What are agent skills?
Agent skills are packaged instructions, examples, reference files, and tools that an AI agent loads to perform a specific kind of task. A skill is usually a folder. Its main file starts with a name and a short description of when to use it, followed by instructions. Other files hold templates, examples, style rules, or scripts.
The agent’s system prompt lists only the name and description of each skill. When a request matches one, the agent reads the full instructions and follows them. Loading on demand keeps the context window small, so one agent can carry many skills. A tool gives an agent an action, while a skill teaches it a type of work and can bundle tools.
Agent skills in practice
A law firm packages its method for reviewing vendor contracts as a skill. It holds the firm’s fallback positions by clause type, two annotated example reviews, the terms that always go to a partner, and a script that formats the redline. Agents load it only when asked to review a contract. When the knowledge management team changes a fallback position, it edits one file, and every agent applies the change on the next review.
What to watch for
- Review third-party skills like code before installing them, because their scripts run with the agent’s permissions.
- Keep skills under version control with a named owner who approves changes.
- Test a skill on real tasks after each edit, since small wording changes can alter agent behavior.