Claude maxed my Amazon subscriptions while I played video games

For a little over three hours, Claude moved back and forth between my Walmart history and Amazon Subscribe & Save. It compared prices, corrected package sizes, calculated how quickly we used things, rebuilt delivery schedules, caught duplicate subscriptions, and checked its work.

I spent most of that time playing Final Fantasy XIV.

Every so often Claude needed me. How many rolls of paper towels did we actually use? Did I care which flavor of dog food got the extra case? Was I really asking for enough Oreos to produce 18 party-size packages every two months? I would answer, look at a product when judgment mattered, and go back to the game.

My total attention was probably twenty minutes.

The Claude session lasted more than three hours without consuming more than about twenty minutes of my Saturday. Claude kept working while I played the game.

If I had done the same job myself, it would likely have taken the whole day. I would have needed to open two purchase histories, identify the things we bought repeatedly, match brands and package sizes, compare unit prices, calculate consumption, learn Amazon's discount rules, rebuild delivery frequencies, and check the result for mistakes.

I also know myself well enough to know that I would not spend a day optimizing Subscribe & Save. Almost nobody would. The subscriptions would have stayed “good enough,” which meant we would keep paying more than necessary, miss higher discounts on poorly grouped deliveries, and occasionally order too much or run short.

Claude made the optimization worth doing because it could stay with the boring parts.

The job began as a narrower question. I asked Claude to use Chrome to inspect my Walmart purchase history, find the things I bought repeatedly, and compare them with Amazon Subscribe & Save. I wanted to move products when Amazon was cheaper and consider them when the price was close enough that automatic delivery would be more convenient. I also told it to consider other brands and package sizes rather than assuming that the exact Walmart item was sacred.

Claude found 485 products in Walmart's My Items history and worked through the 50 most frequent items that could ship. It checked the prices Amazon showed inside my account, where the subscription discount was visible.

The first pass already demonstrated why the job needed persistence. Claude matched the wrong package size for several products. When I asked it to check alternatives and larger sizes, some conclusions flipped. Snacks and dog treats that looked overpriced on Amazon became cheaper once Claude found equivalent packages. Bigger wasn't automatically better either. One paper-towel size had the lowest unit cost. Another larger-looking pack was more expensive per sheet.

Then I widened the assignment. Earlier payment trouble had caused several Amazon deliveries to fail. The payment method was fixed, but we needed the missed products and a better schedule. My priorities were simple to say and annoying to execute: maximize the Subscribe & Save discount first, then space deliveries as far apart as reasonably possible without over-ordering.

That required household knowledge Amazon didn't have. I told Claude how quickly we used paper towels, mouthwash, snacks, pet food, saline, floor cleaner, Roomba supplies, and pet-fountain filters. Claude translated those answers into quantities and delivery intervals.

It found 17 active subscriptions scattered across five delivery dates. One fabric-softener subscription was sitting on a day by itself, so it earned the lower discount instead of joining enough products to unlock Amazon's maximum savings. The first proposed solution was wonderfully neat: put almost everything on one delivery every two months.

Then Amazon's limits interfered. Soda could only be set to five units per delivery, which was not enough for two months of our actual use. Claude rebuilt the plan with faster products arriving monthly and slower ones every two months or longer. Each delivery date still had enough separate subscriptions to qualify for the best discount.

Claude chose a cheap Roomba maintenance kit with a 4.9-star rating. I looked at it and noticed the rating came from only 16 reviews. The kit also duplicated bags we were buying separately. I said I didn't want cheap for the sake of cheap. Claude removed it, compared the alternatives, and returned to the genuine iRobot kit with thousands of reviews.

Amazon also created duplicate soda subscriptions without warning. After an order, Claude reopened the subscription manager and noticed that the number of subscriptions had jumped from 17 to 21. The cart had not improved the existing soda subscriptions. It had quietly created a second Coke subscription and a second Diet Dr Pepper subscription. Left alone, we would have received eight packs of each every month while using about four.

Claude found and canceled the duplicates. I never had to hunt through the manager to discover them.

Dog food needed another checkpoint. Claude divided two Purina ONE flavors into a two-to-one ratio. I asked why foods used at the same overall rate needed different proportions. The answer came down to packaging: one product mixed flavors within a case; the other sold separate flavors, and three cases could not divide evenly between two products. We looked at a longer schedule that would even out the flavors, then decided to leave the cadence alone.

Then Claude showed me the Oreo calculation. I had described our consumption as roughly two packages per week. When Claude converted that into a two-month subscription, the result was 18 party-size packages—more than a quarter of the proposed new-subscription total. Claude flagged it before creation. I removed Oreos from the plan. That was an "Oh my Zeus, what are we doing" moment.

Paper towels moved from every two months to every four after we translated the pack into one roll a week. The fountain filters stayed closer to their old cadence. I knowingly left a four-day dog-food gap rather than add another purchase just to make the calendar perfect.

Claude then created seven new subscriptions through the individual product pages so their first deliveries would join the October bundle instead of shipping immediately. Amazon defaulted most of them to monthly and capped product-page quantities at five. Claude created each subscription, opened the manager, corrected the quantity and frequency, and verified it afterward.

The final manager showed 28 active subscriptions. The October delivery had 17 products and the November delivery had 13. Both cleared Amazon's five-subscription maximum-savings threshold by a wide margin. The failed items were restored. Duplicate soda was gone. Paper towels would arrive every four months, floor cleaner monthly, and the slower maintenance products on longer schedules.

Claude moved more deliveries into the maximum discount, found products that were genuinely cheaper, and worked through consumption math and scheduling combinations I would never have calculated for a shopping subscription. I spent about twenty minutes checking its decisions, then went back to Final Fantasy XIV while it finished a job I wasn't going to spend a day doing myself.