DESIGN.md, Stitch, and the Bet on Machine-Readable Brand

Here is a sentence I did not think I would type with a straight face, and I mean it: it is a beautiful move that Google put DESIGN.md on the table as a first-class artifact—not another slide template, not a Figma file you bless once and forget, but a portable, reviewable, versionable description of how a product should feel in the generative stack.^[1]

Not because a markdown file is magic. Because the hardest problem in the agentic era is not model quality—it is the translation layer between brand, design intent, and the flood of “fast” content. LLMs are happy to produce volume. The failure mode is volume without coherence: a thousand on-brand adjectives and zero machine-checkable constraints. DESIGN.md, in this framing, is a bid to make that intent legible to tools in the same way package.json is legible to the ecosystem.

From Figma to something agents can load

I have been arguing the same thing from the advertising and systems side: creatives should not fight the tools; design systems should behave like code—versioned, composed, reusable. Stitch's bet is the adjacent file: a design document that is explicit enough to travel with a repo, diff in review, and ground automated UI work without becoming a second job for design ops.

That is the through-line to what we built in Brand Lockup: the brand is not a PDF of guidelines; it is a system of primitives and rules you can apply under pressure. When the builder is a human, that means components and checklists. When the builder is an agent, it means structured language the runtime can see—typography, tone, affordances, what “good” means in context. A DESIGN.md that captures that intent is the difference between a prompt that says "make it look premium" and a spec an agent (or a linter) can actually reason about.

Brand Lockup, "determinism," and the Stitch sandbox

The brilliance of the DESIGN.md move is that it nudges design intent into the same social contract as code: review, diff, version, handoff. I am not solving that in the abstract—I am shipping it in production through Brand Lockup: design systems for advertising that behave like code. The goal is not a prettier template library. The goal is reproducible, rule-governed assembly—the same approved primitives and constraints, composed again and again, so the output is not a lottery spin every time a model touches the canvas.

I use the word determinism carefully. Creative work is not a closed-form equation. But governance can be deterministic when the system—not the model—owns the invariants: tokens, structural roles, disallowed phrasing, approval scope, and how far variation is allowed to go. The LLM can propose; therails accept, reject, or re-route. That is the same posture as a compiler: you still design the program; the compiler is merciless about the rules. Brand Lockup is the direction I am taking that seriously for advertising, where the cost of a wrong pixel or line is not embarrassment—it is delay, drift, and audit. If you want the longer arc, the build story is in the Brand Lockup case write-up.

I have also been using Google's Stitch and the public DESIGN.md lane as a test surface, not a religion: can we describe product and brand posture in a way that actually steers automated UI work, not just decorates the brief? A pattern, written clearly enough, should drive decisions for tools downstream—where layout goes, what the hierarchy is, which affordances are allowed, what the brand refuses to do. The win if this sticks is that those patterns become fuel for the next wave of LLM-native design tools: the model reads the spec, the runtime enforces the spec, and humans move up a layer to judgment, taste, and exception-handling.

There is a harsh corollary we should not sugarcoat. If you can describe any brand in words and ship a convincing asset in a weekend, the asset is on its way to commodity. Generative quality is improving faster than most brand guidelines are. So the defensible work shifts: not "we made a banner," but who owns the rules, the lineage, the audit trail, and the stack that keeps outputs inside policy when scale hits. DESIGN.md and systems like Brand Lockup are the bet that orchestrated, reviewable design infrastructure outlasts one-off generations—especially in regulated, global, and agent-mediated markets.

Why this matters for “fast” content and synthetic noise

The buyer is becoming an agent; the producer already is. When everyone can ship a thousand variants overnight, the moat is not production speed alone—it is authentic, verifiable, consistent identity in the feed. The brands that win will treat brand as infrastructure, not flair: tokens, rules, and narrative constraints you can compose and audit.

DESIGN.md does not finish that work by itself. It is a shape of answer: a single, honest place to write down product voice, design posture, and motion in human language—close enough to code culture that the same document can sit next to prompts, skills, and deployment configs. It pairs naturally with the payload discipline I care about in high-density agentic payloads: if you are going to pay in tokens, pay for entropy, not throat-clearing.

Native gen in every wall—and the extraction problem

Here is the uncomfortable complement to the DESIGN.md story: the big ad platforms are all racing to ship generative surfaces inside the buy—headlines, hooks, more variants, faster local tests. For a single team living in one DSP, that is real leverage: production finally talks to the auction. For anyone wholives in more than one garden, the picture breaks unless you have a portable layer of your own.

The work is not only "make a brief machine-readable," though what Google is doing with DESIGN.md helps. The other half is extraction: what did the machine produce, which assets got rewarded, what failed in review, what is safe to repeat in another market or channel, and what should never ship again. Native tools rarely want to answer that in a neutral format. They want you to optimizethere. Your organization needs a discipline—schemas, runbooks, ownership—that lets youpull that signal out and redeploy it on the next platform without running the same ad‑hoc ceremony every time.

I have been up close to that friction in Reddit specifically: the surface punishes un-threaded, generic "performance marketing" voice, subreddit context matters, and creative fatigue comes fast. That is why the product I am building at Creative Patterns is not "another text box that writes ads." It is a pattern engine for paid creative—repeatable creative moves, notebooks, LLM-ready runbooks—so teams can standardize the loop where Reddit is the forcing function, then carry the structured outputs and learnings to whatever comes next. Stitch-style intent files are part of the mental model: what you are allowed tobe in creative. A cross-platform compounding system is the rest: what you learned, encodedso it is not trapped in a single tool's black box.

The strategic question for CMO- and rev-ops-adjacent teams is not whether each platform will have gen (they will). It is whether your brand and performance knowledge accrues in your infrastructure—or gets re-negotiated every quarter in each silo. DESIGN.md is a promising start on thedesign half. The operating half is still: capture, govern, and ship learning like you ship code—so "optimize in Reddit, Meta, Google, elsewhere" becomes a routing problem, not a religion.

Optimistic, not naive

I am not claiming Google solved taste in a file format. Standards are adoption games; teams will need defaults, linting, and leadership that treats design spec as a living contract. The exciting part is directional: a major platform normalizing the idea that design intent should live where engineers and agents can address it—as text, in repo, in review—raises the floor for how fast, generated content stays on brandas volume scales.

We are early. It could be brilliant. If you are still treating brand as a handoff instead of a loadable spec—whether you use Stitch or not—start there: make the intent addressable, then wire the automations. The rest is compounding.

What to own when pixels go cheap

I will say the quiet part with volume: the object of most marketing design—an approved layout with a logo and a line of copy in range—will not stay a rare craft for long. Models and native tools are closing the gap. So stop optimizing for a prettier single shot; optimize for repeatability, auditability, and compounding—a system where every generation is a branch you can defend when someone asks who approved it, why it is safe, and what happens when the next platform ships another gen box.

That is the work Brand Lockup is for. That is the pattern layer I am hardening in paid with Creative Patterns. And that is why a line like DESIGN.md in Stitch matters: not because the file is clever, but because the discipline is transferable. In the next few years, the winners will not be the teams with the shiniest one-off asset—they will be the teams with the tightest spec and the coldest, kindest no when the model steps out of bounds. Everything else is noise. The same bet I have made in Why I back agentic infrastructure applies here: systems that compound when pixels do not.