source-a-skill

source-a-skill

sourcewriteshands-off

Use this when: you know exactly which capability you need

Problem it solves — When you know exactly which capability you need, divergent discovery is wasted effort. This is convergent: go find the best known version of that one skill and bring it back, judged by vet before adoption.

Used in workflows: Grow the harness

Source a skill

You already know the gap. Don't author blind — someone has probably solved it well. This finds that, judges it, and adopts it.

The flow

1. Name the need precisely — the task class, the trigger, the must-haves. A fuzzy need finds fuzzy matches; sharpen it first (or send it through vet/brainstorm).
2. Scan — dispatch the scanner agent (subagent_type: scanner) to shortlist real candidates from current skill catalogs / repos, with links + maturity signals. Keeps the scrape noise out of here.
3. Vet each — run the shortlist through vet: real fit for our five pillars? adopt / adapt / skip + risk (boost collision, drift, maintenance). Convergent doesn't mean uncritical — most candidates should still lose.
4. Adopt the winner — via build-claude-skill: bring it to house convention (name by verb, the contract, tiered confirm, (formerly upstream/<name>) provenance), adapt rather than copy if it's close-but-not-ours, and wire its trigger. Credit the source — add a row to docs/agents/sources.md (what · where it landed · how tweaked) so the lineage survives the PR.

Where it sits

• Not qa-skills — that's divergent ("what am I missing across the catalog?"); this is convergent ("get the best known thing for this named need"). Both end at vet.
• Not build-claude-skill — that authors from scratch; source brings in existing work (and then uses build-claude-skill to land it).
• Sibling source-* skills acquire other things for a known need — packages, agents — same flow, different object.