The Problem
One of our SEO pages said the product could export a video format it can only import. It had been wrong for a while. Nobody noticed, because it was an AI tools page, one of dozens, and nothing about it was on anyone’s dashboard. Then someone bought on the strength of it and asked for their money back. The correction was one sentence. Getting to it is what bothered me. The words on that page are not stored with the rest of the website. They sit in a separate system built for managing text, so an agent told to go fix the page goes hunting through the site’s code for a sentence that was never in it. Doing it properly meant knowing where the words actually lived, knowing where the true answer lived, and having someone available who knew both. For one sentence. On a page nobody had looked at in months.The Approach
I did not want to be faster at that. I wanted the requirement gone. An engineer on the team had already built a small command-line tool that pulls one page’s content straight from the source, so I never had to open that system at all. Before automating anything on top of it, I wanted to run the whole correction by hand once. So we did it live: pull the page, read every claim against the public help center, correct only the fields that were actually wrong, package up the change, verify it in the tool, and then look at the draft page with my own eyes. The tool cannot publish. That is deliberate. It can only ever produce a draft, never touch the live page, so nothing it does can break the site if someone gets it wrong. Putting the change live is a person’s job on purpose. Claude tried to make the work bigger than it was, twice. First it wanted a maintained reference file of what is true about the product, which for something that comes up a few times a month is a cache that goes quietly stale while the help center stays current. Then it shaped the tool around the one page type we happened to start with. Both times the correction was mine to make.What Happened
Doing the fix for real shook four bugs out of the tool itself. Claude wrote all four up. Three fixes were merged that day. The fourth, a missing route from an edit to something ready to publish, went in as a known gap and is still a step I do by hand. Nobody went looking for those bugs. They surfaced because we were doing something that mattered instead of testing. The page went live corrected. By the end of it the loop existed as a reusable workflow: find the right answer wherever it already lives, pull the page, make the change, verify it twice.What I Learned
The page was wrong for a long time and the only thing that surfaced it was a refund. What I wanted was a support person posting what they are hearing from users, an inaccuracy or a gap or something a page explains badly, and me tagging in an agent that goes and finds the truth wherever it lives, makes the correction, and shows its work, without any of it needing a call. That is what runs now. The teammate does the whole thing and stops in exactly the place I stopped: it creates the fix, a person reviews it and merges. A pull request workflow, for marketing pages. If you build something like this, aim it at the pages nobody owns. The ones that matter already have someone watching them.Drafted by my Claude from our working session. I edited and approved it.