People describe their style in abstract aesthetic categories ('casual,' 'minimalist,' 'business casual') that mean wildly different things to different people. Clothing preferences are actually a fit-and-function problem unique to your body, your life, and your tolerance for specific discomforts.
Michael Tiffany
Open your closet and look at the left side, or the bottom shelf, or wherever the clothes you never wear have migrated. Every one of those items was a reasonable purchase at the time, and every one of them is now teaching you something about your preferences that you probably haven't articulated to anyone, let alone to an AI. That grey wool blazer you bought for a conference two years ago and wore exactly once? It wasn't the color or the cut, it was that the lining made your arms stick when you reached for a whiteboard marker and you felt self-conscious every time you raised your hand. You never returned it because it looked great on the hanger and now it lives in textile purgatory, taking up space and quietly reminding you that looking good and feeling good in clothing are two different problems.
This is the central difficulty with teaching an AI about your wardrobe: the information that matters most is the information you've never bothered to verbalize, because clothing failures are experienced physically and emotionally but almost never described. You know instantly when something is wrong, but translating "wrong" into language that an AI agent can learn from requires a kind of attention most people have never practiced. The good news is that your closet already contains all the evidence you need, organized into a natural experiment: the things you reach for constantly and the things you don't.
The simplest way to teach your AI about your wardrobe is to show it what you own and tell it which things you actually wear, which things you avoid, and why.
Pick a category to start with. Tops are usually the most revealing because most people own more of them, which means the contrast between favorites and rejects is sharper. Take a photo of each item or lay out a few at a time and photograph the group. You don't need professional lighting or a flat-lay aesthetic, a quick phone photo against a bed or floor is fine. Then, sort every item into one of four buckets:
Heavy rotation. The items you reach for first, the ones you'd replace immediately if they were ruined in the wash. For each of these, tell your agent what makes them work: "This Uniqlo crew neck in medium is perfect because the cotton is heavyweight enough that it doesn't cling, the sleeves hit right above my elbows when I push them up, and it holds its shape after fifty washes."
Situational. The items that work for specific contexts but not daily life: the interview blazer, the hiking base layer, the one pair of linen pants that only comes out in July. For these, tell your agent the context and the constraints: "I wear this linen shirt when it's over 30°C, but only untucked, and only if I'm not going anywhere that requires sitting for a long time, because it wrinkles catastrophically."
Purgatory. The items you keep but never reach for. This is where the richest teaching data lives, because the reasons for avoidance are almost always about fit, texture, or maintenance rather than aesthetics. "This shirt looked great in the store but the collar is too tight and I spend all day tugging at it." "These jeans fit perfectly for the first hour and then the stretch denim goes baggy in the knees by lunch." "I love this sweater but it pills after one wear and I don't own a fabric shaver." Each of these negative descriptions teaches your agent a constraint it could never learn from a style quiz.
Gone. The items you're ready to donate, sell, or discard. Before they leave, take a moment to tell your agent why: "Bought this at a fast fashion store on impulse. The stitching on the shoulder came apart after three washes. I'm done buying from this brand." Even departing garments carry signal.
When AI styling tools ask you to describe your style, you get a dropdown menu of aesthetic labels: casual, classic, streetwear, minimalist, bohemian. These labels describe how clothes look on a mannequin, and they tell the agent almost nothing about whether you'd actually wear something. The gap between "I like the look of minimalist fashion" and "I will put this specific garment on my specific body and leave the house" is enormous, and it's filled with information about fit tolerances, fabric preferences, temperature sensitivity, body insecurities, maintenance thresholds, and context-dependent dress codes that vary by workplace, social circle, and even mood.
The closet audit bypasses aesthetic abstraction entirely and goes straight to behavioral evidence: what do you actually put on, and what do you leave on the hanger? The reasons you give for each decision are the raw material your agent needs, and those reasons tend to cluster into a surprisingly small number of categories:
Fit. Not your size number, but the geometry of the garment on your particular body: where the shoulder seam falls, how much room you need in the chest to raise your arms comfortably, whether you prefer a rise that sits at your natural waist or your hips. Size numbers are nearly useless across brands; a medium at Uniqlo, a medium at J. Crew, and a medium at Patagonia are three different garments. Your agent needs to learn your fit preferences in terms of dimensions and sensations, not labels.
Fabric and texture. Some people cannot tolerate synthetics against their skin. Others find merino wool unbearably itchy despite its technical virtues. You might love the look of linen but refuse to iron. You might prioritize stretch in every pair of pants because you commute by bicycle. These are hard constraints that override aesthetics entirely, and they're analogous to the food intolerances we covered in the previous article: not dangerous, but powerful enough to eliminate entire categories of otherwise attractive options.
Maintenance burden. A silk blouse that requires hand washing and air drying will never enter heavy rotation for someone with three kids and a single laundry day. Dry-clean-only is a dealbreaker for some and irrelevant for others. Pilling, shrinkage, color bleeding, wrinkle resistance: these are the practical realities that determine whether a purchase becomes a wardrobe staple or a closet ghost.
Context mismatch. You bought the leather jacket because you loved it in the store, but you work in a business-casual office and go to parent-teacher conferences on weekday evenings, so it sits. Your agent needs to know not just what you like but what your actual life requires you to wear on a typical day, and how many of those days involve which contexts.
After the audit, your agent has a corpus of item descriptions sorted by how much you actually wear them, annotated with reasons. That's enough to start testing comprehension.
The simplest test is the shopping filter: browse a store or website you'd actually shop at, screenshot ten to fifteen items you'd plausibly consider, and ask your agent to rank them by likelihood that they'd enter your heavy rotation. Don't tell it which ones you like; let it predict based on what it learned from the audit. Then compare its ranking to yours and see where it diverges.
The divergences are the teaching moments. When it ranks something high that you'd never buy, the correction reveals a constraint it missed: "You put that crewneck first, but look at the fabric composition: 60% polyester. I told you I avoid synthetics against my skin. This should be last." When it ranks something low that you'd actually love, it reveals a pattern it hasn't generalized yet: "You ranked this olive chore coat last, but I have three jackets in that exact silhouette in heavy rotation. You learned my color preferences but missed that I always gravitate toward boxy, structured outerwear."
Each correction refines the model, and unlike the food feedback loop from the previous article, you don't need to wait for the garment to arrive and be worn. You can run this exercise from your couch on a Sunday afternoon, iterating through several rounds of screenshots and corrections in a single session.
The closet audit is a one-time investment that gives your agent a baseline, but your wardrobe preferences evolve with your body, your job, your climate, and your age, so the model needs periodic updates. The lowest-effort way to maintain it is to narrate only at moments of high signal:
When you buy something new, take a photo and tell your agent why you bought it and what you're hoping it replaces or adds. When you reach for the same item three times in a week, mention it, because that frequency is a signal worth capturing. When you try something on in a store and put it back, spend ten seconds saying why: "Loved the color, hated the buttons" is a data point. When the seasons change and you rotate your wardrobe, note what comes out of storage first, because that sequencing reveals your priorities more honestly than any style profile.
Over time, your agent accumulates not just a snapshot of your closet but a longitudinal record of how your preferences shift, which brands hold up, which purchases were mistakes, and which silhouettes you keep returning to even as trends change. That record is something no styling app currently builds, because they optimize for the next purchase rather than for the long arc of your relationship with your clothes.
How many items do I need to audit for this to work? Start with one category: all your tops, or all your pants, or all your outerwear. Twenty to thirty items with honest annotations is more useful than a hundred items with vague descriptions. You can audit other categories later, but a single well-documented category is enough to test whether your agent is learning.
Does my agent need to see photos, or can I just describe the items? Photos help enormously, because multimodal AI models can identify color, pattern, silhouette, and fabric texture from images in ways that complement your verbal descriptions. But if you're using a text-only agent, detailed descriptions work: "Navy cotton crew neck, heavyweight, slim fit, no graphics, slightly faded from washing" is enough to distinguish one shirt from another.
What if I genuinely don't know why I avoid certain clothes? Put the garment on. Walk around for five minutes. The reason will usually surface physically before it surfaces linguistically: you'll tug at a collar, pull at a hemline, notice the fabric bunching at your waist when you sit down. Narrate what your body does, and your agent will extract the preference from the behavior.
Should I share my body measurements with my AI? Measurements are useful context but are not required. If you're comfortable sharing them, your agent can cross-reference fit preferences with actual dimensions, which is valuable when shopping across brands with inconsistent sizing. If you'd rather not, fit descriptions like "I need extra room in the shoulders" and "I prefer a mid-rise that sits two inches below my navel" work almost as well.
How is this different from AI styling apps? Most AI styling apps optimize for purchase recommendations and operate on aesthetic profiles: body type plus style label plus seasonal trends. They are essentially trying to sell you something. The approach here optimizes for prediction accuracy about what you'll actually wear, which is a different objective entirely, and it works with any general-purpose AI agent rather than requiring a specialized fashion app.
Pull out five items you wear constantly and five items you never touch. Photograph them, and for each one, tell your agent why it's in the category it's in. Be specific about fit, fabric, and context, not about aesthetic labels. That's your foundation, and it will take maybe fifteen minutes. Next time you're browsing a clothing website, screenshot a handful of items and ask your agent to predict which ones you'd actually wear. Correct its mistakes, and you've completed the first cycle of the loop.
