Skillsmp Skill Recommender

SkillsMP Skill Recommender

Select high-fit SkillsMP skills for a specific project and return a practical keep / replace / add plan.

When to Use This Skill

• Building a production skill set for a new project
• Replacing low-quality or low-fit skills in an existing stack
• Comparing candidate skills for the same capability
• Auditing whether current skills cover required delivery areas
• Creating a focused install list from a broad project description

Invocation

User invokes this skill by describing the project they want to build.

Examples:

• "Use $skillsmp-skill-recommender to build a production skill stack for my project."
• "Use $skillsmp-skill-recommender and suggest skills even if my stack is not finalized yet."

Input Contract (Required)

Extract these fields before ranking:

• target product type
• stack and frameworks (or enough constraints to select them)
• priority outcomes (e.g., speed, security, UX, CI/CD maturity)
• team level (solo/small team/enterprise-like)
• existing skill list (optional but preferred)

If enough context is present, continue to search.

If context is insufficient, ask clarifying questions before recommending.

Clarifying Questions (If Needed)

Ask up to 5 short questions, only for missing critical inputs.

Prioritize in this order:

• what exactly to build
• required stack/constraints (or whether stack selection is needed first)
• expected production level
• must-have areas (security/testing/release/etc.)
• current skills to replace (if any)

Do not ask questions when the request is already sufficiently specific.

Stack Selection (Required Before Search)

If the user did not provide a finalized stack, you must select one first.

Rules:

• infer the best-fit implementation stack from the product description, goals, constraints, and team size
• prefer practical production-ready defaults over exotic combinations
• state the chosen stack explicitly before skill search
• only after stack selection, run capability discovery and search queries for that chosen stack
• do not ask the user to choose stack unless there is a hard technical constraint conflict

Do not skip stack selection when stack is missing or vague.

Capability Discovery (Dynamic)

Build groups from the user intent, not from a fixed preset.

Method:

• extract nouns (tools/platforms/domains)
• extract verbs (build/test/deploy/secure/monitor/optimize)
• merge overlaps into implementation-oriented groups
• add cross-cutting groups only when needed

Use references/goal-query-map.md only as a seed source. Never force platform- or stack-specific groups unless they are requested or clearly implied.

Search Strategy

Preferred sources:

1. SkillsMP website UI in MCP Playwright browser
2. Public SkillsMP docs pages only for orientation, not API querying

Run 2-6 focused queries per discovered group in the site search UI.

Execution method (required):

• open SkillsMP site with MCP Playwright browser
• navigate to the skills/catalog search interface
• if authentication is required, complete it via browser login/session flow
• type each query manually in the web search input and submit
• collect candidate skills from rendered UI results
• do not request or use API tokens for this workflow
• do not call API endpoints directly from terminal or HTTP client
• after collecting results, close the Playwright browser session before returning recommendations

Hard restriction:

• never use curl, fetch, or direct calls to SkillsMP API endpoints such as /api/v1/skills/search or /api/v1/skills/ai-search (token-protected)

Search quality rules:

• keep queries short and concrete
• avoid broad mega-queries
• collect at least top 5 candidates per group
• deduplicate by id and githubUrl

Filtering and Ranking

Reject candidates when:

• they are repo-internal maintenance only
• they are stack-incompatible
• they are mirrors of a better canonical source
• description is vague and non-actionable

Use relevance tiers:

• Tier A: direct fit for high-priority groups
• Tier B: useful adjacent support
• Tier C: generic fallback

Rank with this priority:

1. relevance tier
2. practical stack fit
3. stars/forks as quality signal
4. recency signal if available

Never let stars override relevance.

Decision Rules

For each current skill slot:

• Replace: clearly better relevance and quality
• Keep: already strong or no meaningful upgrade
• Add: missing capability not covered today

When no strong upgrade exists, explicitly keep.

Output Format (Required Order)

Return sections in this exact order:

1. Discovered Capability Groups
2. Top Install/Replace List
3. Replace Matrix
4. Keep As-Is
5. Coverage Check
6. Source Notes

Each recommended skill must include:

• name
• source repo/author
• stars
• fit rationale
• priority (P0 / P1 / P2)
• source note from web UI search context

Quality Bar

• Prefer canonical upstream sources.
• Prefer workflow-driven skills over vague guidance.
• Prefer reusable cross-project skills over one-repo internals.
• Mark uncertain recommendations explicitly.
• If context is still insufficient after clarification, say so.

Reference

• Seed query map (non-exhaustive): references/goal-query-map.md