The More Feedback, the Less Got Fixed
The Dream Team has a bot that reviews drafts. Whenever a card-news or article draft comes out, it points out the mistakes. The first instruction was simple: 「Find every mistake.」
One Real Problem Among Fourteen
It flagged 14 items per draft on average. But when we opened the list, most were small things like comma placement and tone preferences, while the critical note, 「the date is wrong」, sat buried in the middle. Readers get tired too, so the fix rate stayed around 40%.
Changing a Single Number
We rewrote the instruction like this:
Give me only the three most critical mistakes. One line each on why it is critical.
We compared the two versions over 5 days on 20 drafts.
• Items flagged: 14 on average → 3
• Fix rate: about 40% → about 85%
• Drafts with a missed date or notation error: 4 → 1
Why It Worked
A cap forces the bot to rank importance itself. 「Everything」 produces a long list, while 「three」 produces a judgment. People are the same: the shorter the checklist, the more likely it gets read to the end.
How to Use It
1. Ask for a count limit and a one-line reason together.
2. After fixing the three, run the same request once more.
3. Collect trivial items in a separate 「taste notes」 section.
It was a small experiment, but what the bot needed was not more eyes, only a narrower question.