I just submitted the Sourdough Baker's Assistant for production review! While I wait to see it in the Play Store, I've been reflecting on what I learned designing and building my first app with AI. Here are my biggest takeaways.
Treat AI as a collaborator in early ideation.
I came up with a rough idea, and Claude helped me refine it. It already had enough context on sourdough that I didn't need to catch it up, and it jumped right into asking about the target user, scope, and scenarios I hadn't considered. Was the persona a little generic? Yes. But it was a good prototype to get me started, and I refined it as I talked to more real bakers.
The output is only as good as your prompt.
Clear direction matters with AI just as much as it does with people. When I gave broad instructions for the design system components I needed, the results were pretty far off. When I added detail and cited examples, I got exactly what I needed in a fraction of the time. This is also a good lesson to remember as a leader directing a team or a project.
Know what each tool does best, and make them work together.
Claude was great at initial designs and prototypes. Figma was better for design finesse, and for building a design system I could see and manipulate. Once I connected them, I could keep design and code in sync without ever creating a spec doc. It felt like the most collaborative designer/egineer pairing session I’ve ever been in.
Focus on strategy and hand off production.
Offloading simple tasks to AI, like bulk design and content updates and design system tweaks, saved me an estimated 40+ hours on the design process. That freed me to focus on ideation and creative direction. Design is not dead, but pixel pushing is.
AI is better at extending art than originating it
These models learn from what already exists, so they're strong at remixing and weaker at inventing something new. When I asked for a sourdough loaf illustration, the results were less than ideal. When I asked for variations on the one I drew myself, it exceeded my expectations. My takeaway: when you need something new and novel, create the first version yourself, then let AI extend it.
AI helped me build the app, but humans helped me perfect it.
Working alone with AI can become an echo chamber. It's built to be agreeable and to help you finish the task at hand, like the coworker who says "Great idea!" to everything. You can prompt it to challenge you, but it will usually stay within the confines of what you started with. It's unlikely to flip your concept on its head.
When a tester told me I was thinking about planning backward, it was an aha moment. Anchoring bias had me building a more powerful version of my original spreadsheet, and AI helped me go deeper down that rabbit hole. A real user showed me a better way. Even with the speed of AI-assisted workflows, critique, cross-functional reviews, and talking to real users in discovery and evaluation are still essential for catching blind spots early.
What would I do differently? Talk to real users sooner, ask AI to push back on my ideas, and bring in cross-functional partners to help me see my blind spots. I’m still learning, and I’d love to hear from you: what have you learned building with AI, and what would you do differently next time?