One thing I've realized is that AI doesn't play nice when you say DON'T.

I mean, how many times have you told AI not to do something, then a minute later seen it do that exact thing again? If I had to guess, a lot. But here's what I've figured out. When you write "don't" or "do not," the AI can easily read that as "do," which means it skips right past the "don't."

This is why I avoid negatives and use positives like this:

"Output plain text only"
"Respond in 2 sentences max"
"Limit answers to features and support topics"
"Answer only from the provided context. If it's not there, say 'I don't have that information.'"
"Respond in Japanese"
"Begin your response with the first data field"

These avoid the main issue of the AI misunderstanding you, while making sure it actually does what it's supposed to. Try this next time you're about to write a hard rule. You might get surprised.

12 comments

One thing I've realized is that AI doesn't play nice when you say DON'T.