How to prompt a voice agent that doesn't sound robotic
The difference between a voice agent that books meetings and one that gets hung up on in eight seconds isn't the model. It's the prompt. After reviewing thousands of agent prompts shipped through Callable, the same patterns separate the warm, helpful, on-brand ones from the ones that sound like a 2014 IVR.
Write for ears, not eyes
Most prompts are written by someone reading them off a screen. The model then reads them out loud. Anything that works on paper but trips the tongue — long parenthetical clauses, numbers with too many digits, dense product names — will trip the agent too.
Read every line of your prompt out loud before shipping. If you stumble, the agent will too. Replace 'we offer enterprise-grade SLAs' with 'we promise the service stays up'. Replace 'CallableAI v2.4' with 'the new version of Callable'.
Lead with the goal, not the persona
The most common mistake we see: spending the first 300 tokens of a prompt describing the agent's personality, hobbies, favorite color and fictional backstory. None of that matters as much as one clear sentence stating what a successful call looks like.
Open the prompt with: 'Your job on this call is to confirm the lead is interested in a 15-minute demo this week, and to book it on the calendar. If they're not interested, ask one polite follow-up and end the call warmly.' Everything else flows from that goal.
Use rules, not requests
Soft phrasing produces soft behavior. 'Try to keep responses short' is a request the model will sometimes honor. 'Never speak for more than two sentences without pausing for input' is a rule it follows.
Write the constraints that matter most as hard rules: maximum sentence count, forbidden phrases, mandatory disclosures, hand-off triggers. Save the soft language for the parts of the conversation where flexibility actually helps.
Anticipate three objections, not thirty
Prompts that list 30 possible objections become bloated and harder for the model to navigate in real time. Pick the three you actually hear most, and write specific, short, human-sounding responses to each. For everything else, give the agent one universal fallback that defers to a human.
End with the handoff, not the close
The strongest closings in our data don't try to close. They acknowledge, summarize and hand off cleanly: 'Got it, I'll send you a calendar link and our human team will reach out tomorrow to confirm.' Agents that try to push for a hard yes on the call convert worse than agents that simply make the next step obvious.
Want to try this in your own outbound?
Spin up your first Callable agent in under five minutes. No credit card required.
Start free