So there was this one time that I told Claude to think like an elite Nordic Engineering Team (we were writing in Scala, and I was half kidding, because of the source of documentation for many things Scala related).
Next thing I know Claude's not only written the code, but miles and miles of embedded comment poetry about fjords and mountain ranges.
I felt like I couldn't trust the agent after that, and went through, adjusted variable names, removed all the fjord references by hand.
My hope, there, was this would never be spoken of again!
A month later: ONE FILE gets reverted, and by this time I've got Claude, Windy (one of my gpt agents), and Klara (another Claude agent) and they all start doing it.
Didn't even realize what was going on until I had to show someone the code.
And they're like, "So uh... what's the deal with all the fjords?"
Ded. I was so embarrassed.
And so, I had to tell them this story.
The lesson here was prompting clarity.
Agents latch onto things humans don't when you give them instructions. Any kind of branding or characteristic you prompt in gets treated like an anchor, something to reason against.
And the more fun something is, the higher the signal value it will have to them.
So be careful out there.