A small self-learning loop for drafts, and the limits I put on it so it doesn't wander

I have an assistant that drafts my work messages (status updates, handover notes). I edit and send them myself. The drafts had the usual AI voice, so I added two pieces around it. A checker on every draft. Banned words block, structural tells warn (lists of three, "it isn't X, it's Y", same-size paragraphs, a summary closing a short message). It was calibrated on 163 of my own sent messages, and any check that flagged more than a quarter of them was cut. And a learning loop. Every draft shown, every edit and every "too AI" rejection gets logged. A nightly job reads the before/after pairs and writes short style rules that load into the next draft. The part I spent the most time on was the leash, since a model writing its own instructions can go wrong fast: max 6 new rules a night, 25 total rules can only be about style, never about what needs my approval or who to contact (the job sanitises its own output and drops anything else) one sandboxed call with no tools, run outside the repo the rules file is plain text, I read it and delete lines freely It's been running a few days and the rules it writes are mostly sensible and specific. "When answering feedback, reply only to the points raised and skip the warm sign-off" is a real one. Link to the write-up in the comments, per the sub rules.