Skip to main content
Writing for voice agents is fundamentally different from writing for text chatbots. In a voice conversation, the user cannot “scan” the text; they must listen linearly. Long, complex sentences increase cognitive load and latency.

​
1. Be Concise

Bad:
“Thank you for calling Acme Corp, the leading provider of widgets in the tri-state area. I understand you are calling about an order status. I can certainly help you with that today. Could you please provide your order number?”
Why it fails:
  • Takes 10 seconds to say.
  • The user forgets the question by the end.
  • Increases latency (more tokens to generate).
Good:
“Thanks for calling Acme. I can help with your order. What’s the order number?”
Tip: Add Keep your responses short and conversational, ideally under 20 words. to your system prompt.

​
2. Avoid Special Characters

LLMs sometimes output markdown, lists, or URLs. These sound terrible when read by TTS. Instruction:
“Do not use markdown, bullet points, or emojis. Do not read out URLs; instead say ‘check our website’.“

​
3. Handle Interruptions

Users will interrupt. Design your prompt to handle context switching gracefully. Instruction:
“If the user interrupts or changes the topic, stop what you are saying and address their new point immediately.”

​
4. Pacing and Fillers

To make the agent sound more human, you can instruct it to use fillers, but use them sparingly. Instruction:
“Use natural fillers like ‘hmm’ or ‘let me check’ only when retrieving data via a tool.”

​
5. Phone Number Formatting

TTS engines sometimes read phone numbers as “one million…” instead of “one, zero, zero…”. Instruction:
“When speaking phone numbers, say them digit by digit, grouping them for clarity (e.g., 555-123-4567).”