Months of reading, so you get one second of answer.

Why we built it

AI image generators are brilliant at making pictures and hopeless at art history.

Type “in the style of Monet” into any of them and you’ll get water lilies and soft purple light. The Monet from the tea towel. It isn’t wrong — it’s just the first thing the machine knows about him, which means it’s the first thing everybody else is getting too.

The trouble is that it’s also the least interesting thing about him. Monet painted in the snow. He painted London in fog so thick you can barely find the bridge. Van Gogh drew in reed pen and ink about as often as he painted in oil, and the Netherlands he drew looks nothing like the Provence he painted. Rodin worked in terracotta and plaster and pencil, not only the famous bronzes.

None of that reaches you if the prompt stops at the artist’s name.

So we decided to do the unglamorous thing: sit down with one artist at a time and write down what they actually made, what they made it out of, and what it genuinely looked like — then let the tool build prompts out of that instead of out of a vague impression.

That’s the whole idea. Everything else is just the work it took.

What’s in it

• 261 artists, each one researched and written up individually.
• Twelve prompts at a time — or six, or twenty-four, if you’d rather.
• Only materials they really used. You won’t get an oil painting from someone who only ever drew.
• Only subjects they really made. No seascapes from a painter who never left the city.
• Three platforms, each with the right settings written in for you.
• It runs in your browser. Nothing you type is sent anywhere, to us or to anyone else.
• It admits what it doesn’t know, and hands you the full instructions to use elsewhere instead.

What got in the way

The tool shuffles, and that changes everything

It doesn’t store a description and read it back to you — it mixes and matches, so every sentence we write has to make sense next to every other one. That sounds small. It isn’t.

Early on, one prompt confidently described a huge altarpiece painted for the wall of a church and then, one line later, called it “intimate and cabinet sized”. Another took a mood we’d written about Sargent’s painting of blinded soldiers and attached it to a picture of four children in a hallway. Nothing was factually wrong. Everything was wrong.

So each sentence has to be checked not on its own, but against every situation it could ever turn up in. That is the single reason this took months rather than weeks.

Finding facts worth having

Anyone can repeat what’s already repeated. We wanted the details that actually change a picture, which meant catalogues, conservation reports, exhibition notes, letters and interviews — and a hard rule that a fact only counts if somebody named actually witnessed it, if there’s a real colour rather than a feeling, and if we know the sizes involved. Plenty of days end with a two-hundred-page report read and one usable sentence gained.

Never repeating ourselves

Two artists described in the same words are, as far as the tool is concerned, the same artist. So repeated phrasing gets hunted down and rewritten — the longest repeat left between any two artists is ten words, and the only phrases we let stand are the plain yardsticks for saying how big a thing is, where there are only so many ways to put it. And this gets harder, not easier: the first artist had nothing to be different from, the two hundred and sixty-first had everything.

Knowing when to say no

Not every artist belongs in a tool like this. Living artists have a say in how their work is described. Work made inside a cultural tradition, under community authority, or carrying sacred meaning isn’t raw material for a phrase shuffler. So some artists carry a plain-language note about what their work may not be used for, and some we’ve chosen to leave out altogether. That makes the collection smaller than it could be. It’s the right call.

No shortcuts available

We built a wall of automatic checks, and they catch an enormous amount. But a computer cannot tell good writing from bad. It can only tell us where to look. So every single sentence in this thing has been read by a human being and judged — which is precisely as slow as it sounds, and precisely why it’s worth it.