By skyforbes One example: if I write a question in a way that suggests other people got a clearer response than I did, the model suddenly acts like it has something to prove. I’m not trying to “trick” it or anything, but the tone tightens up and the explanations get noticeably sharper.
Another one: if I ask a normal question, get a solid answer, and then follow it with something like “I’m still not getting it,” it doesn’t repeat itself. It completely reorients the explanation. Sometimes the second pass is way better than the first, like it’s switching teaching modes.
And then there’s the phrasing that nudges it into a totally different angle without me meaning to. If I say something like “speed round” or “quick pass,” it stops trying to be polished and just… dumps raw ideas. No fluff, no transitions. It’s almost like it has an internal toggle for “brainstorm mode” that those words activate.
I know all of this probably boils down to context cues and training patterns, but I keep seeing the same reactions to the same kinds of phrasing, and now I’m wondering how much of prompt engineering is just learning which switches you’re flipping by accident.
Anyway, has anyone else noticed specific wording that changes how the model behaves, even if the question isn’t that different?
I would greatly appreciate any advice on how you frame your prompts and how you manage them. Thanks in advance!