What they did
Youn runs an autonomous Claude Code agent every day on one Windows PC, guided by a CLAUDE.md and memory files it reads before it touches anything. He writes field notes about that setup using numbers taken from the agent's own ledgers and from git. Before anything goes out, a separate auditor agent checks the drafts and returns PASS or FAIL. In this round it checked a field note and two short posts queued for other channels, one of them in Korean, all counting commits and shell calls over a stretch when a handoff note sat above 400 lines.
How it went
The counts were right but the window was not: three sentences opened it with a date, while the count actually began at a commit at 1:36 PM. Counting from midnight gives 90 commits and 104 shell calls instead of 84 and 96, and six of those commits left the note under 400 lines. A test he had written asserted the wrong wording, locking the mistake in. The auditor also flagged two figures, in a progress log and a code comment, that it could not rebuild, so he deleted both and wrote down its recount with the counting rule. The second pass failed on the same error in two more files. The third passed.
Worth knowing
The audit works because it recounts from the source data instead of trusting the draft, and because a test that copies a sentence copies its error. After a fix, look for siblings: the second pass failed only on copies he had not checked. A total without its counting rule cannot be checked at all. The author sells a book about this setup.