Last month Google Labs open sourced the format that powers their Stitch tool. It is called DESIGN.md. A single Markdown file. YAML at the top with your design tokens. Prose underneath explaining the why. Drop it in your project root and your AI coding agent reads it before it writes a line of UI.

If that pattern sounds familiar, it should.

CLAUDE.md does the same thing for code. AGENTS.md does it for general agent behavior. README.md has done it for humans reading your repo for thirty years. Now DESIGN.md does it for visual identity.

You are going to see a lot more of these. Buckle up.

The Pattern

The shape is always the same. Plain text file. Lives in your project root. Versioned in Git. Readable by a human. Parseable by a machine. No platform, no plugin, no lock in.

You can already see where this is headed. SECURITY.md for your security posture. DATA.md for your schema and naming conventions. BRAND.md, TONE.md, VOICE.md for marketing. Whatever the next agent needs to know about your business, somebody is going to write it down in a Markdown file and check it into the repo.

This is good. This is how it should work.

Autonomous Does Not Mean Omniscient

Here is the part people keep missing about AI agents.

Autonomous and smart are not the same thing. An agent can run on its own. That does not mean it knows what you want. It does not know your brand. It does not know your tolerances. It does not know that your contractor client's customers are mostly homeowners over fifty who will not tap a button smaller than their thumb. It does not know that your manufacturer partner has spent twenty years building a quiet, engineered brand and would rather die than ship a screen with a gradient on it.

It cannot know any of that. Not until you tell it.

Most of the complaints I hear about AI output come back to the same thing. The agent did exactly what it was asked to do. It just was not asked the right way. It did not have the context. So it guessed. And the guess looked generic, because the average of everything on the internet is generic.

The fix is not a better model. The fix is better inputs.

Constraints, Guardrails, Variables

That is what these Markdown files actually are. Constraints. Guardrails. Variables. The three things every good engineer and every good designer has always used to make sure the output is not a coin flip.

A DESIGN.md file says: these are your colors, these are your type sizes, this is the why behind them, here is what you do not do. That is a guardrail. It does not slow the agent down. It points it in the right direction so the work it produces is actually useful.

CLAUDE.md does the same thing for code. Here are the conventions we use. Here is the framework we have picked. Here is what we do not want, even if you have seen other people do it. Same pattern. Same purpose.

This is not gatekeeping the AI. This is operating it the way it was meant to be operated.

Now Think About Your Own Business

Here is where I want you to stop reading for a second.

Think about your business. Your craft. The thing you do every day that you are good at. Now ask yourself: what are the rules of how it works that you have never written down?

Every business has them. They live in the head of the founder. They get repeated in the meeting the new hire missed. They show up the moment somebody does it wrong and gets corrected. Tribal knowledge. Unwritten. The stuff "we just kind of know."

That is your DESIGN.md. That is your CLAUDE.md. That is your AGENTS.md. You just have not written it yet.

If you run an HVAC shop, your rules might be: we run service calls in these zip codes and not those. We do not sell on price. We answer the phone in two rings during business hours and within fifteen minutes after hours. We do not install equipment we cannot service ourselves. Our techs wear shoe covers, every time, no exceptions. Those are guardrails. They are why your customers trust you. Right now they live in your head.

If you run a manufacturing brand, your rules are the ones your dealers are quietly bending in ten directions right now. Logos in the wrong colors. Co-op ads that look nothing like your brand. Product photos shot in somebody's garage under fluorescent lights. You have a PDF brand book somewhere. Nobody reads it because nobody can search it, version it, or hand it to an agent.

If you run an agency, your rules are the ones the new project manager learns the hard way. What clients you take. What you charge. What stack you build on. What you say no to. What "done" actually means.

None of that is in a Markdown file yet. It probably should be.

My Version of This

At Red Barn we have been doing this in pieces for years. Our tech stack is documented. Our design system is documented. The way we build SEO for a contractor is documented, because we have done it hundreds of times and the process is the process.

Now we are writing it for the agents. Same content, slightly different format. Tokens at the top. Prose underneath. The why next to the what. So when an AI helps us build a contractor's site, it does not start from the average of the internet. It starts from how we do it. Our colors. Our defaults. Our patterns. Our reasons.

That is the work. Not picking the model. Not chasing the new tool. Sitting down and writing down what we actually do, in a format that a human and a machine can both read.

Most businesses I see have not done this. The rules live in the founder's head and in the team's habits. It runs fine that way until you want to scale, or hire, or hand off, or use any tool that needs to be told what good looks like. Then the absence of written rules is the bottleneck.

AI is exposing this. AI is not the only reason to write your business down. It is just the loudest one.

What This Means If You Pay for Software

If you are a contractor, a manufacturer, or anybody buying digital work from an agency, here is the part to take with you.

The companies that get good output from AI are not the ones with the fanciest model. They are the ones with the best inputs. The clearest brand. The cleanest standards. The most explicit guardrails. If your vendor cannot tell you where their constraints live, where your tokens are defined, where the rules for your project are written down, your output is going to be generic no matter how powerful the model is.

Ask where the Markdown lives. Ask to see it. If the answer is "we just kind of know," you are paying for guesswork.

Where I Land

DESIGN.md is not a moment. It is a marker on a trend that started with README and is going to keep going. The shape of the future is small text files that tell agents what you actually want.

Autonomy without context is just expensive guessing. Constraints are not the enemy of speed. They are the reason speed is worth anything.

Write your business down. Version it. Hand it to the machine.

Simple. Beautiful. Works.