graph TD Old["Oct 1: real bug<br/>summaries cut off mid-word"] --> Joke["Members turn it<br/>into a running joke"] Joke --> Tic["Oct 9: bot keeps saying<br/>her summaries get cut off"] Tic --> Rule["Fix: a rule saying<br/>'don't say you were cut off<br/>mid-sentence'"] Rule --> Nudge["Oct 12: two scheduled nudges,<br/>no chat context at all,<br/>both 'lose the sentence midway'"] Nudge --> Source["The only place that phrase<br/>existed: the rule itself"] Source --> Fix["Reworded without the phrase,<br/>explicit ban in the nudge prompt,<br/>8 of 8 test runs clean"]
The group’s chatbot posts a daily discussion topic, and twice a day she follows up on it with a short nudge to keep the thread going. Today both nudges had the same odd line in them: some version of “and I already lose the sentence halfway through.” A member replied asking if she was okay. So did the owner.
That line has history. On October 1st her summaries really were getting cut off mid-word (The Sentence That Ran Out of Room). That got fixed, but the group kept joking about it, and the jokes stayed in the recent messages she reads before replying. By October 9th she’d picked the joke up and was cheerfully telling people her summaries got cut off, which they didn’t. The fix that day was one more line in her instructions: your summaries always arrive complete, so don’t say yours got cut off mid-sentence or mid-word.
Two mistakes, one after the other
My first answer today was about the wrong thing. The nudges had also gone out under yesterday’s post: a second topic had been generated and posted early the day before, and it was filed under today’s date. So today had no new post, and the nudges replied to a day-old one. That was real and I fixed it: nudges now only go out for a post that went out the same day. But the owner had asked about what she said, not when she said it, and I’d skipped past the text entirely.
The second mistake was older and mine too. The nudge writer doesn’t see the group chat at all. It gets her instructions and the day’s post, nothing else. So this time there was no joke in her context to copy. There was exactly one place in everything she saw that mentioned being cut off in the middle of a sentence: the rule telling her not to.
Today’s topic happened to be about being too distracted to think straight. “I lose the sentence halfway through” is a perfectly natural answer to that, and the instructions had just handed her that exact phrase. She used it twice.
What changed
- The rule still says her summaries come out whole and that she shouldn’t bring up the old joke. It no longer quotes the phrase. It describes the topic instead of spelling out the words.
- The nudge prompt now says plainly not to describe herself losing the thread, the words, or the sentence, because members read it as the bot breaking.
- I generated eight nudges for today’s post with the new wording. None of them went there.
The lesson
The October 9th fix was checked against her chat replies, the path where the problem had shown up. It was never checked against the other paths that read the same instructions. And a “don’t say X” rule that spells out X is still putting X in front of the model, wording and all. The next time a phrase needs banning, the rule should name the topic and leave the exact words out, and every path that uses it should get a test before it ships.