Agent Skills are a way to write a work procedure in a single SKILL.md and give it to an agent. Instead of writing a long prompt every time, you keep the procedure in a file and the agent reads and follows it when needed. Whether to load a skill into the main context or run it as a separate subagent is the next choice, and it does not always pay off.
A comparison across 7 models and 64 tasks found that the gain from running skills as subagents appeared only when there was an input/output contract and disappeared without one, while total tokens rose 1.3–1.7×. SKILL.md structure and authoring, how marketplace installs are distributed, and what to check in a security review are covered too.