Mathew PaquetTemplate · 4 min read

The Brand Rules File

A fill-in Markdown brand file with the five parts an AI can check, plus a five-output pass/fail test sheet.

01 Why a brand guide fails the model

A brand guide is written for people, and people infer. Tell a designer "bold but approachable" and they fill the gaps with what your company made before. A model has none of that. Anthropic's prompting documentation tells people to think of its model as "a brilliant but new employee who lacks context on your norms and workflows." Every task is that employee's first day.

So "friendly but professional" lands on the safest middle there is: the generic voice you were trying to avoid. Pasting the whole PDF in doesn't fix it either. Pages of adjectives and mood boards give the model nothing to check its own work against.

The gap isn't a missing document. It's a missing set of rules it can check.
02 The five-part brand file

Copy this, fill in the brackets, and keep it short. Harbor Roasters is a made-up coffee roaster, and so are its rules.

brand.md
# [BRAND NAME] brand file
Owner: [NAME] Last change: [DATE]

## 1. Exact values (numbers and names, not adjectives)
Headline font: [FONT] Body font: [FONT] Backup fonts: [FONT], [FONT]
Colors: [NAME] [HEX CODE], [NAME] [HEX CODE]
Accent: [COLOR], used [HOW OFTEN, e.g. once per page]

## 2. Voice as examples
Sentences we wrote and loved:
1. [SENTENCE]
2. [SENTENCE]
3. [SENTENCE]
Sentences we would never write:
1. [SENTENCE]
2. [SENTENCE]
3. [SENTENCE]

## 3. Rules, each with its reason
- [RULE], because [WHO OUR CUSTOMERS ARE / WHY IT MATTERS]

## 4. What to do (then a short banned list, last)
- Write like [A PERSON TALKING TO OUR CUSTOMER]: [HOW]
Banned words (five or six at most): [WORD], [WORD]

## 5. Checks (each answered yes or no)
- [CHECK, e.g. Headlines under eight words]
- [CHECK, e.g. Product name first in a subject line]

A worked example: Harbor Roasters

Exact values. Headlines in Fraunces, body in Source Sans, roast green is the accent, used once per page. Voice. Loved: "Our Ethiopia is out. Bright, jammy, gone by Friday." Never: "Experience the magic of our artisanal blends!" A reason. "Our customers are regulars who dislike being shouted at, so we write like a barista who knows their order. One exclamation mark per email, only for a sellout." What to do. Not "don't sound corporate", which gives it nowhere to go: "Write like a barista talking over the counter: short sentences, one idea each." Checks. Headlines under eight words. The product name comes first in a subject line. Alt text describes the cup, not the mood.

Why examples and reasons. Anthropic's documentation calls examples "one of the most reliable ways to steer" tone, and says the model is smart enough to generalize from the reason behind a rule.

03 The five-output test sheet

First hand the file to someone outside marketing and ask them what the brand sounds like. If they would be confused, the model will be too. Then give the same file to the AI and ask for these five things.

A subject line. Mark every rule pass or fail.
A social post. Mark every rule pass or fail.
An ad headline. Mark every rule pass or fail.
A product description. Mark every rule pass or fail.
An image brief. Mark every rule pass or fail.
Rule Subject Post Ad Product Image
[CHECK 1] pass / fail pass / fail pass / fail pass / fail pass / fail
[CHECK 2] pass / fail pass / fail pass / fail pass / fail pass / fail
04 Fix the file, not the output

When a rule fails, find the sentence that was ambiguous, add the example it needed, and run all five again. If you correct the output by hand every time, you teach nothing, and the same miss comes back next time.

Say the headline comes back at eleven words. The file only said "keep headlines short," and short is an adjective. The fix is the checkable version, "under eight words," plus one good headline as an example.

05 Where the file lives
In one place, with one owner. Name the owner and date every change. On a Figma plan that has Specs. Specs is Figma's place to write design rules in plain Markdown that people and AI agents both read; it is in open beta and rolling out gradually. Figma's help center notes guidelines can raise the credit cost of a prompt, depending on how big they are, one more reason to keep the file short. Not on Figma. A plain Markdown file does the same job. Google Labs publishes an open format called DESIGN.md, with sections for colors, typography, components, and do's and don'ts; it is still labeled alpha. Or use whatever your AI tool reads at the start of every session: a project file, a custom instruction, a skill.

Sources

Mathew's LibraryLibrary card
Free access

Keep reading free with your library card

One email opens this template and all 86 resources in Mathew's library. No account, no password.

It opens right here, and a copy goes to your inbox. You'll also get The Operator Notes. Unsubscribe in one click.

Thanks
You finished it
Now do one thing

Find one rule in your brand guide that is an adjective and rewrite it so you could answer it yes or no.

Saving works best in your phone's browser. A copy is also in your inbox.