Writing prompts for a voice agent that works on the phone

Most first voice agents fail the same way. Someone pastes in a long chatbot prompt, the agent answers with three sentences and a list, and the caller says "hello? are you there?" The words were fine. They were written for eyes.

On a phone the caller hears each word once, in order, and cannot look back. Everything in your prompt should serve that one fact.

Write for the ear, not the page

Tell the agent to speak in short sentences, usually one or two per turn. Callers cope best with one idea at a time. A long answer is not more helpful; it is a monologue the caller will interrupt.

Ban anything that only works visually: bullet lists, bold, emojis, URLs, and "as shown above". If the agent must give a web address or an email, tell it to offer to send it by message instead of reading it aloud.

The second version asks one thing. The details follow once the caller has chosen.

One question at a time

Callers answer the last thing they heard. If the agent asks for a name, a date and a reason in one breath, it will get one of the three and then have to work out which. Write the rule plainly in the prompt: ask one question, wait, then ask the next.

Order the questions by what the caller cares about, not by what your form needs. A dental clinic in Pune should ask "what day suits you?" before "may I have your date of birth?"

Confirm what was heard, especially names, numbers and dates

Speech recognition is good but not perfect, and Indian names, mixed Hindi and English, and a noisy street all make it worse. Tell the agent to repeat back anything that will be acted on: the appointment time, a phone number, the spelling of a name.

Teach it a pattern for numbers: read digits in small groups, with a pause, and ask "is that right?" For spelling, ask the caller to say the letters, or to use a word for each letter. Do not let the agent guess a surname and move on.

Write the greeting as a separate, spoken line

The first sentence sets the caller's expectation, so give it its own care. It should say who is calling or answering, for which business, and what the call is about, in one breath. Avoid a greeting that ends in a long menu of things the agent can do.

Write the greeting in the language the caller will hear. A greeting written in English for a Hindi call tends to stay in English, and the caller notices at once. If the agent speaks more than one language, author each version rather than asking the system to translate on the fly.

Say what the agent is. If your business plays a recorded notice that the caller is speaking to an AI, the prompt should never contradict it. See Should your AI agent say it is an AI? for the reasoning.

Keep it on task, and say what it does not know

A voice agent has a job. State it in the first lines of the prompt, then list what is in scope and what is not. A caller who asks a clinic's agent about a neighbour's prescription, or a coaching institute's agent about politics, should hear a polite redirect, not an opinion.

The most important sentence in many prompts is the one about not knowing. Tell the agent what to do when it lacks an answer: say so, offer to take a message or pass the call to a person, and never invent a fee, a date or a policy. An agent that makes up a fee is worse than one that says "I do not have that, let me get someone who does."

Tools: say when to call them, and when not to

If the agent can book an appointment, update a CRM or send a message, describe each tool in plain terms: what it does, what it needs, and the moment to use it. "Book only after the caller has confirmed the day and time" is better than leaving the model to decide.

Keep tools few. Each call to an outside system is a pause the caller sits through, so the prompt should say what to tell the caller while it happens ("one moment while I check the calendar") and what to say if the tool fails. For why those pauses matter, see Why the pause matters.

Guardrails belong in the prompt

Write down what the agent must not do. Three are worth stating every time:

• No pressure. If the caller says no, the agent accepts it once and ends politely. It does not argue or ask again.
• No false urgency. The agent says "only two seats left" or "offer ends Friday" only if that is true today. Put real offers and real end dates in the prompt, and nothing else.
• No denying it is an AI. If a caller sincerely asks, the agent answers honestly.

Invented scarcity is also the quickest way to lose trust, and the rules on automated calls in India are worth reading before you write any sales script.

Test with real calls, then edit one thing

A prompt is not finished when it reads well. Call it yourself, from a real phone, in the same noise your customers have. Then ask a colleague who has never seen the prompt to try to break it: interrupt, mumble, ask something off topic, say nothing.

Listen to the recording and read the transcript side by side. Change one thing, call again, and compare. Changing five things at once tells you nothing about which one helped. You can try an agent before you build one on the live demo.

A starting checklist

• One job, stated in the first three lines.
• Short sentences, one question per turn, no lists or links read aloud.
• A spoken greeting in the caller's language.
• Repeat-back for names, numbers, dates and spellings.
• A clear "I do not know" path that ends with a person or a callback.
• Each tool described with when to use it and what to say if it fails.
• Guardrails: no pressure, no false urgency, offers only if true.
• At least ten real test calls before launch, including bad ones.