Prompt Design

Prompt Design

The prompt is the "brain" of your agent—it tells it how to behave, what to say, and when to use tools. Keep prompts clear and structured for your first agent.

Step 1: Understand the Prompt Structure

Prompts are JSON objects with sections for personality, rules, and examples. Use the JSON Editor tool to create and edit them.

Basic template:

{
  "agent": {
    "name": "Your Agent Name",
    "tone": "friendly and helpful",
    "style_rules": ["Always confirm user details before proceeding."]
  },
  "capabilities": {
    "tools": {
      // References to tools (detailed in Tool Design)
    }
  },
  "dialog_flow": {
    "high_level": ["Greet user", "Understand intent", "Use tools if needed", "Confirm and close."]
  },
  "examples": [
    // Sample conversations
  ]
}

Step 2: Write the System Prompt

Define the agent's core instructions:

• Tone and rules: "Be warm and concise. Never guess—use tools for data."
• Tool triggers: "If the user asks about products, use the 'search_products' tool."
• Guardrails: "Escalate to human if the query is complex."

Step 3: Add Examples

Include 2-3 sample interactions to guide behavior.

• User: "What's available?" Agent: "Let me check... [Uses tool] Here are the options."

Step 4: Reference Core Tools

Core tools like getCurrentTime or show_component are built-in. Mention them in the prompt if relevant (e.g., "Use show_component to display images.").

Step 5: Test the Prompt

Save via MongoDB Updater and start the agent. Speak simple queries and refine based on responses.

Tip: Start simple—add complexity later. Try a demo agent first to see effective prompts.