Skill Capability Matcher

Skill Capability Matcher

Match natural language requirements to available skills, designing an optimal workflow step sequence.

Purpose

Parse user's workflow description, understand their intent, and recommend which skills should handle each part of the workflow. Output includes step IDs, skill names, missions, and data flow.

Process

Step 1: Parse Requirements

Extract from the user's description:

Input Types:

• video transcript
• audio transcript
• article/blog post
• documentation
• raw text
• structured data (JSON/YAML)

Processing Goals:

• analyze (deep understanding)
• summarize (condensation)
• generate (create new content)
• transform (change format)
• extract (pull specific data)
• validate (check correctness)

Output Format:

• JSON structure
• Markdown document
• Summary text
• Structured data

Step 2: Load Registry

Read the plugin registry from registry-loader output (v0.7.0):

{
  "plugins": [...],
  "summary": { "totalSkills": N, "installedSkills": N, "availableSkills": N }
}

Build a capability index from skill descriptions and inferred capabilities. Include installed status when scoring - prefer installed skills for immediate execution.

Step 3: Match Capabilities

Score each skill by:

1. Description match - Does skill description align with requirements?
2. Capability overlap - Do inferred capabilities match processing goals?
3. Input/output compatibility - Can skill handle input type and produce expected output?

Step 4: Design Step Sequence

For each matched skill:

1. Create a step ID (kebab-case, descriptive)
2. Determine dependencies (needs:)
3. Write a mission description (what to accomplish)
4. Define data flow (input → output)

Step 5: Recommend Sequence

Order skills logically:

1. Analysis skills first (understand content)
2. Generation/transformation skills second
3. Assembly/output skills last

Ensure proper data dependencies.

Step 6: Flag Gaps

If requirements can't be fully satisfied:

• List unmatched capabilities as gaps
• Suggest creating custom skills if needed
• Indicate if workflow is partial

Input

Provide:

1. User's natural language description
2. Registry JSON from plugin-registry-scanner

Output Schema

{
  "requirements": {
    "inputType": "video transcript",
    "goals": ["extract key points", "generate outline"],
    "outputFormat": "structured JSON"
  },
  "recommendations": [
    {
      "skill": "media-reviewer",
      "suggestedStepId": "analyze-content",
      "goalId": "analyze",
      "matchScore": 0.92,
      "capabilities": ["content analysis", "theme extraction"],
      "mission": "Deep analysis of video transcript. Extract key themes, quotes, and narrative structure.",
      "rationale": "Primary skill for content understanding"
    },
    {
      "skill": "idea-synthesis",
      "suggestedStepId": "generate-ideas",
      "goalId": "generate",
      "matchScore": 0.85,
      "capabilities": ["idea generation", "hooks and angles"],
      "mission": "Generate hooks, angles, and questions from the analysis. Read user profile for personalization.",
      "rationale": "Creates engaging content ideas from analysis"
    }
  ],
  "suggestedSequence": ["analyze-content", "generate-ideas", "build-output"],
  "dataFlow": {
    "analyze-content": {
      "needs": [],
      "provides": "analysis.json"
    },
    "generate-ideas": {
      "needs": ["analyze-content"],
      "provides": "ideas.json"
    },
    "build-output": {
      "needs": ["analyze-content", "generate-ideas"],
      "provides": "output.json"
    }
  },
  "gaps": [],
  "customSkillNeeded": false,
  "clarificationNeeded": true,
  "clarifications": {
    "sections": [
      {
        "id": "input",
        "title": "Input",
        "completed": false,
        "questions": [
          {
            "id": "content-type",
            "text": "What type of content will this workflow process?",
            "type": "single-select",
            "options": [
              { "id": "video", "label": "Video transcripts", "inferred": true },
              { "id": "audio", "label": "Audio transcripts" },
              { "id": "text", "label": "Text articles" },
              { "id": "web", "label": "Web pages (fetched via search)" }
            ],
            "reason": "Inferred 'video' from description, confirm or change"
          }
        ]
      },
      {
        "id": "goals",
        "title": "Goals",
        "completed": false,
        "questions": [
          {
            "id": "primary-goal",
            "text": "What are the primary goals for this workflow?",
            "type": "multi-select",
            "options": [
              { "id": "analyze", "label": "Analyze and extract key insights" },
              { "id": "summarize", "label": "Create structured summaries" },
              { "id": "generate", "label": "Generate creative content ideas" },
              { "id": "document", "label": "Build comprehensive reports" }
            ]
          },
          {
            "id": "depth",
            "text": "How deep should the analysis be?",
            "type": "single-select",
            "options": [
              { "id": "quick", "label": "Quick overview (1-2 key points)" },
              { "id": "standard", "label": "Standard analysis (5-7 key points)" },
              { "id": "deep", "label": "Deep analysis (comprehensive)" }
            ]
          }
        ]
      },
      {
        "id": "output",
        "title": "Output",
        "completed": false,
        "questions": [
          {
            "id": "format",
            "text": "What output format do you need?",
            "type": "single-select",
            "options": [
              { "id": "json", "label": "Structured JSON" },
              { "id": "markdown", "label": "Markdown document" },
              { "id": "both", "label": "Both JSON and Markdown" }
            ]
          }
        ]
      },
      {
        "id": "review",
        "title": "Review",
        "completed": false,
        "questions": []
      }
    ]
  }
}

Scoring Guidelines

Scoring Rubric

Use these concrete criteria when assigning scores:

Relevance Score (Primary Factor)

• 90-100: Skill's primary purpose directly addresses requirement
• 70-89: Skill clearly capable of addressing requirement
• 50-69: Skill partially addresses requirement
• 30-49: Skill tangentially related
• 0-29: Minimal or no relevance

Completeness Score

• 90-100: Skill fully addresses all aspects of requirement
• 70-89: Skill addresses most aspects, minor gaps
• 50-69: Skill addresses core need, notable gaps
• Below 50: Significant portions unaddressed

Specificity Score

• 90-100: Skill specifically designed for this exact use case
• 70-89: Skill well-suited for use case category
• 50-69: General-purpose skill applicable to use case
• Below 50: Very general skill with broad applicability

Confidence Calculation

confidence = (relevance * 0.5) + (completeness * 0.3) + (specificity * 0.2)

Minimum threshold: 60% - Skills below this should not be recommended.

Semantic Matching Guidelines

Match requirements to skills beyond simple keyword matching:

1. Synonym Recognition

Match conceptually equivalent terms:

• "authentication" ↔ "login", "session management", "user verification"
• "database" ↔ "data persistence", "storage", specific ORMs
• "analyze" ↔ "review", "examine", "inspect", "assess"
• "generate" ↔ "create", "produce", "synthesize", "build"

2. Hierarchical Matching

Understand skill scope and specificity:

• General skill may satisfy specific requirement (e.g., "api-client" for "REST calls")
• Prefer specific skill over general when available (e.g., "stripe-integration" over "payment-api")

3. Intent Extraction

Focus on what user wants to accomplish, not just keywords:

• "I need to send emails" → match email/notification skills
• "Process video content" → match media-reviewer, not just skills with "video" in name
• "Build a report from data" → match assembly/documentation skills

Edge Cases

Mission Writing Guidelines

Each mission should:

• Start with an action verb (Analyze, Extract, Generate, Create, Transform)
• Describe the specific goal for this step
• Mention key outputs expected
• Reference context from previous steps if applicable
• Be 2-4 sentences

Good mission example:

Deep analysis of video transcript. Extract key themes, important quotes with timestamps, and narrative structure. Focus on insights that would interest the user based on their profile.

Bad mission example:

Analyze the content.

Important Rules

1. Skills-first - Never recommend agents, only skills
2. Include missions - Every recommendation must have a detailed mission
3. Score realistically - Don't inflate match scores
4. Complete data flow - Ensure all dependencies are resolvable
5. Flag gaps honestly - Report capabilities that can't be matched
6. Include goalId - Each recommendation must have a goalId linking it to a clarification goal

Clarifications Schema (v0.6.4)

When user requirements are ambiguous, include clarifying questions in the response. The wizard UI uses these to gather additional context before generating the final workflow.

Clarification Fields

Section Schema

{
  "id": "goals",
  "title": "Goals",
  "completed": false,
  "questions": [...]
}

Question Schema

{
  "id": "primary-goal",
  "text": "What are the primary goals for this workflow?",
  "type": "single-select | multi-select | text",
  "options": [
    { "id": "analyze", "label": "Analyze and extract key insights", "inferred": true }
  ],
  "reason": "Optional explanation for why this was inferred"
}

Question Types

Inferred Values

Set inferred: true on options that match keywords in the user's description:

• "video" or "youtube" → infer video content type
• "analyze" → infer analyze goal
• "summar" → infer summarize goal
• "generat" or "creat" → infer generate goal

Recommendation goalId

Each recommendation must include a goalId that links to a goal option:

{
  "skill": "media-reviewer",
  "suggestedStepId": "analyze-content",
  "goalId": "analyze",
  "matchScore": 0.92
}

This allows the wizard to filter recommendations based on selected goals in real-time.

Structured Preferences for Enriched Prompt (v0.6.4)

When wizard answers are included in the enriched prompt (via "User clarifications: Q: ... A: ..."), these become structured preferences that MUST be incorporated into the workflow.

Common Preference Types

Preference Extraction from Enriched Prompt

When receiving an enriched prompt like:

User clarifications: Q: Which platforms? A: twitter, linkedin. Q: How many articles? A: top5. Q: Focus areas? A: llm, adoption.

Extract as structured data:

PLATFORMS: twitter, linkedin
ARTICLE_COUNT: 5
FOCUS_AREAS: llm, adoption

Mission Templates with Preferences

WITHOUT preferences (BAD):

Search for AI news articles.

WITH preferences (GOOD):

Search for top 5 AI news articles focusing on LLM developments and adoption trends.

WITHOUT preferences (BAD):

Compile findings into a report.

WITH preferences (GOOD):

Create engaging social media posts optimized for twitter and linkedin, focusing on LLM and adoption angles.