Teaching AI to See My Inventory
When generative AI got good, my first thought was: where it's strong, put it to work. And the thing that turned out to be a genuine unlock wasn't writing copy, it was seeing. Over the last year or two these models got dramatically better at photos. Not generating six-fingered people, I mean: I feed in pictures of an item I just shot, and the model recognizes it. "Widget A, these specs, this and this." It reads the UPC right off the box, and if it can't identify the item outright, it runs the UPC through an online lookup and gets there anyway. I can't overstate what an unlock that is.
A skill file, and 30 iterations
But it needs guidance. You cannot dump 40 or 100 photos into one of these and expect anything usable back. So around August or September of last year I started writing a skill: a markdown file that's essentially a faux-code logic tree. Do this, then that, if not this then that. It gives whatever model I'm using repeatable instructions.
My first mistake was being too rigid. "Do step A, then A1, A2, A3, then B, B1, B2..." Lock it down that hard and the moment it hits something your instructions didn't foresee, it breaks: sometimes in genuinely funny ways, sometimes in ways that just make you say what? The lesson: leave it room to think. I went through something like 30 iterations getting that balance right.
The Nano Banana detour
I also tried letting AI edit the photos, including a run through Nano Banana, asking it to pick the best variant of a shot and tune brightness, crop, and angle to convert well on mobile. And honestly? Nano Banana is fantastic at it. But it stamps a visible watermark on the image, plus other watermarks reportedly baked into the pixels, and a watermark on a resale photo reads as something to distrust to any buyer who notices. So I cut it from the pipeline. Great tool, wrong context.
Where it landed
Claude looks at my photos, picks the best of the batch, OCRs the packaging and UPC for identifying info, web-searches each item, and writes a markdown file with the retail price, specs, and a description pulled from real listings, plus the UPC. Each item gets its own subdirectory with its photos and that markdown file. That's stage one.
Stage two has it take a first pass at strategy: is this better for eBay, or Facebook/Craigslist? That one's genuinely touchy. Yes, the fees are near zero on Facebook and Craigslist, but they come with a pile of other friction and usually a lower sale price than a marketplace like eBay. Sometimes it says "eBay" when experience tells me "Facebook." So I review after each stage; early on it was right maybe 60 to 70% of the time. Then, like everyone else, I watched the models take a step-change in that November–December window and get exponentially better. That jump was a thing to witness. I've gone from checking every stage of every listing to mostly just reviewing the final product after it's live.