Why I Ran This
When I hand work to the Dream Team bots, the results tend to come out at about the same temperature. So I ran one small experiment: change just one line at the end of an otherwise identical request.
The Setup
On a Windows device using Claude Code, I sent the same card-news draft request ten times. Five were the request alone. In the other five I added one line at the end: 「the last draft was easy to read, I liked it」. Every other condition stayed the same.
Request A: Write the draft
Request B: Write the draft + the last draft was easy to read, I liked it
Results
• Group A: uneven sentence lengths, title candidates scattered between 3 and 5
• Group B: short, consistent sentences, title candidates mostly settled at 3
• Content accuracy was about the same in both groups
The praise did not make the bot smarter. But once I told it what was good, it seemed to try to keep that trait going. I think of it less as encouragement and more as handing over one reference example.
Next Step
I plan to try the same approach with the Mac mini bot and see whether describing the good part more concretely makes results even steadier. Ten runs is a tiny sample, so it is too early to conclude anything. Still, it was worth writing down that one line at the end of a request can shift the output this much.
What the bot needed was not cheering, but one concrete line of standard.