use-skill

# Use Skill A **user-invoked** skill is one your harness hides from you and starts only when a person invokes it by hand, usually with a slash command. A person telling you to use it in plain words carries the same intent. This skill turns that naming into the invocation: find the skill's file, prepare it the way invocation would, and follow it. Run the skill the user named, and only that one. When it is in your list of skills, invoke it yourself instead: your harness then prepares it for you. When your harness refuses it, stop and tell the user it is theirs to invoke, with the command to type. ## Procedure 1. **Find.** A skill follows the [Agent Skills specification](https://agentskills.io/specification): a folder holding a `SKILL.md` whose frontmatter `name` matches the folder's name. Harnesses keep skills in folders such as `.agents/skills/`, `.<harness>/skills/` and installed plugins', in the project and in the home directory, sometimes several levels deep. Search for `**/SKILL.md` with your file-search tool from two roots: the project root and `$HOME`. Match the user's whole naming, including an author or source they give, against what you find: first the `name` and the skill's path; then the `description`, for a skill named by what it does. Copies of one skill (several plugin versions, a marketplace clone) count as one: read the newest. When the project and the home directory both have a skill with that name, the project's wins. Done when exactly one skill matches. - Several different skills match: ask the user to pick, one line per candidate with its name, path and description. - None match: tell the user where you searched. A skill installed only on another machine is not reachable from here. 2. **Prepare.** Read the whole file, and resolve its relative paths against the skill's folder, as the specification does. Beyond the specification, a harness can transform a skill when it invokes it: substitute the arguments into placeholders, insert the output of shell commands. Do each of those yourself, the way your harness would; its documentation is the source for syntax you don't recognize. The **arguments** are what the user would have typed after the slash command ("use the tdd skill on the parser" gives `the parser`). Done when nothing is left in the text that invocation would have replaced. 3. **Run.** Before anything else, tell the user in one line which skill you are running and the full path to its `SKILL.md`. Then follow the prepared skill as if they had invoked it, to the skill's own completion criteria. When the skill declares that it runs in a separate context, hand it to a subagent as its prompt and relay the result. When the skill's work goes to another agent, put the skill's name and full path in the hand-off, so that agent reads and runs the file itself. A summary loses the skill's detail. ## Listing user-invoked skills When the user asks which skills only they can invoke, search the same folders, keep the skills your harness marks as user-invoked, and list each on one line: its name, path and description.