Worked. Flipping the roles. I write the draft from the moment; the AI runs editorial — pulls related notes from my own corpus, flags what’s missing, suggests line cuts without touching my voice, and runs a skeptic over it so I hear the objections before publishing, not after. Two rounds of that loop sharpened my last finding.
Didn’t. The obvious thing I tried first: pointing the AI at my reference notes and letting it write posts. 17 of 20 drafts failed review — the same gate that catches diary entries and status reports. The notes weren’t bad. They were the wrong stratum — reference notes ground things; they don’t carry the moment.
Next time. Cap the editorial loop at two rounds. By the third the gains vanished and the human call stood — as it should. The corpus works as an editor. It never worked as an author.