How to Write Prompts for an AI Phone Receptionist

A good AI phone receptionist prompt has six parts: role, speaking style, fields to collect, what it may answer, what it must never do, and how to end the call. Because callers only hear the agent, voice prompts need much stronger rules about brevity and pacing than chat prompts. And every prohibition should come with an instruction for what to do instead.

Why split the prompt into six parts

One long block of instructions is hard to review and easy to break: change one paragraph and it quietly contradicts another. Separate parts make the prompt auditable by people who aren't prompt engineers (the clinic manager, your support lead) and keep each change in one place.

Part	What it covers	Example (dental clinic)
Role	Who the agent is and what calls it handles	"Front desk at Maple Dental, handling appointment calls"
Speaking style	Tone, sentence length, pacing	"Calm and polite. One or two sentences per turn"
Fields	What must be captured by the end of the call	Name, callback number, reason, preferred time
Allowed answers	Questions it may answer, with the answers	Opening hours, closed days, parking
Never do	Things it must not say or do	Diagnose symptoms, quote treatment costs
Endings	How to close in each situation	"Someone from the clinic will call you back"

Writing the role

Two or three sentences: who the agent is, which calls it handles, and which it doesn't.

Good: "You are the phone receptionist at Maple Dental. You handle calls to book, change or cancel appointments and collect the details we need. You don't give clinical advice; for anything about symptoms or treatment, arrange a callback from a dentist."
Avoid: "You are a helpful, knowledgeable AI assistant. Answer any question the caller has."

"Answer any question" invites the model to answer questions it has no business answering. For a front desk, a narrow scope is the safe default.

Speaking style for the ear

A caller can't scroll back. Put these rules in the speaking-style section:

One or two sentences per turn. Long explanations don't survive the trip through someone's short-term memory.
One question at a time. "Can I get your name, number and preferred time?" reliably loses one of the three.
No lists, symbols or URLs. TTS engines read bullets and asterisks badly, and nobody can write down a URL they heard once.
Group digits when reading numbers. "Five five five, two one three, four four seven one" is far easier to follow than a run of ten digits.
Short acknowledgments. "Got it." "Thanks." Repeating back everything the caller said makes the call drag.
Plain, polite language. Overly formal phrasing makes sentences longer and harder to follow by ear.

A short first sentence also makes the agent sound faster, since speech synthesis can start sooner (see Voice AI latency).

Writing the "never do" rules

State prohibitions as absolute rules, each paired with an alternative action. Soft wording like "try to avoid" gets read as "sometimes it's fine."

Never	Instead
Suggest a diagnosis or treatment from symptoms	Say a dentist will explain, and arrange a callback
Estimate treatment costs	Answer only what's in the FAQ; otherwise offer a callback
Say a booking is confirmed before it is	Confirm with the booking system first; if that fails, offer a callback
Guess at facts it doesn't have (hours, staff schedules)	Say "Let me check and get back to you"
Share anything about other patients	Decline, however the question is framed
Argue with the caller	Acknowledge, apologize and hand off to a person

"Never guess at facts you don't have" belongs in every receptionist prompt. LLMs are very good at producing plausible answers, and without this rule you'll eventually hear the agent invent your opening hours. Choosing what the agent may answer is covered in Designing FAQ answers for a phone AI.

Collecting and confirming details

List only the fields you must have at the end of every call. Each extra field lengthens the call and raises the chance the caller hangs up halfway.

Once everything is collected, instruct the agent to read it all back and ask for corrections. Phone numbers and dates matter most: get them wrong and nobody can call back. Field design is covered in Structured data extraction from phone calls.

Endings and handoff

Write at least two endings:

Details collected: "Thanks, that's everything. Someone from the clinic will call you back."
Resolved by an FAQ answer: "Is there anything else? If not, thanks for calling."
Handing off: the conditions (caller asks for a person, urgent symptoms, a complaint) and what to say.

Handoff conditions get their own post: Designing AI-to-human handoff.

Doing this with voicast

In voicast, the six parts map directly onto fields of the conversation setup: role and style go in persona, the opening line in greeting, fields in fields, allowed answers in faq, handoff conditions in handoff_when, and endings in closing (and closing_answered for calls resolved by an FAQ answer). Here's a dental clinic setup as an example:

bash

curl -s https://api.voicast.jp/v1/agents \
  -H "Authorization: Bearer $VOICAST_API_KEY" \
  -H 'content-type: application/json' \
  -d '{
  "name": "Maple Dental front desk",
  "config": {
    "greeting": "Thank you for calling Maple Dental. How can I help you today?",
    "persona": "You are the phone receptionist at Maple Dental. Speak calmly and politely, one or two sentences at a time, and ask one question per turn. Never suggest a diagnosis or estimate treatment costs; arrange a callback from a dentist instead. Never guess at facts you do not have; say you will check and call back.",
    "fields": [
      { "key": "name", "label": "Name", "required": true },
      { "key": "phone", "label": "Callback number", "required": true, "type": "phone" },
      { "key": "purpose", "label": "Reason for calling", "required": true },
      { "key": "preferred_time", "label": "Preferred date and time" }
    ],
    "handoff_when": ["The caller asks to speak to a person", "The caller describes severe pain or swelling", "The caller is making a complaint"],
    "closing": "Thanks, that is everything. Someone from the clinic will call you back.",
    "closing_answered": "Thanks for calling Maple Dental. Goodbye.",
    "faq": [{ "q": "What are your opening hours?", "a": "Weekdays 9 to 6, and Saturdays until 1. We are closed on Sundays and public holidays." }],
    "domain": "Dental clinic",
    "vocabulary": ["implant", "cleaning", "checkup"]
  }
}'

Note that the phone field type checks for a Japanese number (10 or 11 digits starting with 0); use a text field for other formats. Setups are versioned, and a call in progress keeps the version it started with.

FAQ

Does a longer prompt make the agent more accurate?

No. Longer prompts slow down every reply and make it easier for instructions to contradict each other. Remove anything that doesn't apply to this use case and keep the six parts short; short rules get followed more consistently.

What if the agent keeps breaking a rule?

Check that the rule is absolute and that it says what to do instead. If it still slips in certain situations, make those situations a handoff condition, or limit answers to what's written in the FAQ.

How formal should the agent sound?

Polite and plain. Very formal phrasing produces long sentences that are hard to follow on the phone. Pick the handful of phrases your best human receptionist actually uses and put them in the style section.