Skip to main content
The prompt is the most important part of your agent configuration. It tells the agent who it is, what it should do, and how it should behave during calls. A well-written prompt leads to natural, reliable conversations.

Prompt Structure

A good agent prompt typically includes these sections:

Background

Tell the agent who it is and set the context for the conversation.

Goal

Clearly state what the agent should accomplish on the call.

Instructions

Provide step-by-step guidance on how the conversation should flow.

Behavioral Notes

Add rules that shape how the agent speaks and handles edge cases.

Using Parameters

You can reference dynamic values in your prompt using {{parameter_name}} syntax. These values are passed in when the call is created — either through the API or from a campaign CSV upload.
For more than simple substitution — conditional logic, branching, and randomized A/B variation — see Dynamic Prompts.

Voice-Specific Tips

Voice agents behave differently from text-based AI. Keep these in mind:

Reading Numbers

Numbers need to be read digit by digit for clarity. Instruct the agent to read them slowly:
For prices, instruct the agent to spell them out:

Spelling Names and Emails

Names and email addresses are often misheard on a phone call. Instruct the agent to offer to spell them:

Handling Hold and Transfers

Agents may be placed on hold or transferred. Make this explicit:

Avoiding Unnatural Output

Phone calls should sound like natural speech, not text output:

Full Prompt Example

Here is a complete example prompt that incorporates the patterns above:

Best Practices

  • Be specific — vague prompts produce vague conversations. Tell the agent exactly what to do in each situation.
  • Test with real calls — make a few test calls after writing your prompt to hear how it sounds in practice.
  • Handle edge cases — think about what happens if the person does not answer, asks to be called back, or asks an unexpected question.
  • Keep it focused — one prompt should cover one use case. If you need different behavior for different scenarios, use flow-based agents with separate prompts per node.
  • Iterate — prompts rarely work perfectly the first time. Listen to recordings, spot issues, and refine.