I've reviewed prompts for hundreds of AI phone agents. At least half of them have the same problem: they tell the agent what to do, but not how to be.
The result is an agent that's technically correct. It answers questions accurately. It follows instructions. But it sounds like it's reading from a manual, and callers pick up on that instantly.
Here's how to fix that. These are the prompt techniques that consistently produce agents that sound like real people having real conversations.
Give Your Agent an Identity, Not a Job Title
Compare these two prompts:
Version A: "You are a customer support agent for a SaaS company. Answer questions about pricing, features, and account issues. Escalate billing disputes to a human."
Version B: "You are Jess. You work on the customer team at a small SaaS company. You genuinely enjoy helping people figure things out. When someone calls, your first instinct is to listen, not to solve. Once you understand what they need, you help them get there. If you don't know something, you say 'Let me check on that' instead of making things up."
Version A produces a help desk chatbot. Version B produces a person. The difference isn't the information. It's the frame.
Show, Don't Tell
Most prompts read like lists: "Be friendly. Be helpful. Don't interrupt."
That's telling the agent how to behave. It's better to show it. Include a short example conversation in your prompt:
"Here's an example of how you should handle a frustrated customer:
Caller: I've been on hold for 20 minutes and nobody's helped me yet. You: I'm sorry about that wait. That's frustrating and it shouldn't happen. Let me help you right now. What's going on? Caller: My payment went through but my account still shows as inactive. You: Let me look into that. Can you give me the email on your account so I can pull it up?"
The agent now has a concrete example of what good looks like. It's not guessing what 'be empathetic' means. It has a pattern to follow.
Define What the Agent Cannot Do
This is just as important as defining what it can do. AI agents tend to be overly helpful. They'll try to answer questions they don't have information for. They'll make promises they can't keep.
Explicitly list what's out of bounds:
"You cannot: - Issue refunds or process payments - Make changes to someone's account - Promise a specific timeline for feature requests - Say 'I'll make sure this gets fixed today' unless you can actually do that
If someone asks about any of these, say 'I want to make sure this gets handled right. Let me connect you with someone who can help directly.'"
Handle the Hard Moments
Real conversations have awkward moments. Most prompts don't prepare the agent for them. Add explicit instructions for:
When the caller is angry: "If someone is upset, don't get defensive. Acknowledge their frustration first. 'I understand why you're frustrated. That sounds really annoying. Let me see what I can do.' Then move toward a solution."
When you don't understand: "If someone says something you genuinely didn't catch, ask them to repeat it. Don't guess. 'Sorry, I missed that last part. Could you say it one more time?'"
When you can't help: "If someone asks for something you absolutely can't do, don't string them along. Be direct but kind. 'I wish I could help with that, but it's outside what I'm able to do. Let me get someone who can.'"
When there's silence: "If the line goes quiet for more than a few seconds, gently check in. 'Still there?' or 'Take your time.' Don't rush to fill every silence."
Keep Responses Short
This is the most common prompt mistake: letting the agent ramble.
On a phone call, people don't want paragraphs. They want answers. Add this rule to every prompt:
"Keep every response under 30 words. If you need to say more than that, break it into two shorter responses and wait for the person to respond in between."
This single rule transforms how natural the agent sounds. Long responses feel like a lecture. Short ones feel like conversation.
Include a Strong Greeting
The initial greeting sets the tone for the entire call. A bad greeting gets people on guard before the conversation even starts.
Some good greetings:
"Hey, this is Jess from Acme. How can I help you today?" "Hi there, thanks for calling Acme. What can I do for you?" "Hello, you've reached Acme. Jess here. What's going on?"
Notice what these have in common: they're short, they identify the person and company, and they end with a question that hands the conversation to the caller.
Test and Iterate
No prompt survives first contact with real callers. Write it, test it, and be ready to change it.
The best way to improve a prompt: read the transcripts of calls that went badly. Find the exact moment where the conversation went off the rails. Figure out what the agent should have done instead. Add that to the prompt.
Do this for 10 calls and your agent will be dramatically better than it was at the start.
The Bottom Line
A great AI phone agent prompt is less about instructions and more about identity. Don't tell your agent to be helpful. Tell it who it is and let being helpful follow naturally.
Start writing better prompts at agentline.cloud.