Why I'm building Trove

Earlier this year I was building a shader for a client's brand. I wanted a particular kind of texture. Not a grain. Something that felt retro, more like a halftone but not a halftone. Like a dither, but with a very specific pattern. I knew I had seen it. I knew I had saved it, because it was the reason I wanted the effect in the first place.

I searched for "halftone." Then "dither." Then "retro texture." Nothing. I found it weeks later by accident, while looking for something else, and when I opened it I remembered everything about it except where I had put it.

I have saved things I have no memory of saving.

I save things daily. In my notes, in Notion, on Instagram, on TikTok, on X. The list goes on. I'm constantly fascinated by what I find, and I know it holds value I'll want to come back to.

The thing that doesn't change, and the thing that won't stop

Design principles are mostly fixed. Hierarchy, contrast, color, restraint. The vocabulary shifts, the ideas don't. Mediums have evolved, but outside the medium the problem solving has stayed nearly the same. A designer working in print in the 1900s and a designer working on an interface today are running recognizably the same process, and the same thinking.

What changed is everything around that process.

For most of design's history the hard part was finding good reference. You saw what was in your city, what was in the local library, what your studio happened to own. Access was the constraint, and the constraint did some of the thinking for you. A small pool is a legible pool. Designers took pride in their collection of scraps, in their books.

That flipped. The hard part now is not finding it. Good work arrives constantly, from people you'll never meet, in volumes nobody has a strategy for. So we save. We built a whole category of tools for saving: bookmarks, boards, archives, read-later queues, folders inside folders. They're all very good at the part where the thing goes in, and nearly useless at the part where it comes back out.

They lack action. You save a reference for a type system you admire, and three months later, when you're actually building a type system, nothing brings it to you. It sits there being saved.

Where it actually hurts

The front half of a project almost never changes. Discovery, research, strategy. All of it happens before there's a file open, before there's anything to look at. That's the part I've done the same way for years, and it's the part where my own collection should be doing the most work.

It's also where my collection is hardest to reach, and I think two things are happening at once.

The first is volume. I have consumed far more than I could ever recall. Not more than I can organize. More than I can remember having seen. There's a specific feeling of knowing you've seen the answer before and having no way back to it, and it isn't a filing problem. It's a memory problem.

The second is worse. The moment you need a saved thing is the moment you're worst at describing it. When you save something you're the person who found it, so you file it under how you found it. When you go looking, you're the person who needs it, and you reach for it by what you need it for. Those are different vocabularies. You're searching a system you built for a stranger.

This is true of most work that involves making something. A salesperson reaches for their last deck, or that one metric, or that one insight.

Why now

Two things got cheap at the same time.

The one everybody noticed: recall got cheap. AI is very good at holding a lot of material and finding the right piece of it. A collection that was only ever legible to the person who made it is now legible to something else. Most people stopped the thought there and built better bookmarking.

The other one matters more. Making things got cheap.

Anyone can generate forty directions before lunch. The part of the work that used to take the most time, producing the options, has collapsed, and what's left is judgment. Knowing which of the forty is any good. And then being able to say why precisely enough that the next forty are better.

That second half is the one nobody's tooling for. When generation is free, the scarce inputs are judgment and specificity, and a prompt is a very low-bandwidth way to transmit either. So you sit there trying to compress a decade of looking at things into a paragraph, and what comes back is technically right and completely generic, because you gave it words when what you had was taste.

But you've already recorded that taste. Everything you chose to keep is a record of what you thought was good, built one small decision at a time over years. It's the best description of your judgment that exists anywhere.

And right now it's unreadable. By you, and by anything else.

That's the whole bet. Not that AI can organize your references — that it can finally read the record you've been keeping without realizing it.

Why filing was never going to fix this

Every tool in this category asks you to file at save time. Pick a board, pick a folder, add tags.

That's asking for a decision at the exact moment you're least equipped to make it. You don't know what the project will be. You don't know which of the six things this reference is about will turn out to matter. You saved it because something in it caught, and naming what caught is work you don't want to do right then. So you skip it, the pile grows unsorted, and the tool quietly becomes a graveyard.

So Trove doesn't ask. You write a quest as a sentence in your own words, like "create a shader for the Trove branding," and the library assembles what it already holds for that sentence. Nothing was filed. One reference can serve two quests without being copied into either. The decision moves from the moment you save to the moment you actually have the question, which is the only moment you can answer it.

What Trove is right now

A private library of design and engineering references, with an agent called Mentor living inside it. Everything in it is something somebody else made that I decided was good. This is where it's gotten to, not where it lands.

Things come in from wherever I find them. A model describes each one on arrival, and then I look at them myself, one at a time, in a screen called Triage, where I either vouch for something or throw it away. The thing can be almost anything: a visual specimen kept for how it looks, a tool, a literal reference like a prompt or a type pairing or a spec sheet, a technique I want to try, an essay I want to read, a person I want to keep up with.

Right now I can send any of these straight into Trove. Mentor tags it automatically and extracts attributes I can reference later. It also leaves a place for my own thoughts, in notes. A note is me saying why the thing is there. It's where I record my judgment, the part the machine can't pick up on its own. The process might seem tedious, but it's extremely rewarding. Like flossing. I hate doing it, and it has probably saved me countless dentist visits.

Mentor records people too, so I can get back to whoever made the thing. It's been a wonderful way to keep up with someone's work, and to get a better understanding of what makes it theirs.

Trove also has what I call quests. It's my attempt at activating my resources when I need them most. I give one my thoughts on the directions I want to try, and Mentor pulls together the tools, techniques, and artifacts that relate to it. It even lets me download them as a zip or copy them straight into Figma.

I've also built an MCP that connects Trove directly to Claude. So now Claude has a better understanding of the work I like. It's developing a sense of what I find interesting, and it has a repository it can tap into for techniques, for context, for learning what I mean when I say something is good.

It behaves like a designer's mind — but the part I care about is that it understands me because I told it, not because it inferred me. Every keep, every discard, every note is a decision I already made, recorded once, on my own time. Nothing about me is guessed.

The bet

Taste isn't a category. Any system that can only hold what you can articulate will lose exactly the judgments worth keeping, which are the ones you've never had to put into words. And retrieval isn't instruction. Handing a model twelve examples gives it twelve examples. It doesn't give it "I like restraint in color and none in motion," which is the thing I was actually trying to transmit.

Nobody has closed that gap, including me. But I don't think you close it by getting better at describing your taste. I think you close it by keeping a long enough record of what you decided, and why, that describing it stops being the only way through.

If this works, a set of references stops being a pile and starts compiling into something closer to a brief. The front half of a project arriving with what I already know inside it.

Trove is private, in alpha, and has one user. I'm writing this from the middle of it rather than the other side. At worst, it's a tool I use myself. At best, it's something other people get to use too. I'm building it either way.

What I'm after hasn't changed since the first day I sketched it: access to the things I have collected when it matters most.