Skill Catalog

Find useful context

The filesystem and skill frontmatter are canonical; generated catalogs are human browsing aids. The canonical tree is this directory's parent (../), including cold capabilities omitted from the host's startup list. Skip discovery when a visible skill clearly owns the request or ordinary project source is enough.

Search bounded filenames and descriptions for the active question. From the skill-tree root, rg -l --glob SKILL.md 'migration|codemod' . is one possible candidate search, not a scoring system or required command. Inspect plausible candidates' name, description, kind, invocation and visibility, then load what helps. Combine matches for distinct needs, not every overlapping procedure. Hidden/manual skills remain usable through explicit requests, prompts, project instructions or this search. No Python, generated inventory or model resolver is required; a search miss does not require building a new inventory.

For a foundation, inspect its compact topic map, search reference filenames and headings, then open the likely capsules. Use the index only when discovery remains ambiguous; it is an inventory, not a mandatory context load. Check each selected capsule's own revision, limits and source links. A remembered graph project or tool name is a retrieval hint, not a required integration.

For maintenance, change the owning frontmatter and regenerate the retained operational/foundation catalogs. Promote nothing automatically. Hot exposure needs recurring use, reliable selection, and distinct demonstrated lift. Missing usage telemetry is unknown, not proof that a skill is unused. Keep meaning and promotion decisions with the model; scripts validate metadata and disjoint surfaces. Hiding a description does not necessarily prevent the host from scanning its body.

Verification

Check explicit callers and host visibility after metadata changes. Run the target checkout's applicable publication checks; do not run template scripts in an unrelated project. ../../scripts/skill-catalog.py provides optional list, search, context, invocation-size and generated-view diagnostics. invocation <name> inspects one loader; invocation --limit 10 lists the largest tracked loaders. These are optional inventories, not runtime-cost measurements or publication limits. The model owns the approach, not these diagnostic commands.