Skip to main content

​
Creating Your AI Persona

The most important part of your AI agent is the instructions (system prompt). This defines who the agent is, what they do, and how they behave.

​
Writing Effective Instructions

Your instructions should include:

​
1. Identity & Role

You are Sarah, a senior appointment coordinator at TopView Dental.
You have 5 years of experience and are known for being warm and professional.

​
2. Goal & Context

Your goal is to confirm {{patient_name}}'s appointment scheduled for 
{{appointment_date}} at {{appointment_time}}.

The patient has been a client for 3 years and prefers afternoon appointments.

​
3. Tone & Style

Be warm, professional, and concise. Use the patient's first name naturally.
If they seem rushed, keep it brief. If they want to chat, be friendly but 
redirect to the appointment confirmation.

​
4. Handling Scenarios

- If they want to reschedule: Ask for their preferred date/time and use 
  the 'check_availability' tool to find slots.

- If they cancel: Express understanding, ask if everything is okay, but 
  do not pressure them. Offer to call back if they change their mind.

- If they have questions: Answer directly or use the 'lookup_patient_info' 
  tool if you need specific details.

​
5. Boundaries

- Never discuss pricing or billing (transfer to billing department)
- Never make medical diagnoses
- Always confirm the appointment details before ending the call

​
Complete Example

Here’s a complete instruction set for an appointment reminder: 
You are Sarah, a friendly appointment coordinator at TopView Dental.

Your goal is to confirm {{patient_name}}'s dental appointment scheduled 
for {{appointment_date}} at {{appointment_time}}.

Be warm, professional, and concise. Use the patient's first name naturally.

If they want to reschedule:
- Ask for their preferred date and time
- Use the 'check_availability' tool to find available slots
- Confirm the new appointment immediately

If they cancel:
- Express understanding
- Ask if everything is okay (but don't pressure)
- Offer to call back if they change their mind

If they have questions:
- Answer directly if you know the answer
- Use the 'lookup_patient_info' tool for specific patient details
- Transfer to the front desk for complex questions

Never discuss pricing or billing. Always confirm appointment details 
before ending the call.

​
Voice Selection

​
Realtime Mode Voices

In Realtime Mode, choose from OpenAI’s preset voices. The default voice is alloy, which works well for most use cases. Other voices include echo, shimmer, ash, ballad, coral, sage, verse, marin, and cedar.
Realtime Mode voices are optimized for low latency and natural conversation. They work best for English but can handle other languages with proper instructions. Check the current list of available voices via the API endpoint GET /v1/voices/builtin?provider=openai or contact support.

​
Legacy Mode: Custom Voices

Legacy Mode unlocks unlimited voice customization:

​
ElevenLabs Voices

  • Thousands of voices from the ElevenLabs Voice Library
  • Voice cloning: Upload 1-5 minutes of audio to clone any voice
  • Multiple languages supported
  • Custom voice IDs: Use any voice ID from ElevenLabs
Check the current list of available ElevenLabs voices via the API endpoint GET /v1/voices/builtin?provider=elevenlabs or browse the ElevenLabs Voice Library directly.
{
  "voice": "21m00Tcm4TlvDq8ikWAM",
  "tts_provider": "elevenlabs",
  "mode": "legacy"
}

​
Deepgram Aura-2 Voices

  • 90+ voices across multiple languages with natural breathing and intonation
  • 7 languages: en, es, de, fr, nl, it, ja
  • Ultra-fast TTS (~100ms latency)
  • Pattern: aura-2-{name}-{lang}
Check the current list of available Deepgram Aura-2 voices via the API endpoint GET /v1/voices/builtin?provider=deepgram or contact support for the latest voice availability.
{
  "voice": "aura-2-apollo-en",
  "tts_provider": "deepgram",
  "mode": "legacy"
}

​
Language & Dialect Control

​
Realtime Mode

  • Auto-detection: Works automatically but best with explicit instructions
  • Instruction-based: Tell the AI “Speak only Spanish” or “Respond in French”
  • Limited control: No explicit dialect selection

​
Legacy Mode

  • Full control: Set stt_language for speech recognition
  • 36+ languages: en-US, en-GB, en-AU, es-ES, es-MX, fr-FR, de-DE, etc.
  • Dialect-specific: Choose British English (en-GB) vs American (en-US)
{
  "mode": "legacy",
  "stt_language": "en-GB",
  "voice": "aura-2-electra-en",
  "tts_provider": "deepgram"
}

​
Temperature & Creativity

Control how creative or focused your AI is: 
{
  "temperature": 0.7  // 0.0 = focused, 1.0 = creative
}
  • 0.0-0.3: Very focused, consistent responses (good for confirmations)
  • 0.4-0.7: Balanced (default, good for most use cases)
  • 0.8-1.0: Creative, varied responses (good for sales, engaging conversations)

​
First Sentence

Set the opening line to control how the call starts: 
{
  "first_sentence": "Hi {{name}}, this is Sarah from TopView Dental. I'm calling to confirm your appointment tomorrow at 3 PM."
}
The first sentence is critical—it sets the tone and immediately establishes context. Make it clear, friendly, and specific.

​
Best Practices

​
✅ Do This

  • Be specific: Include exact scenarios and how to handle them
  • Set boundaries: Clearly define what the AI should and shouldn’t do
  • Use variables: Leverage {{variable}} syntax for personalization
  • Test thoroughly: Try different scenarios to ensure the AI behaves correctly
  • Iterate: Refine instructions based on real call transcripts

​
❌ Don’t Do This

  • Vague instructions: “Be helpful” is too generic
  • Conflicting goals: Don’t ask the AI to both sell and not be pushy without clear boundaries
  • Missing context: Provide relevant information about the customer, product, or situation
  • Too long: Keep instructions focused—aim for 200-500 words

​
Advanced: Multi-Turn Conversations

For complex conversations, break down the flow: 
1. Greeting: Introduce yourself and state the purpose
2. Confirmation: Confirm the appointment details
3. Questions: Answer any questions they have
4. Next steps: Confirm what happens next
5. Closing: Thank them and end professionally

​
Next Steps

Custom Voices Guide

Learn how to use ElevenLabs voice cloning and Deepgram voices.

Function Calling

Give your AI tools to interact with your systems in real-time.