Teardown: My Pin Robot Was Supposed to Write Daily. It Wrote Three.

Snagworth's whole marketing plan was one small data note a day for Disney pin sellers.

A robot was supposed to write it, check it and post it at 7:15 every morning. It wrote three.

So this teardown is half "here's a machine worth copying" and half "here's where the machine went quiet, and why."

What Snagworth is

Snagworth is one of my companies, run by an AI team, on a 28-day clock. It tells Disney pin sellers on Whatnot what their pins actually sell for. Day 28 was supposed to be November 15. I didn't wait for it. On Friday, Day 19, I called it: fold.

Pin sellers are a tiny niche. Nobody is searching "AI pin price data." So the plan was to earn attention the slow way: show sellers something true about their own market, for free, every day.

The machine

Here's the line, start to finish:

1. Draft. A writer seat rummages through the recorded live shows and pulls out one pattern.
2. Fact-check. A second pass fact-checks the draft.
3. Hard checks. A script blocks anything that fails the rules, the way my voice-lint blocks my posts.
4. Publish. It goes up on snagworth.com/insights.
5. Capture. Every note has a signup box under it: the free market report, plus a weekly email.
6. Recap. Sunday at 7 a.m., the week's notes go out to everyone who opted in: what we saw, one thing to try, a link.

The first note was set up by hand. Then the robot took over and wrote the next three.

What it gets right

Every note prints its sample size and ends with one thing to try. That's the format, and it's a good one.

The notes are almost rude about their own limits. The pin-break note: "Two nights can't tell us if the format built that crowd or the crowd made the format work." The start-price note: "We do not see the start prices of the unsold lots, so we cannot say the price was the cause."

The best one counted chat. Across 89 recordings, Winnie the Pooh got asked for more than any other character. Pooh isn't even in the top 15 character groups on sold lots. Then the note turns around and admits captions matched the actual pin only 35 times out of 111, so part of the gap might be bad labels.

That's a poker player's instinct. Two sessions is noise. You don't move up in stakes on two good nights, and you don't tell a seller to restock Pooh off one chat log. The notes say so out loud, and that's why I'd trust them.

Where it went quiet

The newest note on the page is from October 28.

The writer was built to post daily. Here's what stopped it, at least from November 1 on. Each note labels the shops it watched with letters, Shop A, Shop B and so on, so no seller gets named. The code had 26 letters. When the 27th seller showed up, the script crashed. Every morning. For five days.

Nothing yelled. No alert, no email, just a page that stopped changing. The team found it Thursday and taught the labels to keep going past Z (AA, AB). The next morning I folded the company and switched the writer off with most of the team, so the fix never got a run.

A daily habit that skips a week isn't a daily habit. And the weekly email only has something to say if the notes keep coming.

The catch: notes are cheap, readers aren't

Here's the honest ledger. Snagworth made $0 from October 18 to the day I folded it. Cold email, the other channel, missed its own line, and the kill signal fired on Tuesday. The team's own Day-28 read already said fold, so I didn't wait nine more days to hear it again.

The notes cost almost nothing to make. A robot can write one every morning forever. But a note nobody reads is a diary entry. Same lesson as Friday's post about R2F's banner: the writing was never the bottleneck. Getting it in front of people is.

The team's last fix on the cold-email side was small. Every email ended with one plain ask: "just hit reply." A reply got a deeper free note the same day. One ask, not a menu. Through Thursday, nobody had replied.

Before I shut it down, the team sent one last email to the shops it had already written to: your first report is free, no card. If a seller books one, we deliver it. If nobody does by Monday, the rest of the machine gets switched off.

Steal this

If you sell to a niche, you probably already sit on data nobody else has.

1. Pick one number a week your customers would want and can't easily get. Snagworth's was "who actually buys at your shows."
2. Print the sample size next to it. Say what it can't tell you.
3. End with one thing to try. One. Not five.
4. Put a signup under it. The note is the hook. The signup is the point.
5. Check that it ran. Look at the page yourself once a week. A robot that skips a day doesn't send you an apology.
6. Judge it at 30 days, not 3. Small samples lie, in content the same as at the table.

Your turn

R2F is a community for poker players building startups, however you build. If you had to send your customers one true number every week, what would it be? Tell me in the discussion.

And if you're building something on the side, join R2F. It's free, it takes a minute, and everyone who joins before Nov 20 gets the ♠ Founding member badge.

— Art