DESIGN SYSTEM / R&D
CHANDRA GARRE · SELECTED WORK
I gave four AI agents the same prompt. The two with no system agreed on 1 colour. The two with Weave agreed on 21. Then I asked one to audit my docs. It found a feature I’d promised and never built.
THE WORKING SYSTEM
Open Storybook
ROLE
Product Designer, solo
TYPE
Design system R&D
TIMELINE
Sep - Oct 2026
STATUS
35 components, 8 known gaps open
TOOLS
Figma, Storybook, React + MUI, Tailwind, Axe-core
At work I maintain enterprise design systems that ship to real users. You don’t test new token architecture on those. Weave is where I test it.

Two runs of the same prompt picked the same arbitrary indigo and the same product name. I chose neither.
The model isn’t the problem. It has nothing to build from, so it decides for itself, then decides again next session. Then someone rebuilds the screen by hand before it can ship.

NO SYSTEM, RUN 1

NO SYSTEM, RUN 2
Left: an illustration I made of a screen where nothing is decided for it. Right: the same screen built with Weave. The real, unedited agent runs are in Results.
Before Weave
After Weave
Rebuild each component by eye, from a screenshot.
Open a component, set its props with live controls, and use it.
Decide in Figma, hand over a file, and find out later what was built.
Decide in Figma, then check the real component in Storybook.
Guess, and guess differently next session.
Read the rules first, then build with what exists.
Figma
Storybook 10
Vite
React + MUI
Tailwind 4
Axe-core
OKLCH
WCAG 2.1 AA
Figma is still where I decide things. It isn't where they become true.
Run for run, the no-system screens looked fine. They just agreed on almost nothing. Two runs each is a small test, but the gap isn’t subtle.
Nothing on hand fixes it. Figma’s MCP server gives layout, not rules. A Storybook is built for people who click. A prompt resets every session.
Open a component, set variant, tone and size with live controls, and use it as you see it. The prop names in the table are the ones you type, and the ones an agent reads.
This is the real Button page, not a recording. Change a control and the button above it changes.
AGENTS.md loads before the first line of code. It tells an agent which tokens to use, what’s off limits, and what to do when something is missing.
Contrast lives in the tokens. Teal-700 carries white text at 6.27:1. Teal-600 is 4.21:1 and fails, so the rule is encoded, not remembered.
Three token layers: raw values, values named by job, values for one component. Components never touch a raw value, so dark mode re-points one layer and everything follows.
Every decision was made in Figma first. Then I rebuilt the file so nothing was implied, and encoded it here.
Built on MUI, not themed from it.
MUI handles focus, keyboard and menus. Weave owns the tokens, the rules and the docs.
Scoped the gate instead of making it strict.
The strict version failed 173 stories, because a lone component has no main landmark. A gate that fails everything gets switched off.
Held the input border, for now.
Every input border sits at 1.75:1. WCAG asks for 3:1. The fix is one token, but it changes every form at once, so I’m checking each one first.
CURRENT, 1.75:1
Email address
WCAG MINIMUM, 3:1
Email address
Two runs, no system
Two runs, with Weave
One dot per distinct value across the pair. Filled means both runs used it. Measured in the browser with the same script on all four screens. Spacing counts padding and gap.
NO SYSTEM, RUN 1

NO SYSTEM, RUN 2

WEAVE, RUN 1

WEAVE, RUN 2

Four runs, one prompt. Layout varies in both pairs because the brief was open. In the Weave pair, almost nothing else does.
“Both controls meet the 44px hit area even though they are drawn smaller.”
FAILURES REPORTED
IMAGINARY
The auditor made the same mistake twice. It checked see-through dark mode colours against a white background and reported contrast failures that didn’t exist. It caught the error, logged it and explained the cause. Next audit, same mistake.
Knowing about a mistake doesn’t stop you repeating it. Turns out that applies to machines about as well as it applies to me.
Agents built a working dashboard with nothing but Weave, and I logged every place it fell short.
Hardcoded widths. Card was fixed at 600px and ListItem at 260px. Both now fill their parent.
Missing pieces. No calendar and no stat tile. Both builds made their own, and they came out different.
Of the 8 still open, six are docs that describe a slightly different system.
01
Drift hides in the docs.
I documented a rule as met and never built it. No review process I’ve worked in would have caught that.
02
Scoping beats strictness.
A gate that blocks everything gets switched off.
03
Automated findings need checking, not trust.
I read every finding against the source before accepting it.
Next: close the 8 open gaps, decide the input border, and hand AGENTS.md to people who didn’t write it.




