ISHA TODKAR / OCTOBER 2026 / 6 MIN READ
The museum
Every brand identity project I’ve worked on ends the same way. Months of thinking go into a brand guideline, the guideline goes to an in-house team, and I leave the room.
Even a thorough guideline, every rule explained, every do paired with a don’t, the rejected explorations turned into examples of what not to do, doesn’t stop one of two things from happening. A year later, the team has either treated it as a museum, repeating what exists until the brand looks exactly as it did on launch day, only tired. Or they’ve treated it as a cage, pushed past it, and the brand has drifted into something nobody intended.
Same document. Both teams doing their jobs. For a long time I couldn’t say what was missing.
Where the idea started
At the end of 2024 I wrote about the pull between consistency and flexibility in a brand system. My answer then was one strong element at the centre of an identity: the thing that stays recognisable while everything around it moves.
I still think that's right. But the strongest element in a system is also the one that depends most on judgment. I know when it should hold and when it's allowed to break, because I made hundreds of small calls while building it. The team inheriting it gets the rules. They don't get those calls.
So the question became: what if the thinking behind a brand could travel with the work, instead of leaving with the designer?
Writing the thinking down
I started by trying to write it down. For every element of a brand, four answers:
Why. The reasoning behind the decision.
Sacred. The principle that has to survive when the execution changes.
Flexible. What can move, and how far.
How I would extend it. The directions I could already see while building it but never got to explore.
Brand OS, Knowledge. The slash, written out as Why, Sacred, Flexible and How I’d extend it, with each rule shown as a do and a don’t, and whether the system can check it.
Sacred versus Flexible was the useful split. A size preset protects legibility. A colour ratio protects hierarchy. The numbers express intentions; they aren't the intentions. Written this way, it felt less like a rulebook and more like a handover note from one designer to another.
What I thought a tool could do
A note only helps if someone reads it at the right moment. So I built a browser prototype called Brand OS, to put the reasoning next to the work while someone designs.
Brand OS has five parts, each built around a moment where a handover usually fails. Knowledge holds the brand’s thinking, element by element. Brief reads a new brief against that thinking before anyone opens a canvas. Explore turns it into directions that take different risks. Canvas is where the work happens, with the reasoning beside every element and a running list of what the system can and can’t check. Decisions keeps every exception and the reason it was allowed.
It's an experiment, not a product. Not yet, anyway. The AI in it is a rule engine with a voice, working from rules I wrote. The brand inside it is Tilt, a fictional quick delivery fashion marketplace with a complete identity behind it. Its signature is a slash: a cut between the plan you had and the plan you now have, which is the whole promise of a brand that gets you dressed in an hour.
The prototype follows Maya, a constructed in-house marketing designer, through a campaign called Plans Changed.
Brand OS, Explore. Three territories: type-led, image-led, and one where the slash becomes the composition and the system notes that it would need a reason.
In the canvas, every element carries its reasoning, and a Checks tab lists what the system can verify and what it can't.
Brand OS, Canvas. The headline is selected; the panel shows its properties and the reasoning for typography underneath.
My bet was that if the reasoning was always in view, someone like Maya would have the context to make calls of her own, including ones I never anticipated. Not better work than a designer. Better decisions when the rules run out.
What worked
Consistency got easy. Off-palette colours weren't available, casing applied itself, partner colours stayed accents. The better surprise was rules that explained themselves: showing which hierarchy a headline breaks is different from saying it's too big. I didn't test that with anyone, but the floor held. Everything the system approved was correct, by every rule I'd written.
Correct turned out to be the problem.
Where it broke
I wrote three briefs and ran each through it: a last-minute Friday plan, the monsoon arriving in Mumbai, a payday weekend. The system varied the work, and could even propose a departure. What it couldn't do was judge whether a brief actually called for one. Every option came from the same approved vocabulary, every risk rounded off before it reached the frame.
Three briefs, what the system approved. Varied layouts, all drawn from the same safe vocabulary.
Nothing was wrong with any of them. Katie Dill, Stripe’s head of design, put it well at Lenny’s summit this year, borrowing from Christopher Alexander: a system can satisfy every rule and still be dead.
Early on I'd built an on-brand score, and it was the first thing I removed. The dullest asset scored 100%. The one I liked, where the slash became the composition, would have been marked down. At Config 2026, Nicole Martinez described this through J Dilla, who pulled beats off the quantized grid so they'd feel human. My system quantized everything.
The guideline behind Tilt was complete. I think the failure was in turning it into things a system could check, and in treating compliance as success.
Trying to fix it
A rule, the principle behind it, and knowing when that principle needs reconsidering are three different things. Scaled up, Tilt's slash stops being an element and becomes the composition. The rule breaks; the intention holds. That's not the same as enlarging it because it looks nice, even though the visible action is identical.
So I rebuilt that part. It's a proposal; nobody but me has used it. Say Maya wants the photo to run full-bleed, past the mask. The system doesn't refuse. It names the rule, says what it protects, and asks whether that reason still holds. She answers in words and sets a scope: this asset, this campaign, or a change to the rule.
Brand OS, Canvas. The photo runs full-bleed and the mask becomes the composition. The panel names the rule, what it protects, and asks for the reason.
Off-palette colours and banned phrases stay locked; changing those means changing the brand system itself. Other rules can be challenged when they protect a principle rather than a fixed requirement, and every exception is recorded for the person responsible for the brand to rule on. The next designer inherits the decision, not just the rule.
Brand OS, Decisions. Each exception with its rule, the reason given, its scope, and a decision record the next designer inherits.
The flow makes judgment visible and traceable. It doesn't transfer it. Maya could write a weak reason for a strong idea, or a convincing one for a weak idea. Someone still has to look at the work and decide.
Where I’ve landed, for now
I went in believing the thinking behind a brand could be translated into a tool. I’m coming out believing most of it can, and that the rest isn’t missing from the tool so much as waiting for a person.
A real model would do far more than my rule engine. It could read a brief without being told which principles matter, argue for a deliberate rule-break, and learn from the decision log which departures get approved. What it shouldn't do is own the ruling. Approving an exception decides what the brand is becoming, and someone has to answer for that next season. The more capable the system gets, the more the work concentrates in that moment of judgment.
Two things I haven't tested yet. An approved exception isn't a successful one, so the log should eventually record what happened next, not just why it was allowed. And a brand whose output is mostly routine, product screens, statements, notifications, spends more of its life on the floor, where a system like this would carry more of the load than it did for Tilt.
So I don't think the tool is a dead end. I had its job wrong. I built it to stand in for me. The version worth building takes the resizing and reformatting, holds the floor, carries the reasoning, learns from every decision, and hands the hard calls to whoever is holding the brand now. They can build that judgment too, if they're given the reasoning, the past calls and the authority to make new ones.