An experiment in whether AI can learn taste without flattening it. Some days I dress myself and explain why. Other days the robot dresses me. Both get the same feedback, and I keep the final say.
By Kelsea Wentz, former product design leader at Netflix. I'm on sabbatical with my three-year-old twins, and this is how I keep up with what AI can do. Everything here is co-authored with AI, from the fashion styling to the social media planning to this website and the product ideas that come next.
outerwear · navy, draped
knit · mohair stripe
denim · wide leg
flat · burgundy
blazer · navy, DB
shirt · pale blue
trouser · split hem
heel · croc slingback
KelseaorRobot?
Fig. 01 · Same closet, two stylists
Three things I'm exploring
Getting dressed is my favorite place to start. It's personal, it's visual and it happens every single morning, which makes it the perfect playground for the question: what happens to taste when AI joins in?
Taste
Can AI learn taste?
Fashion first. The broader question reaches every visual craft, product design included: can a model learn what someone loves, or only what is average?
Closet
Easier mornings
Building my closet, understanding my own taste and refining it, and making it simple to get dressed well from what I already own.
Craft
Prototyping in public
How I concept and prototype products: the prompt library, this site, and maybe a real product, built in the open as the experiment runs.
/benchmark
Taste benchmark
Taste isn't one score. It's four layers, from what the culture says is current down to what works for you this morning. Here's how I'm breaking it down, for anyone getting dressed and for anyone building models.
Trend
Is the outfit on trend? Do I care?
Layer 1 · Society
Trends start with tastemakers and trickle down. Then they swing like a pendulum: once everyone has the new thing, the people setting trends reach for the old one to stand apart. That is why classic can be the most current move. Lean in, stay put, or borrow one piece.
Distance from a dated reference corpus (e.g. Vogue, by season), scored against the target the brief sets, not maximized. Models must track direction as well as position: trickle-down diffusion (Simmel) and status signals that move up from below (Bellezza's trickle-round). Refreshed each season.
Personalization
Does it feel like me? Do I feel confident in it?
Layer 2 · You
A great outfit on a hanger can be wrong on you. This is about your coloring, your shape and your signature moves.
Ground truth is the person. Blind pairwise preference vs. their own pick, plus "felt like me" ratings. On-hanger and on-body judged separately.
Context
Does the outfit work for what I plan to do?
Layer 3 · Today
School pickup is not a date, and a date is not the office. And then the real-life questions: What's the weather? Am I walking? What's even clean?
Mostly pass or fail: hard constraints from the brief, plus "worked for my day." Style dials shift by context.
Execution
Is the outfit well executed based on timeless principles?
Layer 4 · Foundation
Take what's current, what's you and what the day needs, then pull it off. Color, proportion, silhouette, texture: the timeless rules of good design, like why navy and brown work together. Once you know the rules, you can break them on purpose.
Rubric score: color relationships, proportion, silhouette balance, texture, cohesion. Expert raters or a model judge. The most objective layer.
Fig. 02 · Trend, personalization and context set the target. Execution holds it all up.
Style dials
Taste is a position, not a score
There's no “right” answer. It's about developing your personal style over time and tuning it as needed to different contexts.
Each outfit maps to coordinates on six axes. A person's style is a distribution per context. The model has to learn the region, not the mean.
Casual
Elegant
Classic
Trend
Minimal
Maximal
Structured
Relaxed
Tonal
Contrast
Practical
Statement
What counts as a win
I put it on, and I don't change a thing.
My outfits come with a note on why they work. The robot's get a scorecard on all four layers. On Sundays you guess who styled it. If you can't tell, it's a good sign the robot has learned my taste.
Try it on any outfit
Is it current, and do I care today?
Do the colors, shapes and textures work together?
Does it look like me, on me?
Can I actually do my day in it?
Modeled on SWE-bench
Resolved = worn as-is. Swaps are the edit distance.
SWE-bench
Taste benchmark
Repo snapshot
A real closet: cataloged items, taste references, rated history
Issue
A brief: plans, weather, constraints
Patch
An outfit of owned item IDs, plus styling notes
Hidden tests
Three tiers of checks, below
Three tiers of checks
HardEvery item exists in the closet. Outfit is complete. Brief constraints are met. Checked automatically.
InvariantsRules mined from the person's history, like signature moves. Breaking them costs points even when the craft is good.
JudgedCraft rubric, blind pairwise vs. the person's own pick, and trend distance from the brief's target.
Task types
AssembleRepair: one swapRescue a pieceCritiqueBoard fidelityCloset gapAgentic
Open problems
Noisy ground truth: people change their minds. Needs test-retest and a verified subset of stable briefs.
Trend drift: the benchmark refreshes by season, which also limits contamination.
Judges lean safe: diversity and novelty need their own metrics.
Privacy: on-body photos and closet data need explicit consent.
Roadmap
v0 · One closet
This experiment. Every day is a task, my pick is the human baseline, Sundays are the blind eval.
v1 · ~20 closets
Volunteers from the Instagram audience, same task format.
v2 · Public
Published tasks and a leaderboard: models vs. stylists vs. the people themselves.
/experiment
The experiment
What I want to find out, how it works and how I'll know. Findings come later, once there's real data to share.
Hypothesis
The robot can style me as well as I style myself.
How it works
It's a teaching loop, not a contest. Monday, Wednesday and Friday are mine: examples, each with a training note on why it works. Tuesday, Thursday and Saturday are the robot's: attempts, each with feedback on how it went. Sunday is the showdown: you guess who styled it.
Matching me is only the first bar. Over time, I hope the robot outperforms me, putting together outfits I love even more than the ones I'd pick myself. And since style keeps evolving with the wider culture, think of this as a two-fold adventure: honing my own style, then letting it evolve based on how much (or how little) I want to take part in broader trends.
North star
Style Parity KPI
[Love it rate on robot looks as-is] ÷ [Love it rate on my own looks]
1.0 means the robot dresses me as well as I dress myself. Above 1.0, better. Checked every two weeks.
Open questions
How much feedback before the suggestions feel personal?
Does the audience vote track "felt like me", or drift from it?
Which part of this would be worth turning into a product?
Additional key metrics
Hands-off rateHow often a robot look is worn as-is. Every swap or veto is me stepping in.
Verdict mixThe robot's Love it, Fine and Meh compared with mine.
Can you tell?How often people guess the stylist right on Sundays. Lower means the robot learned.
Style I love
In addition to my outfits and notes, the robot also gets outside inspiration: the people and brands below. It reads them before every outfit, alongside my closet and past feedback, so they shape what it picks.
Every outfit so far. My days come with a training note on why it works; the robot's come with feedback on how it went.
Sample Kelsea post
I get dressed, snap the outfit and provide notes. The robot analyzes it, indexes it and creates a board.
outerwear · navy, draped
knit · mohair stripe
bag · black leather
denim · wide leg
flat · burgundy
Training note
I wanted something casual enough for school pickup but nice enough for breakfast out. It's getting colder, so I went for layers. I'm loving my new La Ligne sweater, and the color story works: navy, burgundy and cream, with burgundy flats picking up the stripe.
End-of-day verdict*
*Love it: I'd wear it again. Fine: felt good, probably wouldn't repeat. Meh: wouldn't repeat, and didn't love how I felt.
Sample robot post
The robot creates a board from my closet. I then get dressed, snap the outfit and give feedback.
blazer · navy, DB
shirt · pale blue
bag · brown leather
trouser · split hem
heel · croc slingback
Training feedback
Outcome
The swapHeels → flats
WhyLoved the outfit, but I was walking to school pickup. Heels weren't going to happen.
End-of-day verdict*
*Love it: I'd wear it again. Fine: felt good, probably wouldn't repeat. Meh: wouldn't repeat, and didn't love how I felt.
Week 1 · Establish the baseline
MonKelsea
My first outfit and why I reached for it.
Upcoming
TueRobot
One piece from Monday, styled a different way.
Upcoming
WedKelsea
The small detail that makes an outfit mine.
Upcoming
ThuRobot
"More like me": a revision of Tuesday.
Upcoming
FriKelsea
A busy day. What practical need came first?
Posted
SatRobot
A real brief, no new purchases.
Upcoming
SunShowdown
Who styled it? Guess, then which and why.
Upcoming
Week 2 · Can the robot learn my taste?
MonKelsea
A piece I love but rarely wear.
Upcoming
TueRobot
The same piece, the robot's way.
Upcoming
WedKelsea
An outfit I'm happy to repeat.
Upcoming
ThuRobot
One change to that repeat. Better, or just different?
Upcoming
FriKelsea
Dressing for a mood.
Upcoming
SatRobot
Same mood, plus a practical constraint.
Upcoming
SunShowdown
Blind round: one look, no labels.
Upcoming
/setup
The tools
No custom code yet. A digital wardrobe app, a shareable closet link and an AI assistant were enough to start.
Catalog the closet
Every piece photographed in a digital wardrobe app. I use INDYX. One item per photo.
Share it with the AI
A public closet link lets the assistant read every item and check for new ones daily.
Give it references
People, brands and an editorial source, each with a note on what to take from it.
Brief, board, swap
Real plans and weather go in. A board of my own pieces comes out. I make one edit.
Write it down
Two scores and a sentence after wearing it. That note is the training data.
What the robot actually is
The “robot” is not some secret fashion AI. We are just using a general AI model plus a page of notes (for now).
One modelClaude, the same one every day, so changes in results come from learning, not from switching tools.
A taste profileA living page of my style rules, signature moves, hard nos and lessons, read before every outfit.
My closetAll 259 pieces from INDYX, numbered, so we only style clothes I actually own.
A feedback logEvery outfit's swap, why and verdict. It's also where the Style Parity KPI is calculated.
The model itself never changes. The notes do, and that's how it learns. Next, we'll explore how this data and these feedback loops could train two things: a broader fashion model that understands style in general, and a layer tailored to my own taste and the contexts I dress for.
/prompts
Prompt library
The prompts that worked, ready to copy. Fill in the brackets with your own closet and references.
Define your taste
Map my taste
Here are the people and brands whose style I love, with what I take from each:
[name] for [quality]
[name] for [quality]
Brands: [list]
Find the common threads: palette, silhouettes, fabrics, and the rules they break. Then tell me what my taste is in five plain sentences I can agree or argue with.
Define your taste
Separate taste from trend
Using [editorial source] as the trend reference, tell me which current trends fit the taste profile we wrote and which would feel like a costume on me. Explain each in one line.
Use your closet
Style a real day
From my closet only (no new purchases), style one outfit for this day:
Plans: [plans]
Weather: [weather]
Must-haves: [e.g. walkable shoes]
Name each piece exactly as it appears in my closet. If you are unsure whether I own something in a photo, say so instead of guessing.
Use your closet
Rescue an underworn piece
I like [piece] but rarely wear it because [reason]. Give me three ways to wear it with things I already own, from safest to boldest, and say what problem each one solves.
Give feedback
Score a look
I wore your outfit. Worked for my day: [1–5]. Felt like me: [1–5]. I changed [item] because [reason].
Update your notes on my taste. Tell me what you'll do differently next time, and what you'll keep.
Give feedback
One swap
Here is an outfit I repeat: [pieces]. Change exactly one item. Tell me why, and whether you think the result is better or only different.
Make boards
Build an editorial board
Make an outfit board from my actual closet photos. Use a loose, organic layout with varied composition and no text. Add an inspiration image only if it adds atmosphere, such as a color or texture echo of one piece.
Make boards
Showdown brief
Here is this Sunday's brief: [brief]. Style a look for it from my closet without seeing mine. Give me only the board, so the A/B comparison stays blind.
