Case study2022 → nowBuilt it, now lead it

Bajaj Design Library

I built Bajaj Finserv's design library on a framework I designed, then wrote the rules that let designers and AI read every component the same way.

v1 + v2Built by me, on my framework, taught in company-wide workshops
2 readersDesigners and AI, one set of rules
~600Components in the everyday library
204 · 902Components and variants in the library for CAI, Bajaj's AI platform

One component, two readers

The same header, as a designer sees it and as an AI reads it.

Live · try it
device
state
canvas
01

The problem

AI couldn't read the library. Designers couldn't search it.

Both came from the same gap: the file never said what anything was for.

A · The AI readerThe file misled it.

An AI reads names, properties and structure, and nothing else. A person reads past a sloppy name; a machine takes it at its word.

Frame 47

Says nothing. A default name has no purpose.

Device device-type

One idea, read as two. Drifting names look like two different things.

Navy/Navy-#002953

A colour with no job. It names the shade, never the use.

detached

Cut loose. No link back to the main component.

B · The designerFinding a part took longer than drawing one.

At about 600 components, people gave up looking and drew their own.

  • Duplicatesthe same card, built many times over
  • Driftnames and spacing that no longer matched
  • Low adoptiona library people worked around, not with

One fix for both: a library that explains itself, in words a designer and an AI read the same way.

02

Where the rules came from

The rules came from practice, a failing AI and interviews.

  1. 01 · PracticeYears of building in Figma

    How components, auto-layout and variables actually behave, and what breaks once a whole company builds on them.

  2. 02 · ExperimentAn AI that kept failing

    Before Figma's MCP (its bridge for AI) could describe much, I built a chain of agents to turn Figma into code. Every failure pointed at something the file left unsaid.

    FigJam diagram: an orchestrator routes a new component through validator, documenter, icon maker, coder and tester agents, each fed by guide documents.
    The Figma-to-code experiment. A lead agent passes each component from validator to documenter to coder to tester. The blue column is what they all read: the guides.
  3. 03 · ResearchAsking the people who use it

    Before rebuilding, I interviewed designers, product owners and stakeholders. They kept naming the same problems: duplicates, thin docs and drifting names. Those became rules.

03

The rules

Four rules do most of the work.

The library team builds from a guide. These four are the ones at work in the header at the top of this page, and they're why a designer and an AI read it the same way.

See them at work ↑

  1. Names say what a thing is for.No Frame 47. A name an AI can parse is a name a new designer can find.
  2. Figma and code share one name.A property in Figma is the prop in code, spelled the same.
  3. One device switch.A device variable with mobile and desktop modes; the screen sets it, every part inherits it. Mobile is built at 360px, the width most customers use.
  4. Every component states its job.When to use it, and when not to, in its own description.
04

Tools

A plugin for designers, a skill for AI.

Both search the descriptions rule 4 asks for, so a designer and an AI land on the same component.

For designers · Figma pluginComponent Scout

Designers describe what they need instead of guessing a name. Scout matches that against every component's description and shows the closest existing ones first. No AI involved.

For AI · Claude Code skillLibrary skill

The AI's way into the same library. It reads the live file and the same descriptions, and picks the component a screen needs before drawing anything new.

Component Scout in Figma. Describe what you want to build, and it shows what already exists before you draw it again.
05

The proof

Every CAI delivery is built from the library.

100%

of CAI deliveries are built from the library. None ships without its components.

CAI, Bajaj's AI-first platform, got these rules from day one, along with everything the first library taught me. One naming system for tokens (the named colours, sizes and spacing), a description on every component, and twelve properties or fewer. The library team built 204 components and 902 variants on it. The everyday library is now moving onto the same system.

Component audit of the CAI library: 204 components, 902 variants, split into atoms, molecules and organisms.
The component audit.

Journeys built entirely from library parts.

06

Also in this work

One vocabulary across the company

With the IT team I set up shared procedures and working routines so token, property and component names match everywhere: design, IT, data and product owners.

Partnership with IT
What else the interviews turned up

Rigid base components, no way to retire old ones, banners misused as containers, no feedback loop. Each got its own fix: flexible slots, a staged retirement, a banner framework, an internal satisfaction score.

From the interviews
Review gates

A seven-stage review, from content to live audit, that catches non-standard components before release.

Library process
One Figma licence for every team

The case for Figma Enterprise, made in workshops with Figma, so one library could reach every team.

Proposal
Bajaj Flux

Where the Figma-to-code experiment led: the header rebuilt as a React component with its own docs site.

React · Next.js
Figma variables + modesAuto-layoutComponent propertiesFigJamFigma REST API + MCPClaude Code agents + skillsFigma pluginReact / Next.js