Rethinking LLM Interfaces: From Chatbots to Contextual Applications
Alternatively titled: Why Chat Isn’t the Endgame for LLM Apps (and What Comes Next)
Hello fellow datanistas!
Have you ever wondered if chat interfaces are really the best way to interact with large language models, or if there’s a better, more intuitive path forward? I’ve been thinking a lot about this lately, and I wanted to share some insights that have fundamentally changed how I approach building LLM-powered applications.
In this blog post, I explore why chat interfaces—while a great starting point—aren’t the ultimate destination for LLM applications. Drawing inspiration from a conversation with my colleague Michelle Faits, Alan Pike’s thought-provoking video on AI UX design, and Clayton Christensen’s jobs-to-be-done theory, I argue that the future lies in embedding LLMs into structured, context-aware workflows.
I share my experience building DeckBot, a Markdown slide deck generator, where I moved away from chat as the primary interface and instead focused on integrating LLM capabilities directly into the data models. This approach made the tool more reliable and delightful to use, and it got me thinking about how we can inject AI into business processes in a way that truly augments human workflows—without forcing users to rely on open-ended chat.
If you’re curious about how to design LLM applications that are more like TurboTax and less like ChatGPT, and want to learn some practical principles for building these contextual tools, I invite you to check out the full post.
The key takeaway is that the most effective LLM applications will be those that embed AI directly into well-defined workflows, providing assistance exactly when and where it’s needed. It’s not about replacing humans, but about making our processes smoother and more delightful.
*If this resonates with you, I’d love for you to read the full post (Rethinking LLM Interfaces: From Chatbots to Contextual Applications), and please consider forwarding it to anyone who might benefit from these ideas. Your thoughts and feedback are always welcome!*
Cheers
Eric
greetings,
"Instead of starting with a UI, I began with the data model: defining a Slide as a Pydantic model with title, content, and type. I tested individual slide generation in a Marimo notebook until each component worked reliably. Then I put them together into a SlideDeck Pydantic model. This allowed me to" <----- what was cutoff here? is it germane to the next paragraph? i enjoy your articles & learn much, so this bit of confusion is important to me.