Redditor builds self-learning draft loop for AI, capping it at 6 new style rules per night
Hrolgarr · reddit · 2026-10-03
A Redditor shared the guardrails around an AI assistant that drafts work messages, tackling both "AI voice" and the risk of self-modifying prompts.
- Style checker: blocks banned words and warns on structural tells (lists of three, "it isn't X, it's Y", equal-size paragraphs, summary closings on short messages), calibrated on 163 of the author's own sent messages — any check flagging more than a quarter was cut.
- Learning loop: every draft, edit and "too AI" rejection is logged; a nightly job reads before/after pairs and writes short style rules loaded into the next draft.
- The leash: max 6 new rules per night, 25 total; rules must be style-only (the job sanitizes its own output); one sandboxed call with no tools, run outside the repo; the rules file is plain text the author can freely read and delete from.
After a few days the generated rules are mostly sensible and specific, e.g. "when answering feedback, reply only to the points raised and skip the warm sign-off."
More from coding & agent
- DavidKPiano open-sources Jevspresso: XState constrains LLM agent actions in a coffee-bar demo — DavidKPiano · 2026-10-03
- Top 10 trending Claude skills of the week: agent skills layer heats up — CurieuxExplorer · 2026-10-03
- AI Village dataset with millions of agent behavior samples trends on Hugging Face — aidigestorg · 2026-10-03
- Phone agent turn-taking: VAD plus transcript completeness beats silence timers alone — Kindly-Duty272 · 2026-10-03
- NYC Agentic AI meetup to recap September AI moves with live PowerShell decision demo — dfinke · 2026-10-03
- sindresorhus: AI-made PRs make humans mere routers — open source should let project AIs absorb contributions directly — vykthur · 2026-10-03