The hardest part of building a startup isn't the code anymore; it's knowing what to build in the first place.
That was the thesis Dan Maccarone brought to the NYU Summer Launchpad accelerator, and it landed. With nearly 30 years in tech and six AI startups under his belt in the past few years alone, Dan has seen what happens when founders skip the hard thinking and jump straight to shipping. Spoiler: it doesn't end well.
The New Risk: Building Too Fast
AI has made developers faster than ever. 84% now build with AI tools, up from 72% just last year. But speed without direction is just expensive chaos. Dan's provocation to the room: "Just because I can, should I?" That question should come before every feature, every sprint, every line of code.
Talk to People. The Right People. The Right Way.
Customer discovery sounds obvious. Most founders do it badly. Dan broke it down:
15-20 interviews is enough, not 50. Real patterns emerge after five or six conversations. Don't over-engineer the process.
Don't just hit your network. Your friends will be nice to you. Go where your actual ICP (ideal customer profile) hangs out.
Structure matters. Start broad: daily life, habits, context, and slowly work toward your domain. Never lead with what you're building. The inverted triangle approach keeps people honest before they know what you want to hear.
Ditch the focus groups. One loud personality will dominate, and you'll walk away with that person's opinions dressed up as consensus. One-on-ones only.
Prototypes Over Pixels
When it's time to test a concept, Dan's process is tight:
- Use Claude to code a lo-fi, black-and-white prototype- fast, cheap, functional
- Test with 5 pilot users (not 50, not your co-founder)
- Feed the feedback back into a dedicated Claude project
- Fix it, then hand it off to real development with a proper design system
The black-and-white part is intentional. Color and polish distract people from what you actually need feedback on: does this work? Do people get it?
AI Is a Collaborator, Not a Strategist
Dan uses Claude and NotebookLM throughout his process. To synthesize interview transcripts, writing screeners and discussion guides (used to take a full day, now takes two minutes), and maintaining what he calls a "product brain" across a project.
But he was clear about where AI stops: it cannot conduct the interviews, find the right participants, decide which features to kill, or craft the actual product strategy. The north star has to come from a human who understands the market, the team, and the problem.
One watch-out: AI-generated copy in prototypes. Users notice it. Generic filler text pulls their attention away from what you're actually testing, so make the copy purposeful, even at the lo-fi stage.
Prioritize Like You Mean It
Dan's Now / Next / Later framework is simple and brutal: if a feature doesn't align with your defined strategy and north star, it doesn't go in "Now." One person asking for something? Later. Multiple users raising the same pain point? Move it up.
And the hardest part? Removing features. Technically easy. Politically brutal. Founders get attached. That's exactly why user testing evidence matters, it gives you something to point to when you need to say no.
The Summer Launchpad accelerator workshop series continues throughout the summer. Stay tuned for more recaps from the NYU Leslie Entrepreneurial Institute.
