AI removed the safety net. I rebuilt it inside the agents.
A live glossary that every AI agent at Wix checked before writing a word onto a screen.
- Standard in every AI coding tool at Wix
- Thousands of terms in 20+ languages
- Two weeks to an MVP, two months to every product area
01 · The safety net
For years, people kept Wix's language consistent.
Keeping product terms consistent is always hard, especially when teams work in silos. At Wix, the Writers Guild handled it with workflows built up over years: editing groups, clear owners, standard reviews. Every word passed through a writer.
02 · The gap
Then AI let everyone ship copy.
With AI coding tools, PMs and designers could write product copy and push it to production themselves, without a writer ever seeing it. The safety net was gone, just as everything sped up.
03 · What I did
I saw the gap early and built the fix myself.
I got VP approval, talked to writers, PMs, and developers about how they handled terms, and wrote the spec. Two weeks later I had a working app, followed by two rounds of proof of concept.
- Gap spotted
- Spec and interviews
- MVP2 weeks
- Proof of concept
- A bigger proof of concept
- Every product area2 months
04 · How it works
Every term, caught every time.
Before an agent put copy on a screen, it sent the strings to the glossary. The lookup checks every form of every phrase against thousands of entries, then hands back what the agent needs: the official term, what it means, where it belongs, and which names to avoid. The agent makes the call. The lookup makes sure nothing slips past.
-
1 split
Every one-, two-, and three-word phrase
-
2 expand
Singular and plural forms
-
3 match
Check every form against the glossary
-
4 respond
What goes back to the agent
The agent makes the call
Example entries, not Wix's real glossary: client, invoice, booking page, automation.
05 · Ownership
People and agents suggest. Owners decide.
Anyone could suggest a new term, including the agents themselves. Only the owners of that product area could approve it: team leads, editors, and senior writers, with permissions built into the app. Once an English term was approved, it went to the language writers for 20+ languages.
06 · Rollout
Every product area, in two months.
Within about two months, every product area's glossary was in the app. Some teams dumped their old lists in without reviewing them, which left conflicts behind, so we ran a second round of fixes before it went to production.
“You need a human to actually do the part that a human needs to do.”
07 · Result
It became how products get built.
The glossary check became part of the standard skill set in every AI coding tool at Wix. Agents didn't build product without calling it. Teams adopted it, and the feedback was very positive.
“At the end of the day, we did not get content that was out of the glossary.”
08 · What came next
It became the anchor for what came next.
With a trusted glossary in place, I used it as the anchor for a semantic layer that maps Wix's product to its code, copy, and data.
Next project
Semantic layer & context engineering →
More detail
How it works, step by step
- An agent in Codex, Claude Code, or Cursor is about to write copy into a product screen.
- Its content-writing skill tells it to send those strings to the glossary tool first, on a dedicated MCP server.
- The tool calls a webhook that runs an n8n workflow.
- The workflow splits the copy into one-, two-, and three-word phrases, adds singular and plural forms, and looks them up in the glossary with SQL.
- It returns every match: the official term, its definition, its product areas, and its allowed and prohibited synonyms.
- The agent decides what to change, following its skill.
My role
- Spotted the gap and got VP approval
- Interviewed writers, PMs, and developers, and wrote the spec
- Built the app, including CI/CD, security, and roles and permissions
- Ran the MCP server and helped write the UX content skills agents use
- Led the move of every product area's glossary, and the round of fixes after
Tools
Questions a hiring manager might ask
- Did it block bad copy?
- It made sure agents always had the full picture before writing. The agent fixed its own copy, so problems were solved before anyone saw them.
- How do you know it worked?
- It became a required step in every AI coding tool at Wix, and off-glossary copy stopped shipping. Individual catches only show up in workflow logs and agent reasoning, because they happened before anything reached a screen.
- What was the hardest part?
- The people part. Moving every glossary in was quick; getting teams to actually review what they moved took a second round.