Skills Pattern

---
name: web-search
description: Search the web for current information and facts
version: 1.0.0
author: agent-team

# Triggers help the agent decide when to use this skill
triggers:
  - "search for"
  - "find information about"
  - "what is the latest"
  - "current news on"
  - "look up"

# Tools this skill requires
tools:
  - web_search
  - url_fetch
  - extract_content

# Token cost estimate (for prioritization)
token_estimate: 1200

# Categories for organization
categories:
  - research
  - information-retrieval
---

# Web Search Skill

## Purpose
Use this skill when the user needs current, real-time information
that may not be in your training data.

## When to Use
- Questions about recent events (after your knowledge cutoff)
- Fact-checking claims that need verification
- Finding specific data points (prices, statistics, etc.)
- Research tasks requiring multiple sources

## When NOT to Use
- Questions you can answer from training data
- Hypothetical or opinion-based questions
- Creative writing tasks

## Instructions
1. Formulate a clear search query
2. Execute the web search tool
3. Evaluate result relevance (discard low-quality sources)
4. Fetch full content from top 2-3 results
5. Synthesize information with citations

## Example Interaction
User: "What's the current price of Bitcoin?"
Agent: [Uses web_search with query "bitcoin price USD current"]
Agent: [Fetches top result, extracts price data]
Response: "As of [timestamp], Bitcoin is trading at $XX,XXX USD."

---
name: code-review
description: Review code for bugs, security issues, and improvements
version: 2.1.0

triggers:
  - "review this code"
  - "check for bugs"
  - "security audit"
  - "improve this function"

tools:
  - read_file
  - ast_parse
  - static_analysis

token_estimate: 2500

categories:
  - development
  - security
---

# Code Review Skill

## Purpose
Provide thorough code reviews focusing on correctness,
security, performance, and maintainability.

## Review Checklist
1. **Correctness**: Does the code do what it's supposed to?
2. **Security**: Are there injection, XSS, or auth issues?
3. **Performance**: Any obvious bottlenecks or N+1 queries?
4. **Readability**: Is the code self-documenting?
5. **Testing**: Are edge cases covered?

## Output Format
```
## Summary
[One-line summary of code quality]

## Issues Found
- [SEVERITY] file:line - Description

## Recommendations
- [Priority] Suggestion for improvement
```

## Language-Specific Checks
### Python
- Type hints present and correct
- No mutable default arguments
- Context managers for resources

### JavaScript
- No `var` (use `const`/`let`)
- Proper async/await error handling
- No prototype pollution risks

# Three-stage progressive disclosure

# Stage 1: Metadata only (minimal tokens)
function getSkillList():
    skills = []
    for skillDir in listDirectory("skills/"):
        metadata = parseYamlFrontmatter(skillDir + "/SKILL.md")
        skills.append({
            name: metadata.name,
            description: metadata.description,  # Short summary
            triggers: metadata.triggers
        })
    return skills  # ~50-100 tokens per skill

# Stage 2: Full instructions (on-demand)
function loadSkillInstructions(skillName):
    skillFile = "skills/" + skillName + "/SKILL.md"
    return parseMarkdownBody(skillFile)  # ~500-1000 tokens

# Stage 3: Examples and resources (if needed)
function loadSkillResources(skillName):
    examplesDir = "skills/" + skillName + "/examples/"
    resources = []
    for file in listDirectory(examplesDir):
        resources.append(readFile(examplesDir + file))
    return resources  # Variable, loaded only when needed

# Agent workflow
function handleRequest(userQuery):
    # Stage 1: Quick scan with metadata only
    skillList = getSkillList()
    selectedSkill = llm.selectBestSkill(userQuery, skillList)

    if selectedSkill:
        # Stage 2: Load full instructions
        instructions = loadSkillInstructions(selectedSkill)

        # Stage 3: Load examples only if complex task
        if taskIsComplex(userQuery):
            examples = loadSkillResources(selectedSkill)
            context = instructions + examples
        else:
            context = instructions

        return executeWithContext(userQuery, context)

from dataclasses import dataclass
from enum import Enum
from pathlib import Path

class DisclosureLevel(Enum):
    METADATA = 1    # Name, description, triggers (~50 tokens)
    INSTRUCTIONS = 2 # Full SKILL.md body (~500-1000 tokens)
    RESOURCES = 3    # Examples, templates (~variable)

@dataclass
class SkillContext:
    name: str
    level: DisclosureLevel
    content: str
    token_count: int

class ProgressiveSkillLoader:
    def __init__(self, skills_dir: Path):
        self.skills_dir = skills_dir
        self._metadata_cache: dict[str, dict] = {}

    def get_skill_list(self) -> list[dict]:
        """Stage 1: Return minimal metadata for all skills."""
        skills = []
        for skill_path in self.skills_dir.iterdir():
            if not skill_path.is_dir():
                continue

            metadata = self._load_metadata(skill_path.name)
            skills.append({
                "name": metadata["name"],
                "description": metadata["description"][:100],  # Truncate
                "triggers": metadata.get("triggers", [])[:5]   # Limit
            })
        return skills  # Minimal token footprint

    def load_instructions(self, skill_name: str) -> SkillContext:
        """Stage 2: Load full instructions on demand."""
        skill_file = self.skills_dir / skill_name / "SKILL.md"
        content = skill_file.read_text()

        # Extract body after frontmatter
        body = self._extract_body(content)

        return SkillContext(
            name=skill_name,
            level=DisclosureLevel.INSTRUCTIONS,
            content=body,
            token_count=self._estimate_tokens(body)
        )

    def load_resources(self, skill_name: str) -> SkillContext:
        """Stage 3: Load examples and additional resources."""
        examples_dir = self.skills_dir / skill_name / "examples"

        resources = []
        if examples_dir.exists():
            for example_file in examples_dir.iterdir():
                resources.append(example_file.read_text())

        combined = "\n---\n".join(resources)

        return SkillContext(
            name=skill_name,
            level=DisclosureLevel.RESOURCES,
            content=combined,
            token_count=self._estimate_tokens(combined)
        )

    def _load_metadata(self, skill_name: str) -> dict:
        if skill_name not in self._metadata_cache:
            skill_file = self.skills_dir / skill_name / "SKILL.md"
            content = skill_file.read_text()
            self._metadata_cache[skill_name] = parse_frontmatter(content)
        return self._metadata_cache[skill_name]

# Usage in agent
class SkillAwareAgent:
    def __init__(self, loader: ProgressiveSkillLoader):
        self.loader = loader

    def process(self, query: str) -> str:
        # Stage 1: Select skill from metadata
        skill_list = self.loader.get_skill_list()
        selected = self.llm.select_skill(query, skill_list)

        if not selected:
            return self.llm.respond_without_skill(query)

        # Stage 2: Load instructions
        context = self.loader.load_instructions(selected)

        # Stage 3: Load examples for complex tasks
        if self.is_complex_task(query):
            resources = self.loader.load_resources(selected)
            context.content += "\n\n" + resources.content

        return self.llm.respond_with_context(query, context.content)

public enum DisclosureLevel
{
    Metadata = 1,     // ~50 tokens per skill
    Instructions = 2, // ~500-1000 tokens
    Resources = 3     // Variable
}

public record SkillContext(
    string Name,
    DisclosureLevel Level,
    string Content,
    int TokenCount
);

public class ProgressiveSkillLoader
{
    private readonly string _skillsDir;
    private readonly Dictionary<string, Dictionary<string, object>> _metadataCache = new();

    public ProgressiveSkillLoader(string skillsDir)
    {
        _skillsDir = skillsDir;
    }

    /// <summary>
    /// Stage 1: Get minimal metadata for skill selection.
    /// </summary>
    public List<SkillSummary> GetSkillList()
    {
        var skills = new List<SkillSummary>();

        foreach (var skillDir in Directory.GetDirectories(_skillsDir))
        {
            var skillName = Path.GetFileName(skillDir);
            var metadata = LoadMetadata(skillName);

            skills.Add(new SkillSummary(
                Name: metadata.GetValueOrDefault("name", skillName)?.ToString() ?? skillName,
                Description: TruncateString(
                    metadata.GetValueOrDefault("description", "")?.ToString() ?? "",
                    100
                ),
                Triggers: GetTriggers(metadata).Take(5).ToList()
            ));
        }

        return skills;
    }

    /// <summary>
    /// Stage 2: Load full instructions for selected skill.
    /// </summary>
    public SkillContext LoadInstructions(string skillName)
    {
        var skillFile = Path.Combine(_skillsDir, skillName, "SKILL.md");
        var content = File.ReadAllText(skillFile);
        var body = ExtractBody(content);

        return new SkillContext(
            Name: skillName,
            Level: DisclosureLevel.Instructions,
            Content: body,
            TokenCount: EstimateTokens(body)
        );
    }

    /// <summary>
    /// Stage 3: Load examples and additional resources.
    /// </summary>
    public SkillContext LoadResources(string skillName)
    {
        var examplesDir = Path.Combine(_skillsDir, skillName, "examples");
        var resources = new List<string>();

        if (Directory.Exists(examplesDir))
        {
            foreach (var file in Directory.GetFiles(examplesDir))
            {
                resources.Add(File.ReadAllText(file));
            }
        }

        var combined = string.Join("\n---\n", resources);

        return new SkillContext(
            Name: skillName,
            Level: DisclosureLevel.Resources,
            Content: combined,
            TokenCount: EstimateTokens(combined)
        );
    }
}

// Agent usage
public class SkillAwareAgent
{
    private readonly ProgressiveSkillLoader _loader;
    private readonly ILanguageModel _llm;

    public async Task<string> ProcessAsync(string query)
    {
        // Stage 1: Quick skill selection
        var skillList = _loader.GetSkillList();
        var selected = await _llm.SelectSkillAsync(query, skillList);

        if (selected == null)
            return await _llm.RespondAsync(query);

        // Stage 2: Load instructions
        var context = _loader.LoadInstructions(selected);

        // Stage 3: Load examples if needed
        if (IsComplexTask(query))
        {
            var resources = _loader.LoadResources(selected);
            context = context with {
                Content = context.Content + "\n\n" + resources.Content
            };
        }

        return await _llm.RespondWithContextAsync(query, context.Content);
    }
}

# Traditional approach: All tools in context
# 50 tools × 3000 tokens/tool = 150,000 tokens per request

# Skills Pattern: Progressive disclosure
# Stage 1: Metadata only
#   50 skills × 50 tokens = 2,500 tokens
#
# Stage 2: Selected skill instructions
#   1 skill × 1,000 tokens = 1,000 tokens
#
# Stage 3: Examples (if needed)
#   1 skill × 2,000 tokens = 2,000 tokens
#
# Total: 2,500 + 1,000 + 2,000 = 5,500 tokens
#
# Savings: 150,000 - 5,500 = 144,500 tokens (96% reduction)

# Even better with vector search:
# Skip Stage 1 entirely - query embeddings directly
# Stage 2 + 3 only: 3,000 tokens
#
# Savings: 150,000 - 3,000 = 147,000 tokens (98% reduction)

def calculate_token_savings(
    num_skills: int,
    avg_tool_tokens: int = 3000,
    metadata_tokens: int = 50,
    instruction_tokens: int = 1000,
    example_tokens: int = 2000
) -> dict:
    """Calculate token savings from skills pattern."""

    # Traditional approach
    traditional = num_skills * avg_tool_tokens

    # Skills pattern with progressive disclosure
    stage1_metadata = num_skills * metadata_tokens
    stage2_instructions = instruction_tokens  # Only selected skill
    stage3_examples = example_tokens  # Only if needed

    # Most requests only need stages 1 + 2
    typical_request = stage1_metadata + stage2_instructions
    complex_request = typical_request + stage3_examples

    # Vector search approach (skip metadata scan)
    vector_approach = stage2_instructions + stage3_examples

    return {
        "traditional_tokens": traditional,
        "progressive_typical": typical_request,
        "progressive_complex": complex_request,
        "vector_search": vector_approach,
        "savings_typical": f"{(1 - typical_request/traditional) * 100:.1f}%",
        "savings_vector": f"{(1 - vector_approach/traditional) * 100:.1f}%"
    }

# Example with 50 skills
result = calculate_token_savings(num_skills=50)
print(result)
# {
#     "traditional_tokens": 150000,
#     "progressive_typical": 3500,    # Metadata + instructions
#     "progressive_complex": 5500,    # + examples
#     "vector_search": 3000,          # Instructions + examples
#     "savings_typical": "97.7%",
#     "savings_vector": "98.0%"
# }

public record TokenSavingsReport(
    int TraditionalTokens,
    int ProgressiveTypical,
    int ProgressiveComplex,
    int VectorSearch,
    string SavingsTypical,
    string SavingsVector
);

public static class TokenCalculator
{
    public static TokenSavingsReport CalculateSavings(
        int numSkills,
        int avgToolTokens = 3000,
        int metadataTokens = 50,
        int instructionTokens = 1000,
        int exampleTokens = 2000)
    {
        // Traditional: all tools in context
        var traditional = numSkills * avgToolTokens;

        // Progressive disclosure
        var stage1 = numSkills * metadataTokens;
        var stage2 = instructionTokens;
        var stage3 = exampleTokens;

        var typical = stage1 + stage2;
        var complex = typical + stage3;

        // Vector search (skip metadata scan)
        var vector = stage2 + stage3;

        return new TokenSavingsReport(
            TraditionalTokens: traditional,
            ProgressiveTypical: typical,
            ProgressiveComplex: complex,
            VectorSearch: vector,
            SavingsTypical: $"{(1 - (double)typical / traditional) * 100:F1}%",
            SavingsVector: $"{(1 - (double)vector / traditional) * 100:F1}%"
        );
    }
}

// Usage
var report = TokenCalculator.CalculateSavings(numSkills: 50);
Console.WriteLine($"Traditional: {report.TraditionalTokens:N0} tokens");
Console.WriteLine($"Progressive: {report.ProgressiveTypical:N0} tokens");
Console.WriteLine($"Savings: {report.SavingsTypical}");

# Vector database approach for skill discovery

function indexSkills(skills):
    for skill in skills:
        # Create embedding from skill description + triggers
        text = skill.name + ": " + skill.description
        text += " Triggers: " + join(skill.triggers, ", ")

        embedding = embedModel.encode(text)

        vectorDb.upsert(
            id: skill.name,
            vector: embedding,
            metadata: {
                name: skill.name,
                description: skill.description,
                tools: skill.tools,
                token_estimate: skill.tokenEstimate
            }
        )

function findRelevantSkills(query, topK = 3):
    queryEmbedding = embedModel.encode(query)

    results = vectorDb.search(
        vector: queryEmbedding,
        topK: topK,
        threshold: 0.7  # Minimum similarity
    )

    return results.map(r => r.metadata)

# Hybrid approach: Vector + keyword fallback
function findSkillsHybrid(query):
    # Try vector search first
    vectorResults = findRelevantSkills(query, topK = 5)

    if vectorResults.isEmpty() or vectorResults[0].score < 0.75:
        # Fall back to keyword matching
        keywordResults = keywordSearch(query)
        return mergeResults(vectorResults, keywordResults)

    return vectorResults

import chromadb
from chromadb.utils import embedding_functions
from dataclasses import dataclass

@dataclass
class SkillMatch:
    name: str
    description: str
    score: float
    tools: list[str]

class VectorSkillDiscovery:
    def __init__(self, persist_dir: str = "./skill_vectors"):
        self.client = chromadb.PersistentClient(path=persist_dir)

        # Use sentence-transformers for embeddings
        self.embedding_fn = embedding_functions.SentenceTransformerEmbeddingFunction(
            model_name="all-MiniLM-L6-v2"
        )

        self.collection = self.client.get_or_create_collection(
            name="skills",
            embedding_function=self.embedding_fn,
            metadata={"hnsw:space": "cosine"}
        )

    def index_skill(self, skill: dict) -> None:
        """Add or update a skill in the vector database."""
        # Create rich text representation for embedding
        text = f"{skill['name']}: {skill['description']}"
        if skill.get('triggers'):
            text += f" Triggers: {', '.join(skill['triggers'])}"

        self.collection.upsert(
            ids=[skill['name']],
            documents=[text],
            metadatas=[{
                "name": skill['name'],
                "description": skill['description'],
                "tools": ",".join(skill.get('tools', [])),
                "token_estimate": skill.get('token_estimate', 0)
            }]
        )

    def find_skills(
        self,
        query: str,
        top_k: int = 3,
        min_score: float = 0.5
    ) -> list[SkillMatch]:
        """Find most relevant skills for a query."""
        results = self.collection.query(
            query_texts=[query],
            n_results=top_k,
            include=["documents", "metadatas", "distances"]
        )

        matches = []
        for i, distance in enumerate(results['distances'][0]):
            # Convert distance to similarity score
            score = 1 - distance

            if score < min_score:
                continue

            metadata = results['metadatas'][0][i]
            matches.append(SkillMatch(
                name=metadata['name'],
                description=metadata['description'],
                score=score,
                tools=metadata['tools'].split(',') if metadata['tools'] else []
            ))

        return matches

    def reindex_all(self, skills_dir: Path) -> int:
        """Reindex all skills from filesystem."""
        count = 0
        for skill_path in skills_dir.iterdir():
            if not skill_path.is_dir():
                continue

            skill_file = skill_path / "SKILL.md"
            if not skill_file.exists():
                continue

            metadata = parse_skill_frontmatter(skill_file.read_text())
            metadata['name'] = skill_path.name
            self.index_skill(metadata)
            count += 1

        return count

# Usage
discovery = VectorSkillDiscovery()
discovery.reindex_all(Path("./skills"))

# At runtime
matches = discovery.find_skills("help me analyze this CSV data")
# Returns: [SkillMatch(name='data-analysis', score=0.89, ...)]

using Qdrant.Client;
using Qdrant.Client.Grpc;
using Microsoft.ML.OnnxRuntime;

public record SkillMatch(
    string Name,
    string Description,
    float Score,
    List<string> Tools
);

public class VectorSkillDiscovery
{
    private readonly QdrantClient _qdrant;
    private readonly EmbeddingModel _embedder;
    private const string CollectionName = "skills";

    public VectorSkillDiscovery(string qdrantUrl = "http://localhost:6334")
    {
        _qdrant = new QdrantClient(qdrantUrl);
        _embedder = new EmbeddingModel("all-MiniLM-L6-v2.onnx");

        EnsureCollectionExists().Wait();
    }

    private async Task EnsureCollectionExists()
    {
        var collections = await _qdrant.ListCollectionsAsync();
        if (!collections.Contains(CollectionName))
        {
            await _qdrant.CreateCollectionAsync(
                CollectionName,
                new VectorParams { Size = 384, Distance = Distance.Cosine }
            );
        }
    }

    public async Task IndexSkillAsync(Dictionary<string, object> skill)
    {
        var name = skill["name"].ToString()!;
        var description = skill["description"].ToString()!;
        var triggers = skill.GetValueOrDefault("triggers") as List<string> ?? new();

        // Create text for embedding
        var text = $"{name}: {description} Triggers: {string.Join(", ", triggers)}";
        var embedding = _embedder.Encode(text);

        var point = new PointStruct
        {
            Id = new PointId { Uuid = Guid.NewGuid().ToString() },
            Vectors = embedding,
            Payload = {
                ["name"] = name,
                ["description"] = description,
                ["tools"] = string.Join(",", skill.GetValueOrDefault("tools") as List<string> ?? new()),
                ["token_estimate"] = (long)(skill.GetValueOrDefault("token_estimate") ?? 0)
            }
        };

        await _qdrant.UpsertAsync(CollectionName, new[] { point });
    }

    public async Task<List<SkillMatch>> FindSkillsAsync(
        string query,
        int topK = 3,
        float minScore = 0.5f)
    {
        var queryEmbedding = _embedder.Encode(query);

        var results = await _qdrant.SearchAsync(
            CollectionName,
            queryEmbedding,
            limit: (ulong)topK,
            scoreThreshold: minScore
        );

        return results.Select(r => new SkillMatch(
            Name: r.Payload["name"].StringValue,
            Description: r.Payload["description"].StringValue,
            Score: r.Score,
            Tools: r.Payload["tools"].StringValue
                .Split(',', StringSplitOptions.RemoveEmptyEntries)
                .ToList()
        )).ToList();
    }
}

// Hybrid discovery combining vector + keyword
public class HybridSkillDiscovery
{
    private readonly VectorSkillDiscovery _vectorSearch;
    private readonly KeywordSkillDiscovery _keywordSearch;

    public async Task<List<SkillMatch>> FindSkillsAsync(string query)
    {
        // Try vector search first
        var vectorResults = await _vectorSearch.FindSkillsAsync(query, topK: 5);

        // If low confidence, supplement with keyword search
        if (!vectorResults.Any() || vectorResults[0].Score < 0.75f)
        {
            var keywordResults = _keywordSearch.Search(query);
            return MergeResults(vectorResults, keywordResults);
        }

        return vectorResults;
    }

    private List<SkillMatch> MergeResults(
        List<SkillMatch> vector,
        List<SkillMatch> keyword)
    {
        // Reciprocal rank fusion
        var scores = new Dictionary<string, float>();

        for (int i = 0; i < vector.Count; i++)
            scores[vector[i].Name] = scores.GetValueOrDefault(vector[i].Name) + 1f / (i + 1);

        for (int i = 0; i < keyword.Count; i++)
            scores[keyword[i].Name] = scores.GetValueOrDefault(keyword[i].Name) + 1f / (i + 1);

        return scores
            .OrderByDescending(kv => kv.Value)
            .Take(5)
            .Select(kv => vector.Concat(keyword).First(m => m.Name == kv.Key))
            .ToList();
    }
}

# Skill discovery evaluation metrics

function evaluateSkillDiscovery(testCases, discoverySystem):
    metrics = {
        precision_at_1: [],
        precision_at_3: [],
        recall_at_3: [],
        latency: [],
        false_positives: 0,
        false_negatives: 0
    }

    for testCase in testCases:
        query = testCase.query
        expectedSkills = testCase.relevantSkills

        startTime = now()
        results = discoverySystem.findSkills(query, topK = 3)
        metrics.latency.append(now() - startTime)

        # Precision@1: Is the top result correct?
        if results[0].name in expectedSkills:
            metrics.precision_at_1.append(1)
        else:
            metrics.precision_at_1.append(0)
            metrics.false_positives += 1

        # Precision@3: How many of top 3 are relevant?
        relevant_in_top3 = count(r for r in results[:3] if r.name in expectedSkills)
        metrics.precision_at_3.append(relevant_in_top3 / 3)

        # Recall@3: Did we find all relevant skills?
        metrics.recall_at_3.append(relevant_in_top3 / len(expectedSkills))

        # False negatives: relevant skills not in top 3
        metrics.false_negatives += len(expectedSkills) - relevant_in_top3

    return {
        avg_precision_at_1: mean(metrics.precision_at_1),
        avg_precision_at_3: mean(metrics.precision_at_3),
        avg_recall_at_3: mean(metrics.recall_at_3),
        avg_latency_ms: mean(metrics.latency) * 1000,
        total_false_positives: metrics.false_positives,
        total_false_negatives: metrics.false_negatives
    }

from dataclasses import dataclass
from typing import Protocol
import time

@dataclass
class TestCase:
    query: str
    relevant_skills: set[str]

@dataclass
class EvaluationResult:
    precision_at_1: float
    precision_at_3: float
    recall_at_3: float
    avg_latency_ms: float
    mrr: float  # Mean Reciprocal Rank

class SkillDiscovery(Protocol):
    def find_skills(self, query: str, top_k: int) -> list[SkillMatch]: ...

def evaluate_skill_discovery(
    test_cases: list[TestCase],
    discovery: SkillDiscovery
) -> EvaluationResult:
    """Evaluate skill discovery accuracy and performance."""

    p1_scores, p3_scores, recall_scores = [], [], []
    latencies, reciprocal_ranks = [], []

    for case in test_cases:
        # Measure latency
        start = time.perf_counter()
        results = discovery.find_skills(case.query, top_k=3)
        latencies.append((time.perf_counter() - start) * 1000)

        result_names = [r.name for r in results]

        # Precision@1
        p1_scores.append(1.0 if result_names[0] in case.relevant_skills else 0.0)

        # Precision@3
        hits = sum(1 for r in result_names[:3] if r in case.relevant_skills)
        p3_scores.append(hits / 3)

        # Recall@3
        recall_scores.append(hits / len(case.relevant_skills))

        # Mean Reciprocal Rank
        for i, name in enumerate(result_names):
            if name in case.relevant_skills:
                reciprocal_ranks.append(1.0 / (i + 1))
                break
        else:
            reciprocal_ranks.append(0.0)

    return EvaluationResult(
        precision_at_1=sum(p1_scores) / len(p1_scores),
        precision_at_3=sum(p3_scores) / len(p3_scores),
        recall_at_3=sum(recall_scores) / len(recall_scores),
        avg_latency_ms=sum(latencies) / len(latencies),
        mrr=sum(reciprocal_ranks) / len(reciprocal_ranks)
    )

# Example test suite
test_cases = [
    TestCase(
        query="search the web for recent news about AI",
        relevant_skills={"web-search", "news-aggregator"}
    ),
    TestCase(
        query="review this Python code for bugs",
        relevant_skills={"code-review"}
    ),
    TestCase(
        query="analyze this CSV and create a chart",
        relevant_skills={"data-analysis", "visualization"}
    ),
]

# Run evaluation
results = evaluate_skill_discovery(test_cases, vector_discovery)
print(f"P@1: {results.precision_at_1:.2%}")
print(f"MRR: {results.mrr:.3f}")
print(f"Latency: {results.avg_latency_ms:.1f}ms")