What is Agent Skill?
Also called skill, capability module.
An agent skill is a packaged unit of instruction that teaches an agent how to perform a particular kind of task, typically containing procedures, conventions, and references, and loaded into context only when relevant. Skills keep the base prompt small while making specialized know-how available on demand. The term is used differently across frameworks, so the details vary.
The mechanism is progressive disclosure. Short descriptions of every available skill sit in context, and the full body of one is loaded only when the current task matches it. This keeps the standing prompt small while giving the agent access to far more procedural detail than could be included at once, and it makes each procedure editable in isolation.
The distinction from neighboring concepts matters. A tool is something the agent can execute. A skill is knowledge about how to do something, including which tools to use and in what order. The system prompt holds what applies always. A rule of thumb is that anything applying to only part of the work belongs in a skill rather than the prompt.
Skills work best when narrowly scoped and written as procedures with concrete steps, examples, and failure handling. Overlapping skills cause the wrong one to load. Skills that duplicate what a tool description already says add tokens without value. Because they change behavior, they deserve the same review and version control as code rather than being edited casually.
Terminology is not settled. Some frameworks call these skills, others call them capabilities, playbooks, or instruction modules, and one open format packages them as folders of documents with optional scripts. What they share is the pattern of on-demand instruction loading. Skills relate closely to the system prompt, the persona that scopes them, and the tools they direct.
Key points
- Loaded on demand so the base prompt stays small.
- A skill is know-how; a tool is an executable action.
- Narrow, procedural skills load more reliably than broad ones.
- Naming varies across frameworks; the loading pattern is shared.
In practice
An agent handling company expenses has a skill for processing receipts. Most conversations never load it. When a receipt image arrives, the skill loads and supplies the whole procedure: extract vendor, date, and total, match against the expense policy thresholds, flag anything over the limit, and file it under the right category. Once the task ends, that instruction leaves the working context.