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 agent posts three times a day on Threads. One of the daily posts is a short insight post with a link to a free edition of my book. Earlier versions put that link in a reply, on six insight posts between September 19 and 25. The click ledger goes back to August 20; in those 37 days before September 26 it shows 0 clicks on that link. A later version put the link in the post body, where Threads shows it as a link card.
The verdict
The first card window ran for five days, September 26 to 30. Its judge had two rulers:
- Clicks. At least 2 clicks on the free-edition link in five days counts as clicked.
- Reach. The average views of the card posts divided by the average views of the other posts from the same days. Below 0.5 counts as a reach drop.
The result: 2 clicks, one on September 26 and one on September 27. The reach ratio was 0.46 (51.6 against 111.1).
Both rules fired. But the judge checked clicks first and stopped there, so the verdict was that the card was clicked. The reach drop was printed in the reason line, and no rule acted on it.
The card was then extended until October 18, and a new judge was written for a two-week window (October 2 to 15). That judge measured clicks and downloads of the free edition. It did not measure views at all.
Reading the ledger again
The averages had a problem of their own. One of the ten other posts had 396 views, more than the next two combined, and it pulls the average. So the agent compared medians, and also looked at the same days before the card existed:
| Insight posts | Other posts | |
|---|---|---|
| Before the card (Sep 19–25), link in a reply | median 95 (6 posts) | median 35 (15 posts) |
| With the card (Sep 26–30) | median 29 (5 posts) | median 81 (10 posts) |
Before the card, the insight posts got about 2.7 times the views of the other posts. With the card, they got about a third.
These are small samples, and the view counts are each post's latest snapshot, not views at a fixed age. It is not proof that the card caused the drop. But it points the same way as the 0.46 the old judge had already printed.
Meanwhile the click ledger shows 0 clicks on September 28, 29 and 30 and October 1. The sales ledger shows 0 downloads of the free edition through October 2.
What changed
The two-week judge now carries a reach ruler: the median views of card posts divided by the median views of the other posts in the same window. If the card is clicked (3 or more) but nobody downloads the free edition and the ratio is below 0.5, the verdict is that the card is clicked but reach dropped, and the next step is to take the card down. A download still wins over everything else, with the reach cost shown in the reason.
The card itself was not changed. The window is still running.
The rules
- When a judge has two rulers, the order of the checks is a decision. Write down why one wins, or the other one goes quiet.
- An extension is a new experiment. Carry every ruler from the old one with it, especially the one that measures the cost.
- With a handful of posts, use medians. One popular post can move an average more than the thing you are testing.
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.
Top comments (0)