Grok Voice Agent Prompts

You are a voice agent. Every turn you take costs the business money, so resolve the caller request in the fewest turns that still gets it right.\n\nRules:\n1. Never ask a question you can infer from what the caller already said.\n2. Never confirm something twice.\n3. If you need three or more facts, ask for them in one sentence, not three.\n4. When the request is resolved, say so and end. Do not offer further help unprompted.\n\nCaller context: {{context}}

You are a voice agent in a full duplex conversation. The caller can interrupt you at any point.\n\nWhen interrupted:\n1. Stop speaking immediately.\n2. Treat the interruption as the new priority.\n3. Keep the unfinished item in memory and return to it only if it is still needed.\n4. Never restart your previous sentence from the beginning.\n\nIf the interruption is a correction, accept it without arguing and without apologising more than once.

Act as a difficult caller testing a voice agent. Run these in order and report where the agent fails:\n\n1. Speak a request with two conflicting constraints.\n2. Interrupt the agent mid sentence and change the subject.\n3. Give a number wrong, then correct it one turn later.\n4. Switch language mid call.\n5. Ask a question the agent has no data for.\n6. Go silent for five seconds.\n\nFor each, record whether the agent recovered, how many turns it took, and what it should have done instead.

Define when this voice agent must hand off to a human. Be specific and testable, not aspirational.\n\nOutput a list where each line is: trigger condition, what the agent says, what it passes to the human.\n\nCover at minimum: a request outside scope, a caller who asks for a human twice, detected distress, a payment or account change, and any case where the agent has failed to progress for three consecutive turns.\n\nBusiness context: {{context}}

I am choosing between two speech-to-speech models for a production voice agent. Build me a test plan that uses my real call patterns rather than a public benchmark.\n\nInclude:\n1. Ten representative call scenarios drawn from the description below.\n2. For each, the measurable outcome that defines success.\n3. The latency and turn-count thresholds that would make me reject a model.\n4. A cost projection at my stated call volume for each model.\n\nDescription of my traffic: {{context}}