Case study · In production across two products
Design OS: Scaling design quality without scaling headcount
Two designers support a 200-person technology organization. That math only works with infrastructure, so I built it: standards written as machine-readable skills, plus an AI design agent that reviews craft before merge. It runs across two products today, with developers shipping against it daily.
My role & the room
Role: Creator and owner. Strategy, system design, agent development (built with Claude Code), rollout and adoption.
Partners: engineering leads across two product teams, with CTO and head-of-engineering sponsorship
Timeline: built across 2025 and 2026; the design agent shipped spring 2026 and is iterating now
2 designers
supporting a 200-person technology organization, flat for three years
82%
icon consistency on the new set, eight months post-migration, measured by an agent-run audit
The brief vs. the problem
There was no brief. The problem was math: a growing product surface, a 200-person technology organization, and a design team of two that hasn’t grown in three years. Every quality decision routed through one person. Me. Documentation alone wasn’t the answer, because documentation is where standards go to be ignored. The standards had to live where the work happens: in the repo, in the review, in the tools engineers already use.
What I built
Four skills, in GitHub, in production. Visual and interaction standards, an accessibility guide, component usage standards (MUI v6 plus our custom library), and UX writing standards. Authored as machine-readable skill files so developers and AI coding tools consume the same source of truth — helping the team move faster while keeping up with standards, and improving them as we go.
An AI design agent that reviews craft before merge. A deterministic pre-filter plus a multi-dimension judge covering layout, typography, color, component usage, and accessibility, gated by an aesthetic brief. It assesses work against both design standards and front-end code specs, and attaches recommended fixes. The initial version is in the repo for developers to pull today. In a validation run against 34 production files, it surfaced accessibility issues, layout and typography corrections, and design variations worth shipping.
Adoption, not just artifacts. I built locally for the first few weeks, then took it to engineering the way engineers take work to each other: opened a PR with the engineering leads and refined the skills through their review. Later, I demoed working software to our CTO and head of engineering, not documentation. Green-lit in the room. Engineering leadership publicly framed it as step one of a new design and engineering workflow. Two product teams onboarded.
[IMAGE SLOT]
Skill file structure
[IMAGE SLOT]
Agent review output
[IMAGE SLOT]
Before-and-after of a flagged screen
In a developer’s review
The clearest sign it works is where it shows up: inside PR review. A developer opens a pull request and calls the skills to validate their own work — UX writing, accessibility, component usage, styling — before a designer ever sees it. The accessibility skill gets the heaviest, most consistent use, catching issues at review that used to ship.
That’s the shift. Design quality stopped being a gate one person holds and became a check developers run themselves.
Business tie
Design headcount has been flat at two for three years while the product surface grew. Supporting a 200-person technology organization at this quality bar would normally mean hiring. Instead, standards apply themselves, review is no longer bottlenecked on one person, and quality debt gets found by an agent instead of by customers. Eight months after we migrated icon libraries, an agent-run audit showed 82 percent consistency on the new set and converted the remaining debt into backlog tickets.
Next
Playwright integration is underway, adding visual regression and automated accessibility checks to the same gate. And the agent layer is extending beyond review: a research agent, built with Claude Code, synthesizes customer feedback into strategic recommendations.
← Previous