How to write an AI voice agent prompt that works on a real call
Separate facts, conversation rules, and permitted actions. Keep replies speakable, make uncertainty explicit, and test corrections instead of only greetings.
Vistrow Voice team
Product, speech, and customer experience
Write for a listener, not a reader
A paragraph that works on a web page can be exhausting over the phone. Ask for short replies, one question at a time, and an answer to the caller’s question before another qualification step. Avoid spoken bullet lists, unexplained acronyms, and a repeated company introduction.
Do not confuse a warm voice with a verbose agent. A brief acknowledgement and a relevant question usually sound more attentive than a polished sales monologue. Voice style should support the task: calm for a complaint, clear for a date, and conversational for an exploratory enquiry.
The prompt works best when it has reliable facts to draw from; the knowledge base overview describes that source of approved answers. If you are designing your first flow, begin with the practical guide to AI voice agents before writing provider-specific speaking instructions.
Separate the prompt into four jobs
Give the agent a role and a clear boundary. Supply approved facts separately from instructions about how to talk. Define the information to collect and the actions it is actually able to perform. The prompt cannot create a booking integration that does not exist.
For example: “Answer using the approved project facts. If availability is not present, say it needs confirmation. Ask one requirement question at a time. A visit is only confirmed after the booking action succeeds; otherwise record it as a request.” This is an illustrative instruction, not a complete production prompt.
- Role and scope: who the agent represents and what it handles.
- Facts: current approved information, with an uncertainty rule.
- Conversation: concise replies, corrections, and language choice.
- Actions: tool prerequisites, success checks, and failure wording.
Use expressive cues carefully
Emotion controls and nonverbal tags depend on the speech provider and model. Do not assume a tag supported by one voice works with every voice. If a provider does not support the format, the caller may hear the tag read aloud or the synthesis request may fail.
Keep provider-specific formatting outside the business facts. Use a tested adapter to translate supported delivery cues, preserve the selected voice, and strip unsupported markup. Test the actual call path as well as the preview button; a preview can use a different provider or settings from the live agent.
Test the moments that break a script
Try a correction, a question outside the knowledge base, a caller switching language, silence, and a direct request to stop. Include a failed tool action. Check whether the spoken answer matches what was saved.
Keep a small regression script and rerun it after changes. If an issue is caused by missing data or a broken tool, fix that system rather than adding another paragraph to the prompt. A longer instruction set can conceal the real fault without making the conversation dependable.