Prompting voice AI agents
Your AI agent works off of prompts, written instructions that tell it how to operate. PeakSend has several built-in prompts that handle common sales scenarios, but if you have specific logic you need your agent to follow, you’ll need to change the prompts.
AI agents are smart in some ways, but surprisingly dumb in others. They don’t have common sense in the way that a human does, so you’ll often need to spell things out in more detail. It can take time to get a prompt right, but once it’s working, your AI will execute the same way consistently and at scale. Because of this, taking the time to write good prompts is well worth it in the long run.
General best practices
Section titled “General best practices”Be thorough. Outline every scenario you can think of. The AI has no social awareness, so write your prompts as if you’re explaining the job to someone who has never had a phone conversation before.
Be specific. “Make small talk” means nothing to an AI. “Ask the prospect where they’re calling from” shows it exactly what you want it to do.
Think about what your best rep does in each situation and write that down as instructions. If your top closer handles pricing objections a certain way, put that exact approach in the prompt.
After you change a prompt, call the agent and test it. If it doesn’t behave correctly, rewrite the prompt a different way and use the Annotate and Eval tools to re-run that same call scenario. If you’ve tried 2-3 different approaches and it’s still not working (even after applying the tips below), reach out to support.
Prompting tips
Section titled “Prompting tips”Use quotation marks for scripted lines
Section titled “Use quotation marks for scripted lines”Put anything you want the AI to say word-for-word in double quotation marks. Instructions to the AI (things it shouldn’t say out loud) go outside the quotes.
"Hi, this is Sarah from Acme Corp. How are you doing today?"Enforce behavior with MUST NEVER and MUST ALWAYS
Section titled “Enforce behavior with MUST NEVER and MUST ALWAYS”Sometimes you tell the AI not to do something and it does it anyway. Or it jumps ahead in the script before finishing an earlier section. When normal instructions aren’t working, write MUST NEVER in all caps:
You MUST NEVER say x.
You MUST NEVER jump to stage B before completing stage A.This works the other direction too. If the AI isn’t saying something you need it to say, use MUST ALWAYS in all caps:
You MUST ALWAYS give the following disclaimer here:
"We do not offer...."All caps won’t fix everything, but it does make a difference for specific behaviors you really need to lock down.
Wait for the prospect to respond
Section titled “Wait for the prospect to respond”The AI doesn’t automatically know when to stop talking and listen. You need to tell it. Insert [wait for response] at every point where the agent should pause and let the prospect speak:
[wait for response]Handle silence with NO_RESPONSE_NEEDED
Section titled “Handle silence with NO_RESPONSE_NEEDED”Sometimes the prospect says something like “hold on, let me grab something” and the right move is to say nothing. By default, the AI will try to fill that silence anyway. To prevent this, tell it to output a silent placeholder:
When the user says X, always respond with exactly the following: "NO_RESPONSE_NEEDED"Branch on prospect responses
Section titled “Branch on prospect responses”It’s often useful to ask questions that your prospects could respond to in different ways. This can help your AI agent better sell to specific pain points, which can increase conversion rates. For every question the AI asks, map out the likely responses and tell it what to say for each one:
"Would you like to find out more about how we can help you achieve X?"
[wait for response]
- If no: "No problem, I hear that from time to time. That's why..."
- If yes: "Great. First..."Any response you don’t account for is a gap the AI will fill on its own.
Use XML tags for multi-step branches
Section titled “Use XML tags for multi-step branches”One-line - If responses work for simple branches. But sometimes a branch needs the AI to say several things in a row, wait for replies in between, then move to a different stage. For those cases, wrap the steps in XML tags. Each tag needs a unique name with an opening (<tag_name>) and closing (</tag_name>):
"Would you like to find out more about how we can help you achieve X?"
[wait for response]
-If no, say what's in the 'resolve_no' XML tag:
<resolve_no>
"No problem, I hear that from time to time. That's why..."
[wait for response]
"Of course,..."
[wait for response]
"Let me schedule an appointment..."
Continue to stage 4
</resolve_no>
-If yes: "Great. First..."The XML tags keep multi-step branches organized and stop the AI from skipping ahead.
What you control vs. what support handles
Section titled “What you control vs. what support handles”You own the script: what the AI says, when it says it, and how it responds to different scenarios (objections, FAQs, conditional paths). You also control settings in the script editor sidebar: agent name, voice, company, and calendar.
Some things live outside the script. Tool calls are anything the AI does rather than says: checking availability, booking an appointment, updating a CRM field, sending an email, transferring a call. Support configures those. Support also sets the max call length before the agent hangs up, and background sounds (call center ambience, coffee shop, etc.).
If you’ve tried multiple prompting approaches and can’t get the behavior you need, reach out. The support team can look at your script and suggest changes.