Voice and the Call Center: The Prompt Designs the Conversation
A text assistant answers in paragraphs, a voice one in real-time turns, and that changes the prompt. The answer can't be scanned by eye, the user interrupts, the audio is sometimes unclear, and latency is audible. So turn-taking rules matter more than a nice persona.
# Role and objective
A delivery-service voice assistant. Help check status, change the time,
or hand the conversation to an agent.
# Language and voice
- Answer in the language of the current turn unless the user explicitly chose another.
- Short phrases, natural conversational wording, no markdown or lists.
# Turn behavior
- One question per turn.
- Don't repeat already confirmed data.
- Before a tool call say briefly: "Let me check the order now".
- After a tool result, state the fact first, then propose one next step.
# Unclear audio
- Don't guess numbers, names, dates, or addresses.
- Ask to repeat only the unclear part.
- For a critical entity, repeat the recognized value and wait for confirmation.
# Interruptions
- Stop the current answer immediately on barge-in.
- Handle the new turn first; don't continue the old text automatically.
# Escalation
- Offer an agent on an explicit request, two failed recognitions,
a policy conflict, or the unavailability of a required tool.Every voice aspect needs both a model instruction and runtime support.
- Latency. In the prompt - a short preamble before a slow tool. In runtime - streaming and low effort for simple turns.
- Barge-in. In the prompt - don't continue an interrupted answer. In runtime - cancel playback and the active response.
- The order number. In the prompt - don't guess, repeat it for confirmation. In runtime - an entity parser and a checksum.
- A write action. In the prompt - state the change in plain words. In runtime - an explicit confirmation token.
- Handing off to an operator. In the prompt - briefly explain the reason and the next step. In runtime - handoff state and a conversation summary.
And about the persona - it is needed, but not first.
The persona comes after the process. Google advises defining the persona first, then the conversation rules and guardrails; OpenAI - short named sections: tools, unclear audio, entity capture, escalation. But the product flow and tool availability still matter more than the timbre of the voice.