Matt Murphy
All posts
No. 03
  • AI
  • Corridor Cards

What went wrong when I automated my eBay listings with AI

Six mistakes from the first month of my listing pipeline, and the three rules that came out of them. Useful before you automate anything.

Part one was the tour: ten steps, six tools, and a clear split between what I do, what code does, and what the AI does.

It reads a lot tidier than it was. Getting there took about a month of mistakes, and honestly those are the useful part, so this post is six of them and the three rules I ended up with because of them.

The procedure is a list of scars

My written procedure started at version 1.0 on August 20, and as I write this it’s on 2.11. Every bump between those two numbers is something that went wrong once and that I didn’t want to see again.

Here’s where each one showed up in the run, if you want to follow along on the map:

The pipeline, start to finish

  • MeA person decides
  • CodeA fixed program decides, the same way every time
  • AIThe AI uses its eyes or its judgment

Before the run

  1. Step 1: Sort and photographMe

The run, at 2:00 a.m.

  1. Step 2: Check the folderCode

  2. Step 3: Give every card an addressCode

  3. Step 4: Publish the photosCode

  4. Step 5: Identify each cardAI

  5. Step 6: Check the AI’s workCode

  6. Step 7: Research the market priceAI

  7. Step 8: Build the eBay fileCode

  8. Step 9: Upload, then read every draft backAI

After the run

  1. Step 10: Review and publishMe

What went wrong

The titles with no team names. Step 8. One of the early batches came out with titles that were missing the team, and nobody searches for a card without the team name in it. So that became a hard failure: if there’s no team, there’s no draft.

The upload eBay refused. Step 9. eBay rejected an entire file with the message “couldn’t identify your template,” and it took a while to find out that the cause was one stray pair of quotation marks on the first line. That’s all it was.

The price that looked saved. Step 9. eBay’s editor showed a price sitting in the box, and the price hadn’t actually been saved. The rule that came out of that one is that typing a price into the box proves nothing, so now every draft gets saved, reopened, and read back before the pipeline believes it.

The black photo. Step 4. A resize command reported success and produced a completely black image. That tool is banned now, and more generally, “it said it worked” stopped counting as evidence of anything.

The card I almost listed wrong. Step 5. After one batch closed with fifty clean drafts, I checked a card by hand and found it was a rare parallel, which is a scarcer version of the same card, and in this case one with only five printed. The draft had the wrong version. Fifty out of fifty had looked like a perfect run, and it wasn’t, which is why I still check cards by hand.

The batch that humbled it. On the fifth batch, 26 of the 50 drafts passed review and the other 24 needed me. I don’t count that as a failure, though. That was the pipeline doing exactly what I’d told it to do, which is to stop and ask when it isn’t sure.

It’s worth noticing what kind of mistakes these were. Four of the six were mechanical things, a title rule, a file format, a save button, a photo tool, and only two were about what a card actually is.

Rule one: never guess

My written procedure puts it bluntly: “never guess a card identity.”

There are dozens of variants of the same sports card, and a wrong title on eBay isn’t just a typo, it’s selling someone a card you don’t have. So when the AI isn’t sure which variant it’s looking at, that row stops and waits for me, and when it can’t find enough sold listings, the price stays blank. I’d much rather have a blank I can fill in than a confident guess I have to catch.

One blocked card doesn’t hold up the other forty-nine, but it doesn’t sneak through either.

Rule two: it can’t publish

This one isn’t a setting I toggled. No command in the pipeline is allowed to publish a listing, and the publish actions are named and banned in its instructions, so there’s no way for it to drift into doing it.

The way I think about it is that I’m happy to hand off the typing, but I’m not handing off my name.

Rule three: the SKU is an address

Every card gets one code, and the code is where the card physically lives: the box, the section, and the slot. That same code follows the photos, the spreadsheet row, the draft, and eventually the listing.

So when a card sells, I don’t have to search for it. I know exactly where it is because of the SKU system.

What I’d tell another seller

  • Write the rules before the automation. Mine were a document before they were code, and I think that’s the only reason the code came out right.
  • Keep the publish button. Automate up to the decision, not through it.
  • Prefer a blank to a guess. You can always fill in a blank, but you can’t un-send a wrong card.
  • Write down every correction. The changelog is where the lessons actually live.
  • Don’t trust “success.” Read it back.

Was it worth it?

That’s really a math question, and part three does the math: what it cost to build, what each run costs, and how fast it paid for itself.

If you’ve got a process that keeps going wrong in the same places, tell me about it.