Show It Before You Speak — #7
In March, I wrote that “AI cannot receive design.”
Figma MCP, Google Stitch, Pencil, Figma Make. I tried them all and concluded the volume of design work had not decreased. The reason: “Design surfaces as image before language.” AI works the other way around. Words first, image after. The order is reversed.
A month later, Anthropic released Claude Design. I tried it.
It was the answer.
With design.md, the Best Tool There Is
Claude Design has a mechanism for feeding it design.md. Color, type, spacing rules, prohibitions. Hand it your design system spec in Markdown, and it generates within those constraints.
This works ferociously well.
If your design.md is complete, Claude Design is the best tool you can hold. The 1→N stage hums. Variations, responsive adaptations, component derivatives. A human builds one, the AI multiplies it into ten. The speed difference is in another league.
But 0→1 is impossible.
Without design.md, ask Claude Design for “a cool card design” and you get the same March result. Something that looks plausible but cannot be touched by hand. Of course. Without constraints, AI can only return the average.
What Was Proven
Claude Design did not prove the performance of AI design tools. It proved a structure: the presence or absence of design.md decides everything.
Same tool, same model, same prompt. The only difference is whether design.md exists. The results are night and day.
What I wrote in March — “there is a moment where the cost of putting it into language exceeds the cost of just making it” — was the experience of the side without design.md. On the side that has it, the language is already done. So it can be passed to the AI. So it is fast.
Where Does design.md Come From
This is the heart of it.
design.md is not something you “write.” It is what remains after you have made things.
You place a color. You adjust the spacing. You change a font weight. There is a moment when something feels right. At that moment, you cannot put it into words yet. But keep moving your hands and patterns surface. “Cormorant Garamond fits this brief.” “Margins at 1.618× feel good.” “Accent color goes only on headings and CTAs.”
The thing that summarizes those is design.md.
Going from 0→1 by language is slower than just moving your hands. design.md is born as a byproduct of the work.
Our L0 Pipeline systematized exactly this process. Phase 0 product understanding through Phase 9 quality gate — across nine phases, design.md is generated as a byproduct. The goal is not to “write design.md.” The goal is to “make.” design.md remains as the result of having made.
In March this was just “our methodology.” Claude Design arrived and proved from outside that this methodology is structurally correct. The era caught up a little.
The Order Problem, Again
In March I wrote: “In design, an image surfaces in the head before the hands move. AI is reversed. It demands words first.”
Claude Design does not solve this problem. To be precise, it is a tool for people who do not need it solved.
For people who hold design.md, the language is already done. The pain of converting head-images into words is in the past. So they can hand it to Claude Design. So 1→N hums.
For people who do not hold it, language work is just beginning. “Dark tone, Cormorant Garamond, clock image on the right, rounded card” — fragmented instructions like this never become design.md. Between fragments and a system, there is no shortcut around hand-time.
The “moment where language cost exceeds making cost” I wrote about in March was exactly the state of not holding design.md. Back then it was a general impression of AI tools. Claude Design arrived and proved the impression was a structural problem.
The Haves and the Have-Nots
The world will figure this out: design.md is required.
The better AI design tools become, the more the presence of design.md decides outcomes. It is not a model gap. It is an input quality gap. Same Claude Design, but what comes out of someone with design.md and someone without is different.
And design.md is not produced by prompt engineering. It exists only on the far side of moving hands, placing colors, feeling spacing, repeating “this is it” and “this is not it.”
Our L0 Pipeline systematized that process. Across the nine phases from Phase 0 to Phase 9, design.md is generated as a byproduct. The goal was never to “write design.md.” It was to “make.” design.md remains as the result.
A person who can write design.md is a person who can make without writing it.
The reverse does not hold. A person who cannot write design.md cannot transmit design to AI. The “head-image does not become a prompt” line from March was pointing at this.
The gap is not decided by AI performance. It is decided by accumulation that precedes language.
For now, the hands keep moving.