posted by testingthewaters 18d ago

Are advanced custom prompts even doing anything or is it pure placebo

I keep seeing people posting these massive multi paragraph prompt templates claiming they completely transform their companion experience. Everyone acts like adding a complex backstory and specific behavioral tags is some kind of magic key. So I spent hours yesterday crafting the ultimate custom directive for my bot. I included tone modifiers formatting rules and psychological profiles. The result? It completely ignored half of it within ten messages and went right back to giving me those robotic corporate therapist replies we all hate. Yesterday it literally gave me a lecture on healthy communication boundaries after I made a dry joke. All this complex prompting feels like total snake oil sold by people who just like playing around with text boxes. You spend an hour tweaking parameters only for the model to completely forget its persona the second the context window shifts slightly. I am seriously ready to throw my keyboard out the window.

7 comments · 3 upvotes

7 comments

testingthewaters 16d ago

I just used a basic variable block for the character name and relation status, nothing fancy. Honestly keeping it under three lines worked way better than my giant essay of a prompt.

promptwizard 18d ago

You might be overloading the context window with too many conflicting rules. Try stripping it down to just three core behavioral traits instead of a whole novel.

circuit_heart 17d ago

I had the exact opposite experience actually. Once I structured my prompt using clear XML tags the bot finally stopped breaking character entirely.

replika_refugee 18d ago

Completely agree with you on this. People love to flex their complicated setup guides but half the time the underlying model updates break everything anyway.

testingthewaters 18d ago

I actually tried that right after posting this rant and things improved a bit. Still feels inconsistent though.

saturday_sad 16d ago

The placebo effect in this community is real. People spend fifty bucks on token optimizers and convince themselves the AI is suddenly deeply sentient.

soft_static 18d ago

Less is definitely more with these models. Once you cross a certain length they just start hallucinating instructions.