How to write a prompt for an AI voice agent: structure, template and testing

Four things change when a language model speaks instead of writes:

• Everything is spoken. Formatting such as bullets, bold text and links means nothing to a voice, and some voices read symbols out literally.
• Length is latency. A long reply takes longer to generate and to say, and callers interrupt long monologues.
• Turn-taking matters. The caller can talk over the agent at any time, so the prompt has to keep each turn short and easy to recover from.
• There is no scroll-back. A caller cannot re-read what the agent said, so one idea per sentence and one question per turn.

Vapi's prompting guide puts it well: treat the prompt as something re-run on every turn, so it should be structured, unambiguous and short.

1. Identity and personality: the agent's name, role, company and tone. Include a line that keeps it in character if a caller tries to make it something else.
2. Response guidelines: how to speak. One or two sentences per turn, one question at a time, spoken numbers, no lists.
3. Guardrails: what it must never do, written as plain conditions and actions. These override everything else.
4. Context: facts it needs on this call: today's date, who it is calling and why, business hours, prices it may quote.
5. Workflow: numbered steps for each scenario, with the branches spelled out, including when to call each tool.
6. Examples: two or three short sample exchanges: a clean booking, an objection, and a tool failure.

Replace the bracketed parts. Keep the section headings; models follow labelled structure more reliably than a block of prose.

Most robotic-sounding agents are caused by text written for a screen. Spell out what the caller should hear.

• Write prices, dates and times in words, or tell the agent to say them that way.
• Add a pronunciation note for brand names and acronyms the voice may get wrong. Many voice providers also have a pronunciation dictionary.
• Use full stops and commas for pacing. Vapi's guide warns that dashes and markup can behave inconsistently across voices.
• Replace lists with “first… then…” or with two options in one sentence.
• Never have the agent read a URL. It should offer to text the link.

Tools are how the agent checks a calendar, books a meeting or looks up a contact's history. The prompt and the tool descriptions together decide whether it uses them well.

• Say when to call each tool in the workflow: “Check the calendar before offering any time.”
• Say what to do with each result: if no times are free that day, offer the next day; if the booking fails, apologise once and try another slot.
• Refer to tools by what they do, not by an ID, so the agent never reads an ID aloud.
• For slow tools, give the agent something to say while it waits, such as “Let me check the calendar.”
• Give the agent the current date and time, or a tool that returns it, so “tomorrow at 4” means the right day.

Write each rule as a plain condition and action. Vapi's guide notes that the prompt is not a security boundary, so anything that must never happen, such as booking outside business hours, should also be enforced by the tool itself.

1. Write ten to twenty test scenarios: a clean booking, a not-now, each objection, a wrong number, a caller who asks if it's a robot, a request to stop, and a tool failure.
2. Call the agent yourself on a real phone. Listen to pacing and interruptions, not just the transcript.
3. Change one thing at a time, and rerun the whole scenario set after each change; fixes for one scenario often break another.
4. Measure the share of scenarios the agent completes correctly without a person stepping in.
5. After launch, read the first twenty real calls before sending more leads.

• A wall of prose with no sections. The model loses track of rules buried in paragraphs.
• Long, polite openers. Say who you are and why in one sentence.
• No instruction for “stop calling me”. Every outbound prompt needs one.
• Letting the agent invent availability instead of checking the calendar.
• Hiding that it's an AI. Say it in the first sentence.

In SuperCampaign, the prompt and the first message go into the agent studio, alongside the model, voice and transcriber from your Vapi account. Attach the calendar tools for Calendly, Google Calendar or iClosed, plus current time and contact history.

For per-campaign details, use the prompt field on the AI call node. It is added to the agent's instructions for that campaign, and merge fields such as {{contact.firstName}} are filled in there. They are not filled in in the agent's saved first message, so keep names out of it. Do-not-contact requests and callback times are captured from every call automatically.

Scripts to adapt into your prompt