A prompt is a one-off instruction. A skill is the same instruction made durable: a SKILL.md file with a name, a description of when it applies and the detailed procedure, next to the examples, templates and scripts it needs. The description is how it gets found - an agent, or a router, reads it and decides the skill matches the task - and the body is loaded only then, so a library of a hundred skills costs nothing until one is used.
Skills sit between prompts and agents in a useful way. A tool does one thing when called with arguments. An agent decides and acts autonomously, calling tools. A skill shapes how a model or an agent approaches a class of task - how this team writes release notes, how to review a contract for this jurisdiction, how to structure a root-cause analysis - without being a program. Most expertise organisations want to capture is of that kind.
A router can select skills the way it selects models. Each skill is a target in the catalogue with its description, domains and the actions it serves; a request that matches is routed to the skill, which then runs on a model the router also chose. Skills can be chained into plans - a persona, then a skill, then a model - and a rule can pin a kind of request to a skill outright, as a translation glossary or an incident-review procedure usually should be.
Because a skill is a folder of text it is easy to version, review and publish. The marketplace indexes skills from public repositories alongside the ones publishers list directly, with the SKILL.md rendered and rated; the open capability manifest describes a skill with the same fields as a model or an agent, so one catalogue holds all of them. The repository's own coding playbooks are skills, and the router can select them like any other.
