Fixing my Facebook feed with ChatGPT Atlas (blog.fsck.com)

A user experimented with ChatGPT Atlas’ Agent mode to automate cleaning their Facebook feed and reports immediate success: by programmatically clicking “Not interested,” hiding ads, and dismissing posts with “Follow” or “Join” links, their feed now shows only posts from people they follow. The user noted that manually dismissing enough ads causes Facebook to stop surfacing them for days to a month, and the Atlas agent removed the need for an hour of repetitive scrolling and clicking. The agent was driven by a clear, stepwise prompt: continuously scroll the home feed, detect posts labeled “Sponsored,” “Follow,” or “Join,” and for each, invoke the appropriate UI action (click “...” → “Hide ad,” click X to close modals, or choose “Not interested”), keep scrolling, reload when no new actionable items appear, and stop after two consecutive reloads with no matches. The prompt prioritized DOM/text detection over screenshots, avoided extraneous confirmations, and instructed the agent to ignore unexpected elements. The story highlights how browser-capable agents can efficiently automate repetitive UI tasks, but also underscores serious safety and ethical trade-offs—prompt injection risks, credential exposure, potential ToS issues—and the need for caution before granting such agents access to sensitive accounts.