Building Somara with a Team of AI

If you're a product designer, a design engineer, or someone who lives where design meets build, you've been told a hundred times this year that AI changes everything. What's harder to find is someone showing their actual work: which tool did which job, what broke, and where the AI quietly fell short and I had to step back in. That's this.

Quick orientation. Somara is a conceptual iOS app that helps people get more out of the clothes they already own. You bring in your wardrobe, an AI stylist builds outfits from what you have, and shows them on a photorealistic version of you. I'm building it with my cousin, Joey Schroeder. It's early, and it's a portfolio project, not a funded startup. Everything below is how I actually built it.

That unworn 80% is the opportunity: a problem people feel constantly, attached to an emotion no product currently owns. When I went looking, the market was crowded with wardrobe apps that restate the problem instead of solving it. They help you catalog what you own, then hand the hard part, the styling decision, right back to you. Human-stylist marketplaces solve that, but on a timescale of days and a price that doesn't fit a daily need. The well-funded AI players are mostly pointed at selling you the next thing, which my research said is the fastest way to lose this user.

So the goal is narrow and a little stubborn: don't build another feed, don't build a store with styling bolted on. Build the thing people actually asked for (help using what they already own) and make that the entire product. Shopping comes second, and only when there's a real gap.

One is my researcher, one my synthesizer and writing partner, one ideates visuals, one writes production code. My role changed accordingly: I'm not typing every line anymore, I'm the person who knows what good looks like, briefs each specialist, checks the work, and stitches it together. The tools do the heavy lifting. The judgment stays mine.

That phrase is doing real work. The strong version of this is human-in-the-loop: AI accelerates the task, but review and the final call stay with a person. The failure mode is the opposite, where the output is fast and fluent and subtly wrong, and you ship it because it sounded right.

There's a bigger shift underneath, and it took me a while to feel it. Early on, the skill that mattered was writing a good prompt. Increasingly, the skill is orchestration: deciding which specialist handles which step, in what order, and how the outputs hand off. More on that below, because there was a specific moment it clicked.

Gemini for breadth. Market size, where the category is heading, who's funding what, what the existing apps actually do. It's strong at sweeping a lot of ground fast and organizing it, so I used it for the industry pass and again for data synthesis. When people ask why I don't just use one model, this is the honest answer: I reach for different tools for different shapes of work.

Claude for the careful parts. The interview guide was one of those. Before talking to anyone, I needed a guide that wouldn't lead the witness, which is harder than it sounds. We built a 60-to-75-minute conversational guide in eight sections, structured so I wouldn't pitch the idea until the very end, with reminders to let silence sit, because the real answer often lives in five seconds of quiet.

Then the unglamorous work: cleaning each recording into a transcript and synthesizing it into a structured field guide. Doing that 18 times by hand would have buried me, so I built a repeatable process in Claude. The synthesis is where the insight lived.

A few findings did most of the work. “Nothing to wear” is almost never an empty-closet problem, it's a combination problem: people own the pieces and can't see how to combine them freshly. Styling what you own ranked as the number one user value, above shopping. And there was a near-universal trust red line: selling body data was a delete trigger for nearly everyone.

The one thing nearly everyone shared was the hard moment: occasions. Not the daily routine, which most people handle, but the dinner or date where you'll be seen and don't have an answer. Even people who rarely feel “nothing to wear” named occasions as the exception. That's the painkiller the whole app should lead with.

From there it splits. Inspiration-driven users (Dani, Jac) want to recreate a look from what they own. Utility-first users (Connor K, Megan) want a fast answer and nothing more. Gap-hunters (Connor B) really want to know what to buy next. Designing for the average of those people would have fit none of them.

Two specifics changed the product:

Avatar dignity: David, the most product-articulate participant, described his ideal app almost exactly as a video-game character creator, then added that the avatar should never appear in its underwear. The digital you deserves the same dignity as the real you. That's now a rule.

The trust line: because the app touches your body and your image, any whiff of selling that data ends the relationship, which turned privacy into a design requirement, not a footnote.

“Habit-forming” is the exact language behind the feeds my interviewees said they were tired of. But the book reframed it: a habit is simply a behavior gone automatic, and the question that matters is whether it serves the person. Eyal is blunt that these tools only earn the right to exist if they make lives better. I held onto that.

I mapped Somara onto Eyal's four-part loop without kidding myself. The trigger isn't a notification I invented, it's an emotion that already exists, the drop in confidence before you're seen. The action is as small as I could make it: open, say where you're going, done, under thirty seconds. The investment is the quiet engine: rate an outfit, log what you wore, add an item, and the next suggestion gets more like you. Eyal calls that stored value.

The reward deserves its own line, because it came straight from a participant. Instead of one “correct” outfit, Somara offers three: Safe, Frisky, Risky. You don't know which will feel right today, and that uncertainty is what makes opening the app feel alive instead of mechanical.

The app is React Native and Expo, written in Cursor. Because the tokens are the single source of truth, a screen asks for a color or spacing value by name and gets the right answer in light or dark mode. The build inherited the design instead of reinterpreting it.

Accessibility is the part that meant the most to me, and a good example of what AI is genuinely great at. I ran a programmatic WCAG 2.1 AA audit with Claude Code, not a vibe check. Real contrast math against every token pair, screen, and component, each measured against the 4.5-to-1 threshold for readable text. This is exactly the work AI should do: exhaustive, rule-based, the kind of pass a tired human skips at 11pm.