What Building FitWardrobe Taught Me About Product Discovery

FitWardrobe is a fashion assistant that classifies clothing on the user's phone - no cloud, no login. I built it because on-device inference was the interesting engineering problem, and I wanted to prove I could ship a real TFLite model to a real user without a GPU bill.

The first person I showed it to looked at the screen for a few seconds, said 'oh cool,' and put her phone down. I kept talking for another minute before I realized she was done.

V0 could tell you a shirt was a shirt. It could not tell you what to wear on Tuesday. I had built a capable classifier and called it a product, and the first five people I showed it to were polite about it in the way that friends are polite about a bad haircut.

The actual pain - the reason people spend ten minutes staring at a closet at 8 a.m. - wasn't 'I don't know what this garment is.' It was 'I don't know what these garments look like together, given the weather, given the meeting, given yesterday.' My classifier answered a question nobody was asking.

The scale of this problem is bigger than I realised at the time. A Marks & Spencer consumer survey of 2,000 adults found the average person spends 17 minutes every morning deciding what to wear - that's 103 hours a year, more than two full work weeks (marksandspencer.com). The problem isn't a lack of clothes. It's a memory and retrieval problem.

I stopped shipping for two weeks and ran seven interviews - nothing formal, just fifteen minutes and one prompt: 'walk me through the last time you couldn't decide what to wear.' I did not show them the app. I did not defend the app. I only listened.

• Six of seven mentioned weather within the first minute.
• Four described a specific "safe outfit" they fall back on when tired.
• Three said they take photos of outfits that "worked" and never look at them again.
• Zero mentioned wanting a computer to identify what type of garment something was.
• One person spent eight minutes trying to navigate the catalog tab before giving up. I hadn't built it to be useful. It just existed.

The zero on that fourth point is the one that stung. My headline feature had no organic pull at all.

I had to choose between two products. Product A kept the classifier as the hero and layered outfit-building on top - protecting my sunk cost. Product B demoted the classifier to a silent helper and made the hero surface a single button: 'what should I wear today?'

I chose B. It meant most of the visible AI I had built became invisible. The demo would be less impressive to other engineers and more useful to actual people. I wrote 'make it boring' at the top of the new spec. I hated writing it. I shipped it nine days later.

The scoped-down v1 shipped in nine days. Day-7 retention with the same test group moved from roughly one in five to closer to three in five. The classifier still ran on every photo - but nobody talked about it, because they were too busy telling me the app 'just knew' what to suggest.

The engineering had not changed. The product had.

Six months later, with zero paid ads, FitWardrobe reached 20,800 impressions and 315 clicks (1.5% CTR, avg position 14.9) through organic search alone - reflecting organic search demand for solving daily wardrobe retrieval rather than another image classifier.

The mistake I keep re-learning: I fell for my own cleverness. On-device inference was a real advantage, but it was an implementation detail, not a value proposition. Users don't care where the model runs. They care whether the answer is right, fast, and doesn't make them feel stupid.

I still open editors too fast. The two-week pause before FitWardrobe's pivot was the only time I stopped myself - and I had to make myself do it.