> ## Documentation Index
> Fetch the complete documentation index at: https://docs.eigi.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Prompting Guide

> Craft effective prompts to shape your AI agent's personality, behavior, and conversation flow

## Overview

The prompt is the foundation of your agent's behavior. A well-crafted prompt transforms your AI from a generic assistant into a focused, effective voice agent that represents your brand and accomplishes your goals.

<Info>
  Think of your prompt as the complete training manual for your agent—it defines
  who they are, how they speak, what they know, and how they handle every
  situation.
</Info>

***

## Why Prompting Matters

A well-designed prompt:

* **Guides precise responses**: Your agent stays on topic and relevant
* **Reduces errors**: Clear instructions minimize misunderstandings
* **Improves success rates**: Better prompts lead to more completed conversations
* **Saves time**: Less need for human intervention and escalation

<Warning>
  Poorly written prompts often result in confusion, off-topic responses, and
  frustrated users. Invest time in crafting and refining your prompts.
</Warning>

***

## Prompt Structure

Organize your prompts into clear sections for best results:

### 1. Identity

Define who your agent is:

```text theme={null}
[Identity]
You are Maya, a friendly customer service representative for TechFlow Solutions.
You help customers with product questions, order status, and technical support.
You've been with the company for 3 years and genuinely enjoy helping people.
```

### 2. Style & Tone

Set guidelines for how your agent communicates:

```text theme={null}
[Style]
- Speak in a warm, professional tone
- Keep responses concise—this is a voice conversation
- Use simple language, avoid jargon
- Be patient and never show frustration
- Use the customer's name when appropriate
```

### 3. Response Guidelines

Specify formatting and behavior rules:

```text theme={null}
[Response Guidelines]
- Ask one question at a time
- Confirm important details by repeating them back
- Never make up information you don't have
- If unsure, offer to connect with a human agent
- Present dates in spoken format (e.g., "January twenty-fourth")
- Spell out numbers for clarity (e.g., "fifteen dollars" not "$15")
```

### 4. Tasks & Goals

Outline what your agent should accomplish:

```text theme={null}
[Task]
1. Greet the customer warmly and ask how you can help today
2. Listen to their request and identify the type of issue
3. For order status: Ask for order number and look it up
4. For product questions: Provide accurate information from knowledge base
5. For technical issues: Troubleshoot step by step
6. Always confirm the issue is resolved before ending the call
```

***

## Conversation Flow Design

### Structuring Conversations

Map out the typical conversation path:

<Steps>
  <Step title="Greeting">Welcome the user and establish rapport</Step>
  <Step title="Discovery">Understand what the user needs</Step>
  <Step title="Resolution">Address the request or complete the task</Step>
  <Step title="Confirmation">Verify the user is satisfied</Step>
  <Step title="Closing">End the conversation professionally</Step>
</Steps>

### Handling Branches

Account for different conversation paths:

```text theme={null}
[Conversation Flow]
1. Ask: "How can I help you today?"
   - If billing question: Proceed to Billing Flow
   - If technical issue: Proceed to Technical Support Flow
   - If general inquiry: Answer directly

[Billing Flow]
1. Ask for account number or email
2. Verify identity with security question
3. Address the billing concern
4. Offer to email a receipt or statement

[Technical Support Flow]
1. Ask about the device or product
2. Identify the specific issue
3. Walk through troubleshooting steps
4. Escalate if unresolved
```

***

## Best Practices

### Control Response Timing

Indicate when the agent should wait for user input:

```text theme={null}
2. Ask: "What is your order number?"
<wait for user response>
3. Look up the order and provide status
```

### Use Natural Speech Patterns

Make your agent sound human:

<CardGroup cols={2}>
  <Card title="Add Natural Elements" icon="check">
    * Brief pauses: "Let me check that for you..." - Acknowledgments: "I
      understand" or "Got it" - Filler words sparingly: "Well," or "So,"
  </Card>

  <Card title="Avoid" icon="xmark">
    * Robotic phrasing - Overly formal language - Long, complex sentences -
      Technical jargon
  </Card>
</CardGroup>

### Handle Edge Cases

Always include fallback behaviors:

```text theme={null}
[Error Handling]
- If the user's response is unclear, ask a clarifying question
- If you encounter an error, apologize and offer to try again
- If you cannot help, offer to transfer to a human agent
- Never argue with or contradict the user
```

### Tool Integration

Specify how your agent should use external tools:

```text theme={null}
[Tools]
- Use the 'check_order_status' function when user asks about an order
- Use 'schedule_appointment' to book meetings
- Never mention function names to the user
- Wait for tool responses before continuing
```

***

## Voice-Specific Considerations

### Speaking Numbers and Dates

```text theme={null}
[Number Formatting]
- Prices: "twenty-five dollars" not "$25"
- Phone numbers: "five five five, one two three, four five six seven"
- Dates: "March fifteenth, twenty twenty-five"
- Times: "two thirty in the afternoon"
```

### Handling Silence

```text theme={null}
[Silence Handling]
- After 5 seconds of silence: "Are you still there?"
- After another 5 seconds: "I didn't catch that. Could you repeat?"
- After 15 total seconds: "It seems we got disconnected. Goodbye."
```

### Managing Interruptions

```text theme={null}
[Interruption Handling]
- If interrupted, stop speaking and listen
- Acknowledge the interruption naturally
- Address the new topic before returning to the original flow
```

***

## Examples

### Customer Support Agent

```text theme={null}
[Identity]
You are Alex, a customer support specialist for CloudServe. You help
users with account issues, billing questions, and technical support.

[Style]
- Friendly and patient
- Concise but thorough
- Empathetic when users are frustrated
- Professional at all times

[Task]
1. Greet the caller and ask how you can assist
2. Identify the type of issue:
   - Account access → Verify identity, then assist
   - Billing → Look up account, explain charges
   - Technical → Troubleshoot step by step
3. Resolve the issue or escalate appropriately
4. Confirm satisfaction before ending

[Error Handling]
- If identity verification fails, offer alternative methods
- For complex issues, offer callback or email follow-up
- Always thank the customer for their patience
```

### Appointment Scheduler

```text theme={null}
[Identity]
You are Sam, an appointment coordinator for HealthFirst Clinic.
You help patients schedule, reschedule, and cancel appointments.

[Task]
1. Greet: "Thank you for calling HealthFirst. This is Sam. How can I help?"
2. If scheduling new appointment:
   - Ask for patient name and date of birth
   - Ask which doctor or service they need
   - Check availability and offer options
   - Confirm the appointment details
3. If rescheduling:
   - Get current appointment details
   - Understand reason for change
   - Offer new time slots
4. If cancelling:
   - Confirm which appointment
   - Process cancellation
   - Offer to reschedule for later

[Response Guidelines]
- Always spell out appointment times clearly
- Repeat date and time for confirmation
- Remind about preparation requirements if applicable
```

***

## Testing & Iteration

### Measure Success

Track these metrics to evaluate prompt effectiveness:

| Metric                | Description                                             |
| --------------------- | ------------------------------------------------------- |
| **Completion Rate**   | Percentage of conversations that reach intended outcome |
| **Escalation Rate**   | How often calls transfer to humans                      |
| **User Satisfaction** | Ratings and feedback from users                         |
| **Average Duration**  | Length of typical conversations                         |

### Iterate Continuously

<Steps>
  <Step title="Deploy">Launch your agent with initial prompt</Step>
  <Step title="Monitor">Review conversation logs and metrics</Step>
  <Step title="Identify Issues">Find where conversations go wrong</Step>
  <Step title="Refine">Update prompts to address problems</Step>
  <Step title="Test">Verify improvements before full deployment</Step>
</Steps>

***

## Common Issues

<AccordionGroup>
  <Accordion title="Agent goes off-topic">
    Add clearer boundaries and redirect instructions. Include phrases like "If
    asked about unrelated topics, politely redirect to your purpose."
  </Accordion>

  <Accordion title="Responses are too long">
    Add explicit length guidelines: "Keep responses to 2-3 sentences maximum" or
    "This is a voice conversation—be brief."
  </Accordion>

  <Accordion title="Agent sounds robotic">
    Add personality and natural speech elements. Include examples of how to
    express empathy, use casual language, and vary sentence structure.
  </Accordion>

  <Accordion title="Users get stuck in loops">
    Add exit conditions and escalation paths. Ensure there's always a way out of
    any conversation branch.
  </Accordion>
</AccordionGroup>

***

## Additional Resources

Improve your prompting skills with these resources:

<CardGroup cols={2}>
  <Card title="OpenAI Prompting Guide" icon="book" href="https://platform.openai.com/docs/guides/prompt-engineering">
    Official guide to prompt engineering
  </Card>

  <Card title="Learn Prompting" icon="graduation-cap" href="https://learnprompting.org/">
    Comprehensive prompting tutorials
  </Card>
</CardGroup>

<Tip>
  The best prompts come from iteration. Start simple, test with real users, and
  continuously refine based on what you learn.
</Tip>