300 Articles, One Obsession

The Problem With AI Art Prompting

When generative AI image tools went mainstream in 2022, most people approached them the same way: type something descriptive, see what comes out, iterate randomly. The results were impressive by accident and inconsistent by design.

The underlying problem was vocabulary. Most people didn't have a precise language for visual style. They knew what they wanted it to look like — they just couldn't say it in terms the model could work with.

I'd been building that vocabulary for years before the tools existed.

A Decade of Art Historical Research

Starting around 2012, alongside my early deep learning research, I began systematically studying the history of visual art — not as an academic exercise, but as a technical one. I wanted to understand how visual style actually works: what makes an Impressionist painting look like what it is, what specific formal choices define Art Nouveau, how the Bauhaus movement's design philosophy manifests in visual decisions.

Over the following decade I published more than 300 articles on my Medium blog examining art movements, artists, and visual styles through a technical lens. Each article asked the same core question: what are the specific, nameable properties of this style? Not "it feels soft and dreamy," but: short directional brushstrokes, broken color, visible texture, high-key palette, atmospheric perspective de-emphasized.

That kind of decomposition — identifying the discrete formal properties of a visual style — turns out to be exactly what effective AI image prompting requires.

The Vocabulary Becomes a System

When Stable Diffusion and Midjourney arrived, I had something most early users didn't: a structured vocabulary for describing visual output with precision. Prompts weren't guesses — they were specifications. I could say chiaroscuro lighting, sfumato edges, Flemish-master color temperature, compressed perspective and understand what I was asking for technically, not just aesthetically.

The articles became a reference system. Each one mapped an art movement or style to the specific prompt vocabulary that would invoke it reliably: which terms hit the model's training distribution well, which combinations produced the intended effect, which approaches were model-specific versus transferable.

From Research to Consulting Practice

This body of work became the foundation of StyleGuideAI. When clients came with a visual identity problem — we want our AI-generated imagery to feel like this, not like that — I wasn't guessing at solutions. I was drawing on a systematic library of style knowledge.

The consulting practice developed around a core insight: AI art direction is an art history problem as much as a technical one. The artists and studios getting the most consistent, distinctive output weren't the ones who had mastered the tools fastest — they were the ones who could describe what they wanted with precision.

The 300+ articles weren't research in the academic sense. They were the slow construction of a technical vocabulary that the field needed but didn't have yet. By the time generative AI made that vocabulary commercially valuable, I'd been building it for a decade.

That's the part that doesn't compress into a quick pivot.