cross-vendor skill search

SKILL.md

---
description: find the best skill for a task across the entire cross-vendor index, ranked alongside the user's personal and team library. trigger when the user asks "is there a skill for X" or "find me a skill for Y".
---

# cross-vendor skill search

the right way to answer "is there already a skill for this". searches all sources at once, ranks them on the same scale, returns the top match with a one-line fit reason. saves the user from building something that already exists.

## intent

before building anything new, check if it already exists. the cross-vendor index has tens of thousands of skills across 9 sources, plus the user's personal and team library on top. the fastest way to a working solution is almost always "find the existing skill" not "write a new one from scratch".

## inputs

- a natural-language description of the task ("summarize a pdf", "generate cold emails", "review a typescript pr")
- optional: a source filter if the user prefers a specific registry
- optional: top-N (default 3) for how many candidates to show

## procedure

### step 1, run /implexa:run with the task description

```
/implexa:run <task description>
```

this fires the unified recommender. it hits two paths in parallel:
  - apply_org_skill against the user's personal plus team library
  - recommend_skills_for_context with skipGates=true against the full aggregated_skills index

both result sets get ranked on the same cosine-similarity scale, with a small score boost for implexa-curated and team-saved entries. the top-N return as a single list, each tagged with its source.

### step 2, read the source tags

every result carries a source tag in square brackets. interpret them:
  - [personal] or [team]: from the user's own library. cheapest to invoke, lowest friction.
  - [implexa]: founder-curated canonical pattern. high signal, opinionated.
  - [anthropic], [smithery], [clawhub], [skills.sh], [agentskills], [github], [cursor], [continue]: cross-vendor third-party. install one-click via apply_recommended_skill.

### step 3, apply the best fit

if the top match is from the user's library, run it via apply_org_skill. if it is a cross-vendor result, install it via apply_recommended_skill. the install is one-shot; the next time the user wants this they can run it directly.

### step 4, if nothing fits, record

zero good matches means the user has found a gap. invoke /implexa:record-skill to capture the workflow as it happens. now the next person who asks the same question gets a hit.

## decision points

- **multiple matches in the same score band (within 0.05 cosine)**: prefer in this order: team > personal > implexa > anthropic > the rest. familiarity beats novelty.
- **the top match is a [github] result with no install count**: open the source SKILL.md and read before installing. github-sourced rows are less curated than the registry sources.
- **the user explicitly wants a specific vendor**: pass sources=[<vendor>] to scope the search. do not silently expand.

## output contract

a ranked list of up to 3 candidates. each row:
  - rank number
  - skill name plus one-line description
  - source tag
  - similarity score (rounded to 2 decimals)
  - "apply" hint with the exact command to run

## outcome signal

success means the user runs one of the surfaced skills and either marks it useful (via submit_skill_feedback) or installs it permanently. a search that returns 3 candidates and none of them get used is a relevance miss, log it for tuning.

## notes

- this is the canonical "before you build, check what exists" pattern. teach new users to run /implexa:run before /implexa:record-skill.
- skipGates=true is the right flag for explicit user queries. it bypasses the ambient recommender's high-confidence threshold so even moderate-similarity matches surface.
- avoid filtering by source unless the user asked. the whole point of cross-vendor search is to let the best result win regardless of origin.