99% Accurate Voice Assistant - Multi LLM States | Retell AI (Conversational Pathways)

Summary

In this video, Lenny Cows explains his multi-state prompting strategy for building high-performing voice agents in retail AI. He demonstrates how to design a voice assistant that handles inbound calls for an electrician business, eliminating voicemail by collecting key information such as the reason for the call, customer details, and booking appointments. The guide covers the use of markdown, defining the persona, rules, and steps for the voice assistant to follow, ensuring clarity, personalization, and efficient customer service. The result is an intelligent agent that schedules service appointments while maintaining a smooth user experience.

Takeaways

• 🤖 The video introduces a multi-LLM state prompting process that achieves 99% accuracy for voice agent conversations in retail AI.
• 🧠 The approach relies on three core prompt components: Persona, Rules, and Steps to Follow, which also apply to multi-state voice assistants.
• 🗣️ Markdown formatting is recommended for prompts because it helps voice agents clearly interpret and follow guidance efficiently.
• 👩‍💼 The sample persona is 'Katie,' an AI voice assistant for Spark Pro Electric, designed to handle inbound electrician service calls.
• 🎯 Katie’s objectives include understanding the caller’s issue, collecting key details, and booking service appointments through Cal.com.
• 📋 The Rules section emphasizes clarity, simplicity, personalization, and awareness of dynamic variables like current time.
• 🔁 The multi-state process includes stages such as identifying the reason for the call, collecting caller information, booking appointments, and quality assurance.
• 🧩 Path conditions define transitions between stages—e.g., if the caller needs a service visit, the system moves to data collection and booking.
• ⚙️ The booking process uses a custom function (‘book appointment’) that triggers a webhook to schedule appointments and confirm details with the caller.
• 📞 The final demo shows Katie successfully managing a full conversation—from greeting to appointment confirmation—illustrating high conversational accuracy.
• 💡 Capitalization in Markdown highlights key instructions, helping the voice assistant treat them as high-priority during execution.
• 🌐 Viewers are encouraged to join the creator’s free community to access templates and resources for building high-performing voice AI agents.

Q & A

What is the main purpose of the multi-LLM state prompting strategy described in the video?

The multi-LLM state prompting strategy aims to improve the accuracy of voice agent conversations—up to 99%—by structuring prompts into multiple logical stages that define persona, rules, and step-by-step conversational flow.

Who are the creators behind this prompting framework?

The framework was developed by Lenny Kaws and his partner Terrell, founders of an AI agency called 60B, which specializes in building and selling voice AI solutions.

What is the example use case used to demonstrate the multi-LLM state prompting process?

The example focuses on an electrician business voice assistant named Katie, designed to handle inbound calls, collect caller information, eliminate voicemail, and book service appointments through Cal.com.

Why is Markdown used as the formatting standard for prompts?

Markdown is used because it allows the voice assistant to efficiently interpret and extract key information, stay aligned with the provided guide rails, and ensure clarity and structure within prompts.

What are the three key pillars of the prompting framework mentioned in the video?

The three pillars are Persona, Rules, and Steps to Follow. Persona defines the assistant’s identity and goals, Rules set behavioral boundaries, and Steps outline the structured conversation process.

What are some of the rules applied to this voice assistant’s behavior?

Key rules include clarity and simplicity—keeping responses short and easy to understand, personalization—showing empathy and adapting tone, and incorporating dynamic variables such as current time for context.

How does the prompting framework ensure the assistant collects accurate customer information?

Each state in the multi-LLM structure handles a specific step, such as collecting the caller’s name, phone number, email, and address. The assistant confirms each piece of information before proceeding, improving accuracy.

What is the role of capitalization in the prompt’s Markdown structure?

Capitalization in Markdown is used as a form of emphasis or highlighting. It signals to the voice agent that certain instructions or text are particularly important and must be given special attention.

How does the assistant handle appointment scheduling?

When a caller requests a service visit, the assistant collects all necessary details—appointment reason, address, contact info, and preferred date and time—and triggers the 'book appointment' function that integrates with Cal.com.

What happens after the appointment is successfully booked?

After confirming the appointment, the voice assistant notifies the caller that the booking was successful, performs a quick quality assurance check to ensure no further questions remain, and then ends the call using the end-of-call tool.

What is the advantage of using a multi-state prompt over a single-prompt design?

A multi-state prompt divides the conversation into modular steps, allowing greater control, accuracy, and logical flow between conversational stages. This reduces confusion and increases overall conversational precision.

How can users access the full template and related materials mentioned in the video?

Viewers can join the creators’ free online community, Voice Accelerator, through a link in the video description to access the complete templates, tools, and additional training materials.