A brand check is the moment your guidelines stop being advice. Something is about to go out, and Brandbook tells you whether it follows your rules, what doesn’t, and how to fix it. Here’s what actually happens inside one.
What gets checked
Pretty much everything. Copy you paste into Check. A design you upload to the Library. Every asset Create makes. Text an agent sends over MCP before it shows you. A post on one of your monitored channels. They all go through the same checker, so a draft gets the same answer wherever it came from.
Rules first
Every check starts with a rule engine. No model, no guessing. It tests the things that can be tested exactly:
- Words on your avoid list, with the word to use instead.
- Exclamation marks in headlines, if your brand doesn’t use them.
- Colours that aren’t in your palette, measured by how different they look, not just by hex value.
- Text and background pairs with too little contrast to read.
- A logo on a photo without an overlay.
- Your own written rules: banned words, required words, word limits, no emoji, no shouting in capitals.
The colour test is worth a word. Brand colours drift in small ways: a green that’s a few shades off because someone eyedropped a compressed JPEG. Brandbook measures the visual distance between colours (ΔE, if you like the jargon). Close enough to be the same colour passes. A near miss gets flagged as off-palette with the exact brand colour to swap to. Something far away gets flagged as not a brand colour at all.
Then, optionally, a model
If AI is turned on, a model reviews the draft after the rules have run. It can notice things rules can’t: a tone that’s off, a photo that doesn’t match your imagery description, a headline in an image. It reads uploaded designs, too.
But the model only ever adds findings. It can’t overrule the rule engine, and it can’t wave something through that the rules flagged. That’s deliberate. A check should never disagree with itself, and it should still work when there’s no AI key or the model is having a bad day.
The rules decide. The model adds notes.
How the score works
Every check starts at 100. Each error takes off 15 points and each warning takes off 8. Errors are the clear breaks, like a banned word or an exclamation mark in a headline. Warnings are things worth a look, like an off-palette colour or weak contrast.
- 85 and above: On-brand. Good to go.
- 50 to 84: Needs a look. Fix the findings, or send it for review.
- Below 50: Off-brand.
We kept it simple on purpose. You should be able to look at a score of 77 and know it means one error and one warning, not wonder what a model was thinking.
Fixes, not just findings
A finding that only says “wrong” isn’t much help at 4pm on a Thursday. Where there’s a clear fix, the finding carries it: replace this word with that one, strip the exclamation marks, swap this red for your brand red, add an overlay behind the logo. On Check, applying them is one click, and the score updates as you go.
After the check
Checks don’t disappear. Each one is logged against the brand, so you can see what’s been checked, by whom or by which tool, and how it scored. If you want a person to sign off, send it to the review queue, and you can set a score above which work is approved without anyone needing to look.
Over time the checks tell you something about your guidelines, too. When the same finding keeps coming up, or people keep rejecting the same kind of thing, Brandbook suggests a rule on your brand’s overview. You decide whether it goes in. The brand gets a little sharper every week, and the checks get more useful because of it.
