skill-search
Find the best skill for the current task across the entire skill library, including the dormant majority whose descriptions are never loaded into context. Use this before starting any nontrivial task
Files1
View on GitHubSKILL.mdskills/skill-search/SKILL.md
GitHubskill-search
Read _index.md, then one shard, from ./.claude/skill-atlas/catalog/
(cwd-relative). Missing → say search is unavailable, suggest /skill-atlas,
work without a skill; never use another directory's catalog.
1 — category
Index lines are name(count) — description — tokens. The description decides;
its …; see <other> carve-out is the strongest signal — treat it as routing.
Tokens only corroborate: the list truncates, so a hit favours a category and a
miss proves nothing. Counts overlap; a right bucket suffices.
2 — shard
Entries are ## <id> [enabled|searchable] (stale), description, path, and
also in: — its other categories, where a sharper sibling may live. Prefer
specific over broad; between near-identical entries take [enabled], then the
sharper description, silently.
A second shard only if the winner's also in: names an unread category, or
nothing rose above weak. Never a third. uncategorized.md counts; serving from
it, say so and suggest /skill-atlas.
3 — name it, then work
One line, then continue the task in the same turn:
Using <id>. — or No skill fits — proceeding without one.
Nothing else, ever: never narrate the search, name a category or shard, list
runners-up, echo a path, announce intent, ask permission, or summarise after.
Reading a [searchable] SKILL.md is internal work.
Using it
[enabled] → Skill tool. [searchable] → Read that path and follow it as
plain instructions; allowed-tools and auto-triggering do not apply. Never
look outside catalog/ — absent from every shard means disabled by choice, and
"no skill fits" is a first-class answer. Never force one.