Internal Tooling · AI-Assisted UX Writing, 2025

From Rules to Runtime: Building an AI-Powered UX Writing Checker in Figma

Our Voice & Tone guidelines lived in a document nobody had time to check copy against — so I built a Figma plugin that does it automatically, reading intent with Claude instead of just matching keywords.

Role

Solo — identified the gap, wrote the rules, built and shipped the plugin.

Scope

Two enterprise products, one continuously growing Figma library, 10+ stakeholders — one shared design system.

Stack

Figma Plugin API, Vercel proxy, Claude Sonnet 4.6.

Status

Live — checking real files across both products today.

The rule examples and screenshots below are the plugin's own output, run against representative frames built in the product's design system.
00

The problem

By the time this started, our design system spanned a large, constantly growing set of Figma files across two enterprise security products, touched by 10+ product owners and engineers, each reading UI copy a little differently.

Our Voice & Tone guidelines were solid on paper, but they lived in a document: nobody had time to check every button, toast and error message against them by hand, so inconsistency only surfaced after it had already shipped.

01

My approach

There wasn't an existing tool for this — enforcing a specific internal writing standard, semantically, inside a design file, isn't a problem a spell-checker solves.

So I split Voice & Tone into two layers: Claude API for the judgment a linter can't make — does this line blame the user? — and structured JSON for feedback a designer can act on inside Figma, not in a report nobody opens.

02

Architecture

The plugin reads text nodes straight off the Figma canvas, sends them through a small proxy for the actual model call, and writes structured feedback back into its own panel — three moving pieces, one continuous loop.

Figma Canvas code.js Figma plugin Vercel Proxy (CORS) Claude API Sonnet 4.6 text nodes fetch request API call 🔒 API key → Vercel env variable (not in code) 🛡️ Plugin token → blocks unauthorized access JSON feedback → ui.html panel
→Request path: canvas text to model, straight through.
┄Return path and the two points where access is actually gated.
03

The rules system

A growing set of rules came out of turning Voice & Tone from prose into something a plugin could actually test. A few examples:

Voice & Tone doc
"Don't blame the user."
Plugin rule
Flags user-blaming phrasing in errors and validation copy.
"You entered the wrong password." "Your password is incorrect."
Voice & Tone doc
"Be specific about errors."
Plugin rule
Flags vague error copy with no stated cause or next step.
"Something went wrong." "Error occurred."
Voice & Tone doc
"No punctuation on buttons."
Plugin rule
Flags trailing punctuation on button and action labels.
"Save changes." "Submit."
Voice & Tone doc
"Keep alert tone calm."
Plugin rule
Flags exclamation marks and alarmist phrasing in system alerts.
"Warning!" "Suspicious activity!"
04

Plugin in action

Run against representative frames built in the product's own design system. Click any screen to expand it.

05

Outcomes

840
text nodes scanned in a single pass, in the run shown above
118
issues flagged in that same pass, before anyone reviewed by hand

Works on any Figma file, any frame — not just the ones I made.

Voice & Tone documentation itself got formalized as a byproduct of writing the rules.

Production-ready security: API key in Vercel environment variables, plugin token auth, private GitHub repo.

06

What I learned

01

Where AI actually helps, and where it doesn't. Semantic judgment — tone, blame, clarity — needs a model; deterministic checks — casing, punctuation — don't. Treating them differently kept the plugin fast and cheap to run.

02

A CORS-safe proxy architecture, built from the start. A Figma plugin can't call an external API directly, so the Vercel proxy became the real security boundary — not an afterthought bolted on later.

03

Turning documentation into executable logic. The hard part was never the API call — it was making "don't blame the user" into something a machine could actually test.