Getting started

With the client skill installed you don't drive the tools yourself — you ask your agent for work, and it runs this loop invisibly:

The consumption loop

1. Search — skill_search(query) returns lightweight matches (name, description, scores) and a covered verdict. Bodies are never returned by search, so your context stays small.
2. Get — skill_get(name) downloads the full skill. It arrives as a compact CRUX document your agent reads in place (see Reading CRUX). The load is recorded so feedback is accepted later.
3. Apply — the agent follows the skill's workflow, edge cases, and output format.
4. Feedback — skill_feedback(name, "worked"|"failed"), optionally with measured tokens/seconds (see Objectives & telemetry).
5. Request — when search says covered: false, the agent briefs the platform via skill_request and a new skill is authored (see Requesting new skills).

Try it

Prompts to give your agent after install:

"Find a skill for summarising long documents and use it on report.pdf."

"We keep re-explaining our release-notes format. Search the corpus; if
nothing covers it, request a skill for it with a proper brief."

"Check whether any of our pinned skills have newer versions."

What your agent reports back

After using a corpus skill, the client skill has your agent state: the skill slug and version used, whether it was served compressed, the action taken, the feedback it reported, and any pending update notice — e.g.:

Used summarise-document@1f3a9c2b v2 (CRUX), reported worked.
Skill update available: summarise-document v3. Say "accept skill updates"
to adopt.