A Case Study · 2025 to 2026

I built the front door for a design system serving four products.

A single source of truth for the Intuit Design System, by AI, for AI and humans. Fifteen months from sprint to shipped.

Role Design Manager & Design IC · IDS
Company Intuit
Shipped March 30, 2026
Stack Claude, Cursor, Figma MCP, Docusaurus
The Problem

Information lived everywhere.

Docs in one repo, components in another, tokens in a Figma library, contribution rules in Coda. The starting point depended on who you asked. Around 70% of designers were dissatisfied with our existing developer documentation.

The Reframe

The real user wasn't only a human. It was an LLM, increasingly.

That shift changed everything downstream. How we wrote, how we structured, how we stored content. The bet wasn't a better doc site. The bet was authoring for both audiences from one source. Markdown, in one GitHub repo.

Design sprint whiteboard in Mountain View
From disconnected sources to a single connected one
The Approach

One designer. Three new tools.

With zero engineers, $0 budget, and an on-prem requirement, the project hit a hard stop in November 2025. Then Claude arrived inside Intuit, and the economics of the project changed rapidly. A durable POC, built one prompt at a time.

01 · Reasoning

Claude Opus

Architecture, content modelling, accessibility review, and stakeholder writing.

02 · Authoring

Cursor

Inline edits, refactors, repo orientation, and paired writing of every file.

03 · Bridge

Figma MCP

Lifted last year's designs into the new build without redrawing.

The Product

Six surfaces. One source.

A coordinated launch. Visual refresh, brand foundations, migration paths, and refreshed docs using MDX. Built using our own IDS components and patterns

IDS Front Door product preview
Reflections

What I learned as both manager & IC.

01

Pair growth with strengths

Amrit shipped visual rigor I wouldn't have. The hardest part was resisting the urge to take the repo back.

02

Resilience is key

Every "no" wasn't a setback. It was a chance to sharpen our strategy.

03

Lower the cost, not the standards

AI didn't replace the engineers we didn't have. It made one designer cheap enough to bring eng a POC worth landing.

04

Pick the fight at architecture

The markdown in one repo bet held every time a UI decision wobbled. Architecture decisions compound.

Metrics · Outcomes

What shipped on March 30.

Source of truth
1

Where there were 9+ scattered surfaces before.

Time saved
60%

reduction in slack support due to self-serve measures.

Useage
250 ~

DAU. Engagement was through the roof.

Vendor spend
$0

Built and maintained using existing resources: Our design team, Claude & Cursor

Outcome 01

One unified home for IDS.

Documentation, tooling, components, tokens, and contribution rules under one URL. Discoverable for designers and for AI.

Outcome 02

An LLM ready design system.

Markdown source means models can index, retrieve, and reason about IDS the same way humans do.

Outcome 03

Engineering wired it in.

Landed on the paved road inside AppFabric. Routing, observability, SSO, and CI/CD already in place.