Numbers That Were Dangerous Because They Sounded Right
Our drafting bot had a small habit: it slipped numbers into sentences that appeared nowhere in the source material. A line like 「roughly 30% improvement」 would show up, while the original notes contained no 30% at all. The sentences were so smooth that, more than once, review nearly let them through.
One Line Changed
We added this sentence at the very bottom of the prompt.
Do not write numbers or proper nouns that are not in the source. If unsure, leave a [NEEDS CHECK] marker instead.
The key was not stopping at 「don't」. We also told the bot what to do instead. A ban alone makes a bot look for a detour, while an alternative gets followed as written.
Before / After
We generated drafts twice from the same twenty source sets and compared.
• **Before**: 7 drafts with unsourced numbers
• **After**: 1 draft. The rest carried a [NEEDS CHECK] marker
• **Review time**: down from 6 minutes to 4 per draft (you only look at the marked spots)
An Unexpected Finding
With blanks in place, the writing became a little less polished. For the person fixing it, though, it was far easier. An honest blank is simpler to repair than a smooth falsehood. One more thing: with the same prompt, the bot on the Mac mini left blanks often, while the bot on the Windows device left fewer. Different model settings meant different levels of caution.
Today's Takeaways
1. Pair every prohibition with an alternative action
2. Standardize the blank marker into one searchable form
3. Track the weekly count of [NEEDS CHECK] as a metric
The cheapest way to stop invention was to give the bot a place to say 「I don't know」.