I have 218 items in my wardrobe today.
I know, because I’ve entered them all into a collection of Claude skills I built around personal styling.
Shirts, pants, dresses, shoes, accessories. All of it.
I can open a terminal and ask it what I should wear today. My system knows what I own. I can tell it what kind of day or event I’m dressing for. It knows what I like. It’ll even check the weather.
And when I don’t like a suggestion, I can say so.
“Less structured.”
“Too much denim.”
It learns from the responses.
It’s one of the most trustworthy AI tools I’ve built for myself.
It’s probably useless to you.
A passion for fashion
This particular idea was partly inspired by Clueless.
Cher Horowitz has a computerized wardrobe so she can assemble her outfits on a screen. It tells her when things match.
I can do the same. Although Cher’s interface is more fun than the command line.
What makes the system work is its context.
It knows what I own. It doesn’t make up items I don’t have. And I’ve spent time teaching it frameworks for fashion, best practices, and my own preferences on how I dress. It usually gives me pretty good advice.
And I can judge everything it recommends. I just need to look in the mirror.
If something doesn’t work, I can tell it why. The feedback comes from outside the model.
But I can’t really offer this to anyone else.
I love fashion. I’m obsessive about the things I build. And I’m willing to spend time entering 218 items into my database.
Most people won’t do this.
There are plenty of wardrobe management applications out there - some of them AI-driven. And they all promise a better-organized wardrobe. It’ll be easier for you to make decisions about what to wear.
But you have to do all of the organizing first.
That’s a lot of work until you can get value. And it’s not very entertaining work.
My system works for me because I’m silly enough to willingly do all these things. It’s not a foundation for a more widely used product.
An experiment in invention
Earlier this year I was really diving into building Claude skill suites. Practical tools for my own use. Automating parts of my day-to-day work. Managing my balcony garden!
One was an invention pipeline.
It’s a series of skills to approach invention from different directions. Relevant frameworks. Conceptual collisions. Scoring hypotheses. I wrote about it in more detail a few weeks ago.
To test it, I wanted to see if it could take something useful for me, and turn it into an idea other people could use.
And could I go from that idea to an actual deployed thing within a day.
I fed it my own Claude Code wardrobe system.
I even referenced Clueless in my starting prompt. And an ambition that nobody using the application would be “ensembly challenged.”
The invention process generated a bunch of concepts. I collided personal styling with DJ set curation. There were ideas around dressing for different chapters of the day.
It was a fun exercise.
But all the research and evaluation kept running into the same problem.
Cataloging.
The more information any application needs about your wardrobe, the harder it is to get anyone to really use it.
I ended up with two recommendations called “Mirror Mode” and “ZeroCloset”. Neither of them required a user to catalog anything.
I combined them into a single experience.
It’s called Style Grammar.
The invention pipeline’s outcome had taken some of the things I liked about my personal tool, and discarded everything that made it high-friction to use.
So I decided to build it
I took those invention results and fed them into another suite of skills, based around software delivery.
We’d produced a product definition, and now we could apply a delivery process to turn it into something real.
On one day, in a single session, the pipeline took forty-one invention concepts. It distilled them from exploratory stages, evaluated the hypotheses, and ended up with a working application.
It was one of those “whoa” moments around the capabilities of AI.
But when I tested the application, I still found nine bugs.
Wrong model identifiers. Mismatched APIs. Interface issues. A narrative stream that hung. All of the kind of things you discover when you actually try to use something you’ve built.
I fixed them. Deployed it. Tested payments.
More issues. Price identifier mismatches. Blank streams.
It’s still getting from concept to deployment in a day, and that’s impressive. It wasn’t quite concept to commercially viable product (not that that was really my intent).
They system had built something pretty solid. But that’s not the same as achieving the desired outcome.
Is this right?
There’s a random B-movie from thirty years ago that’s stuck in my head. Sketch Artist II: Hands That See is a pretty dumb thriller about a police sketch artist working with a blind witness.
In the climactic courtroom scene, a defense attorney is challenging the sketch artist’s reliability.
How can he know that the drawing he made represents the person the witness is describing.
The sketch artist tells him. “I show them the drawing.”
The defense attorney points out that the blind witness couldn’t see the drawing. What a twist!
Our AI models are getting very good at drawing pictures, literally and figuratively.
I’m not sure if they’re very good at telling if they’re right.
Does it feel right?
Consider my wardrobe system.
Claude recommends an outfit to me, using items I own. There’s some stuff we can test.
Did it reference my actual clothes?
Did its recommendations match established principles of color and styling?
If it tells me a combination matches my preferences, does that match what I’ve told it?
These are good tests. They’re useful.
But then I put on the clothes and I look in the mirror.
And I decide whether I like what I see.
That’s not an objective assessment about good style. It’s my judgment. But really that’s the only judgment that matters for a tool I’m using to help dress me.
There’s a source of truth outside the system. It’s me.
Style Grammar presents a different problem.
A user finds the site, answers some questions, and receives an assessment of their style.
I can test if the application works. Follows the rules I’ve defined, and gives recommendations correctly.
The user might think it’s right, or they might think it’s wrong.
Both those outcomes pass the system’s tests. But even a user agreeing with the result doesn’t necessarily mean they’ve found it useful. We could only really discover that by observing what they do with the result.
That’s not a problem unique to AI, of course. We’ve always confused completion with success. Correlation with causality.
The man in the mirror
Agents optimizing against an evaluation will learn to satisfy that evaluation. But the evaluation is a proxy for what we want.
It’s especially difficult when we’re evaluating something subjective.
Like whether someone looks good in an outfit.
Or whether a customer finds a product genuinely useful.
We use multiple agents to challenge each other. We keep dissent in the loop. We have adversarial review.
It’s all worthwhile.
But independent reviewers aren’t independent from the assumptions they evaluate against.
A group of sketch artists could all agree that a portrait was an accurate reflection of a witness’s description.
They still need the witness to tell them whether they’ve drawn the person they remember.
Building Style Grammar was a way for me to examine that problem. I took something demonstrably useful to me, and tried to turn it into something fun and useful to other people.
I automated an extraordinary amount of the work. Invention processes challenging the original idea. Delivery processes building an application. Reviewers finding problems. Automated testing.
I still don’t have any evidence that people want what I’ve built.
That can’t be evidence that the model could make it. And it’s not evidence that another model approved that first build.
It’s evidence from the people who are meant to use it.
Which is exciting, not disheartening.
The more we automate work, the more we can focus on which parts of that work can be evaluated from inside our system, and which require something outside it.
My own styling assistant has a big advantage over Style Grammar.
Every day, I can look in the mirror and decide if it gave good advice. I’m the person who’s actually going to wear the clothes.
Style Grammar doesn’t have that advantage.
I can ask the machine if it thinks the picture is right.
But eventually someone else has to look at it.
Further reading:
Kim, M. The Clueless Effect: How One Cult ‘90s Movie Made a Lasting Impact on Fashion. Vogue, Jun 2015.
Haynes, B. Can it work? Does it work? Is it worth it? British Medical Journal, Sep 1999.
Northoff, G. & Gouveia S. Does artificial intelligence exhibit basic fundamental subjectivity? Phenomenology and the Cognitive Sciences, Feb 2024.
Article photo by Vytenis Malisauskas on Unsplash.