We paid a lot of money for a lifestyle brand
We sold professional services automation software. What came back from the rebrand was a lifestyle brand.
It was a big agency and a big invoice, and the result was beautiful in isolation and wrong for what we did and who we sold to. Then the brand team was cut, and the rest of us picked up the pieces and reworked it for new leadership and a new board.
That left 73 pages of brand guideline nobody used, and we wondered why.
Showing people what good looks like is fine. But unless they know what it takes to build like that, you’re asking for inconsistency. Pages of examples, no instructions.
Hand that to a contractor or an offshore team and every deliverable comes back feeling like a different brand. You’re relying on a designer to study the examples, work out the rules behind them, and carry the vibe on their own. That also assumes the examples were the right ones. Brand guidelines are full of billboard mockups and tote bags. We were making one-pagers, ebooks and presentations.
The brand itself needs about 10 pages. Split the rest into instructions for the people who use it, and give each of them what they need rather than what makes the document look impressive. Without the people using it, the brand doesn’t exist.
Then we paid a different agency to build the design system
This one was for the website replatform, and they came in through an existing relationship on the marketing side.
The colour options were picked on taste, with nothing underneath. Nobody had asked whether these were our competitors’ colours, whether they set us apart, or what they were meant to make a prospect feel. One weight per colour, no scale, no semantic roles. Then the high-fidelity wireframes arrived looking like a website from 20 years ago, every component slightly different on every page and everything spaced at random.
So I built the colour and type system myself. Colours chosen to sit near Salesforce and away from everyone we competed with, and type that was free, readable and ours. They were designing with no tokens, so I built them a starter design system in Figma to work from. They took the build on from there, and an Optimizely partner agency finished it and built it into Optimizely.
The half that was already solved
Design systems worked this out years ago. Tokens, scales, components, semantic roles, written so a machine could hold them and a person could pick them up without a workshop. Brand mostly didn’t get the same treatment, because brand’s reader was always a person, and people fill in the gaps with what they know about the company.
Then we started handing brand to machines, and the gaps got filled with the model’s defaults instead.
You can spot the version with no rules instantly
Everything looks the same. It looks like every vibe-coded homepage made in one prompt.
Container city. Boxes inside boxes inside boxes, every section wrapped in a bordered rounded rectangle. Low-contrast text, bad spacing, no hierarchy. One accent colour, light or dark mode, no colour ramps.
That isn’t a model problem. It’s what you get when the thing doing the making has been handed examples of good instead of instructions for good.
What AI can and can’t do for a brand
AI can do brand, but not well without clear guidance, and bad branding at the start just means inconsistent, mis-messaged content at scale. There’s a bigger problem coming too. You can ask Claude or ChatGPT to build you a brand, and if lots of new businesses do exactly that, we’ll end up with more same-y brands than we had before AI, all built from the same training data and the same handful of frameworks.
It knows a lot of brand theory. What it can’t do yet is taste.
Where it works is when you take your strategy and put it into a system the AI can use. The thinking stays yours, and the AI becomes your strategist, assembler and first reviewer. Your job moves from production and policing to creative direction and strategy.
That’s the real prize. Brand stops being the queue and starts being the thing that helps.
What a rule looks like once a machine has to check it
At Dremio our guidelines said the usual things. This is the voice. This is how we want to sound. Active, not passive. All reasonable, and enough for a copywriter who already knows how to write.
Hand the same thing to a model and it does follow the rules. It writes actively. Then it brings in everything the guideline never thought to mention, because it never had to. Every general AI tell. Banned words. Non-standard ways of writing our own product name. British spellings where we needed American ones.
The guideline wasn’t wrong. It was written for a reader who turns up with defaults, and a copywriter’s defaults are not a model’s. The guidelines relied on a writer’s judgement about things like exclamation marks. A model needs that guidance made explicit.
So the fix isn’t a better sentence. It’s a check.
brand_lint reads the copy and runs it against banned phrases, discouraged terms, AI tell words and patterns, and British spellings where we need US English. Then the structural rules, like sentence construction and exclamation mark counts. Then it works out what’s fine and what needs changing, scores it and hands back a grade.
The grade is the part that matters, because it isn’t a gate that only says no. The model takes the score, rewrites what failed and runs again. When it passes, a person reviews it before it goes out.
Same on the design side. Colour tokens, spacing, typography, all checked against what actually got made. Which catches the more interesting failure. Output that passes because every colour is a brand colour, but leans far too heavily on one of them, or puts the brand colour somewhere it was never meant to go.
It passes the colour check, but using the right colours isn’t enough to make it feel like the brand.
Automated accessibility checks sit on top. They catch the repeatable failures a reviewer starts to miss once the volume goes up, like contrast and missing alt text, and the person reviewing still judges the rest.
What to systematise, and what to leave alone
This is where I’d push back on my own argument, because it’s the bit people get wrong in the other direction.
Make creative too much of a system and you lose the freedom that produces anything with taste in it. Systematise it completely and it’ll be perfectly consistent and not at all different, which puts you back at everything looking the same, just by a more expensive route.
Deciding where to draw that line takes judgement. Spacing, typography, component structure, the whole design system layer, should be locked down. Ad creative and campaign work stay freeform, because they need fresh ideas, and a rule set that’s too specific leaves no room when creativity is actually required.
Get that boundary wrong and you’ve automated your way to mediocrity. Hand brand to an agency without keeping the creative direction yourself and you get there too, as we found out twice.
Creativity plus AI is scale. AI on its own is volume, and those aren’t the same thing.
If you own a brand or a design system, here’s the Monday version
You’ve probably already done most of the hard part and don’t realise it.
You’ve got components in Figma, patterns in Storybook, a tone of voice doc, a positioning deck. It’s all written down somewhere. So how do you connect it to an AI? How do you tell it this is this component, and when to use it?
And the harder half, how do you give it the strategy and the thinking behind it? A book of components is fine for a designer, who knows how to use it. So is a tone of voice page, for a writer who already knows how to write. Give either to a model with no context and no idea why it should reach for that component in that situation, and it won’t use it correctly. It’ll produce something structurally valid and strategically meaningless.
Two questions worth sitting with. How would an AI actually read this? And what context does it need that a designer supplies from their own head without noticing?
Then try it on one thing. Pick a task your team does every week, like a sales deck. Write down what a good designer supplies from experience when they make it, the things your guidelines never say. Give that to the AI with the task, check what comes back, and fix the guidance wherever it failed.
Here’s what that looked like for us. One brief asked for a prospect overview deck, under ten slides, using content from our website. The agent built it, rendered every slide and looked at its own output. It found an overlapping subtitle and a white logo sitting invisible on a white card. Neither would show up in a check of the copy. Both showed up once the slides were rendered, and it fixed them before anyone saw the deck.
The grade tells you something failed. The rendered file tells you what to fix. That’s where content at scale stops being a threat and starts being useful.
I build Sidenote, which turns a brand into context that agents and people can both work from, over MCP and an API. It exists because of the argument above.
From the workCertinia refreshDremio