Primary Hook:
"AI deleted all my files."
[Visual: "AI" gets strikethrough animation]
"Actually, I asked AI to delete all my files."
Alternative Hook (if testing a softer open):
"I told Claude to delete everything in my tmp directory. It refused."
AI deleted all my files.
[strikethrough "AI" animation]
Actually, I asked AI to delete all my files.
I told Claude to delete everything in my tmp directory so I could retry an experiment with new data.
But I did something you probably don't do.
I gave Claude seemingly random context.
I said, "I've got a tmp directory with my last dataset in it. There are 2 files plus a large zip. Delete everything in tmp and grab the new dataset."
Claude didn't immediately delete my files. It was thinking. A lot.
C'mon, this is `rm -rf tmp`. Is Opus overthinking?
Then I read its thoughts.
It looked in the tmp directory first. And noticed I was wrong about what was in there.
It came back with: "Hey, there's a bunch of marketing stuff in here that looks really important. You told me we were deleting a dataset."
Yeah. That data was really important. Not saved anywhere else.
Claude saved me hours of work because of one thing.
I overexplained myself. I was verbose. I was loquacious. I was garrulous. I gave Claude context so it could make better decisions.
The directory contents were a small example, but the principle scales.
Give your LLMs more context and you'll get far better results.