Skip to main content
These are practical prompt snippets you can drop into your agent’s prompt_config.prompt to handle situations that come up on almost every voice agent. Each one solves a specific problem with tested wording.

Pronouncing phone numbers

Models tend to read phone numbers as large integers (“four billion, one hundred fifty-eight million…”) or rush through digits without pauses. Add this to your prompt to force digit-by-digit pronunciation:
This works across all TTS providers. The dash characters create natural pauses in speech.

Spelling out email addresses

Email addresses are tricky because they mix letters, symbols, and domain names. Instruct your agent to spell them out clearly:

Pronouncing website URLs

Speaking times and dates

Models default to written formats like “3:30 PM” which TTS can mangle. Be explicit:

Handling silence and holds

When a caller says “hold on” or goes quiet, you usually don’t want the agent to keep talking. How to handle this depends on your LLM provider.

Standard approach

Add this to your prompt:
Pair this with your prompt_config settings — set ai_speak_after_silence to true and ai_speak_wait_time to a comfortable duration (3–5 seconds) so the agent gently re-engages if the silence goes on too long.

Adjusting silence behaviour

The tts_config.responsiveness setting (0–1) controls how quickly the agent jumps in after the caller stops speaking. Lower values give callers more space; higher values make the agent respond faster. For calls where callers might need time to think (scheduling, providing account numbers), set this around 0.3–0.5.

Handling interruptions

The tts_config.interruption_sensitivity setting (0–1) controls how easily callers can cut in. Higher values mean the agent yields more readily. For conversational agents, 0.7–0.8 works well.

Confirming important information

When your agent collects critical details — booking times, phone numbers, spelling of names — always confirm:

Handling “I don’t know” gracefully

Keeping responses concise

The single most common problem with voice agents is responses that are too long. Add this as a hard constraint: