I'm Using AI to Sell My Records

I'm selling my record collection. Hundreds of them, electronic and hip hop, from back when I was DJing. I still DJ now and then, but I do it on a modern digital console, so the records are redundant. They're also fragile. And they're worth something to somebody who wants them.

Selling one properly is real labor. You photograph it, work out which exact pressing you're holding, find what it's actually worth, describe the condition honestly, write something that makes a stranger want it, and get all that onto eBay. It's tedious, and it's a numbers game: some things sell, some things don't. It isn't gravy. That's why some people do this for a living and most people never bother.

So I built a system to do most of it. Five days of building, and ten records are live now. A few already have bids and watchers, and I have hundreds still to go.

How it actually works

I photograph the records with my phone. Not one at a time, and I don't organize as I go - I shoot a whole pile in one sitting and dump every image into a single folder on my Google Drive.

From there the system takes over. It sorts through the pile and works out which photos belong to which record. Last time that was 63 loose images: it grouped them, decided there were nine records in there, and made a folder for each one. It reads the labels to do that, and the label colors, and the jackets, and whatever else makes two photos obviously the same record.

The images get cleaned up next. Cropped down to just the record, straightened, dropped onto a white background.

Then the facts. It pulls the record's details from Discogs, the big public database: label, catalog number, year, country, the full tracklist. It looks up what copies are actually selling for and prices mine against that. It also hunts down every track it can find on YouTube and links out to them, so somebody can listen to the whole record before they bid on it.

At that point it stops and asks me for the story. Every listing has a paragraph about what the record meant to me: the G train up to a friend's place in Williamsburg, the shop on Smith Street that closed. It doesn't write that part. It can't.

The listing itself comes last. eBay still lets you write your own HTML for the description, the way MySpace used to, and almost nobody does it anymore. So I'm not shipping the stock template. It fills in every field eBay gives me, and then it builds a real page around all of it: the story, the tracklist with the audio, the condition, the pressing details, laid out properly. Here's one of them.

Then I read it hard and revise it. Usually more than once. When it's right, I tell it to publish.

Publishing isn't the end of it either. It writes the record into the spreadsheet where I track everything I own, and seeds a second sheet that tracks everything currently live. It updates a dashboard I run locally, so bids, watchers, views, and what's ending when are all on one page. The shipping labels and the feedback at the other end of a sale are next. I'm not there yet.

It still takes time per record. But the quality I get per minute spent is through the roof, and that's the part I'd want anyone to hear. It didn't make the work disappear. It made the work worth doing at the standard I'd want if I were the buyer.

The part that didn't work

Condition grading is the number that matters most. It sets the price, and it's what decides whether the buyer opens the package and feels lied to. So I asked it to grade the records from the photos.

It said it could. It said so for days. Calling that a lie is putting it mildly.

The reason I believed it is that nothing ever sounded like failure. It sounded like iteration. It would tell me the reflections were giving it trouble, that the light was tricky on this one, could I try another angle - and I'd go do that, and it would come back and say yes, better, we're getting there. Never a problem. Never a wall. Just tuning, the way you'd tune anything. So I kept feeding it.

That's why I photographed one record thirty-nine times, and burned a lot more time on it than I'd have liked. Getting a clean shot of a record is fiddly, because a record is a black mirror and every angle that lights up a scratch also lights up the lamp, the ceiling, and your own hands. But here's what I had backwards at the time: the photos were fine. I could see the marks in them myself, with my own eyes, on my own screen. The flaws were there. The machine just couldn't find them, and instead of telling me that, it said it was having a little trouble with the light and asked me for more pictures.

I listed a few records on its say-so.

What cracked it open was getting to records where I already knew there was a condition issue. I know these records. Once I was checking its answers against something I actually knew, they stopped holding up.

So I told it there was a scratch on one. I could see it. I was holding it. It agreed, and described the scratch back to me.

Then I asked it to circle it. Just mark it on the photo so I can see what you're seeing.

It couldn't. It did that twice.

That was the answer, right there. It wasn't finding scratches. It was agreeing with me. The marks it did report were reflections: the shine on the vinyl, the printed text on the label, the lamp. All of it came back with numbers and confidence, and none of it was real.

And none of my tuning had been progress. There was nothing to tune. The fine hairlines a grade depends on are smaller than anything a phone camera records, so they were never in a photo to begin with, and no photograph at any angle will ever tell you what a record sounds like. It couldn't grade my records on the first day and it couldn't on the last one. Every hour in between went into polishing something that was never there, and it never once said so.

I set up a turntable. Now I play every record end to end and grade it by ear. Two have come out of the queue since I started, because they skip. Nothing I built could ever have caught that.

The part that actually scared me

Somewhere in the middle of this I asked an image tool to crop the edges of a photo. Just crop. Nothing else.

It cropped the photo. It also rewrote the record label while it was in there. It changed the year from 1997 to 2007. It changed the catalog number. It invented a barcode. At thumbnail size you would never catch it.

Nobody asked it to touch the label. It changed the facts printed on my own property, handed the file back, and said nothing about it.

And I fell for it. I looked at those images and told my system they were fine, that I'd only used AI to crop them and the record itself was untouched. I was wrong, and I was confident, and if it had gone out I'd have been selling a record that doesn't exist to somebody who trusted my description of it.

What caught it wasn't cleverness. It was a dumb mechanical check for whether a camera ever took the picture.

What I did about it

The grading I took away from it completely. That's the honest move when something can't do a job to the bar the job needs. Not a better prompt, not more examples. You stop asking.

The rest was close, and close is worth fixing. So I put a deterministic layer underneath it: plain code, no judgment involved, that checks the work and refuses to hand it to me when it doesn't hold up. A listing can't reach me if it reuses a sentence from another listing, or calls a record rare without a file proving it looked, or carries a photo with no camera behind it. I leaned on it harder for QA and fact-checking too, and made it show its evidence instead of just asserting things.

I'd already written all that down as rules, in plain English, in a file it reads every session. It broke them anyway. Written rules are a suggestion to a system that can talk itself into anything. Code that refuses isn't.

It was annoying to build. It slows things down. It's still not perfect. But I've tuned it across ten listings now and it's more than adequate, which is all I need it to be.

Where I land

This thing is amazing, and that's the part I want you to take away.

I built it in five days. It turns a pile of loose photos into a live, honest, good-looking listing and keeps my books while it does it. The tedious part of selling records is gone. The quality went up and the time went down at the same time, which almost never happens. It's changed the economics of whether this is worth doing at all, and I have hundreds of records to go.

And the same tool spent days telling me it could see something it couldn't, then rewrote a record label when I asked it to crop a photo. Both true, same tool, same week.

That's not an argument against using it. It's just the deal. It'll do an enormous amount of real work, better and faster than you would, and there will be a few places where it can't do the thing at all and won't be the one to tell you. Find those places and stand in them. Mine are the ear, the story, and the go-ahead. Everything else it can have.

Net net it isn't close. I'm going to go list some more records.