An agent with tools but no guidance does the minimum: one quick search, then a hedged guess. A skill fixes that: a written method the agent can read, like a manual you look up when the situation calls for it. Attach one, and the same run follows the procedure: every claim sourced, every gap flagged.
What you will learn
A skill is a written method
Markdown with a name, a description of when it applies, and the steps: a repeatable procedure, not a saved prompt.
Loaded on demand, mid-run
Only the skill's name and description sit in the agent's context. The full method loads like a tool call, exactly when the agent decides it's relevant, no context crowding.
The effect on behavior
The same question, same tools: without the skill, a vague unsourced guess; with it, three targeted queries and an answer where every claim carries its source.
Every integration ships skills
You don't have to write them, integrations come with ready, battle-tested skills, and you can add your own to share across the workspace.
Here is the agent from the video with the skill attached:
The same run, with and without the skill
The video runs the same lead-qualification task twice. Without a skill, the agent searches once and guesses: "probably Series B, maybe a hundred people." With Exa's answer-with-citations skill attached, it loads the method mid-run, turns each claim into a specific factual question, pulls a sourced answer for every one, and returns a verdict where each figure carries its citation.
Why not just a longer prompt?
Skills keep procedure out of the prompt. Only the name and one-line description stay in view; the full method loads only when the agent judges it relevant: so you can attach many skills without bloating context, and the agent picks the right manual for the moment.
Ready to use
Adding a skill is the same motion as adding a tool: pick it in the block's Skills section and it shows on the block. Integrations ship curated skills out of the box, and skills you write are shared across your workspace.