The Pattern Layer
Everyone is creative. Design is about provoking that creativity and giving it form.
I don’t think traditional roles are going away for most orgs, but I do think some walls can come down. The walls between roles were never about expertise. Designers know things product managers don’t, and that always stays true. The walls were about who was allowed to make things. At some point, we all decided we needed to separate doing the work from thinking about the work.
Design docs are finally having their day. Coding tools like Claude Code are more than happy to read specs for design system components and build a layout using them. However, three people using the same prompt may get three different layouts. The outcome will technically comply with the component library, likely using reasonable patterns and common conventions. Still, with many patterns and conventions to choose from, the model may choose one that breaks convention. It’s not wrong; it’s just wrong in context.
Designers make these kinds of decisions using best practices, intuition, and knowledge gained from collaborating with their partners day in and day out.
What’s missing is the composition layer. Sitting on top of the component library is a set of composition specs — layout rules, decision trees, spacing values, class names — written as explicit, structured, machine-parseable files. The organizing frame is Brad Frost’s Atomic Design, working up from components to the patterns they form.
These specs live in a directory alongside a skill that loads a pattern index, identifies the relevant specs, reads them fully, and then builds. A cross-page consistency check also runs in the self-review step, catching things like a status badge using outline on one page and default on another.
As a result, everything built in this codebase works from the same source of truth. So PMs running the latest Claude model will get the same result as a designer running a GPT model in Figma Make.
It’s not that the composition layer was never considered. Designers will happily sit in a room for days deciding the right pattern for an edge case. The gap is carrying those decisions forward, adding them to a collective “brain” rather than the individual memory of the group. The goal is to make them searchable and readable for models so the next person who comes along doesn’t have to rehash that days-long conversation only to come out with a different answer.
We’re all learning out loud how to operate in the AI-first world. I’ve found the composition layer has shifted design critiques away from pattern policing. Necessary work, but a poor use of the most expensive recurring hour on a team’s calendar. Now that work happens before the artifact reaches the room, because the conventions are enforced by the system generating it. What's left is the conversation critique was always supposed to be: whether this is the right solution, whether we understand the problem, what we're trading away. Less policing, more judgment.
The second outcome is more subtle. Writing composition rules precise enough for a machine to follow surfaced ambiguities that had been sitting in the design system for a while. Designers knew how to work around them quietly. We had rules that contradicted each other, and conventions nobody could actually articulate. Documenting for a reader who can’t infer is the most thorough design system audit I’ve ever run.
I strive to build things that get people excited about exploring their creativity and impact, then make it easy to tell that story. That’s how the best idea wins.