Agent Skill
Ajan Yeteneği
D3
In the Agent Skills open format, an agent skill is a reusable package of instructions and supporting resources, represented by a directory containing at minimum a SKILL.md file with metadata and instructions.
Review status: 2026-12-04
Technical explanation
The format permits scripts, references, and assets in addition to SKILL.md. The file’s initial YAML metadata section identifies a skill and explains what it does and when to use it; its Markdown body can provide operational instructions and supporting material can be loaded when needed.
Conceptual boundaries
Here ‘skill’ means this packaged open-format artifact. It is not a general claim that an agent or model has an innate, learned, or reliably executable capability.
Provider-neutral example
Illustrative: a team packages a SKILL.md that tells an agent how to validate a CSV, alongside an optional script and a short data-dictionary reference.
Limitations
A valid directory structure does not prove that the instructions are correct, safe, available in a given environment, or approved for a specific task.
Related concepts
Atomic claims and evidence
1.1Agent Skills: Specification
- Source
- Agent Skills: Specification
- Source role
- Authoritative source
- Exact locator
- Directory structure
- Supported claim
- The Agent Skills specification defines a skill as a directory that contains at minimum a SKILL.md file.
- Last verification
- Review due
- Scope limitation
- This is a statement about the Agent Skills format, not a universal definition for every product’s extension or capability system.
2.1Agent Skills: Specification
- Source
- Agent Skills: Specification
- Source role
- Authoritative source
- Exact locator
- ‘SKILL.md format’, ‘Body content’, and ‘Optional directories’
- Supported claim
- The format requires an initial SKILL.md YAML metadata section with name and description, followed by Markdown content; it also defines conventions for scripts, references, and assets.
- Last verification
- Review due
- Scope limitation
- The specification permits these contents but does not guarantee that every host loads, runs, or trusts optional files.
3.1Agent Skills: Specification
- Source
- Agent Skills: Specification
- Source role
- Authoritative source
- Exact locator
- ‘Progressive disclosure’
- Supported claim
- The specification recommends progressive disclosure: metadata is available before an agent loads the full instruction file and references needed for the task.
- Last verification
- Review due
- Scope limitation
- This is a format-design recommendation, not evidence that a particular agent performs a skill accurately or has access to every declared tool.