I’ve been designing language-centric experiences for more than two decades, from simple grammar-based systems accessed over the phone (IVR), to natural-language-understanding (NLU) systems, to the latest generation of Large Language Model (LLM)-based technologies. As the technology has evolved, I’ve adapted my design practice, never losing sight of how human beings interact via speech and language.

My design work has included significant features for Amazon Alexa (including weather, reminders/to-do lists, and calendaring); for Microsoft’s Xbox Kinect; for Ford SYNC (in-car navigation and traffic features); and for a variety of customer-service systems (e.g. phone-based award booking for American Airlines).

My design process always begins by understanding the customer needs the solution is trying to meet; I work closely with user researchers, product managers, and end users to understand the problem and the solution. I continue by sketching sample interactions as sample dialogs.

In the pre-LLM days, I would then create detailed flow diagrams indicating the desired conversational flow. With LLMs, this step is generally no longer useful or actionable; rather, I spend the time creating additional sample dialogs that reflect various desired and potentially non-desired paths through the experience. I also create detailed evaluation criteria for the conversation to help fine-tune the model’s behavior on the conversation.