Most voice agents fail for one reason: a lazy prompt. A blob of "you are a helpful assistant for Bob's Plumbing, be nice" will demo fine and then fall apart the moment a real caller does something unexpected. A production voice prompt has six required sections. Skip any one and you've built a liability. Here's what goes in each.
1. Identity
Who the agent is, in two or three sentences. Name (or "the virtual assistant for"), the business, the vertical, and whether it handles inbound or outbound. Include the disclosure rule: identify as a virtual assistant when asked. This section sets the frame the whole call runs on.
You are the virtual assistant for Northside Plumbing, a home-services business. You handle inbound calls. If asked, say you're a virtual assistant.
2. Context & knowledge
What the agent is allowed to know. Hours, services, policies, and — critically — the pricing rule, not a price list to recite. This is where you scope the knowledge so it can't wander. The single most important line in the whole prompt: if the information isn't here, do not guess.
Hours: Mon–Fri 8–6. Never quote a firm price on jobs that need a site visit — offer to schedule instead.
3. Guardrails
The hard rules that keep the agent safe. No inventing information. Confirm every important detail out loud (name, number, address, time). No medical, legal, or financial advice. One question at a time. These aren't suggestions — write them as absolute rules, because the model treats them that way.
4. Task / call flow
The actual job, as a numbered sequence. Greet → understand → drive to the goal → collect and confirm → close. Be specific about the goal ("book a service appointment", "qualify and transfer"). Without an explicit flow, the agent meanders and callers hang up.
- Greet. 2. Find out what they need. 3. Book the appointment. 4. Read back name, number, address, time. 5. Confirm what happens next.
5. Escalation
When and how to hand off to a human — the section that prevents disasters. Trigger on: emotional callers, "let me speak to a person", the word "emergency", and anything out of scope. Specify the destination (warm transfer with context, or take a message + callback window). An agent without a defined escalation path is a trap waiting to spring in front of a customer.
6. Style
How it sounds. Tone, sentence length, pacing. Short sentences. Speak like a person, not a brochure. Light acknowledgments ("sure", "got it") so the caller feels heard. Keep responses tight to stay low-latency — a monologuing agent feels robotic no matter how good the voice is.
The mistakes that break agents
- No "don't guess" rule → confident hallucination on the first edge question.
- No confirmation step → booked appointments with wrong numbers and no-shows.
- No escalation triggers → the agent argues with an angry caller instead of transferring.
- Everything in one paragraph → the model can't prioritize; rules get ignored under pressure.
- A price list instead of a pricing rule → it quotes a number it shouldn't and the client is furious.
How to use this
Write all six sections, in this order, as clearly labeled blocks. Then test it adversarially (ask it something out of scope, get angry, give a wrong detail then correct it). Every failure points back to a thin section — usually guardrails or escalation. Tighten and re-test.
A good voice prompt isn't clever. It's complete. Six sections, each doing its job, is the difference between an agent that survives real calls and one that embarrasses your client the first busy afternoon.