SKILL.md file describing
a capability, optionally accompanied by reference documents and scripts in the
same directory. Skills are indexed into a dedicated vector collection and reached
by two core tools: search_skills(query) runs a semantic search and returns
matching skills along with the reference files each one offers, and
load_skill_ref(skill_name, filename) pulls a specific reference or script into
context only when it is actually needed. That two-step design keeps a large
library affordable: the index rides in the prompt, full bodies do not.
It will also cover indexing (python -m skills.loader walks the skill roots and
rebuilds the collection), the three-root discovery chain, and the source labels
Ziro attaches to every reachable skill: bundled for shipped skills, workspace
for a project’s own ./skills/ directory, and plugin for skills surfaced by an
installed plugin. The /skills command shows the label next to each name.
Subagents can be scoped to an explicit skill allowlist in their frontmatter, in
which case searches over-fetch and keep only the declared skills.
This page is in progress. Until it lands, the
SKILL.md files under the
repository’s skills/ directory are the best format reference.