Skills: Giving an Agent Know-How on Demand
Learn how skills give an agent detailed know-how only when a task needs it, keeping its context lean.
Picture an office assistant agent that should follow your brand’s style guide, build slide decks the way your team likes, apply the expense rules, write SQL reports, and run a legal checklist. Each of those needs pages of detailed instructions, maybe with templates or small scripts.
The obvious approach is to put all of it in the system prompt. It works, badly: tens of thousands of tokens are sent with every single request, most of them irrelevant to the task at hand. That’s slow, expensive, and, as the last few lessons showed, distracting.
Skills take a smarter route, often called progressive disclosure. Each skill is a folder: a short name and one-line description, detailed instructions, and optionally templates or scripts. Only the descriptions are always in the agent’s context, a few dozen tokens each. When a task matches, say “make a slide deck in our brand style”, the agent opens just those skills and reads their full instructions, then gets to work.
It’s like a new employee with a well-labelled shelf of manuals. They don’t memorise every manual on day one; they know which ones exist and pull the right one off the shelf when needed.
Skills complement the other pieces you’ve met. Tools and MCP give an agent abilities (things it can do); knowledge from RAG gives it facts; skills give it know-how: the steps, standards and tricks for doing a kind of task well. And because they’re just folders of text, teams can write, share and improve skills without retraining anything.
One caution: a skill may include scripts the agent can run, so, like MCP servers, skills should come from sources you trust.
Key takeaways
- Putting every instruction in the system prompt is slow, expensive and distracting.
- Skills are folders of instructions (plus templates or scripts) loaded only when a task needs them.
- Only each skill’s short description is always in context: progressive disclosure.
- Tools give abilities, RAG gives facts, skills give know-how; trust the source of any skill with scripts.
Check your understanding
Go deeper (more technical)
Anthropic introduced “Agent Skills” in October 2025: a folder with a SKILL.md file whose short header (name and description) is always visible, and whose body plus any extra files are read only when needed. Similar ideas appear in other agent tools as instruction files or reusable “recipes”.
Skills can be nested: a skill’s instructions can point to further files that are loaded only if the task goes that deep, keeping even large bodies of know-how cheap until they’re used.
The simulation’s skills, sizes and costs are illustrative.