System Prompt

A System Prompt (also called an MCP Prompt) is your primary way to customize how your MCP client interacts with the MCP Server and LLM. You define your own System Prompt to align the agent's behavior with your application's brand, business logic, and custom toolset.

You have full control over the System Prompt to tailor the agent's behavior to your specific needs. The example template provided by Gracenote is only a starting point. Here are some things to customize:

• Persona and Tone: Define your agent's identity (e.g., "polite concierge," "technical expert," "casual friend").
• Tool Orchestration: Tell the Agent which custom APIs it has access to, in addition to the Gracenote tools, and the rules for calling them. For example, you may have your own APIs for user history or entitlements that your user must access before performing a query.
• Business Rules: Implement specific logic, such as prioritizing content available on your own streaming service or applying content rating filters before making recommendations.
• Format: specific output format, such as compliance with a desired JSON schema. LLM output formatting can increase latency, and is often specified to best effect via langchain or similar.

The example below is a simple example that tells the LLM what it is (an expert in recommending entertainment) and how it should format results (JSON, to be consumed by a CTV UI). You can learn more about system prompts from the Anthropic MCP Server Prompts documentation.

Code
 
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 the mentions of entertainment entities:

- People (actors, acctresses, directors, writers, etc.)

- Works (movies, TV series/episodes)

- Characters

- Franchises

- Use the MCP tool 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.

FINAL OUTPUT FORMAT

- Your final response must be a single JSON object matching the agent's output schema.

- Do NOT include any natural language outside of JSON.

- If the tool result is ambiguous and contains multiple candidates, choose the single best candidate using the user's question as context.

Return ONLY the JSON object as final output, no markdown, no explanations.