One-person Amazon seller runs daily ops on Claude scheduled tasks

The setup spans six recurring jobs, from order intake and shipping requests three times a day to a weekly FBA reimbursement audit and a daily check of Coconala requests and Rakuten campaigns. Execution happens in the cloud, and only tasks that need a browser hunt in order through a laptop, a VPS and the browser inside the desktop app. That division of labor is where the design discipline comes from: a job that cannot find a usable browser is told to send an urgent notice and stop rather than improvise.

The principles read as a reaction to specific failures. The inventory incident, where Amazon's count did not fall when the warehouse hit zero, became a written rule with a formula: target equals warehouse free stock minus unregistered orders minus a safety stock of one, floored at zero. Browser selection is pinned to device IDs because display names change on every connection, and data is pulled through JavaScript rather than screen rendering because a sleeping PC narrows the browser and times out screenshots.

Cost control follows the same logic of matching effort to stakes: Opus where an error hits the account, Sonnet for fixed procedures, Haiku for read-only checks, with product-name matching handed to the Gemini API. Reports are kept to one line on quiet runs so real alerts are not missed. His summary of the benefit is about mental load rather than hours saved: the bigger gain was losing the burden of remembering what must not be forgotten.