Skip to content

Project Blueprint

Luxury Presence · LUX design system

Turning a legacy migration nobody could size into a measured plan, and giving the design system a way to work with the rest of the company.

Role
Design system owner
Duration
August to November 2026
Skills
Strategy, Roadmapping, Measurement, Governance
Tools
Linear, Notion, GitHub, Claude Code, Slack workflows
larger than the original ticket described
17×
legacy usage in the dashboard so far
−12%
legacy buttons in the dashboard
−46%
office hours held, plus 13 weekly updates
23

The problem

The product dashboard was built on a legacy component library. Every import of it is a place where the product does not get design system updates, theming or accessibility fixes. Retiring it had been on the board for months, but nobody knew how big the job was. The estimate on the board was never reproducible, and neither was the first audit.

The system also had no regular voice. Requests arrived as pull requests or Slack messages, tickets lived in three tools, and reviews were slow enough that at least one engineer merged without one.

First, a number we can trust

I fixed the method before reporting anything: named symbols imported from the legacy library in the dashboard, tests excluded, verified against a specific commit. The same method, measured again later, gives a trend instead of a guess.

  • Board estimate3,600 · never verified
  • Verified baseline2,708 · August
  • Same method2,392 · November
The board estimate, the verified baseline, and the same measure two months later.

The audit changed the plan more than the number did.

  • The job was about 17 times larger than the ticket described.
  • Six symbols, mostly layout and text primitives, were 55% of it.
  • There were 36 screens to migrate, not 19.
  • 51 imports were dead code. Deleting them left the production bundle byte identical.
  • Most of it was owned by product teams, not by the design system.
  • 60%Mechanical, for a codemod or an agent
  • 32%Needs an engineer
  • 8%Needs design first
How the work splits. Most of it does not need a person at all.

One launch, two workstreams

With real numbers, I reframed the cycle as one launch with a name: Project Blueprint, the foundation the design system runs on afterwards. It had two workstreams, clear the legacy and establish governance, and a written definition of done instead of a list of tickets.

The charter changed too. Most front end code at the company is now written by people working through agents, and agents only produce correct output cheaply against a consistent codebase. So the system's primary consumer is AI assisted development, and people are the second. That one sentence explains why the cleanup matters: every legacy import is a place where an agent learns the wrong pattern.

The plan was just as clear about what we would not do: no full Figma parity, no opening contribution wide before the checks existed, and no mobile, which has its own system.

What moved

BeforeAfter
Legacy imports in the dashboard2,7482,423 (−12%)
Files importing it636570
Legacy buttons15886 (−46%)
Screens with no legacy1316

The rule that worked best was to migrate what we were about to change anyway. To change the button color, we migrated the buttons first. The dashboard screen itself came off the legacy library in November.

How the system works with the company

  • A weekly update, under five minutes, crediting whoever contributed that week. Thirteen so far.
  • Office hours, twice a week, with notes posted after. Several decisions in the system trace back to them.
  • One way to ask: a Slack workflow for bugs, proposals and reviews, triaged daily.
  • A contribution path, with published guidelines, a review time written into the repo, and explicit code ownership per package.
  • A baseline survey of designers and engineers. Seven of ten said the system got better this year, none said worse, and every designer admitted to working around it, which is the next problem.

Reflection

The most useful thing I did was refuse to report a number I could not reproduce. A verified baseline turned a vague cleanup into a plan with a definition of done, made it obvious which work an agent could take, and gave every later update the same yardstick. It is still in progress, and the total leaks: other work keeps adding new legacy imports while we remove old ones. Measuring that honestly is part of the job.