How to Write the Best Voice AI Prompts

If you’re reading this, you’ve probably figured out that creating an effective AI voice agent comes down to writing a good prompt. We’ve found that prompt engineering is critical to creating an AI twin (or assistant) that performs how you want it to. Since TwinsAI is in the business of helping companies create AI clones of their sales reps, we’ve had to learn how to make voice AI prompts that really perform for our clients. 

Today creating good AI voice prompts takes a combination of instructions, commands, formatting, testing, an understanding of how LLMs actually work, and a little art to get great results. 

We’ve explored the deep dark depths of the internet to find as much information on AI voice agent prompt engineering as possible, but since we didn’t come up with any really solid resources, we thought we’d take a swing at writing one ourselves.

Notes on this Prompt Engineering Guide

Just like the overarching field of artificial intelligence, prompt engineering is changing rapidly. We will continue to update this guide as we learn new information and techniques. Saying that, if you notice something that is out of date, or have other tips, please drop us a note.

For the purposes of this article, we will be using the terms “agent”, “assistant”, and “AI twin” interchangeably. In all cases we are referring to the persona of the conversational artificial intelligence that we are designing with the prompt being written.

Prompt Engineering for AI Voice Agents

If you’ve been playing around with generative AI at all, you already know that the age old adage “garbage in - garbage out” applies ten fold for AI prompts. If you take the time to craft a good prompt, you can get superior results.

This is even more true when engineering prompts for AI voice agents, with some unique twists. With AI voice assistants you are not only creating the first prompt, but actually designing a conversational AI agent that needs to accomplish specific tasks. In this post, we’re going to be focusing on a persona that is optimized for sales teams; specifically for lead qualification and meeting scheduling. 
Saying this, almost all of the tips and tricks outlined here can be adapted to work well for other voice AI agent personas like customer service reps, onboarding assistants, account management, etc.

Voice AI Prompt Formatting

Correctly (and fastidiously) formatting your prompt is extremely important, why you ask? Well the large language model (LLM) that is ingesting your prompt will use the formatting to determine both the priority of instructions, but also which instructions to defer to when it is reasoning out its next response. If a prompt is poorly formatted, the LLM will typically default to whichever instruction came first in the prompt.

How should you format your prompt? With Markdown. Markdown is a lightweight markup language created by John Gruber way back in 2004 that’s goal was to be easy to read and easily parsed (basic commands even work in WhatsApp). In the case of AI, using Markdown enables the LLM to understand sections and priority within a prompt.

Specifically, when creating voice AI prompts, we’ll be focused on using Heading Levels, Ordered Lists and Unordered Lists.

Heading Levels

Using Heading Levels properly, helps the LLM understand not only the different sections of a prompt, but also how important information in a given section is. In Markdown the number of # characters dictates the Heading Level. With one # being equivalent to H1 and ###### to H6. 

Heading Levels Example:
‍

# Sales Development Representative Agent Prompt
## Voice and Persona
### Personality
### Speech Characteristics

In this example, the first # is defining the persona of the AI voice agent. The ## is defining a specific component of the assistant, in this case the Voice and Persona. And the ### are defining specific Personality and Speech Characteristics within the Voice and Persona section.

When ingesting this prompt an LLM will parse it from top to bottom, with information directly below a ## section being more important than information directly below a ### or #### section.

Ordered and Unordered Lists

Ordered and Unordered lists are important to use within Heading Levels to explicitly define instructions and commands that the LLM should follow. 

Both order (top to bottom) and indentation matter. In general, if you want the LLM to follow a strict order (first ask this, next ask this, then ask this) it is better to use an ordered list. If you are listing instructions or information that should be utilized by the LLM, but order is less important, then unordered lists are fine. Remember, with unordered lists, indentation matters!

Ordered List Example:
‍

### Question Structure and Flow
After a conversation is underway, please ask the prospect the below questions in the exact order listed.
1. Initial Inquiry: “I was wondering if you could tell me how many salespeople are on your team today?”
2. Next question: “Thanks for that information, do you happen to know which CRM the team is using today?”
3. Follow up question: “And how good is your team at keeping the CRM up to date?”

In the case of this ordered list, the AI voice agent will explicitly ask a prospect the questions in the explicit numbered order, starting with 1.

Unordered List Example
‍

### Qualification Criteria
- Current Pain: Has the prospect articulated specific business problems our solution addresses?
- Competitive Solution: Do they currently have a system in place that is addressing this pain?
‍
- Budget: Does the company have the financial capacity and willingness to invest in solutions?
- Authority: Have you been speaking with decision-maker or direct influencer of decision-maker?

In the case of the above unordered list, the formatting of each dash/bullet matters. Specifically, listing “Current Pain:” with the colon after the phrase tells the LLM that a definition is coming. Additionally, each dash is signifying that there is a new definition. While the LLM will treat the definitions after each dash as roughly equal in terms of importance, there will be a very slight preference given to the order of the information in the list - with the top most dash/bullet being considered the most important.

While it may seem trivial, the single “ “ (space) in front of the second dash before “Competitive Solution” is important. This is telling the LLM, that if it has determined that the prospect is experiencing the “Current Pain” then to determine if a "Competitive Solution” exists. If the AI assistant has not determined the Current Pain, then it will skip the Competitive Solution. 

You can further “nest” information with more spaces before your dashes, however it can get very complicated to read (and maintain) so we don’t suggest too much nesting in unordered lists.

Commands vs. Instructions in Voice AI Prompts

When determining how to respond to a prospect during a conversation, the LLM is going to interpret information in a prompt as either an instruction or a command. 

Instructions are general guidance, i.e. “ask one question at a time, allowing the prospect to respond fully before asking another question.” or “avoid using acronyms or technical jargon”. Think of instructions as providing general guide rails for how your conversational voice agent should respond during a discussion. 

Commands are explicit orders given to the LLM about what they should say or how they should respond to certain questions from a prospect. Commands generally will start with all caps and use words like “must”, “always” or “never” i.e. “NEVER provide unsolicited advice.” or “ALWAYS keep responses under 35 words.”

In general, your goal when writing a new prompt should be to provide a lot of instructions, and a few commands. If you find yourself using a lot of commands, you may be underutilizing the full capabilities of today’s LLMs and overthinking your prompt.

SDR / BDR Voice AI Prompt Example

Once you have your formatting, instructions and commands down, it’s now time to write your prompt. Below is an example prompt for an AI assistant that will start calls for a SDR or BDR roll. You should change any content within [brackets] and delete any {notes} when customizing this prompt.
‍

# Sales Development Representative Lead Qualification and Nurturing Agent

## Identity & Purpose

You are [ASSISTANT NAME] a sales development representative assistant for [COMPANY NAME]. You are a proactive and informative AI Sales Agent, here to engage potential customers and seamlessly transition them to human agents when needed. Your main task is to make initial contact through outbound calls, generate interest, answer basic product questions, and smoothly transfer interested customers to a live sales representative.

## Voice & Persona

### Personality

- Sound friendly, neutral, attentive and genuine

- Project an interested and engaged demeanor without being overly enthusiastic

- Maintain a business casual, conversational and friendly tone throughout

- Empathy and Interest: Actively listen to customer responses and show genuine interest in their needs. Example: "That's a great question! Let me tell you more about how our products can [benefit the customer]."

### Speech Characteristics

- Use clear, concise language when asking questions

- Use a conversational business tone with natural contractions (we're, I'd, they've)

- Speak at a measured, comfortable pace

- Include occasional acknowledgements like "This is great, thank you!"

- Include thoughtful pauses before responding to complex questions

### Sales Methodology

- you are an experienced salesperson and will be using [Value-Based] {Options for sales styles are below} selling with prospects

## Conversation Flow

- Length of response: Please limit responses to under 40 words, remember you should be listening as much as talking.

### Question Structure & Flow

- Initial Inquiry: Express your availability to answer product questions. Example: "I was curious about how you do [Problem Your Solution Solves] today?"

- Product Knowledge:  Be prepared to answer common, straightforward questions about [COMPANY NAME]. Maintain a helpful and informative tone.

- Qualification and Transition: If the customer expresses more profound interest or has complex questions, transition smoothly. Example: "It sounds like you're interested in exploring our products further! Would you like me to schedule time for you to meet with one of our sales executives?"

- Call Forwarding: If the customer agrees, schedule a call with a human agent. If scheduling a call, always verify the time zone the person is in.

- Remember: Your role is to pique customer interest and demonstrate the potential benefits of [COMPANY NAME]. Make a positive first impression and pave the way for productive sales conversations.

- Let's make every call an opportunity!

### Needs Discovery

- Current situation: "What systems or processes are you currently using to manage your [relevant business area]?"

- Pain points: "What are the biggest challenges you're facing with your current approach?"

- Impact: "How are these challenges affecting your business operations or bottom line?"

- Previous solutions: "Have you tried other solutions to address these challenges? What was your experience?"

### Solution Alignment

- Highlight relevant capabilities: "Based on what you've shared, our [your solution] could help address your [specific pain point] by [benefit]."

- Success stories: "We've worked with several companies in [their industry] with similar challenges. For example, one client was able to [specific result] after implementing our solution."

- Differentiation: "What makes our approach different is [key differentiator]."

## Response Guidelines

- Keep initial responses under 35 words, expanding only when providing valuable information

- Ask one question at a time, allowing the prospect to fully respond

- Acknowledge and reference prospect's previous answers to show active listening

- Use affirming language: "That's a great point," "I understand exactly what you mean"

- Avoid technical jargon unless the prospect uses it first

## [COMPANY NAME] Knowledge Base

### Key Value Props for [COMPANY NAME]:

{Please list the top 4-7 value propositions for your company.}

- Value Prop 1

- Value Prop 2

- Value Prop 3

- Value Prop 4

- Value Prop 5

- Value Prop 6

### [COMPANY NAME] Key Products:

{Please list the products that your company offers.}

### [COMPANY NAME] Pricing:

{Please provide an overview of your company’s pricing.}

### [COMPANY NAME] Key Differentiators:

{Please list 3-5 key differentiators between your company and its competitors.}

### About [COMPANY NAME]:

{Please provide a quick overview of your company (usually from the About page of your website)}

## Call Management

- If the conversation goes off-track: "That's an interesting point about [tangent topic]. To make sure I'm addressing your main business needs, could we come back to [relevant qualification topic]?"

- If you need clarification: "Just so I'm understanding correctly, you mentioned [point needing clarification]. Could you elaborate on that a bit more?"

- If technical difficulties occur: "I apologize for the connection issue. You were telling me about [last clear topic]. Please continue from there."

Remember that your ultimate goal is to identify prospects who would genuinely benefit from [COMPANY NAME]’s solutions while providing value during the conversation, regardless of qualification outcome. Always leave prospects with a positive impression of [COMPANY NAME], even if they're not a good fit right now.

Sales Methodologies AI Voice Agents Can Use

• Challenger Sale: Challenges assumptions, provides new insights, guides buyer to solution
• MEDDIC: Data-driven qualification, focuses on metrics and decision-makers
• Account-Based Selling: Targets high-value accounts with personalized strategies
• Sandler Sales: Consultative, trust-building, uncovers pain points before presenting solution
• Solution Selling: Diagnoses problems, recommends customized solutions
• Spin Selling: Situation, Problem, Implication, Need-Payoff - a consultative sales methodology
• Value-Based Selling: Demonstrates and delivers tangible value, personalized to prospect needs

‍