skill-discovery（技能发现）

Skill Discovery Metasystem

Holographic system for discovering skills from the claude-plugins.dev registry (36,500+ skills) using progressive disclosure, Markov-chain intent classification, and Pareto-optimal ranking.

Quick Start (Depth 0)

When invoked, follow this 4-step workflow:

1. Classify Intent - Determine user's discovery pattern (DirectQuery/TaskBased/Exploratory/MetaDiscovery)
2. Execute Search - Use appropriate strategy based on intent
3. Rank Results - Apply Pareto optimization to surface best matches
4. Present Progressively - Show top 3 skills, expand on request

Intent Classification

Analyze user input to determine search strategy using Markov chain classifier.

State Space

Classification Logic

Extract features from user input:
- has_explicit_keywords (skill names, categories)
- describes_task (verb+noun patterns)
- is_exploratory (browse/show/list keywords)
- is_meta (questions about discovery itself)
- is_synthesis (combine/synthesize/meta-skill keywords)

Apply confidence thresholds:
- DirectQuery: ≥ 0.70
- TaskBased: ≥ 0.60
- Exploratory: ≥ 0.50
- MetaDiscovery: ≥ 0.75
- SkillSynthesis: ≥ 0.65
- SkillSynthesis: ≥ 0.65

Select highest confidence state above threshold

For detailed state transition matrix and probabilities, see [references/markov-chains.md]

Search Execution

Direct Query (Keywords Provided)

When: DirectQuery state, confidence ≥ 0.70

Execution:

Use WebFetch:
URL: https://claude-plugins.dev/skills?q={keywords}
Prompt: "Extract all skills with names, identifiers (@owner/repo/name), descriptions, and download counts. Return as structured list."

Alternative (deterministic):

./scripts/search_skills.py "{keywords}" --sort=relevance

Example:

Query: "testing debugging"
→ Search: "testing debugging"
→ Returns: test-driven-development, systematic-debugging, etc.

Task-Based (Semantic Expansion)

When: TaskBased state, describes problem/goal

Execution:

1. Expand user's task description with domain synonyms
2. Execute search with expanded query

Expansion Examples:

"debug" → "debugging troubleshooting error-handling diagnosis"
"frontend" → "frontend ui interface client-side web-design"
"build" → "building construction development creation"

See [references/search-strategies.md] for complete expansion algorithms.

Example:

Input: "I need to build a frontend interface"
→ Extract: ['build', 'frontend', 'interface']
→ Expand: "building creation frontend ui interface client-side web-design"
→ Search with expanded query
→ Returns: frontend-design, ui-builder, architecture-patterns, etc.

Exploratory (Browse Mode)

When: Exploratory state, user wants to see options

Execution:

Load: references/popular-skills-cache.md

Or fetch fresh:
URL: https://claude-plugins.dev/skills?sort=downloads
Prompt: "Extract top 20 skills sorted by downloads with names, identifiers, descriptions, and download counts"

Categories to present:

• Development (testing, debugging, refactoring)
• Design (frontend, architecture, patterns)
• Workflows (planning, execution, review)
• Meta (skill-creation, prompt-engineering)
• Deployment (ci-cd, release-automation)

Example:

Query: "show me popular skills"
→ Load popular-skills-cache.md
→ Present top 5 by category:
  Meta: skill-writer (96.1k), using-superpowers (7k)
  Design: frontend-design (45.1k), architecture-patterns (21k)
  Development: systematic-debugging (13k), test-driven-development (8.5k)

Meta-Discovery (Self-Reference)

When: MetaDiscovery state, questions about skill discovery

Execution:

1. Include this skill (skill-discovery) in results
2. Search: "skill discovery" | "finding skills" | "skill search"
3. Add complementary meta-skills:
   - skill-writer (@pytorch/pytorch)
   - using-superpowers (@obra/superpowers)
4. Optionally explain search process (Depth 3)

Example:

Query: "how do I find skills?"
→ Present:
  1. skill-discovery (this skill) - "You're using it now!"
  2. skill-writer (96.1k) - Guide for creating skills
  3. using-superpowers (7k) - How to find and use skills
→ Offer Depth 3 explanation of search strategies

Result Ranking

Apply Pareto optimization to balance relevance, popularity, recency, and diversity.

Ranking Formula

score(skill) = α·relevance(skill, query)
             + β·popularity(downloads)
             + γ·recency(last_update)
             - δ·redundancy(skill, presented_skills)

Default Configuration (Balanced):
  α = 0.60  (relevance - primary)
  β = 0.25  (popularity - secondary)
  γ = 0.10  (recency - tertiary)
  δ = 0.05  (redundancy penalty)

Component Functions

Relevance: Cosine similarity (bag-of-words) between query and skill description Popularity: log(downloads + 1) / log(100000 + 1) - logarithmic scaling Recency: exp(-days_since_update / 180) - exponential decay Redundancy: Max semantic similarity with already-presented skills

Deduplication

Remove semantically similar skills (threshold: 0.85):

./scripts/semantic_similarity.py results.json --threshold=0.85

For alternative configurations (High Precision, Popularity Mode, Exploration Mode), see [references/ranking-algorithms.md]

Pipeline Synthesis (Hypergraph Aggregation)

When the user asks to go beyond individual skills or requests "synthesis", aggregate multiple skills into a single pipeline that transcends any one skill.

Goal: Build a homoiconic, holographic, Pareto-optimized "meta-skill" that composes the best parts of multiple skills into one reusable workflow.

Synthesis Layers:

1. Pattern extraction from each skill
2. Capability matrix and complementarity analysis
3. Architectural role mapping (discover → decide → design → build → verify → deliver)
4. Pareto optimization across relevance, popularity, recency, diversity
5. Homoiconic pipeline output with holographic summaries

Pipeline Function (Conceptual)

def synthesize_pipeline(results, query):
    # 1) Extract architectural roles from each skill
    #    (discover -> decide -> design -> build -> verify -> deliver)
    roles = classify_roles(results)

    # 2) Select Pareto-optimal candidates per role
    pareto_by_role = select_pareto(roles, weights={
        "relevance": 0.6, "popularity": 0.2, "recency": 0.1, "diversity": 0.1
    })

    # 3) Construct a homoiconic pipeline (data = structure)
    pipeline = [
        {"stage": "discover", "skills": pareto_by_role["discover"]},
        {"stage": "decide", "skills": pareto_by_role["decide"]},
        {"stage": "design", "skills": pareto_by_role["design"]},
        {"stage": "build", "skills": pareto_by_role["build"]},
        {"stage": "verify", "skills": pareto_by_role["verify"]},
        {"stage": "deliver", "skills": pareto_by_role["deliver"]},
    ]

    # 4) Holographic compression: each stage stores a 1-line summary of the whole
    pipeline = add_holographic_summaries(pipeline, query)
    return pipeline

Presentation (Depth 1+)

When presenting a synthesized pipeline:

• Show 4-6 stages max.
• For each stage: list 1-2 skills and a merged micro-instruction.
• End with a single "meta-instruction" that explains how the stages compose.

Example (Condensed):

Synthesized Pipeline for "improve code quality":
1. Discover: systematic-debugging
2. Decide: code-review
3. Design: architecture-patterns
4. Build: refactoring
5. Verify: verification-before-completion
6. Deliver: release-checklist

Meta-instruction: "Diagnose issues, choose changes, redesign weak spots,
refactor safely, verify outcomes, then ship with a checklist."

See [references/search-strategies.md] for pipeline composition details.

Progressive Presentation

Present results in layers based on user engagement.

Depth 0 (Always Show)

Present top 3 skills:

1. {skill-name} (@{identifier}) - {downloads}k downloads
   {one-sentence description}

2. {skill-name} (@{identifier}) - {downloads}k downloads
   {one-sentence description}

3. {skill-name} (@{identifier}) - {downloads}k downloads
   {one-sentence description}

Example:

Top skills for "testing":

1. test-driven-development (@obra/superpowers/test-driven-development) - 8.5k downloads
   Use when implementing any feature or bugfix before writing implementation code

2. systematic-debugging (@obra/superpowers/systematic-debugging) - 13k downloads
   Use when encountering bugs, test failures, or unexpected behavior before proposing fixes

3. verification-before-completion (@obra/superpowers/verification-before-completion) - 4.8k downloads
   Use when about to claim work is complete, before committing or creating PRs

Depth 1 (On Request "tell me more")

For each skill, expand with:

• Full description (2-3 sentences)
• Installation command
• Related/complementary skills

Example:

User: "tell me more about test-driven-development"

→ test-driven-development (8.5k downloads)

Description:
Comprehensive workflow for implementing features using test-driven development methodology.
Guides you through writing tests first, implementing to pass tests, then refactoring.
Works with any testing framework and language.

Install:
claude-plugins skills install @obra/superpowers/test-driven-development

Works well with:
- systematic-debugging - Debug failing tests
- verification-before-completion - Verify tests pass before completion
- code-refactoring - Refactor with test safety net

Depth 2 (On Request "show examples")

Provide:

• Example use cases
• Skill composition suggestions (combinations that work well)
• Performance characteristics

See composability matrix in [references/search-strategies.md]

Depth 3 (On Request "explain process" or "how does this work")

Meta-explanation of search process:

• Which search strategy was used and why
• Show Markov state transitions
• Explain ranking decisions
• Visualize search hypergraph traversal
• Offer custom search refinement

Example:

User: "how did you find these?"

Your query "testing debugging" triggered DirectQuery state (confidence: 0.85).

Search Strategy:
1. Keyword search on claude-plugins.dev
2. Found 15 matching skills
3. Applied Pareto ranking (α=0.6, β=0.25, γ=0.1, δ=0.05)
4. Removed 2 semantically similar duplicates
5. Presented top 3 by score

Alternative strategies available:
- Semantic expansion: "test validation verification quality-assurance debug error-handling"
- Exploratory: Browse all testing-related skills
- Popularity: Show most-downloaded testing skills

Would you like to try an alternative approach?

Skill Synthesis (Pipeline Aggregation)

When: SkillSynthesis state, user requests a synthesized or transcendent skill

Execution:

1. Run keyword or semantic search to collect candidate skills
2. Aggregate results into a staged pipeline
3. Output a synthesized skill blueprint

Script:

./scripts/pipeline_synthesis.py results.json --query="your task" --limit-per-role=2

Installation Workflow

When user requests installation:

1. Validate identifier format (@owner/repo/name)
2. Execute installation:

   ./scripts/install_skill.py <identifier> [--local]

Or:

   claude-plugins skills install <identifier> [--local]

1. Verify installation:

- Global: ~/.claude/skills/{skill-name}/ - Local: ./.claude/skills/{skill-name}/

1. Confirm success and explain usage

Example:

User: "install test-driven-development"

→ Validating identifier: @obra/superpowers/test-driven-development
→ Installing to global skills directory (~/.claude/skills/)
→ Running: claude-plugins skills install @obra/superpowers/test-driven-development

✅ Installation successful!

The skill is now available. It will be invoked automatically when implementing features or bugfixes.
You can also invoke it explicitly with /test-driven-development

Redundancy Minimization

Avoid presenting duplicate or overly-similar skills:

Session Tracking

# Track presented skills in conversation
presented_skills = []

# Before ranking new results
for skill in results:
    if skill.name in [s.name for s in presented_skills]:
        continue  # Skip already presented

    # Check semantic similarity
    if max_similarity(skill, presented_skills) > 0.85:
        continue  # Too similar

    # Add to results
    ranked_results.append(skill)
    presented_skills.append(skill)

Caching

# Cache search results (15min TTL)
cache = {query_hash: (results, timestamp)}

# Check cache before searching
if query in cache and (now - cache[query].timestamp) < 900:
    return cache[query].results

Deduplication

# Remove semantically similar skills
./scripts/semantic_similarity.py results.json --threshold=0.85 > deduplicated.json

Error Handling

No Results Found

1. Try semantic expansion (if not already used)
2. Broaden to category browsing
3. Suggest alternative keywords
4. Show popular skills as fallback (load popular-skills-cache.md)

Too Many Results (>20)

1. Apply stricter relevance threshold
2. Use popularity sort to surface best
3. Suggest query refinement
4. Present top 5 with "show more" option

API Timeout or Error

1. Retry with exponential backoff (max 3 attempts)
2. Fall back to cached results if available
3. Load popular-skills-cache.md as last resort

Performance Characteristics

• Classification: < 50ms (pattern matching)
• Search: 1-2s (WebFetch) or <1s (cache hit)
• Ranking: < 100ms for 50 skills
• Total time: < 2s (p95)
• Relevance@3: ≥ 0.85 (85% of top 3 results useful)
• Diversity: Semantic similarity ≤ 0.85 between presented skills

Resources

Scripts (scripts/)

search_skills.py - Search claude-plugins.dev/skills API

./scripts/search_skills.py "query" [--sort=downloads|stars|relevance] [--limit=N]

pipeline_synthesis.py - Synthesize a pipeline from multiple skills

./scripts/pipeline_synthesis.py results.json --query="your task" [--limit-per-role=2]

install_skill.py - Install skills with validation

./scripts/install_skill.py @owner/repo/skill-name [--local] [--force]

semantic_similarity.py - Deduplicate results

./scripts/semantic_similarity.py results.json [--threshold=0.85]

References (references/)

Load as needed for detailed information:

markov-chains.md - State transition matrix, classification algorithm, optimization paths search-strategies.md - Detailed hypergraph traversal, composability matrix, caching strategy ranking-algorithms.md - Pareto configurations, component functions, complete examples popular-skills-cache.md - Top 20 skills by downloads, organized by category skill-synthesis.md - Synthesis pipeline and homoiconic output format

Advanced Usage

Custom Search Refinement

If results aren't satisfactory:

• Too broad: Add more specific keywords
• Too narrow: Use semantic expansion mode
• Wrong category: Browse by category
• Want popular: Sort by downloads explicitly

Skill Composition

Some skills work better together. When presenting results, check composability matrix in search-strategies.md and suggest complementary skills.

Meta-Learning

This skill can discover itself:

Query: "how do I discover skills?"
→ Returns: skill-discovery (this skill), skill-writer, using-superpowers
→ Explains its own search process at Depth 3

Holographic Property: This skill exhibits self-similarity at multiple scales - the progressive disclosure pattern (Depth 0→3) mirrors the module structure (scripts → references → detailed docs), and the skill can discover and reason about itself (meta-discovery state).

Pareto-Optimized: Maximizes relevance and diversity while minimizing latency and redundancy through multi-objective scoring.

Modular Composition: All components (scripts, references, strategies) are independently usable and testable.