What they did
The author runs ads for a small travel business. They pointed Claude Code, driving their own logged-in Chrome through Claude in Chrome, at the company's Meta Ads Manager account. The jobs were ordinary ad work: pull results for chosen date ranges, divide the location targeting between two ad sets, publish the changes, and gather the links to the resulting ad posts for the rest of the team. Partway through, the author changed how the agent worked the page: instead of taking screenshots to decide where to click, it read the page as text and through the accessibility tree, found controls by their name and clicked them by reference.
How it went
Screenshots were the weak point. In one session the screenshot call hit its 30 second timeout six times on the heavy Ads Manager editor. Reading the page as text kept working: the agent located the Remove button for one location by name, clicked it, read the targeting list back to check it, and published. The author then packaged the approach as an open-source Claude Code plugin with recipes for other dashboards. Their own benchmark on five public pages found the text read used 2.6 to 37 times fewer tokens than screenshots.
Worth knowing
Dumping the whole accessibility tree can cost more tokens than a screenshot on large pages (10k to 70k), so read the text first and then search for one element. The author built in a confirmation step before anything that spends money or goes live. Only the Meta Ads Manager recipe was tested live; the others were waiting for testers.