Brand OS
Creative Director, designer, and sole developer on Brand OS, an internal AI tool that turns a written brief into on-brand Strategy assets in seconds. Built as a technical proof of concept to test a question I kept running into as brand custodian: whether the brand system could be encoded tightly enough that anyone in the company could produce work on system, without a designer in the loop.
Gallery
Owning a brand system means watching it get diluted at the edges. Guidelines get read once. Templates get stretched. The gap between what the system specifies and what actually ships is where brand equity leaks out, and it widens with every new team, region, and deadline.
Brand OS was my attempt to close that gap with software instead of policing.
The premise
Brand guidelines are a document that asks people to make correct decisions. A tool can make those decisions structurally impossible to get wrong.
So the brand rules are not suggestions in Brand OS, they are constraints in code. Inter only, at weight 500 for headlines, never bold. The orange is #FA660F and it is the only accent. Orange type never sits on an orange field, so the orange theme flips the logotype and headline to black automatically. Over photography, text goes white with a scrim, and if the image is bright the whole stack inverts to dark for legibility. There is no colour picker. There is no font menu. You cannot produce an off-brand asset because the tool has no way to represent one.
What it makes
You describe what you need in plain language and it returns a finished, editable design.
Five social formats: square, landscape, wide, story, and a 4:1 banner. Then A4 documents, which run as a different class of output entirely: one-pagers, case studies, newsletters, reports, and a digest format. The document variants are full multi-section layouts with mastheads, stat rows, pull quotes, use cases, and footers, and they export as press-ready PDF rather than image.
Four themes across all of it. White, dark, orange, and a full-bleed image mode. Three logos, with the Strategy mark, the Mosaic product mark, and the World 2027 conference mark, each swapping to its white variant automatically when the background calls for it.
The layout engine
This is the part I care most about.
Every position and size in Brand OS derives from the canvas dimensions. Nothing is hand-tuned per format. The margin is 5.5% of the shorter edge. The type unit is the geometric mean of width and height, so type reads at consistent optical weight whether the canvas is a 9:16 story or a 4:1 banner. The scale runs on the golden ratio, with the subhead and CTA sitting at the headline divided by phi squared. Headlines anchor to the upper rule-of-thirds line.
The rule I held hardest: headline length changes wrapping, never size. Most template systems shrink type to fit, which is how a system degrades into mush. If a headline is too long here, it wraps, and if it still does not fit, the copy is wrong and the tool should not hide that.
The practical result is that adding a new format is a ratio and nothing else. The layout falls out of the formula.
The brief, and four ways to answer it
The interesting design problem was not generation, it was the first screen.
A blank prompt is a bad brief in every case. So the entry point asks about the project, offers an enhance pass that sharpens a rough sentence into a usable brief, and carries a set of pre-written starting points for people who would rather edit than write.
From one brief the tool returns four distinct directions, varied by layout archetype, theme, and copy angle, rendered as live scaled frames rather than thumbnails. You choose one and it becomes the working canvas. The other three stay reachable. This mattered more than any single output quality improvement, because it moves the user from judging one answer to choosing between four, which is the decision designers are actually good at making.
Full-quality image generation only fires after selection, which keeps the cost of exploring near zero.
Built as a single file
The whole application is one HTML file. No framework, no build step, no dependencies beyond three CDN scripts. State goes into one object, layout is a pure function of that state, and rendering is a redraw.
That was a deliberate constraint rather than a shortcut. A proof of concept earns its keep by being easy to hand over, easy to host, and impossible to bit-rot through a toolchain nobody maintains. It runs on a static host with a thin serverless layer for the model calls, Claude for brief interpretation and copy, a hosted image model for photography.
The same layout function drives the screen and the export, so a PNG at 3x is the design, not an approximation of it.
Working as one
I designed and built this alone, which is unusual for me and was the point. Directing an agency or an in-house team means specifying intent and reviewing output. Building it meant every brand decision had to survive contact with an actual implementation, and several did not survive in their original form. Rules I had written into the guidelines turned out to be ambiguous the moment they had to be executed by something that could not infer taste.
The tool improved the guidelines as much as the guidelines built the tool.
Where it is now
Brand OS runs as a working proof of concept inside Strategy. It is not a shipped product and it was never scoped as one. Its job was to prove that the brand system is coherent enough to be executable, and to give the business a concrete thing to react to rather than a deck about AI tooling.
That question came back answered. A brand system that can be run as software is a system that has no soft edges left, and the parts that resisted encoding turned out to be exactly the parts that needed sharpening.