A skill is a way to turn a repeatable procedure into a product. Technically it is a folder with a required SKILL.md file, whose header requires a name and a description, plus optional directories for scripts, reference material and supporting files. Cursor looks for them in several standard places, including compatible directories of other agent tools, and a nested project skill automatically gets the scope of its subdirectory. The point of the mechanism is not storing text but that the procedure loads when it is needed and takes up no room until it is.

The naive alternative is to explain the same sequence of actions anew in every conversation. It works but scales badly: the wording differs a little each time, steps get lost, and the result depends on how thoroughly you wrote the request today. This shows worst when work is handed over: what lives in your habit of typing a request does not travel to a colleague along with the repository. A skill fixes the order, and from then on the procedure runs identically, while its changes go through ordinary review like code.

Progressive loading is the main engineering idea here. In the initial context only the skill's name and description are visible; the full content is pulled in on activation, and heavy material - references, templates, scripts - only when it is actually reached. That is why the main file is kept short: the trigger, the prerequisites, ordered steps, stop conditions and the required evidence. Everything bulky moves into separate directories. The saving is not decorative: a dozen skills each loading in full would eat the context before the work even starts.

It helps to see a skill's skeleton in full once. Below are a header with a name, a description and path patterns, and a body of four steps where the last step is an explicit stop before an action with external consequences. It is exactly the stop condition that separates an instruction skill from an automaton: the procedure brings the work to the point of decision and hands it to a human. Note the demand for evidence in the last item as well: without it a report turns into a claim that all is well, and there is nothing to check that claim against.

The description in the header carries more weight than it seems. The model uses it to decide whether the skill is relevant to the current task, so it must name the moment of application rather than an abstract capability. The wording validate a release candidate and produce a checklist works; the wording helps with releases does not. There is also the reverse case: if a skill should be started only by a human, automatic selection is disabled with an explicit field. That is done with procedures whose cost of a false trigger is higher than the benefit of switching on by itself.