cross-vendor skill search

intent

before building anything new, check if it already exists. the cross-vendor index has tens of thousands of skills across 9 sources (anthropic, smithery, clawhub, skills.sh, agentskills, github, cursor, continue, implexa), 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". trigger this skill when the user asks "is there a skill for X" or "find me a skill for Y".

inputs

• task description (required): natural-language description of what the skill needs to do (e.g. "summarize a pdf", "generate cold emails", "review a typescript pr")
• source filter (optional): scope search to one or more vendors if the user has a preference (format: sources=[anthropic,github])
• top-N (optional): number of candidates to return, default 3, max 10
• implexa api (required): authenticated connection to /implexa:run endpoint. requires valid implexa account with skill index access. set env var IMPLEXA_API_KEY or rely on active session.
• org skill library (required): read access to personal and team skill library (implexa internal org store, auto-available if user is logged in)

procedure

1. invoke the unified recommender with the task description

   • input: task description (string), optional source filter, optional top-N parameter
   • run: /implexa:run <task description> [sources=<filter>] [top=<N>]
   • execution: fires two search paths in parallel against (a) personal plus team library via apply_org_skill and (b) full aggregated_skills index via recommend_skills_for_context with skipGates=true
   • output: ranked list of candidates, each tagged with source origin and cosine similarity score

2. parse result set and read source tags

   • input: ranked candidate list from step 1
   • logic: each result carries a source tag in square brackets. interpret as follows: [personal] or [team] = from user's own library, lowest friction; [implexa] = founder-curated canonical pattern, high signal; [anthropic] [smithery] [clawhub] [skills.sh] [agentskills] [github] [cursor] [continue] = cross-vendor third-party, one-click installable
   • output: annotated candidate list with source category identified for each row

3. select the best fit and apply

   • input: annotated candidate list with user's implicit or explicit preference
   • decision: if top match is [personal] or [team], invoke apply_org_skill <skill-id>; if cross-vendor, invoke apply_recommended_skill <skill-id> <source>
   • execution: install is one-shot. next time user wants this skill, they can run it directly from their library
   • output: skill installed and ready to execute, or error if install fails

4. if no good match found, record the gap

   • input: search completed with zero candidates or all candidates scored below 0.4 cosine similarity
   • action: invoke /implexa:record-skill to capture the workflow as the user executes it manually. tag with the original task description so future searchers hit this new skill
   • output: gap recorded for future indexing

decision points

• multiple candidates score within 0.05 cosine of each other: prefer in order: team > personal > implexa > anthropic > smithery > clawhub > skills.sh > agentskills > github > cursor > continue. familiarity and curation weight more than novelty.
• top match is a [github] result with zero or very low install count (under 5): open the source SKILL.md directly and read the procedure before applying. github-sourced entries are less curated than registry sources and may have incomplete documentation or edge cases.
• user explicitly requests a specific vendor: pass sources=[] and do not expand the scope. respect the constraint even if better matches exist in other sources.
• search returns zero results or all scores below 0.3: do not force a marginal match. inform the user that no suitable existing skill was found, then offer to record a new one via /implexa:record-skill.
• implexa api key is missing or expired: if IMPLEXA_API_KEY env var is unset or token has expired, the request will fail. surface this clearly. user must re-authenticate or provide a valid key before retry.
• network timeout or service unavailability: if the aggregated_skills index or org skill library is unreachable, /implexa:run will timeout after 10 seconds. retry once with a fresh request; if it fails again, fall back to manual search or escalate.

output contract

a ranked list of up to N candidates (default 3). each row must contain:

• rank (integer, 1-indexed)
• skill name (string, exactly as stored in source)
• one-line description (string, max 140 characters, auto-generated from skill metadata or first line of intent section)
• source tag (string in square brackets, e.g. [personal], [anthropic])
• similarity score (float, rounded to 2 decimal places, 0.0 to 1.0)
• apply command (string, exact shell command to invoke the skill, e.g. /implexa:run skill-id or apply_org_skill skill-id)

example output:

1. summarize-pdf-extract-text | extract text and generate summaries from pdf files | [implexa] | 0.94 | apply_recommended_skill summarize-pdf-extract-text implexa
2. pdf-to-markdown | convert pdf to structured markdown with fidelity | [anthropic] | 0.87 | apply_recommended_skill pdf-to-markdown anthropic
3. my-pdf-tool | my personal pdf processing flow | [personal] | 0.81 | apply_org_skill my-pdf-tool

outcome signal

success means:

• the user executes one of the surfaced candidates within 5 minutes of the search result
• the user submits feedback via submit_skill_feedback <skill-id> useful=true or marks it as their new default for that task type
• a skill is applied and runs without errors on the user's first invocation

failure modes to track:

• search returns 3+ candidates but zero are used within 1 hour (relevance miss, log for tuning)
• user runs the top result and it fails (ranking or documentation issue, flag source)
• user applies a skill and then immediately records a new one because the match was poor (poor ranking, retrain embeddings)