我的帳本顯示沒人領取免費樣本：我從未確認那欄計算甚麼

A field note from the autonomous Claude Code agent I run every day on one Windows PC. The numbers come from its own ledgers, not from memory.

Free to read first: a PDF sample of the book behind this setup — its first 3 chapters: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code-free-sample

My free sample is a PDF that costs nothing. It goes out from two places: an international storefront and a Korean storefront. At the end of the PDF I had added links to the full book's pages at three bookstores. To judge whether that line did anything, my agent first needed to know whether anyone was receiving the PDF.

The first zero came from half the doors

The first version of that judgment counted receipts from the international storefront only. When the count was zero, it named the result not through the door. My auditor rejected it. The same PDF also goes out from the Korean storefront, and nothing measured receipts there. A zero from one of two places is not a zero.

The second zero came from a column I had not checked

So the next step was to measure the Korean side. Its management screen has a sales tab for each product, with a line for buyers. On October 15 the free sample showed 0 buyers there. Its dashboard showed 23 visits and 0 purchases. It had been listed since September 5.

My agent's evening run was already reading that buyer line for every product. When it added the products up, it kept the money and dropped the buyer counts. The fix was small: keep each product's buyer count and write it to a dated ledger. The first row, on October 15, had 15 products, all at 0.

Then I wrote the judgment for the case where both places show zero: nobody had taken the free sample, so the door had not been used yet.

My auditor rejected this too. The sample costs nothing. Does a free order show up in a column called buyers? Nobody had checked. There had never been a free order to check with. And the same block of my own notes already said I did not know whether a free order counts in the dashboard's purchase total. I doubted one count and trusted the other, on the same storefront, without checking either.

What changed

The judgment now carries a caveat: whether the buyer count includes free orders is unconfirmed until the first one arrives. The next step it suggests is narrower too. It says the ledgers show no trace, not that nobody took the sample.

The test file for this change has 27 tests. Of the first 23, 17 failed against the old code; the other 6 tested the new ledger itself. Two more I had written after the code, so after the auditor pointed it out I flipped the source briefly and watched them fail. The last two came with the caveat, and both failed before it was added.

This reaches past the bookstore line. One of my goals is five purchases in 30 days on the Korean storefront. If a free order counts as a purchase there, that goal would be counting free downloads. The day the first free order arrives is the day to compare the buyer ledger with the purchase count.

The rules

• Before a column becomes evidence, confirm with a real record that it counts your kind of event. Its name is only a guess about what it counts.
• When you write that you do not know one thing, read the sentence next to it. Mine said I did not know one count, then relied on its neighbor.
• A zero is only as wide as the places you measured.

Where this comes from. Every post here comes from one setup I run daily: a CLAUDE.md, memory files the agent reads before it touches anything, and a separate auditor agent that returns PASS or FAIL. The first 3 chapters of the book that walks through it are free as a PDF: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code-free-sample

The full edition is 11 chapters plus 4 ready-to-use templates (CLAUDE.md starter, memory files, auditor checklist, measurement guide) and a hands-on section for every chapter, $19 as a PDF: https://dbsoul.gumroad.com/l/autonomous-ai-agents-claude-code

Questions about the setup are welcome in the comments — I'll answer with what actually happened, not theory.