II. Implementation

System Prompt

Overview

A System Prompt is your primary way to customize how your LLM interacts with the MCP Server. It defines your agent's behavior, persona, and rules for using MCP tools. The system prompt is a set of instructions sent to the LLM that:

• Defines the LLM's persona and tone
• Sets business rules and constraints
• Determines output format

You have full control over the system prompt to tailor the agent's behavior to your specific needs.

How to Use Your System Prompt

Your system prompt is passed to the LLM when making requests. For example:

Code
 
system_prompt = """
You are a helpful assistant that provides clear and concise responses to user queries.

Your job:
- Read the user's request and understand the intent
- Identify mentions of entertainment entities (movies, TV shows, actors, etc.)
- Use the MCP tools as the SINGLE source of truth for IDs and metadata

RULES FOR TOOL USE:
- If there is at least one entity mention, you MUST call the MCP tool before producing a final answer
- Always verify information against Gracenote data

FINAL OUTPUT FORMAT:
- Provide clear, natural language responses
- Include relevant metadata (year, cast, availability)
- Be concise but informative
"""

llm_response = llm.generate(
    messages=[
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_query}
    ],
    tools=mcp_tool_definitions
)

See Connecting to the MCP Server for complete implementation.

What to Customize

1. MCP Server Usage Instructions (Required)

Always include these instructions to help your LLM format requests correctly:

Code
 
MCP Server Instructions:
Use this server to search for movies and TV series from Gracenote ON data. The canonical input for resolve_entities is: {'request': [{'title': 'Star Trek', 'entityType': 'Show', 'yearOfRelease': 1966, 'language': 'en', 'countryCode': 'USA'}]}. Use same format, but with entityType 'Movie' for a movie and entityType 'Episode' for an episode. Convert root ids to tmsIds, search for availabilities by tmsId. See the individual tool descriptions.

This helps your LLM understand:

• A suggested format for resolve_entities requests
• When to use 'Movie' vs 'Show' vs 'Episode' entity types
• That root IDs need conversion to tmsIds for availability lookups
• To check individual tool descriptions for detailed parameters

2. Persona and Tone

Define your agent's identity:

Code
 
You are a friendly entertainment concierge helping users discover content.
Be conversational, enthusiastic, and helpful.

Or:

Code
 
You are a technical assistant providing precise entertainment metadata.
Be factual, concise, and accurate.

3. Tool Orchestration

Tell the agent how to orchestrate tools, such as tool sequence if needed:

Code
 
Always use resolve_entities first to identify content before checking availability.

4. Business Rules

Implement specific logic for your application:

Code
 
Business Rules:
- Prioritize content available on [Your Streaming Service]
- Filter out content rated above PG-13 for family mode
- Always include "where to watch" information in responses
- Prefer HD quality streams when available

5. Output Format

Specify how responses should be structured based on your use case:

For user-facing applications, use natural language responses. Example applications include:

• Chatbots and conversational interfaces
• Search interfaces where users see results directly
• Recommendation engines with explanations
• Voice assistants
• Customer-facing web/mobile apps

Prompt Example:

Code
 
Provide responses in natural, conversational language.
Include title, year, and brief description.
End with streaming availability if relevant.

For data processing applications (e.g., catalog enrichment, API endpoints), specify a structured output format. Examples include:

• Batch processing and ETL pipelines
• Data harmonization and catalog enrichment
• API endpoints that return structured data
• Database updates and synchronization
• Microservices and backend integrations
• Any application that parses the response programmatically

Prompt Example:

Code
 
You must respond with a single JSON object and nothing else.
Do not include any text, explanation, markdown formatting, or code fences before or after the JSON.
Your entire response must be parseable by json.loads() with no preprocessing.

Required JSON schema:
{"tmsId": "string", "title": "string", "year": number, "genres": ["string"], "cast": ["string"]}

Example of a correct response:
{"tmsId": "MV000722830000", "title": "The Matrix", "year": 1999, "genres": ["Action", "Sci-Fi"], "cast": ["Keanu Reeves"]}

Claude-specific note: Claude models may wrap JSON in markdown code fences (```json ... ```) or add introductory prose even when instructed not to. If your application parses structured JSON from Claude, include defensive parsing that strips code fences and extracts JSON from surrounding text. See the Catalog Enrichment example for a robust implementation. Other LLMs (GPT-4, Gemini, Llama) may not require these refinements, especially when using native JSON mode features.