How a Skill Gets Used

A "skill" is a packaged bundle of instructions — and sometimes scripts or templates — for doing one specific kind of task well: filling out a PDF form, building a slide deck, following your team's code-review checklist. The whole point is that an agent can have dozens of these available without any of them costing context until the moment they're actually needed.

That only works because skills load in stages instead of all at once. If every skill's full instructions sat in context on every single turn, a handful of skills would burn the whole context window before the conversation even started. So the system is deliberately lazy: the agent sees just a name and a one-line description for everything it could do, and only pulls in the full instructions for the one skill that actually matches what you asked.

This is the same "lost in the middle" problem as a crowded context window, one layer up — except here the fix is structural rather than something you have to manage by hand. The agent doesn't decide what to forget; it decides what to load in the first place.