One of the curious things about starting a company, particularly in the early days, is it requires lots of storytelling. Because honestly, that’s all you have - an idea. It’s a story that you’re not even sure you believe at first. But as you tell it more and more, you refine it and hopefully it becomes more convincing to both you and other people.
So let’s say (hypothetically of course), through all this, you secure some pre-seed funding. Now, people are expecting things from you and you’ve gotta build that story into something real.
So, how do you know what to build? Presumably, you’ve developed a solid hypothesis by now. I’m going to build x thing for y customer which will change their world for z reason. Once you have that, the goal is to design a series of experiments to quickly falsify your hypothesis so you can move onto alternative hypotheses that are closer to the “ground truth” (aka what customers want).
Discovery as hypothesis-testing
When you start building, you may think people will want my product for xyz reasons. But you’ve probably only got a few data points at this stage - your experience and maybe some friends’. How do you know which specific features to focus on? What will bring people back to use it multiple times a day? You need lots of data to answer those questions. I think of customer discovery as hypothesis-testing. And if there’s anything you learn in grad school, it’s how to beat your head against a hypothesis.
Like most early grad students, I was maybe a little too ambitious. I focused on some grand behavioral neuroscience theories and did countless experiments using in-vivo optogenetics. I was pretty confident it would yield that sacred CellNatureScience paper. But nothing worked. What’s worse, it often took months to get enough data to understand results.
After almost three soul-crushing years, and at the encouragement of some amazing mentors, I started focusing on patch-clamp electrophysiology - recording intracellular electrical activity from individual neurons. It’s a notoriously non-scalable approach. But it allowed me to do multiple experiments a day - a big difference from behavioral experiments. More importantly, I could see the results in real-time and design new experiments on the fly. Long story short, patch clamp basically unlocked grad school for me and I finished less than two years later.
In customer discovery, your goal is similar - rapid hypothesis testing. Every customer conversation yields a bunch of data. Rather than waiting for a large sample size, you iterate quickly. In the early phase, this is not a tightly controlled experiment - you might be changing multiple things at once. Update your hypothesis/design for the next customer call. Get more data, incorporate into design. In an endless cycle of talky talky.
So how to do rapid iterations in customer discovery? I don’t have a magic bullet, but I’ve been playing around with some tools over the last couple days.
Tools for customer discovery
I know it’s very meta (maybe cliche) to use AI to build AI/ML tools. But hear me out. I like using “wireframes” for design. They’re basically a super lo-fi way to demonstrate functionality. At this stage, I don’t want to distract customers with my UI. I don’t even have a logo. And I don’t need it to look fancy, I just need it to convey an idea.
Wireframes are probably pretty easy to create if you know what you’re doing. The problem is I don’t really know what I’m doing. I’ve got lots of high-level ideas and amazing design friends that I lean on for strategy, but I have to execute it for now (with the help of a co-founder). So how do I speed up my iteration cycles?
I found a tool called WireGen AI. It’s a Figma plugin that makes it really easy to create wireframes. This isn’t an advertisement - it’s definitely helpful but there are also flaws. Basically I can use natural language to do a first pass on WireGen, like “I want to build a five page mockup. Page 1 is a typical login page. Page 2 is a website with xyz features placed here, there, everywhere, etc”.
For prompt engineering, I use ChatGPT to take a first stab as I bumble through terminology. It also helps me improve my prompts (e.g., to fit within a character limit, tell me the names of features, etc). I actually use ChatGPT for most of my prompts across various GenAI tools. I’ll ask it to mimic certain design features from various websites where I’ve drawn inspiration. I’ll do multiple iterations with each page in WireGen until it gets close enough to be usable.
Then I go in and do the rest myself in Figma. There’s always a level of customization required. In my experience, most GenAI tools are helpful for getting me to the consultant 80/20 level of work (maybe more like 60/40 and I do the additional 20).
I’m definitely still learning best practices as I go. If you have suggestions, please let me know!
What I’m learning
A decent part of my job as a founder is learning ways to speed up iteration cycles. Being aware of new tools and how to integrate them can be game changing.
Maybe I’ll find something better than WireGen. I was only half-joking when I recently told a friend that we need an LLM that tells you which LLMs to use for different tasks and how to optimally stitch them all together. (Again, let me know if this exists! I like Perplexity AI for discovering tools).
Not an earth-shattering insight, but finding ways to automate repeatable processes has definitely given me more time to think creatively and do more.
One last thing about customer discovery - you’re basically trying to portray an ideal view of the world in discovery calls. This doesn’t always translate into real-world functionality, so there’s an interesting back-and-forth when it comes to engineering. It’s always hard to find the line between being visionary and realistic (more on that in a future post maybe). The discovery process at least gives you some sort of north star you can use to build a roadmap.