If I Started FitWardrobe Again Today

FitWardrobe is live and used daily. It works. That doesn't make it the right version. The temptation with these posts is to be either humble-brag ('I nailed it, only tiny tweaks') or performative-flagellation ('I got everything wrong'). Neither is useful. The honest answer is that the frame was right and the sequence was wrong.

The frame was validated by more than just my interviews. Movinga's global audit of 18,000 households across 20 countries showed that the vast majority of closet items go unworn for over a year (movinga.com). The product bet was right. The build sequence was wrong.

Model before users. Features before metrics. Impressiveness before trust. Each cost about a week. All three were avoidable with habits I only formed later.

The evidence I'd lean on today is the evidence I generated the hard way - the seven-interview retention swing, the stopwatch session that killed three features, the trust affordances from the CRM. I already have the receipts. Starting again means using them from Monday, not month three.

Here is the version I would build.

• Week 1: seven discovery conversations before any model work. A Figma clickthrough that mocks a single 'what should I wear?' answer. Measure one thing: whether people tap again without being prompted.
• Week 2: pick the model architecture based on the four-axis exercise - latency, privacy, cost, recovery story - with the answer already leaning on-device but written down, not assumed. Week 2 would probably slip. It always does. I'd build something in week 1 that the interviews partly invalidated, and I'd need time to adjust.
• Week 3: build the trust surface - read-back, undo, 'not right? tap to fix' - before the recommendation is good enough to be worth trusting.
• Week 4: instrument two numbers - daily active outfits shown, and correction rate - and write the kill condition on day one.

The plan isn't really about weeks. It's about the order of bets, and the most important bet is that talking to people before building beats building before talking.

This ordering would have shaved several weeks off v1 and made the retention curve rise earlier. The prediction I'll take: it would have made the retro conversations shorter, because the metrics would have been in the plan.

The sequence I've described is the one I'm most confident in. But I keep coming back to one assumption I haven't fully tested: that the week-1 Figma mock works. It assumes users will engage meaningfully with a prototype of a product that doesn't exist yet. Sometimes they do. Sometimes you need a rough real thing before the conversation becomes real.

I keep this list in a note. Some of it I'm confident in. Some of it I've already started to doubt.