No Figma. No Canva. Design with code.

PROJECT MUSE — an experiment in programmatic design.

Somewhere in the last year, AI got better at making things than I am at explaining what I want. That sentence bothered me enough that I decided to test it for real: in one day, talking to an AI agent, I built a small creative studio. No Figma. No Canva. No design tool at all.

This article came out of that studio. So did the carousel version of it you might be swiping instead. Both were laid out by a 611-line program. Nobody dragged anything, because there was nothing to drag.

The mouse was the middleman

Here's the thing I only understood by building this: design tools were never really about design. For forty years, describing what you wanted in code was harder than dragging it with a mouse — so the industry bolted a mouse onto everything and called the dragging "design." Figma and Canva are the peak of that workaround. Brilliant ones.

AI flipped the cost. Describing now beats dragging, almost everywhere. The mouse turns out to be what it always secretly was: a middleman.

I tried automation first. It was meh.

Muse is my second attempt at this. The first version I built the obvious way — automation: a pipeline that produced content on its own, end to end. It worked, technically. And the results were meh. Competent, complete, and hollow — nothing in them was mine.

So this version inverts the philosophy: augmentation. The AI doesn't replace my understanding — it turns my understanding into things. I bring the evidence and the taste; it brings the making. Every artifact in this experiment started as something I actually thought, said, or chose. That single inversion is the entire difference between the version that got deleted and the one that made this page.

A studio in a folder

The whole experiment is embarrassingly simple. A project is a shared folder. I drop evidence in — raw notes, voice-to-text dumps, typos included (my voice-to-text is a crime scene). An AI agent writes artifacts next to it: an article, a post, a deck.

The entire studio is one formula:

raw notes + design system = artifacts.

The notes own the truth — nothing gets claimed that isn't in them. The design system owns the look. The agent just multiplies the two. Which means the real work of this experiment wasn't the platform at all. It was the second term.

Taste, compiled

A design system usually lives in Figma, or in a PDF, or in a designer's head. Mine is a program — and building it was most of the actual work.

It happened in two layers. First the taste: I spent weeks hand-building the craft with AI — what makes a slide sequence work, where the energy of a deck rises and falls, what a slide may never do. All of it written down as plain text rules the agent reads before every job. Taste that used to be my gut is now a file it can obey.

Then the look: measured, not eyeballed. The navy is the darkest pixel of our live site. The paper field is flat because the site's is flat. One rust accent, fixed, never a second color. All of it frozen into the 611-line renderer — a headline is the same size on every slide because it physically cannot be anything else.

A design system that lives in code has one more gift: it keeps itself. I never proofread brand compliance. I can't ship an off-brand slide even when I'm rushing, because the renderer won't draw one. Guidelines ask politely; a program just keeps the standard.

One note in. A campaign out.

Here's what the formula feels like in practice. From one messy voice note, the studio produced a 1,352-word article with every figure traced back to the source — 31 of 31 — a post, and a branded deck. Same folder, same day.

The convenience is hard to overstate. I don't open a tool and start assembling. I talk, I drop files, I point at what's off — and finished, on-brand artifacts accumulate in the folder like the work is condensing out of the conversation. Rebranding is editing a file: change one line, and every slide rendered after it obeys.

Now what?

What this experiment left me with isn't a conclusion — it's a bigger appetite. If a brand can compile, what else can?

Script-based video editing — cutting a video by editing its transcript, the way I edit this article by editing markdown? Decks that assemble from meeting notes? A whole studio where every craft is a renderer and every renderer is a file? I don't know yet. That's the point.

For forty years the bottleneck was writing the code. That bottleneck is gone, and I keep noticing how much of what we call "workflow" was just scaffolding around it. Take the scaffolding away and the real question surfaces, the one that was always underneath: not what can the tools do — what do I actually want?

I'm finding out in the open at DF Labs, where we keep experimenting, sharing, and figuring out what AI means for how we work. Follow along.