Agent Skills format¶

Definition¶

Agent Skills is an open format — originated by Anthropic, released as an open standard — for packaging agent capabilities as a folder containing a SKILL.md file (name and description metadata plus task instructions), optionally bundling scripts, templates, and reference material, portable across any skills-compatible agent.

Explanation¶

The format answers a context problem: agents are capable but lack the procedural and organisational knowledge to do real work reliably. Skills package that knowledge as version-controlled folders loaded through progressive disclosure in three stages — discovery (at startup only each skill's name and description load, enough to spot relevance), activation (a matching task pulls the full SKILL.md into context), and execution (the agent follows the instructions, running bundled code or reading referenced files as needed). Because full instructions load only on demand, an agent can hold many skills with a small standing context footprint. The design trades eager capability for context economy, and its portability claim — write once, use in any compliant client — depends on the format staying minimal: a folder and one markdown file.

Key Properties¶

A skill is a folder with SKILL.md: name + description metadata and instructions
Three-stage progressive disclosure: discovery, activation, execution
Standing context cost is names and descriptions only; instructions load on demand
Open standard with multi-product adoption beyond its originator

Relationships¶

No relationships recorded yet.
Code Mode for MCP — both apply progressive disclosure to agent capability: skills load full instructions only when a task matches, Code Mode loads tool schemas only when model-written code requests them

Applications¶

Encoding a team's review checklist, deploy procedure, or house style as a reusable agent capability. Distributing domain expertise across different agent products without per-product rewrites.

Sources¶

https://agentskills.io/home