Your brand just became installable
Not a style guide your team is supposed to remember: a file your tools are unable to ignore.
Something quietly changed in the last year: AI tools stopped being where your team drafts things and became where it produces things. Landing pages, sales decks, emails, product UI, built in minutes by Claude, Cursor or Codex, by whoever needs them, designer or not.
Which surfaces the real problem. The bottleneck was never can we produce this? It’s does any of it look like us? Your marketing team knows the drift: three tools, four freelancers, and a brand that fragments a little more with every asset shipped. The standard fix, a PDF brand book and a review bottleneck, was already failing when humans did the producing. Against AI-speed production it has no chance.
So we ran an experiment on ourselves: what happens if the brand isn’t a document people consult, but law the tools load before they build?
What we built
We took Design Catalyst, the service that installs an in-house AI designer inside your company, and pointed it at client zero: ourselves.
Worth being precise about what that means, because “AI designer” is not a chatbot that suggests colour palettes. Paired with agents like Claude Code or Codex, it’s a designer that ships: graphic design, web pages, UI, motion, video, briefed in plain language by anyone on your team, producing finished work in your brand. The codification of your brand is the core that makes it trustworthy; the agents are what make it a designer rather than a document.
First, the study. We audited our own live site and design system and forced every implicit preference into an explicit, written decision. Why this blue and no other. Where the warm accent is allowed to appear, and the longer list of places it never may. Which two typefaces, at which weights, and why nothing else. That produced a Decisions Log: twenty numbered rulings, each with its reasoning attached. Not moodboard adjectives: law.
Worth being honest about the head start: as client zero we were codifying a brand that already existed: a live site to audit, a design system to read. Though even that head start was the same method in embryo: the site itself was built by one developer working from the codified design system through Claude Code, before any of it had a product name. With a client, Design Catalyst starts one step earlier. We work alongside your designer, yours, or one we bring, to put the brand book together first: identity, palette, type system, the components your team actually ships with. Every choice is made deliberately and written down as it’s made, so the Decisions Log isn’t reverse-engineered from finished work: it’s drawn up with it. Either route ends in the same place: your brand as numbered, reasoned law.
Then, the packaging. The log was distilled into a developer pack: a machine-readable BRAND.md, a design-tokens file byte-aligned with the design system, the logo suite, a nine-brief Prompt Playbook for non-technical teammates, and a skill that teaches the AI both how to build on-brand and how to review work against the rules.
Finally, the install. The pack went into the website’s own repository, wired so that every AI coding session loads the brand law before it writes a line, automatically, not as a reminder someone has to paste in. The files are locked down server-side, so the law ships with the code without being public.
Where this lands commercially
This is the part that matters for your commercial team. The experiment isn’t about design tooling: it’s about what happens when expert judgment is installed rather than consulted.
With the pack in place, brand consistency stopped being a review step and became a default. Ask the AI for a new landing page section, a profile page, an email template: it comes out in the right type, the right blue, the right radii, the right voice, because the rules were in its context before the first token. The person prompting doesn’t need to know a single one of the twenty decisions. What your company has effectively hired is an in-house AI designer, graphic, web, UI, motion, that anyone on the marketing or sales team can brief in plain language, and that never forgets the rules. Your human designer isn’t a bottleneck anymore; the founder isn’t the last line of QA at 11pm.
And brand evolution changed shape entirely. Mid-experiment we retired our combined logo lockup, a decision that in most companies decays over months of stale decks and forgotten templates. Here it was one ruling in the log, one versioned release of the pack, and every tool that loads it was building by the new rule the same day. Four versioned releases shipped in the first 48 hours, each one propagating brand-wide the moment it landed.
What worked
Three things, concretely:
- A full page, dark hero, editorial grid, responsive, live in production eleven minutes after the first prompt, on-brand without a designer touching it. The very page this post lives on: we recorded the AI building the Field Notes home from prompts containing zero brand instructions, and about nine of those minutes were actual working time. The pack’s Brand QA earned its keep on camera, too: it failed the first pass on two invented greys that existed nowhere in the theme, named the on-palette replacements, fixed them, and passed.
- The Brand QA skill catches drift, not just builds. Pointed at existing work, it flagged rule violations (a wrong emphasis colour, a retired logo still in use) the way a code reviewer flags bugs.
- Non-technical teammates get the same result. The Prompt Playbook means someone in Claude chat, no repo, no code, produces on-brand documents from the same law.
What didn’t, and where we wouldn’t point it yet
Honesty section, because this is where it gets useful.
The law only governs tools that load it. Assets built in a project where the pack wasn’t installed kept an old emphasis colour and the retired logo for days after the rules changed. Coverage is binary: installed, governed; not installed, drifting. Rolling the pack into every surface is real work the codification doesn’t do for you.
The AI can’t make the decisions, only enforce them. Distilling twenty rulings took genuine judgment about what our brand is, and the system only held because every change went through the log first, in one direction. If your brand decisions live in someone’s head or shift by mood, there is nothing to codify yet. That discipline is what makes the AI designer trustworthy; the pack is just its container.
Copies rot. Within days we had a local copy of the site quietly out of sync with production, and had to write an arbitration rule, the live site wins, into the log itself. Any system of record needs a rule for what happens when reality and the record disagree.
The takeaway
The shift for your company to act on is this: written, versioned expert judgment is now infrastructure. AI tools will produce whatever volume your marketing and sales teams ask of them: the only question is whether your standards ride along.
Getting them to ride along is what Design Catalyst is. Not a tool you configure: a finished package: senior brand and marketing strategists design your brand with you (or refresh the one you have), every decision is codified into law, the result is installed as your in-house AI designer in Claude, Cursor, and Codex, and your team is trained to run it: four weeks from start to handoff. Everything in this post, including where it broke, is that package tested on its hardest critic: us.
And notice the premise generalises past brand: pricing rules, sales qualification, onboarding steps: anything you currently enforce by reviewing output can instead be loaded before the output exists. That’s the idea behind the whole Catalyst series. Brand was the experiment we could run on ourselves first.
If you want to see where it would land in your company, download the brochure for what’s inside and what you own at handoff, or book a discovery call. We’ll also tell you honestly if design isn’t your binding constraint.
Field Notes is our build log: each edition, one AI capability we actually tested, led by its business application, including where it broke.