What are Agent Skills?

An agent skill packages a procedure, checklist or set of conventions as a folder. At its center is a SKILL.md file holding YAML frontmatter and markdown instructions. The folder can also carry reference documents and scripts that the instructions point to. Claude Code documents skills as a way to add repeated instructions without writing them into every prompt, and a skill can be triggered by the model when it looks relevant or called directly with a slash command.

Skills are loaded in stages to keep the context window small. The name and description of each skill stay in context so the model knows what exists. The full SKILL.md body enters the conversation only when the skill is invoked, and supporting files are read on demand. Scripts are run rather than read into context.

Claude Code's docs state that its skills follow the Agent Skills open standard, which is meant to work across several AI tools, and that Claude Code adds its own extensions on top, such as invocation controls and running a skill in a subagent. Skills can live in a personal directory, a project directory, a plugin, or an organization-managed location, and the location decides which sessions can use them.

One caveat from the docs: fields beyond the standard set work in Claude Code but are rejected in some other upload paths, so portable skills should stick to the standard fields.