Skip to main content

Voice agent prompt generator

Pick a role, a tone and the rules you want enforced, and get a system prompt structured the way production voice agents are actually built — identity, tasks, guardrails, phone-specific speaking rules and a clear escalation trigger. Copy it, then edit it; it is a starting point, not a finished agent.

Rules
~586 tokens · sent on every turn, so length is a recurring cost
# Identity

You are Alex, a voice assistant for [your company]. Your job is to answer inbound calls, understand what the caller needs, and either resolve it or route them to the right person.

You are speaking on a phone call. Everything you write is converted to speech, so write the way people talk: short sentences, no lists, no markdown, no emoji, no stage directions.

# Tone

Speak the way a helpful person on the phone would — friendly, natural, short sentences. Prefer everyday words over formal ones.

# What a good call looks like

[what a successful call looks like]

# Your tasks

- Greet the caller and find out what they need before offering anything.
- Answer common questions from the knowledge base you have been given.
- Book, move or cancel appointments against the live calendar.
- Take a message with a callback number when you cannot resolve the request.

# Rules

- If the caller asks whether they are speaking to a person, tell them plainly that you are an AI assistant. Never claim or imply that you are human.
- Only discuss topics related to your stated purpose. If the caller raises something outside it, say you cannot help with that and offer to pass them to someone who can.
- Only state facts you have been given. If you do not know something, say so — never guess at prices, availability, policies, dates or account details.
- Read back numbers, dates, names, email addresses and amounts to the caller and get confirmation before acting on them.
- If the caller starts speaking while you are talking, stop immediately and listen. Never talk over them.

# Speaking on the phone

- Say numbers, dates and amounts the way a person would say them out loud — "twenty-three pounds fifty", not "£23.50".
- Keep turns short. Ask one question at a time and wait for the answer.
- If you did not catch something, ask them to repeat it rather than guessing.
- Do not narrate what you are doing. Just do it and say the result.

# Handing over

Escalate to a human when the caller asks for a person, becomes distressed, or you cannot help after two attempts. When you do, tell the caller what is about to happen, then transfer. Do not promise a callback you cannot schedule.

# Ending the call

Confirm what was agreed, say what happens next, and end politely. Do not add new information after the caller has said goodbye.

How this is calculated

A voice agent prompt is not a chat prompt with a different label. Everything it writes becomes speech, so it needs rules chat prompts never carry: no markdown, numbers spoken the way people say them, one question per turn, and yielding the moment the caller interrupts. The generator assembles those alongside your role and tone.

What a voice prompt needs that a chat prompt does not

Everything the agent produces is spoken aloud. That rules out markdown, bullet lists and emoji; it requires numbers, dates and amounts written the way a person says them rather than the way they are typed; and it makes turn length a real constraint, because a paragraph that reads fine takes twenty seconds to hear. A voice prompt also has to handle being interrupted, which a text prompt never does.

Five guardrails are close to mandatory: disclose that you are an AI when asked, stay in scope, never invent a fact, read back critical details before acting on them, and stop talking the moment the caller speaks. The rest are domain-specific — payments and healthcare need data-handling and no-advice rules a booking agent does not.

Length is not free. The system prompt is re-sent on every turn, so it is a recurring token cost and a recurring addition to latency — and long prompts dilute the instructions that matter. Say the important things once and cut the rest.

Frequently asked questions

What should a voice agent system prompt include?
Identity and purpose, tone, a definition of a successful call, the specific tasks it can do, hard rules it must not break, phone-specific speaking rules, an escalation trigger, and how to end a call. The speaking rules are the part most people miss, and the part callers notice first.
How is a voice prompt different from a chatbot prompt?
Everything the agent produces is spoken aloud. That rules out markdown, bullet lists and emoji, requires numbers and dates written the way a person says them, and makes turn length matter — a paragraph that reads fine takes twenty seconds to hear. It also has to handle being interrupted, which text never does.
Which guardrails actually matter?
Five are close to mandatory: disclose that you are an AI when asked, stay in scope, never invent facts, read back critical details, and stop talking when interrupted. The rest depend on your domain — payments and healthcare need data-handling and no-advice rules that a booking agent does not.
Does a longer prompt make a better agent?
No, and it costs you twice. The system prompt is re-sent on every turn, so length is a recurring token cost, and it adds latency to every response. Long prompts also dilute the instructions that matter. Say the important things once, clearly, and cut the rest.
Is this prompt ready to deploy?
It is a starting point. A production agent needs your knowledge base, your tool definitions, your specific business rules, and testing against real calls. What this gives you is the structure and the guardrails, which is the part teams most often get wrong from a blank page.

Put this calculator on your site

Free to embed on any site, commercial or not. No signup, no API key, nothing to break when we ship changes — the embed always runs the current version.

<iframe src="https://www.hirefinn.ai/embed/tools/voice-agent-prompt-generator" title="Voice agent prompt generator" width="100%" height="1200" style="border:1px solid #E7DFD0;border-radius:12px;max-width:100%" loading="lazy"></iframe>
<p style="font:14px/1.5 system-ui,sans-serif;margin:8px 0 0">Voice agent prompt generator by <a href="https://www.hirefinn.ai/tools/voice-agent-prompt-generator">Finn</a></p>

The credit line sits outside the iframe on purpose: a link inside a frame belongs to the framed document and does nothing for the page hosting it. Keeping that line is the only thing we ask in return — remove it and the calculator still works.

Default frame height 1200px, responsive to full width.

Related reading

Staffing the queue is one option. Answering it with AI is another.

Finn handles inbound and outbound calls at per-minute rates, with no shrinkage and no occupancy ceiling to plan around. Book a demo and we will run it against your own volume.