Work / Callstack client, enterprise software team

Keeping coding agents on an enterprise design system

Shared web and mobile component library plus an AI toolkit that blocks off-system code. Screen prototyping went from about an hour to about ten minutes.

Organisation
Callstack client, enterprise software team
My role
Acting tech lead for the React and Storybook side and the AI toolkit; top committer on the repository
Period
Apr 2026 - Sep 2026

Context

An enterprise software team ships digital products on web and mobile. Callstack was brought in to give them one design system across both platforms and a reliable way to build screens with AI coding agents.

Problem

Teams had built their interfaces separately, and the existing design system covered only part of the products. Coding agents connected to Figma produced screens that looked like the design but used generic React and Tailwind instead of the design-system components and tokens. A reviewer looking at a screenshot could not see that, and nothing compared the rendered screen with the frame it came from.

Approach

  • Built a shared component library for web (React, Tailwind v4) and mobile (React Native, NativeWind), documented in Storybook, with Playwright visual-regression and axe-core accessibility checks on every pull request.
  • Packaged an AI toolkit: MCP servers that expose the component catalog, design tokens and live Figma data, plus Code Connect mappings from Figma properties to React props.
  • Added an ESLint rule and a pre-write hook that block imports the design system replaces, before the file is written.
  • Built a Figma pixel-diff MCP server so the agent compares its rendered screen with the design and retries up to three times.
  • Made one install set up the whole workflow for Claude Code and Cursor.
  • Brought Callstack React Native evals into the repository and added visual scoring, so agent changes are measured in a sandbox instead of judged by eye.

Key decisions

Block at write time, not at review
A screenshot review cannot see an off-system import. A lint rule and a pre-write hook stop it before it exists.
Record the Figma-to-props mapping explicitly
Compiled bindings tell the agent which React prop a Figma property means, so it never guesses what a label like "xl - mobile only" maps to.
Compare against a visual threshold with a retry cap
The agent gets a measurable stop condition instead of declaring itself done. We also documented where screenshot comparison misleads, so the number is not over-trusted.
Stay vendor-neutral
The loop is built from MCP servers, a pre-write check and a verification tool, so it runs with any agent that can call MCP tools. Only the hook wiring is agent-specific.

Outcomes

  • Mapped 60+ components with full Code Connect coverage.Source: Callstack public case study
  • Cut prototyping of a typical enterprise screen from about an hour to about ten minutes (about 6x faster).Source: Callstack public case study
  • Generated screens are checked against a visual-diff threshold below 2%.Source: Callstack public case study
  • One install sets up the component library and AI workflow; the initial rollout covered 30+ engineers.Source: Callstack public case study
  • Shipped 72 web and 60 native components across three packages.Source: Project history
  • Found and fixed a WCAG AA contrast failure in dark mode (4.48:1 raised to 4.69:1).Source: Project status report

The client is not named. Numbers marked "Callstack public case study" are published by Callstack with the client's approval.

Stack

  • React
  • React Native
  • NativeWind
  • Tailwind CSS v4
  • Storybook
  • Playwright
  • axe-core
  • MCP
  • Figma Code Connect
  • ESLint
  • TypeScript