is distributed under the CC-BY-4.0 license.
Things to check before applying
If you plan to use this format, you should first check whether your AI tool or agent client supports Agent Skills. The README directs you to refer to the 'Client Showcase' rather than providing a specific list of supported products, so it is advisable to check compatibility through that material. Additionally, it is recommended to review the safety of scripts or external resources included in the skill before loading them into the agent.
