What's this about?
Chinese AI educator Baoyu answered a common confusion about agent Skills: people assume the model is called once to choose a skill and again to execute it. In fact every skill's name and description is loaded into the system prompt from the first turn, so the model reads your request and picks the skill in the same single call, then pulls in the skill's full instructions on demand.
What exactly is a "Skill"?
A Skill is a packaged instruction set an agent can invoke for a specific kind of task — a recipe card in a binder. The binder's table of contents (names and one-line descriptions) is always in view, so choosing a recipe costs nothing extra; only when a card is picked does its full text get read into working memory.
Why does it matter?
Mental models of agent internals lag the products badly, and pricing/latency intuitions built on wrong models lead to wrong architecture choices. This is the kind of one-tweet correction that quietly fixes a thousand system designs.