Implementation Patterns

Implementation Patterns

Common design patterns and best practices for skill systems.

Skill Selection Patterns

When to Use File-Based Skills

Best for:

• Large, stable skills shared across projects
• Public/open-source skills
• Complex resource files (forms, templates, docs)
• Team collaboration with centralized organization

from pydantic_ai import Agent
from pydantic_ai.toolsets.skills import SkillsToolset

# File-based approach: organize skills in directory structure
# ./skills/
#   ├── data-analysis/
#   │   ├── SKILL.md
#   │   ├── reference.md
#   │   └── scripts/
#   │       └── analyze.py
#   └── web-research/
#       ├── SKILL.md
#       └── scripts/

toolset = SkillsToolset(directories=['./skills'])
agent = Agent(model='openai:gpt-4o', toolsets=[toolset])

When to Use Programmatic Skills

Best for:

• Runtime-generated skills based on configuration
• Skills requiring runtime dependencies (databases, APIs)
• Dynamic resource generation that changes per session
• Application-specific skills tightly coupled with logic

from pydantic_ai import RunContext
from pydantic_ai.toolsets.skills import SkillsToolset

class MyDeps:
    database: Database
    config: Config

skills = SkillsToolset()

@skills.skill()
def database_analysis() -> str:
    """Analyze data in database."""
    return "Use database resources and scripts to analyze data"

@database_analysis.resource
async def get_schema(ctx: RunContext[MyDeps]) -> str:
    """Get schema from actual database."""
    schema = await ctx.deps.database.get_schema()
    return f"## Current Schema\n{schema}"

agent = Agent(
    model='openai:gpt-4o',
    toolsets=[skills],
    deps=MyDeps(database=my_db, config=my_config)
)

Mixed Approach

Combine both for flexibility:

from pydantic_ai.toolsets.skills import SkillsToolset

# Mix file-based and programmatic skills
toolset = SkillsToolset(
    directories=['./skills'],              # Stable, reusable skills
    max_depth=2                            # Reasonable discovery depth
)

# Add runtime-specific skills
@toolset.skill(
    name='runtime-monitor',
    metadata={'version': '1.0.0'}
)
def monitoring() -> str:
    return "Monitor application runtime metrics"

@monitoring.resource
async def get_metrics(ctx: RunContext[MyDeps]) -> str:
    metrics = await ctx.deps.monitoring.current_metrics()
    return f"## Metrics\n{metrics}"

agent = Agent(model='openai:gpt-4o', toolsets=[toolset])

Resource Parameter Patterns

Static Resources

Use for reference documentation and fixed content:

from pydantic_ai.toolsets.skills import Skill, SkillResource

skill = Skill(
    name='reference-skill',
    description='Provide reference documentation',
    content='Main instructions...',
    resources=[
        SkillResource(
            name='api-reference',
            description='Complete API reference',
            content='''
## API Reference

### GET /users
Retrieve all users...

### POST /users
Create a new user...
'''
        ),
        SkillResource(
            name='examples',
            description='Code examples',
            content='''
## Examples

```python
# Example 1: Basic usage
client = APIClient()
users = client.get_users()

''' ) ] )

### Dynamic Resources with Context

Use for resources that depend on runtime state:

```python
from typing import TypedDict
from pydantic_ai import RunContext
from pydantic_ai.toolsets.skills import SkillsToolset

class MyDeps(TypedDict):
    database: Database
    cache: Cache

skills = SkillsToolset()

@skills.skill()
def data_skill() -> str:
    return "Database and cache management"

# Resource that fetches current schema
@data_skill.resource
async def get_current_schema(ctx: RunContext[MyDeps]) -> str:
    """Get latest schema from database."""
    schema = await ctx.deps.database.get_schema()
    # Could also cache this
    await ctx.deps.cache.set('schema', schema, ttl=3600)
    return f"## Current Schema\n{schema}"

# Resource that lists available tables
@data_skill.resource
async def get_tables(ctx: RunContext[MyDeps]) -> str:
    """Get available tables."""
    tables = await ctx.deps.database.list_tables()
    return "## Available Tables\n" + "\n".join(f"- {t}" for t in tables)

Parameterized Resources

Resources can accept parameters for dynamic content:

from pydantic_ai.toolsets.skills import SkillsToolset

skills = SkillsToolset()

@skills.skill()
def query_skill() -> str:
    return "Execute database queries"

@query_skill.resource
def documentation(topic: str = "general") -> str:
    """Get documentation for a topic."""
    docs = {
        "general": "General query documentation...",
        "aggregation": "Aggregation query documentation...",
        "joins": "Join query documentation..."
    }
    return docs.get(topic, "Topic not found")

Script Execution Patterns

Simple Synchronous Scripts

For simple, stateless operations:

from pydantic_ai_skills import SkillsToolset

skills = SkillsToolset()

@skills.skill()
def text_processing() -> str:
    return "Process text data"

@text_processing.script
def count_words(text: str) -> str:
    """Count words in text."""
    words = len(text.split())
    return f"Word count: {words}"

@text_processing.script
def reverse_text(text: str) -> str:
    """Reverse text."""
    return text[::-1]

Stateful Scripts with Initialization

For scripts that manage shared state through dependencies:

import sqlite3
from dataclasses import dataclass, field
from pydantic_ai import RunContext
from pydantic_ai_skills import SkillsToolset, Skill

@dataclass
class DataAnalyzerDeps:
    """Dependencies managing database connection state."""
    db_path: str = ':memory:'
    db: sqlite3.Connection | None = field(default=None)

    def get_db_tables(self) -> list[str]:
        """Get list of tables currently loaded in the database."""
        if self.db is None:
            return []
        cursor = self.db.cursor()
        cursor.execute("SELECT name FROM sqlite_master WHERE type='table';")
        return [row[0] for row in cursor.fetchall()]

# Create skill with stateful management
skill = Skill(
    name='data-analyzer',
    description='Analyze data using SQL queries on in-memory database',
    content='''Use this skill to:
1. Call `load_data` script to initialize database
2. Use `run_query` script to execute SQL analysis
'''
)

@skill.script
async def load_data(ctx: RunContext[DataAnalyzerDeps], csv_path: str) -> str:
    """Load CSV data into in-memory SQLite database.

    Initializes database connection if needed (idempotent).
    """
    if ctx.deps.db is None:
        ctx.deps.db = sqlite3.connect(ctx.deps.db_path, check_same_thread=False)

    loaded_tables = []
    # Load CSV file into table
    import pandas as pd
    df = pd.read_csv(csv_path)
    table_name = Path(csv_path).stem

    if table_name not in ctx.deps.get_db_tables():
        df.to_sql(table_name, ctx.deps.db, if_exists='replace', index=False)
        loaded_tables.append(table_name)

    if loaded_tables:
        return f'Data loaded. Tables created: {", ".join(loaded_tables)}'
    return 'Data already loaded. All tables available.'

@skill.script
async def run_query(ctx: RunContext[DataAnalyzerDeps], query: str) -> str:
    """Execute SQL query on loaded data and return formatted results."""
    if ctx.deps.db is None:
        return 'Error: No data loaded. Run load_data script first.'

    try:
        cursor = ctx.deps.db.cursor()
        cursor.execute(query)
        rows = cursor.fetchall()
        columns = [description[0] for description in cursor.description]

        if not rows:
            return 'Query executed successfully. No rows returned.'

        # Format results as aligned table
        col_widths = [max(len(str(col)), max(len(str(row[i])) for row in rows))
                      for i, col in enumerate(columns)]
        header = ' | '.join(col.ljust(col_widths[i]) for i, col in enumerate(columns))
        separator = '-+-'.join('-' * width for width in col_widths)
        result_lines = [header, separator]

        for row in rows:
            result_lines.append(' | '.join(str(item).ljust(col_widths[i])
                                           for i, item in enumerate(row)))

        return '\n'.join(result_lines)

    except sqlite3.Error as e:
        return f'SQL Error: {e}\n\nEnsure table names are correct and query syntax is valid.'

toolset = SkillsToolset(skills=[skill])

Asynchronous Scripts with Security

For scripts requiring dependencies with validated parameters:

from pydantic_ai import RunContext
from pydantic_ai_skills import SkillsToolset

class AnalyticsContext:
    database: Database
    logger: Logger

skills = SkillsToolset()

@skills.skill()
def analytics() -> str:
    return "Perform data analytics with security controls"

@analytics.script
async def analyze_user_data(
    ctx: RunContext[AnalyticsContext],
    user_id: int,
    metric: str = "sales"
) -> str:
    """Analyze user data by metric.

    SECURITY: Uses whitelist to prevent SQL injection.
    """
    await ctx.deps.logger.info(f"Analyzing user {user_id}")

    # SECURITY: Enforce whitelist of allowed metrics to prevent SQL injection
    allowed_metrics = {
        "sales": "sales",
        "revenue": "revenue",
        "total_compensation": "total_compensation",
    }

    if metric not in allowed_metrics:
        raise ValueError(f"Unsupported metric: {metric!r}. Allowed: {', '.join(allowed_metrics.keys())}")

    column_name = allowed_metrics[metric]

    # Safe query: column_name is from whitelist, user_id is parameterized
    data = await ctx.deps.database.query(f"""
        SELECT {column_name} FROM users WHERE id = ?
    """, (user_id,))

    result = sum(data)
    await ctx.deps.logger.info(f"Analysis complete: {result}")

    return f"Analysis result: {result}"

Chaining Scripts with Sequential Dependencies

Design scripts that build on each other, where agents call them in sequence:

@analytics.script
async def prepare_data(ctx: RunContext[AnalyticsContext], dataset: str) -> str:
    """Prepare raw data for analysis.

    Prerequisite for run_analysis and export_results scripts.
    """
    raw = await ctx.deps.database.fetch_raw(dataset)
    cleaned = await ctx.deps.database.clean(raw)
    await ctx.deps.logger.info(f"Prepared {len(cleaned)} records")
    return f"Prepared {len(cleaned)} records"

@analytics.script
async def run_analysis(ctx: RunContext[AnalyticsContext], dataset: str) -> str:
    """Run analysis on prepared data.

    Expects prepare_data to have been called first.
    """
    prepared = await ctx.deps.database.load_prepared(dataset)
    results = await analyze_data(prepared)
    return format_results(results)

@analytics.script
async def export_results(
    ctx: RunContext[AnalyticsContext],
    dataset: str,
    format: str = "json"
) -> str:
    """Export analysis results in requested format.

    Args:
        dataset: Dataset name to export results for
        format: Output format (json, csv)
    """
    results = await ctx.deps.database.load_results(dataset)

    if format == "json":
        import json
        output = json.dumps(results, indent=2)
    elif format == "csv":
        output = convert_to_csv(results)
    else:
        return f"Unknown format: {format}. Use 'json' or 'csv'."

    return output

Error Handling

Catching Skill Errors

from pydantic_ai.toolsets.skills import SkillsToolset
from pydantic_ai_skills import (
    SkillNotFoundError,
    SkillResourceNotFoundError,
    SkillScriptNotFoundError,
    SkillScriptExecutionError
)

toolset = SkillsToolset(directories=['./skills'])

# Handle missing skills
try:
    skill = toolset.get_skill('non-existent')
except SkillNotFoundError as e:
    print(f"Skill not found: {e}")

# Handle missing resources/scripts in tools
try:
    resource = toolset._find_skill_resource(skill, 'unknown-resource')
    if resource is None:
        raise SkillResourceNotFoundError(...)
except SkillResourceNotFoundError as e:
    print(f"Resource not found: {e}")

Graceful Degradation

from pydantic_ai import Agent, RunContext
from pydantic_ai.toolsets.skills import SkillsToolset
from pydantic_ai_skills import SkillNotFoundError

toolset = SkillsToolset(directories=['./skills'])

agent = Agent(model='openai:gpt-4o', toolsets=[toolset])

@agent.tool
async def safe_load_skill(ctx: RunContext, skill_name: str) -> str:
    """Load skill with graceful fallback."""
    try:
        skill = toolset.get_skill(skill_name)
        return f"Loaded {skill_name}: {skill.description}"
    except SkillNotFoundError:
        available = list(toolset.skills.keys())
        return f"Skill '{skill_name}' not found. Available: {available}"

result = agent.run_sync('Use the arxiv-search skill or let me know what skills are available')

Timeout Handling

For long-running scripts:

from pydantic_ai import RunContext
from pydantic_ai.toolsets.skills import SkillsToolset
from pydantic_ai_skills import SkillScriptExecutionError

class MyDeps:
    database: Database

skills = SkillsToolset()

@skills.skill()
def long_running() -> str:
    return "Execute long-running operations"

@long_running.script
async def expensive_query(
    ctx: RunContext[MyDeps],
    query: str
) -> str:
    """Execute expensive database query.

    Note: This script has a default 30-second timeout.
    For longer operations, increase timeout or break into chunks.
    """
    try:
        result = await asyncio.wait_for(
            ctx.deps.database.execute_complex(query),
            timeout=25  # Leave 5s buffer
        )
        return str(result)
    except asyncio.TimeoutError:
        return "ERROR: Query exceeded time limit. Try with filters or smaller dataset."

Dependency Management

Structured Dependency Types

Use TypedDict for clear dependency contracts:

from typing import TypedDict
from pydantic_ai import Agent, RunContext
from pydantic_ai.toolsets.skills import SkillsToolset

class MyDeps(TypedDict):
    """Dependencies available to skills."""
    database: Database
    cache: Cache
    logger: Logger
    config: AppConfig

skills = SkillsToolset()

@skills.skill()
def db_skill() -> str:
    return "Database operations"

@db_skill.resource
async def get_status(ctx: RunContext[MyDeps]) -> str:
    """Get database status with full type hints."""
    # Type checker knows db_skill.database is Database
    status = await ctx.deps.database.health_check()

    # Log the check
    await ctx.deps.logger.info(f"Health check: {status}")

    return f"Database status: {status}"

# Initialize agent with deps
agent = Agent(
    model='openai:gpt-4o',
    toolsets=[skills],
    deps=MyDeps(
        database=my_database,
        cache=my_cache,
        logger=my_logger,
        config=app_config
    )
)

Optional Dependencies

Handle optional dependencies gracefully:

from typing import Optional

class MyDeps(TypedDict, total=False):
    """Dependencies with optional fields."""
    database: Database      # Required
    cache: Optional[Cache]  # Optional
    logger: Optional[Logger]  # Optional

@db_skill.resource
async def get_cached_status(ctx: RunContext[MyDeps]) -> str:
    """Get status with optional caching."""
    status = await ctx.deps.database.health_check()

    # Cache if available
    if 'cache' in ctx.deps and ctx.deps['cache']:
        await ctx.deps['cache'].set('db_status', status)

    return f"Database status: {status}"

Lazy Initialization

Defer expensive setup:

from pydantic_ai.toolsets.skills import SkillsToolset

skills = SkillsToolset()

@skills.skill()
def expensive_skill() -> str:
    return "Use expensive resources"

@expensive_skill.resource
async def get_expensive_resource(ctx: RunContext[MyDeps]) -> str:
    """Lazy-load expensive resource only when needed."""
    # First call initializes, subsequent calls reuse
    if not hasattr(ctx.deps, '_expensive_resource'):
        ctx.deps._expensive_resource = await initialize_expensive_resource()

    return str(ctx.deps._expensive_resource)

Testing Patterns

Unit Testing Skills

import pytest
from pydantic_ai import RunContext
from pydantic_ai.toolsets.skills import Skill, SkillsToolset

class MockDeps:
    def __init__(self):
        self.data = {"test": "value"}

def test_skill_resource():
    """Test a skill resource independently."""
    skill = Skill(
        name='test-skill',
        description='Test skill',
        content='Test'
    )

    @skill.resource
    async def my_resource(ctx: RunContext[MockDeps]) -> str:
        return ctx.deps.data['test']

    # Test the resource
    mock_deps = MockDeps()
    ctx = RunContext(deps=mock_deps)

    result = asyncio.run(my_resource(ctx))
    assert result == "value"

def test_skill_in_toolset():
    """Test skill integration with toolset."""
    skill = Skill(
        name='test-skill',
        description='Test skill',
        content='Test content'
    )

    toolset = SkillsToolset(skills=[skill])

    # Verify skill is registered
    assert 'test-skill' in toolset.skills
    assert toolset.get_skill('test-skill').description == 'Test skill'

Integration Testing

import pytest
from pydantic_ai import Agent
from pydantic_ai.toolsets.skills import SkillsToolset

@pytest.fixture
def test_skills():
    """Create test skills."""
    skills = SkillsToolset()

    @skills.skill()
    def test_skill() -> str:
        return "Test skill for integration testing"

    @test_skill.script
    def test_script(value: int) -> str:
        return f"Result: {value * 2}"

    return skills

@pytest.mark.asyncio
async def test_agent_with_skills(test_skills):
    """Test agent interaction with skills."""
    agent = Agent(
        model='openai:gpt-4o',
        toolsets=[test_skills]
    )

    # Agent can access skills
    result = await agent.run('List available skills')
    assert 'test-skill' in result.data

See Also

• Advanced Features - Decorator patterns and custom executors
• Programmatic Skills - Creating skills in code
• Creating Skills - File-based skill creation