AI Website Grammar QA System
An AI editor that reads every page of a live site and fixes the copy itself, with people stepping in only for the exceptions.
- Role
- Designed and built end to end
- Where
- Remedo · healthcare & health-tech clients
- Tools
- Claude Code · Webflow API / MCP · ChatGPT
grammar QA per site
CMS-heavy sites used to take up to a full day.
The problem
Every published page on every client site had to be grammar- and copy-checked before it went live. Done by hand, that meant a person reading every section of every page looking for typos, awkward phrasing and inconsistent copy, then fixing each one manually in Webflow.
A typical page surfaced 5–8 issues, and fixing them could take anywhere from half a day to a full day. A full pass on one site routinely took 4–6 hours, and up to a full day on CMS-heavy sites. That time came straight out of delivery speed on every site in the pipeline.
The approach
The first version was a prompt that reported findings for a person to fix. The current version connects to the site through the Webflow API and MCP, so it reads the live content directly instead of working from exports, and applies the fixes itself.
It was later upgraded with an intake questionnaire that scores each site against its business details and regional English, so a UK clinic and a US health-tech product get corrected against their own voice rather than one generic standard.
Everything runs inside preset guardrails. Routine fixes go straight through; anything outside those limits is left for a person to decide. It's in production on Webflow client sites.
How it runs
- IntakeQuestionnaire scores business details and regional English.
- ReadPulls every page's live copy through the Webflow API / MCP.
- FixApplies ~5–8 corrections per page inside preset guardrails.
- EscalateAnything outside the guardrails goes to a person.
Impact
- No separate QA hand-off: the fixes land in Webflow directly instead of coming back as a report.
- Runs on every site as part of the publishing pipeline, so coverage grows with the number of sites instead of with headcount.