The UX Writing Agent

I'm a UX writer, not a developer. I built it by talking.

I've never shipped a Figma plugin, never set up a Postgres database, never written vanilla JavaScript for production. The plugin was built by talking — me describing what I wanted, Claude Code writing the code, both of us looking at the result, me reacting to what was off, Claude Code adjusting.

That's what eToro means when it says "be a Builder." Not the AI builds for you. Not you supervise the AI. It's conversational engineering: the human supplies taste, judgment, and the ability to know when something is wrong without being able to articulate why. The AI supplies the part that's actually hard for non-developers — writing the code, knowing where files live, executing the change without breaking adjacent things. Both halves are necessary; neither is sufficient.

• It grew by accretion, one annoyance at a time. The first version was a single sidebar with a "Loading…" spinner. After the third 90-second wait I asked for parallelization — a worker pool at concurrency eight, progressive rendering, a Stop button wired through an AbortController.
• The feedback infrastructure cycled through three dead paths before landing on the fourth. Power Automate (no Premium license), Cloudflare (eToro IT enforces SSO so I couldn't sign up), Vercel (Hobby tier is non-commercial). Finally landed on Render — Node + Postgres on the free tier.
• The mandatory-rules pass took one session. Claude Code pulled up the style guide JSON and presented every rule one at a time with Mandatory / Suggested buttons. About 30 minutes, 150 rules, 117 mandatory.
• The card redesign happened in 20 minutes. I said the styling looked weird. Claude Code wrote a design brief from first principles, built a side-by-side HTML preview with three states each and a written rationale, sent it back with "sign off before I touch live code." I opened it, said ship it.

What it does

Three tabs — each replacing a piece of the job a single UX writer can't do at company scale.

The editorial-control layer

Some style-guide rules are preferences. Others — Title Case on buttons, no exclamation marks in product UI, regulated financial terms — are firm. The plugin treats them differently.

• 117 mandatory rules replace Reject with Override (give reason). A textarea requires a written justification before the finding can be dismissed.
• 40 suggested rules keep the lightweight Accept / Reject pattern.
• Every override is logged — POSTed to a Node + Postgres feedback backend I shipped on Render. A private dashboard lets me pattern-match where rules get bypassed too often.

That's the part that makes this a content-architecture artifact, not just an AI toy. Editorial control survives the handoff to PMs — and produces data on what to change next.

Where it fits

The plugin is the first concrete piece of the AI-based content system I pitched to the Head of Solutions (~600-person R&D division) in February. The standards I wrote into the style guide are now enforceable in the place where PMs and designers actually work — rather than only at the post-hoc localization chokepoint. It sits earlier in the lifecycle, catching issues at design time before they ever reach a key/value sheet.

The journey from July 2025 to today goes from "here's a mess, let us know when you've fixed it" to a working AI agent that enforces my standards at design time, across the whole org, with editorial oversight and a measurable feedback loop. Built by one person who was told there would be no team.