← All articles

Founder

Your Brand Book Needs a Machine-Readable Source

A brand book still has a job. But when AI tools help create copy, images, interfaces, and reviews, the PDF cannot be the only source. The underlying rules need structure, sync, and checks, with judgment kept human.

Brendan McWeeney6 min read
A single ordered lattice of light on the left radiating several synchronized parallel streams of glowing particles to the right on a dark background

I used to make brand books for a living. Years at an agency, turning a company into forty careful pages: color swatches, type specimens, a tone-of-voice section with do's and don'ts. A person opened the PDF and felt the brand. It worked, and it still works, because a good reference document teaches a human reader judgment.

Then I started building BRAWLER AI alone, and the population applying my brand changed shape. The reader did not disappear. The executor multiplied.

BRAWLER AI is a training journal for Brazilian Jiu-Jitsu, and I am the whole team. To ship at all, I lean on software: a token generator that pushes the design system into Figma, a media generation setup that draws assets for the app and the marketing site, AI reviewers that check copy before I do. Every one of those executors needed to be told what the brand looks and sounds like. So I told them, one paste at a time. A paragraph in one prompt, a markdown file beside another, a doc for a third. A month in, the brand was described a dozen times in a dozen places and no two descriptions agreed.

The forty-page book had not failed. It was doing its original job for its original reader. It had simply become incomplete: most of what now executes the brand cannot sit with a PDF and absorb it, and the human copies made from it drift out of sync. What was missing was not a better document. It was a source underneath the document, with my judgment still on top.

A brand is more than its reference document

A brand is not colors and words. Those are outputs. A brand is the answer to one question: what job do I do for my customer, and how should every touchpoint feel while I do it? BRAWLER AI serves a grappler years into training who optimizes everything except their jiu-jitsu, who can quote their deadlift but cannot say whether their guard passing improved. Every color choice, every sentence, every safety rule derives from serving that person. Anything that does not is noise.

When that answer lives only as prose, duplicated across files and prompts, it drifts. I caught my own drift in the act: an emotional-tone list with four adjectives in one file and five in another. Nobody decided to change it. Prose slides when you edit one copy and trust memory for the rest, and a dozen quiet discrepancies add up to a brand you cannot describe the same way twice.

Untangling that taught me to separate three things I had been blurring. Brand intent is the judgment: the job, the feel, the lines we do not cross. The structured source is that intent written down as explicit, versioned rules. And the brand book becomes a generated output of the source, still human-readable, still valuable, no longer the lonely master copy.

Three layers, not two axes

I first modeled the source as two axes, visual and verbal. Real use broke that model; the brand needed three layers.

Visual rules: color values plus their semantic names, perceptual descriptions, and usage rules; contrast requirements; composition and art direction. A hex value alone does not reliably tell a tool the things that make the color right: what it is for, what material quality it should carry, how it sits in a composition. "Warm cream, paper-like, matte, always the background" travels across tools in a way six hex digits do not.

Verbal rules: voice principles, tone ranges, approved terminology, banned language with the replacement attached to each entry, naming, punctuation. Broad prose can genuinely guide generation; it sets direction. But deterministic validation needs explicit rules. "Confident, not corporate" steers a draft. A closed list of banned phrases is what a check can actually pass or fail.

Behavioral rules: safety prohibitions, depiction rules that span media, interaction expectations, claim boundaries, accessibility requirements, and the named cases where a human must review. This third layer is where the brand's ethics live, and it is the layer a two-axis model has no room for.

The mechanism

The operating model is plain, and stating it plainly matters:

  1. People author and approve the rules.
  2. The rules live in structured source files.
  3. Builds generate the human-readable references and the tool-readable ones from the same source.
  4. Tools receive the context relevant to their task, not the whole book.
  5. Validators test the deterministic requirements.
  6. Release gates block known violations from shipping.
  7. A person reviews for truth, quality, judgment, and exceptions.

Be clear about step zero: structured data does not enforce anything by itself. A JSON file is as inert as a PDF. Enforcement comes from consumers that read the source, validators that check what can be checked, and release gates that refuse known violations. Judgment comes from the person at step seven, and nothing below that step replaces them.

The rule that stopped repeating

One rule earned this whole system, and it started as a rejection.

Early drafts kept producing a scene I refuse to publish: the practitioner journaling from behind the wheel on the ride home. It read as convenient. It was unsafe, and I never want to suggest it, not in a caption, not in a single frame. So I rejected it and rewrote it. Then it came back. In a carousel. In an ad draft. In an image prompt, the model cheerfully painting the same scene again. The same correction, over and over, across every medium, because the rule lived only in my reactions.

Now it is one behavioral rule, authored once: never depict or describe using the app while operating a vehicle, with the approved alternatives attached, the locker room, the walk out of the gym, home after training. The same rule flows to copy guidance, image guidance, and video guidance, and reviewers check for it. Generators start from the safe scene instead of being corrected into it. The checks do not guarantee detection, so my eyes stay in the loop. But the mistake stopped being fresh each time, and the judgment stays mine.

What changes for a team

I run this alone, where the payoff is not re-explaining the brand to every tool I touch. For a team, the interesting change is distribution, and it is not about replacing anyone.

Brand expertise stays exactly as valuable; it is what authors the source. What changes is access. A new hire on day one, a team three time zones away, and the AI tool drafting a first pass all read the same rules the brand's keepers wrote, the moment they need them. First passes arrive closer to on-brand because the rules were present at the start instead of applied at the end. And review gets better, not smaller: with the checkable parts checked by validators, expert attention concentrates where it is actually needed, on judgment, taste, and the exceptions no rule anticipated.

The brand book becomes an output

The brand book still matters. It becomes stronger when it is generated from the same source used by the systems helping execute the brand. People define the judgment. Structured rules distribute it. Validators protect the parts that can be checked. The final call remains human.

Here is the exercise I would hand any founder or brand leader reading this: find the correction you have made more than twice this quarter, the same note on the same kind of mistake. Then decide where it belongs. In the structured source, if it is a rule. In a validator, if it can be checked. In human review, if it is judgment. The correction you keep repeating is your brand asking to be written down somewhere machines and people can both read it.

BRAWLER AI

Voice-first training intelligence for Brazilian Jiu-Jitsu. Talk about your training. We'll find the patterns you missed.