Copy as a governed system, not a queue of tickets.
Good UX writing cleans up after design. Great UX writing shapes it before production.
I build the frameworks, pattern libraries, and tooling that put it there — voice consistent and dark patterns out, across iOS, Android, and web.
Copy governance systemA closed loop from Figma to a living canon — two tools with a self-updating source of truth between them.
The problem
The product ships copy across three platforms from one Figma source that doubles as dev handoff. At ~1,600 strings, three things kept breaking:
- Drift. The same string worded differently across platforms; regressions reintroduced in new versions.
- Canon rot. The voice & tone guide lived in Notion as a doc nobody re-read; review decisions never made it back in.
- Bottleneck. Every “is this on-brand?” routed through one writer. Copy QA didn’t scale past one person.
What I built
copy-parityNode.js · Figma REST API
- Pulls every text node, pairs equivalents across platforms, lints each against
rules.yaml— the copy canon as code, where adding a rule means one regex in a config file. - Posts findings back onto Figma frames as node-anchored comments. Read + comment only — never mutates the canvas.
- Catches the machine-checkable class across 1,600+ strings in one pass: casing, broken tokens, drift, banned constructions.
ux-copy-auditAI skill · Notion via MCP
- Anyone — PM, designer, writer — drops a screenshot or Figma frame. No briefing; the skill carries the context.
- Reads the live authority stack from Notion: the UX Copy Patterns DB (
Status = Active), tone-of-voice pillars, naming rules — then judges the screen against canon. - Returns a before/after with rationale. Approved calls graduate back into the Patterns DB and land in Field Notes with per-writer attribution.
Impact
Next: merge the deterministic and judgment halves into one LLM-in-the-loop linter.
pair + lint emit| Canon | iOS | Android | Web | Status |
|---|---|---|---|---|
| Not now | Not now | Not now | Not now | identical |
| See psychics | See psychic | See psychics | See psychics | drift |
| Chat longer and save | Chat longer & Save | Chat longer and save | Chat longer and save | style |
| Recommended for you | Recomended for You | Recomended for You | — missing | typo · gap |
| [Name] | [Name] | [ExpertName] | [Name] | token |
Caught as a pattern, fixed once
The payoff, in one number — an unverifiable “most accurate” accuracy claim had quietly shipped across the product:
One claim, shipped to 81 screens. Plain text-similarity linting misses this class — it isn’t visible “drift,” it’s a canon issue. The system surfaced every instance; one ruling fixed them all.
What was live in productionBefore the audit system existed — now caught automatically. The pile the linter clears so I can spend judgment on the hard calls.
Trust by designResonance without manipulation — in a product built on spiritual hope, for people who arrive at vulnerable moments.
People come to a psychic-chat product mid-breakup, in grief, or facing a hard decision — and the business monetizes that moment. The frictionless copy manipulates: a guaranteed soulmate, a countdown, a struck-through price. I held the line between resonance and manipulation on every surface a user meets — the conversation, the promos, and the money flow — and encoded it into the system so the line holds without me.
The brief was “here’s the flow, fix the copy” — a ~220-string chatbot funnel steering new users toward a trial psychic chat. I reframed the problem before touching a word.
I ran two passes: a string-level pass, and a path-level pass — programmatically assembling seven end-to-end transcripts and reading them as real conversations for rhythm, repetition, and cumulative tone. The audit framework had no conversational lenses, so I built five and added them as a reusable pattern family:
The lead moment — give each line a job
After the PM signed off, the real question surfaced: how are these responses even formed? The whole corpus ran on one generator — mirror the user, weigh the feeling, bridge to the psychic — so the whole thing read like a catalog of captions. The assistant never responded — it evaluated. The fix wasn’t another polish pass. It was giving authors a palette of conversational jobs and a quota per path.
Being good at what you do and still not being seen — that chips away at you.
That’s an exhausting way to work. I’m glad you said it straight.
More people carry this one than ever say it out loud.
Noted — word for word. This goes to your psychic as is.
The conversational work set the principle; these show it applied at product breadth — keep the emotional pull, remove the certainty the product can’t honestly claim.
1
1
2
In a money flow, the frictionless copy manipulates the choice. I took three patterns out — two the governance system now catches on its own, and one that needed a person to name.
confirmshaming rule; the free-week offer stays as a separate button, minus the guilt.urgency-scarcity rule → a plain, informative label.Neither was spotted one screen at a time — confirmshaming and urgency-scarcity are rules I encoded in rules.yaml, so the linter flags every instance on every screen automatically.
The one no rule could catch
Some risks aren’t pattern-matchable — they need a person who knows the law.
A struck-through “original” manufactures a discount — but there was no record of that higher price ever being the standing one. I flagged it as an FTC compliance risk and escalated it to PM and legal — the kind of judgment a linter can’t make. And I set Crowdin as the localization source of truth, so fixes propagate instead of drifting.
One line runs through all of it — the conversation, the promos, the paywall: a product built on hope can’t grow by spending users’ trust, and the honest choice should never be the harder one.
Let’s talk.
I’m after senior/lead content design where copy is owned as a system.