Case study Design infrastructure · AI-augmented · 2026

The design system that ships itself.

How the ZenDocs design system became a machine-readable AI skill: one encoded source of truth that any teammate, or any AI tool, builds on-system with. The 14-surface redesign and this portfolio are both outputs of it.

RoleUX / Design Lead, Ruby Labs
WhatDesign system encoded as an AI skill
StatusIn use org-wide at Ruby Labs
Proof14 redesigned surfaces + this site
1Source of truth, readable by people and machines
4Layers encoded: foundations, components, patterns, rules
14ZenDocs surfaces built with it, in days
2Products running on it, including this portfolio

Counts are verifiable from the artifacts on this site: the redesigned surfaces appear in the ZenDocs case study, and this page itself renders on the system's tokens. Org-wide adoption metrics are internal and not yet published.

In 30 seconds
  • The move: the design system stopped being documentation and became an executable skill: tokens, component specs, page patterns, and interaction rules encoded so AI tooling produces on-system output by default.
  • The proof: 14 interactive ZenDocs surfaces prototyped in days rather than weeks, and this portfolio, running on the same encoded system.
  • The honest part: a skill is only as good as its governance; what is measured, what is not yet, and what breaks without an owner.
01 · The problem

Design systems decay as documentation.

Every design system I have built or inherited had the same failure mode: the artifact was documentation, and documentation is passive. It waits to be read, it drifts from production, and every handoff is a retelling. The Part I audit of ZenDocs found exactly this: an off-token primary blue running in production was the highest-severity visual finding. The system existed; the surfaces had stopped listening to it.

AI tooling changes the failure mode. If the tools that generate interfaces can read the system directly, the system stops being a reference and becomes a constraint. The question was how to encode it so that output is on-system by default rather than by discipline.

02 · What the skill is

The system, written for two readers.

The skill is the ZenDocs design system condensed into a structured, machine-readable document with four layers: foundations (type scale, color tokens, radii, shadows, the glass recipe, spacing rails, motion), component specs (buttons, inputs, navbar, cards, modals, the TOC rail), page-level patterns (the tool funnel order, directory structure, form landings), and interaction rules that travel with everything built on it.

A short excerpt, verbatim:

### Glass (signature effect)
- Recipe: background: rgba(242,242,242,0.62) + backdrop-filter: blur(25px)
- Used for: nav dropdown panels, mobile menu sheet, modal backdrops

## 4. Interaction rules
- Every overlay closes on outside click and Escape
- Active/selected state = blue text/border + soft blue fill
- Buttons that aren't wired show a toast rather than doing nothing
- Copy tone: plain, benefit-led, sentence case; no exclamation marks

The last layer matters most. Encoding copy tone and interaction behavior next to color means voice and behavior cannot drift independently of the visuals, which is where most systems quietly die.

03 · How it runs

One loop: govern the source, generate the surfaces.

Design system governed, one owner Encoded skill tokens, specs, rules AI tooling + team generate on-system Surfaces funnels, prototypes, this portfolio AUDITS OF SHIPPED SURFACES FEED THE NEXT SYSTEM REVISION
FIG. 01 The loop. The system stays governed by a person; the skill makes it executable; every audit finding routes back into the source instead of into a one-off fix.
04 · Evidence

Two products you can inspect right now.

ZenDocs redesign

Fourteen interactive surfaces, from the home to the PDF editor, prototyped in code in days rather than weeks, every one on-token. The full walkthrough is in the ZenDocs case study.

This portfolio

The site you are reading runs on the same encoded system: same tokens, same interaction rules, same copy tone. The unwired-button toast rule from the skill file is live in the ELIAS Studio reconstruction.

The team loop

Two UX engineers on my team build design work in code against the skill; I direct and sign off their output, so review happens against the system, not against taste.

Token drift, closed

The audit finding that motivated the encoding (an off-token primary blue in production) is now structurally hard to reintroduce: tooling that reads the skill does not know the wrong blue exists.

Redesigned ZenDocs homepage generated on the encoded design system
FIG. 02 Output, not mockup: the redesigned ZenDocs home, generated on the encoded system and inspectable as a working prototype.
05 · Limits, honestly

What this is not, yet.

Adoption is qualitative for now. The skill is used across the org, but I have not yet published time-saved or consistency metrics with clean baselines. The next revision ships with instrumentation: surfaces audited per quarter, token-violation counts, time from brief to on-system prototype.

A skill without a governor rots like documentation does. Encoding does not remove the ownership problem, it concentrates it. The system stays healthy because sign-off still runs through a person; the skill just makes the approved decision the cheapest one to reproduce.

It constrains; it does not decide. The skill guarantees on-system output. It does not choose what to build or whether the funnel framing is right. That remains design judgment, which is the part I am actually hired for.

Next case study

ZenDocs: the system, applied at funnel scale