Building the operating model
From AI prototype lab to production platform.
Nick Johnson and I created a place for rapid product experiments. I led the design and frontend system that made quality part of the build environment—not a review after it.
Britton · design + frontend foundation Nick · backend + CI/CD
02 The field test
Before choosing a frontend foundation, I made the AI prove it.
The frontend foundation would become part of every builder’s working context. The right choice could improve what AI produced before I ever stepped in to review it.
So I tested the leverage instead of choosing on familiarity or taste. I compared multiple foundations across five distinct product problems—not five repetitions of one brief.
One foundation kept taking the work further.
Across five distinct product problems, its output showed a consistent directional advantage: more developed, focused, and cohesive concepts with less prompting.
Comparison method I compared multiple component foundations—including Shadcn and Mantine—while keeping the feature problem, model, tools, and core instructions comparable within each test.
Evidence standard Qualitative evidence used to choose a foundation—not a scored benchmark or a claim of causality.
03 The environment
I kept the foundation AI understood—and rebuilt the visual language around us.
The selected system gave the model a component language it handled well. I then customized its tokens, typography, spacing, states, and component styling so we kept that leverage without inheriting the same look as everyone else using the base.
Nick and I turned that tailored foundation into a shared building environment. I led design and frontend; Nick led most backend and CI/CD. Together, we created specialized skills and agents so each job began with relevant context. I also created a global API reference and standards document so agents could discover available services and use the expected input and output shapes.
AI-readable structure, a BENlabs-specific visual language, and explicit API contracts gave agents enough context to build real product work without guessing at services or data shapes. Automated multi-agent PR reviews then surfaced issues from complementary quality perspectives, improving code quality while shortening the review loop.
Workflow shift
Working software became the design material.
With real behavior and real data in front of me, I could complete the design pass directly in code. Automated PR reviews caught issues earlier; engineers still owned approval and release. The result was a faster loop without lowering the quality bar.
04 The production proof
A test environment earned the right to ship.
The portal began as a low-risk place to make ideas tangible. It ultimately became the company’s main environment for building new customer-facing features.
The adoption signal was not AI-tool usage. It was trusting the environment with production work.