An agent skill is a reusable package that teaches an AI agent how to handle a particular job. Instead of pasting the same long prompt every time, a team can bundle its instructions, examples, templates, reference material, and sometimes executable scripts into a capability the agent loads when the task calls for it.
In systems that follow the Agent Skills format, the entry point is usually a SKILL.md file: a short description says when the skill applies, while the body explains the workflow. Supporting files can hold checklists, brand examples, document templates, or validation scripts. Loading that material only when needed keeps it from occupying the context window on every conversation.
A skill is not the same thing as a tool call. A tool gives the agent an action it can take, such as searching a database or editing a file. A skill teaches the agent how and when to combine instructions, tools, and judgment for a repeatable outcome. An agent hook, by contrast, fires automatically at a defined event.
Why it matters at your desk. A marketer might create a campaign-brief skill containing the approved voice, required fields, and review checklist. An engineer might encode the team's release procedure for work in Cursor. A writer using a persistent workspace such as Claude Projects can turn a recurring editorial routine into a named capability rather than rebuild the prompt from scratch.
The practical test is repeatability. Create a skill when the same multi-step instructions have proved useful more than once and the output can be checked. Keep one-off facts in ordinary project context, and keep permissions outside the skill: reusable instructions should not silently become reusable authority.