Beyond CUJs: Why "Example Prompts" Are the New North Star for AI-First Products
In the age of AI, defining core user journeys is evolving. Here's how to rethink product development by starting with what you want your AI to do.
For years, I've sworn by the importance of defining core user journeys (CUJs) in product development. But as I've been working on new AI-first products, I've noticed a fundamental shift: CUJs, while still relevant, are taking a backseat to what I'm calling "example prompts" or "example tasks."
Think about it. In traditional software development, we map out the steps a user takes to achieve a goal. When I worked on Outlook, a key CUJ was the ability to rearrange file folders. We meticulously documented the user's journey: dragging files between folders, creating subfolders, renaming them. We'd think through all the edge cases. It was all about understanding how the user would interact with the existing technology.
But AI changes the game. We're no longer just designing around the limitations of the technology. Instead, we're starting to explore a much wider range of what's possible, often things that users haven't even imagined yet.
From User Journeys to Example Prompts
With AI-first products, the initial focus shifts from "What steps does the user take?" to "What do we want the AI to be able to do?" And to answer that, we need to define the "example prompts" or "example tasks" that represent these core capabilities.
Let's say you're building a new AI-powered writing tool. Instead of starting with traditional CUJs like "user logs in" or "user creates a new document," you might start with example prompts like:
"Summarize this 10-page research paper into a single paragraph."
"Generate 5 creative social media captions for this product launch."
"Rewrite this email to be more persuasive and engaging."
These prompts become your north star. They define what success looks like for your AI.
Your Eval Set is Born
Here's where it gets really interesting: these example prompts don't just guide development; they also form the foundation of your initial evaluation dataset. By testing your AI's ability to handle these core tasks, you're directly measuring its ability to deliver on the core value proposition of your product.
This creates a powerful feedback loop. As you iterate on the AI system – perhaps by experimenting with different prompting techniques, model calls, or even entirely different architectures – you can continuously evaluate its performance against these core tasks. You're not just building; you're building to learn.
Prompt-First Product Development: A New Paradigm
This shift has led me to rethink my entire approach to product development. I now think in terms of "prompt-first product development." Instead of starting with detailed specs, I start with the core user tasks or prompts that I want the AI to handle. Then, I figure out what's possible. This involves a lot of experimentation. As I mentioned in my last post, using AI every day has become a critical part of this discovery process.
Take Project Mariner, for example. We knew we wanted to build a browser extension that could automate complex web tasks. But instead of immediately jumping into traditional CUJs like "user logs in" or "user accepts terms," we started by defining the core tasks we wanted Mariner to perform. These became our initial "example prompts": things like, "Find me three highly-rated hotels in San Francisco for under $200 a night" or "Add the items from this recipe to my shopping cart."
This "prompt-first" approach allowed us to quickly test the limits of what was possible and identify areas where the AI system needed improvement. We could then iterate on the underlying architecture, experiment with different prompting strategies, and even define new CUJs based on what we learned. It also forced us to think about unique design challenges. For instance, because Mariner directly interacts with the user's browser, we needed to build in "pause" and "stop" functionality – something that wouldn't be necessary for a more passive AI research tool.
The Takeaway
The rise of AI is forcing us to rethink the fundamentals of product development. While traditional CUJs still have their place, they're no longer the sole starting point. "Example prompts" are becoming the new cornerstone of AI-first products, serving as both a guide for development and the foundation for evaluation.
This "prompt-first" approach is about more than just defining tasks; it's about embracing a new mindset. It's about understanding that with AI, a lot more is possible than we might initially realize. But it's also about recognizing that achieving those possibilities requires a new way of thinking, a new way of building, and a new way of evaluating. It requires us to be explorers, constantly pushing the boundaries of what AI can do and then using that knowledge to build products that are truly transformative. And all that starts with asking: "What do we want our AI to do today?" Then using that answer as the building blocks to define the product around it.
One might even call this “prompt-first product development”.