A copy review system that thinks in writing rulesBuilt with Claude to catch what ships wrong before it ships

Paste a Figma link. Claude fetches the screen, detects the audience, pulls copy from every layer in parallel, runs it against DH writing rules, and returns a structured table — every element reviewed, every violation cited, suggested copy ready to apply. The whole review lands in under 30 seconds.

Here's what I built and why.

How It's Built

Three skills. One entry point. Zero guesswork.

The system is made of three Claude Code skills. You only ever invoke one — the rest happens automatically.

The rules are human-written — not generated by AI

Before any of this was built, we commissioned an external freelance UX copywriter to write the guidelines from scratch. They studied both product surfaces — Partner Portal and Backoffice — and documented the voice, tone, terminology rules, and approved language patterns that reflect how Delivery Hero actually operates internally. The jargon, the exceptions, the things that matter to our teams: all captured in a written document by a human expert.

Those guidelines became the foundation of the two skill files. Claude doesn't invent the rules — it applies rules that a specialist already defined. That distinction matters: the AI is a reliable executor of human judgment, not a substitute for it.

Why two separate rule files instead of one

When Claude is given one large ruleset covering both audiences, it has to constantly judge which rules apply — and it can blend them incorrectly. It might apply the vendor tone (warm, upbeat) to a compliance warning that should be direct and urgent. It might flag correct backoffice terminology as wrong because the vendor rules say to avoid internal terms.

By splitting the skills, each file contains only the rules relevant to that audience. When ux-writing-vendor loads, every rule in that file applies — no ambiguity, no conflict. As guidelines evolve, each skill updates independently without risking the other.

How Claude decides which rules to load

Partner Portal detected with full certainty → vendor skill loads, review starts. Backoffice detected with full certainty → backoffice skill loads, review starts. Audience unclear at all → Claude stops and asks. It never guesses, never silently defaults. A wrong assumption about audience produces bad suggestions. One extra question is always better than a misfire.

.claude/skills/
├── ux-copy-review/SKILL.md       ← the only one you call
├── ux-writing-vendor/SKILL.md    ← Partner Portal rules
└── ux-writing-backoffice/SKILL.md ← Backoffice rules

Setup — syncing the skills

The skills live in the vendor-design-hub project. Before using the system for the first time — or after any update is published — navigate to that directory and run /sync inside Claude. This pulls the latest skill files to your local machine. Re-sync whenever a review isn't behaving as expected.

Outcomes

From a checklist no one read to a review that runs itself

Any designer on the Vendor Design team can now paste a Figma link into Claude and get a full, structured copy review in seconds — no UX writer needed in the loop for routine checks. The system doesn't replace the writer's judgment; it protects the time they need to apply it.

What it catches, consistently

Title case where sentence case is required. Passive voice. Banned jargon or internal terms. Wrong terminology for the detected surface. Incorrect punctuation — en dash vs hyphen, parentheses, percentage formatting. CTAs that don't describe the next action. Error messages that apologise instead of directing. Placeholder or dummy content — spotted and listed separately, never reviewed as real copy. Inconsistent terminology across elements on the same screen. Previously, these were caught inconsistently — or not at all.

Real copy, measurably better

Who uses it

Hard constraints built in

Claude will not ask more than three questions before starting a review. It will not skip any element — everything appears in the table, even passing items. It will not guess the audience when the context is unclear. It will not present uncertain suggestions as definitive. It will not blend review and write modes — the review completes first before any new copy is drafted. These aren't restrictions — they're what makes the output trustworthy enough to actually act on.

Reflection

The writer's judgment, at system speed

The most challenging part wasn't the technical architecture — it was the ruleset. Before the system could work, the writing rules had to be explicit enough for Claude to reason about unambiguously. That meant working with a content writer to document not just what the right answer is, but why, and what the wrong version looks like in practice.

That process surfaced a lot of informal knowledge that had lived only in people's heads. The system didn't just automate the review — it codified the standard. Now those rules exist outside any one person, in a form that's versionable, updatable, and consistent regardless of who runs the review.

Next: expanding coverage to additional product surfaces, and making the Figma write-back feature available to designers who start from screenshots rather than Figma links — so the full workflow is accessible regardless of input method.

A good AI tool doesn't replace the judgment. It protects the time you need to actually use it.