AI Painting Teachers Are Mostly Slop. Here's a Version That Isn't.

Ask any current chatbot how Vermeer built up his light. You'll get a confident answer in twelve seconds. It will mention lead-tin yellow, ultramarine glazes, and the camera obscura. It will sound like it knows what it's talking about. It will be partly right, partly wrong, and structurally untraceable — there's no way to tell which part is which.

That's the problem with AI painting content as a category. Not that it's stupid. That it has no idea where its own claims come from.

We built Apprentice's recreation guides to be the opposite of that, and we want to tell you exactly how, because it's the thing we're betting the project on.

What goes wrong by default

Run a typical large language model with a prompt like "give me step-by-step instructions for recreating Constable's View of Salisbury Cathedral" and you'll get something fluent. It will look like a study guide. It will name specific pigments, give brush sizes, suggest a session structure. A working painter will read it and notice three things:

1. The specifics aren't anchored to this painting. They're a blend of generic 19th-century-landscape advice and whatever the model has indexed about Constable. There's no way to know which sentences are about this canvas and which are about an aggregate Constable that doesn't exist.
2. The confidence is uniform. Whether the model is repeating a well-documented technique or guessing about the underdrawing, the prose feels equally sure. The reader cannot calibrate.
3. The sources are gone. Even if the model is paraphrasing a real book — Solomon's The Practice of Oil Painting, say — that linkage is lost. You can't follow up. You can't disagree. You can't learn the book.

The output is not wrong, exactly. It's unfalsifiable. And unfalsifiable instruction is entertainment, not training.

What changes when you build a corpus

The fix is unglamorous: stop asking the model to be a teacher and start asking it to compose evidence.

Our corpus is ten classical art-instruction texts spanning 1390 to 1920 — Cennini, Vibert, Solomon, Speed, Ruskin, Dow, Chevreul, Goethe, Vanderpoel, Bridgman — plus ~2,200 artist Wikipedia biographies, ~200 famous-artwork pages, and Wikipedia entries for movements, techniques, and pigments. All of it chunked into roughly 14,000 passages, embedded with nomic-embed-text, indexed for nearest-neighbor lookup. The full list is on the sources page.

For each artwork, the pipeline runs four retrieval queries — composition, color, drawing, materials — pulls the top passages, then has a small model (qwen2.5:7b-instruct) rerank them for whether they would actually help someone paint this specific work. Embedding similarity finds passages that share keywords with the query; the rerank step weeds out passages that share keywords but don't carry usable knowledge. Only after both passes do the surviving passages become context for the generator.

The full pipeline is documented on the methods page. The short version: by the time the generator sees the prompt, it has been handed a labeled block of source passages and told that every concrete claim must trace back to one of them.

The discipline that actually matters

The prompt is the second half of the work. The retrieval gives the model real material; the prompt tells it what to do with that material. Three rules:

Specific visual details — objects on the wall, clothing patterns, facial expressions — may only be mentioned if a source passage explicitly describes them. No filling in from training-data stereotypes of the artist's other paintings.

General artist practice — a painter's documented palette, signature methods — may be stated without an exact quote, but must be phrased as such. "Vermeer characteristically…" not "in this painting…".

When uncertain, hedge with "likely", emit null, or list the gap explicitly in knownGaps. Confident specifics that paper over uncertainty are not allowed.

The knownGaps field is the discipline made visible. Every recreation guide on Apprentice has one. It is what the corpus didn't cover. It is on the page on purpose. We would rather show you what's missing than hide it.

What the numbers say

We ran the pipeline against a priority batch of 2,500 famous artworks. Of those:

Two of those numbers deserve a hard look.

95.4% sourced steps. Of the ~14,000 individual painting steps in the corpus, more than thirteen thousand cite a specific book chapter or Wikipedia page that supports them. The other 4.6% are the cases where the model judged that no single source supported a step strongly enough to cite — and so it labelled the step uncited rather than fabricating one. That's the discipline working.

100% non-empty knownGaps. Every guide admits at least three things it doesn't know. The Piranesi entry openly says that specific visual details of his oil paintings are not described in the sources, because Piranesi is primarily known for prints. The generator wrote that itself. It's allowed to. It's expected to.

What this doesn't fix

A grounded LLM guide is a scaffold, not a substitute for a working painter. The recreations are best understood as a way to focus attention — on the brushwork phase, the palette choices, the historical method — and as a cross-reference to what the classical books actually say. The corpus is the source of truth. Read it.

We also still don't know how factually correct each individual guide is on close reading by a domain expert. The numbers above are about process — that claims have sources, that uncertainty is surfaced, that the pipeline doesn't invent reference labels. The next layer of verification is humans. When users flag issues, we'll log them on a public errata page and update the affected records. That's not built yet. It will be.

What you get instead of slop

Pick any artwork in the collection. The Constable case, since we started there. You'll see:

• An overview that names which of Constable's documented working principles apply, with Source 8 pointing at his Wikipedia bio.
• A grisaille underpainting step citing Source 1 — a real chapter in Solomon's The Practice of Oil Painting.
• A palette listing the actual pigments documented from his period.
• A knownGaps block saying specifically what isn't in the sources for this work.
• A footer listing every cited book, linked to where you can read it.

A working painter can read this and either agree or check the citation. That's the whole bet. Not that AI replaces a teacher — that AI, pointed carefully at a corpus that already exists, can surface the body of teaching the books no one reads anymore have been carrying for centuries, and put it next to a specific painting.

Whether that's actually useful to a painter at an easel is something we'll only find out from painters at easels. We're listening.