TL;DR
- ·Situation & Research: McKinsey data shows global clothing purchases jumped 60% (mckinsey.com), Movinga's 18K household audit found items sit unworn for 12+ months (movinga.com), and M&S data showed users spend 17 mins/day (103 hrs/yr) deciding what to wear.
- ·Task & Action: Framed the core problem - apps solve closet organization instead of daily retrieval. Built an on-device AI testing version, identified cold-start logging friction during pilot testing, and iterated to an instant recommendation loop.
- ·Result & Next Step: Polishing the experience based on test feedback and launching the Beta/Early Access version very soon. Early users can claim access at https://fitwardrobe.me (validated by 20.8K organic impressions in 6 months).
Overview
FitWardrobe is an AI fashion assistant that classifies clothing on-device and recommends daily outfits in under 30 seconds. Built following a complete S.T.A.R. product cycle: analyzing industry research, framing the retrieval problem, building a pilot testing version, solving real user logging friction, and preparing the imminent Beta launch at fitwardrobe.me.
Problem Statement
Wardrobe apps fail because they solve organisation instead of retrieval. While McKinsey research showed clothing buying jumped 60% (mckinsey.com), Movinga's 18,000-household audit proved that people forget what they own (movinga.com), and M&S survey data showed users spend 17 minutes every morning unable to decide (marksandspencer.com). The answer isn't a better catalogue. It's a system that does the remembering for you.
Market Need
Research by McKinsey & Company showed global clothing purchases jumped 60% over a 15-year period (mckinsey.com), creating a visibility crisis in personal closets. The market problem was hiding in plain sight: not a lack of clothes, but a memory and retrieval bottleneck. On-device vision had just become fast enough to classify garments in under two seconds without a server. The moment that technical constraint lifted, the product became worth building.
Target Users
I assumed I was building for fashion-conscious people. Discovery interviews said otherwise. The people who engaged most were time-constrained and mildly anxious about social presentation - not particularly interested in fashion, but very interested in fewer decisions per morning.
Research
Before building, I ran 20+ discovery interviews and 6 diary studies - where participants texted me every morning with what they wore and why. This matched what academic research from PLATE 2025 found: clothing retention is a retrieval and memory challenge, not a storage problem (plateconference.org). When pilot users tested early builds, most stopped logging by day 4 - revealing high cataloguing friction. That feedback reshaped our solution to focus on automated single-photo recommendations rather than manual entry.
Insights
- ·Users don't want more outfit ideas - they want fewer, and confident ones.
- ·'Fit' is emotional: the same shirt gets rejected on a Monday and picked on a Friday.
- ·Cataloguing feels like homework. If it takes more than 10 seconds per item, it doesn't happen.
Opportunity
The right surface isn't a better closet - it's a morning ritual. Get a user a good outfit in under 30 seconds after opening the app.
Scope
- ·Photo-to-item vision pipeline (on-device background removal + attributes)
- ·Daily outfit generator with weather + calendar signals
- ·Feedback loop: keep / swap / never-again
Out of Scope
- ·Marketplace / shopping integration for v1
- ·Social sharing feed
- ·Multi-user closets
Assumptions
The riskiest assumption was that people would accept 3–5 recommendations instead of an infinite scroll. Everything I'd built before had given users more options, not fewer. I wasn't sure the constraint would feel like confidence rather than limitation until I watched a pilot user tap 'keep' without scrolling.
Success Criteria
- ·Target: D7 return rate meaningfully above the onboarding cohort baseline.
- ·Target: kept-outfit rate above half of daily recommendations - the model earning trust, not just generating variety.
- ·Kill condition: if correction rate rises above swap rate, the model is creating cleanup work, not saving time.
Expected Behaviour
Open app → see today's outfit in under 5 seconds → tap keep or swap → move on with the morning.
North Star Metric
Weekly Kept Outfits per Active User - chosen over DAU because it captures actual model trust. A user can open the app daily and reject everything. That's a failing product that looks good in the wrong metric.
Supporting Metrics
- ·Time to First Outfit
- ·Swap Rate
- ·Correction Rate
- ·D7 / D30 Retention
- ·20.8K Impressions / 315 Clicks (1.5% CTR, avg position 14.9) - 6 months organic
Prioritization
RICE across the backlog, with Confidence replaced by an evidence-class label. Vision quality and outfit-generation confidence won every round; social features consistently lost to retention bets.
Wireframes
A Figma clickthrough showing a single 'what should I wear?' response - no catalog, no UI chrome, just an answer. I measured one thing: whether the test user tapped again without being asked. Most did. That was enough to start building.
PRD
One-page per release. The non-goals section was the one I took most seriously - it had to name the things I was explicitly choosing not to build. If I couldn't articulate why something was out of scope, it wasn't actually out of scope.
Prototype
A clickable mock shown to a handful of people before any model code existed. The biggest thing I learned: users engaged with the recommendation before they engaged with the catalog. They cared about the answer, not the archive.
Final Solution
A vision-first onboarding, a daily home screen that opens on today's outfit, and a feedback loop that quietly trains the model on every keep or swap.
Results & Impact
20,800+ organic Google impressions and 315 clicks (1.5% CTR, avg position 14.9) over 6 months - zero paid spend. Following pilot testing - which surfaced cataloguing friction - and subsequent UI refinements, we are preparing the Beta launch at fitwardrobe.me.
Reflection
By analyzing industry research (McKinsey, Movinga, M&S) and gathering pilot feedback, we identified cataloguing friction and iterated on product improvements, generating 20,800+ organic impressions without paid acquisition. We are now preparing to open early access for users at https://fitwardrobe.me.
Lessons Learned
- ·The classifier - my original headline feature - became invisible in the final product. Not because it didn't work, but because the product worked better when it wasn't visible. That was a strange thing to accept.
- ·Discovery before building is a habit, not a phase. I keep having to remind myself of this on every new project.
- ·The catalog was the wrong hero. Users abandon apps that make them do upfront data entry - UX researchers call this the cold-start problem (plateconference.org). The fix was flipping the responsibility: the app remembers, the user adds one photo. Once that clicked, the whole product got simpler.