How to Write a Voice AI Receptionist Prompt on Retell AI (The Structure I Use)

TL;DR: A voice AI receptionist lives or dies on its prompt, not its voice. The agents that book appointments and never embarrass the client share the same skeleton: a tight identity, a hard scope, a step-by-step flow for the two or three calls that actually matter, explicit escalation rules, and guardrails against the model improvising. I build production voice agents on Retell AI for US clients and run VoiceDash, a white-label client portal for voice AI agencies, so this is the exact prompt structure I use and how I tune it from real call transcripts.

Most people building their first Retell agent write the prompt like they are briefing a smart human. They type a paragraph like "You are a friendly receptionist for a dental office. Answer questions and book appointments." Then they are surprised when the agent invents prices, promises things the business does not offer, or freezes when a caller says something off-script.

The fix is not a longer paragraph. It is structure. A voice AI prompt is closer to a spec than a description, and the best ones read like a checklist the model cannot wander away from. Here is how I build them.

Why voice prompts are different from chatbot prompts

Before the structure, one thing that trips up almost everyone: a voice prompt is not a chatbot prompt with the word "say" added.

Voice is real time and unforgiving. There is no scrollback, no way for the caller to re-read, and every pause feels like the line went dead. That changes what a good prompt has to do:

• Answers must be short. A chatbot can dump three paragraphs. A receptionist that talks for 20 seconds straight gets interrupted or hung up on. I instruct agents to answer in one or two sentences and then ask a question.
• The agent has to drive. Callers do not know what information you need. The prompt has to make the agent lead the conversation to an outcome, not wait to be told what to do.
• Numbers and spellings need special handling. Phone numbers, dates, and names spoken aloud get mangled by speech-to-text. The prompt has to tell the agent to confirm them back.

Keep those three realities in mind and the structure below makes sense.

The five-part structure of a receptionist prompt

Every reliable agent I ship has the same five sections, in this order. Retell lets you put all of this in the general prompt, and the order matters because the model weights earlier instructions heavily.

1. Identity and role

Two or three sentences, no more. Who the agent is, what business it answers for, and the single job it exists to do.

You are Riley, the virtual receptionist for Bright Smile Dental in Austin, Texas. You answer inbound calls, book and reschedule appointments, and answer basic questions about the practice. You are warm, brief, and efficient.

Notice it names one primary job. An agent told to "do everything" does nothing well. If a business needs booking, FAQs, and lead intake, those are still one agent, but the identity should rank them so the model knows what to prioritize.

2. Scope and hard boundaries

This is the section that separates a demo from a production agent. You have to tell the model, explicitly, what it does not do. Voice models are eager to be helpful, and left unbounded they will quote prices, give medical or legal opinions, and promise callbacks nobody scheduled.

You do NOT give clinical advice, quote exact prices, or confirm insurance coverage. If asked, say a team member will confirm details and offer to book a consultation. Never invent information. If you do not know something, say so and offer to take a message.

"Never invent information" is the single most valuable line in a voice AI prompt. Everything a business hates about AI receptionists traces back to a model that guessed instead of deferring.

3. The conversation flow

Here is where beginners under-specify. For the one or two call types that matter most, write the actual steps, in order. Do not describe the goal and hope the model figures out the path.

For a dental receptionist, booking is the money flow, so I spell it out:

When a caller wants an appointment:

1. Ask if they are a new or existing patient.
2. Ask what they need (cleaning, checkup, emergency, specific concern).
3. Ask for their preferred day and time.
4. Check the calendar and offer the two closest matching slots.
5. Collect full name, phone number, and date of birth.
6. Confirm the appointment details back to them before booking.

Steps five and six are non-negotiable. Confirming details back catches the phone number that speech-to-text heard as "five one three" when the caller said "five nine three." I make the agent read critical data back every single time.

4. Escalation and handoff rules

An honest agent knows its limits. The prompt has to define exactly when the agent stops trying and hands off, and how. This is what earns a client's trust, because the failure mode they fear most is the agent confidently mishandling an angry or urgent caller.

Transfer to a human immediately if: the caller has a dental emergency (severe pain, trauma, bleeding), explicitly asks for a person, or becomes frustrated. To transfer, say "Let me connect you with someone who can help right away" and trigger the transfer. If no one is available, take a detailed message and confirm a callback.

I cover the operational side of this, phone numbers and porting and go-live, in how to onboard a voice AI client. The prompt is only half of a safe handoff; the routing has to exist too.

5. Style and guardrails

Last, the delivery rules. Short because the model already got its identity up top.

Keep every response to one or two sentences. Spell out and confirm all phone numbers, names, and dates. Do not use jargon. If interrupted, stop and listen. End calls politely once the caller's need is met.

Write for the ear, not the page

A prompt can be structurally perfect and still produce an agent that sounds like a robot reading a form. A few habits fix that.

Tell the agent to use contractions and everyday phrasing. "I can get you in Thursday at 2" beats "I am able to schedule an appointment for you on Thursday at 2:00 PM." Give it one or two acknowledgment phrases so it does not repeat "Great!" on every turn. And explicitly allow it to handle small talk briefly before steering back, because a caller who says "how's your day" and gets ignored feels like they are talking to a machine.

The thing you are optimizing is not intelligence, it is naturalness under constraint. The smartest-sounding agents are usually the most tightly scoped ones, because they never wander into territory where they sound uncertain.

Tune the prompt from real transcripts, not guesses

Here is the part almost everyone skips, and it is where the real quality comes from. Your first prompt is a hypothesis. The calls are the data.

Once an agent is live, I read transcripts, not to admire them, but to hunt for the exact moments the agent went off the rails. A caller asked something the scope section did not cover. The agent quoted a range it should have deferred on. It booked without confirming and got the number wrong. Every one of those is a one-line prompt fix, and the transcript tells you exactly which line to add.

This is why I do not treat call recordings and transcripts as a client-facing nicety only. They are my debugging tool. Every voice AI prompt I have that works is version 15, not version 1, and the versions came from reading what actually happened on the phone. I made the broader case for why this visibility matters in voice AI client reporting.

The catch is that Retell's own dashboard is built for a developer poking at one agent, not for reviewing calls across every client at a glance, and definitely not for letting the client themselves flag a bad call. That is a big part of why we run VoiceDash. I connect Retell once, and every agent's calls, recordings, and searchable transcripts land in one branded place. When a client says "the agent messed up on Tuesday," I open the transcript, find the turn, and ship a prompt fix, often before the next call comes in. The client sees a polished portal on their own domain; I get the fastest possible loop from "that went wrong" to "fixed."

A prompt checklist you can steal

Before I mark any agent ready for real calls, I run it against this list:

• Identity names one primary job, not five.
• Scope explicitly states what the agent does NOT do.
• "Never invent information" is in the prompt, verbatim.
• The main outcome flow is written as numbered steps.
• Names, numbers, and dates get confirmed back to the caller.
• Escalation triggers and the transfer script are defined.
• Every response is capped at one or two sentences.
• The agent has been tested with three deliberately awkward callers.

If the platform choice underneath ever changes, the prompt discipline does not. I compared the engines in Retell AI vs VAPI vs Bland, and the structure here ports to any of them, because the hard part was never the vendor. It was writing a spec the model cannot wander away from, and then tuning it against reality.

A good voice AI receptionist is not a clever prompt written once. It is a tight, boring, well-structured spec that you sharpen from real calls until the failure modes disappear. Get the structure right, then let the transcripts finish the job.

Want the fastest loop from a bad call to a fixed prompt? Start free on VoiceDash or book a demo and I will show you how every agent's transcripts land in one branded portal you and your clients can actually read.