Took this from reddit
"Tell ChatGPT you'll tip $200 and it performs 10x better."
"Threaten AI models for stronger outputs."
"Use psychology-framed feedback instead of saying 'that's wrong.'"
These claims are everywhere right now. So I tested them.
200 tasks. GPT-5.2 and Claude Sonnet 4.5. ~4,000 pairwise comparisons. Six prompting styles: neutral, blunt negative, psychological encouragement, threats, bribes, and emotional appeals.
The winner? Plain neutral prompting. Every single time.
Threats scored the worst (24–25% win rate vs neutral). Bribes, flattery, emotional appeals all made outputs worse, not better.
Did a quick survey of other research papers and they found the same thing.
Why? Those extra tokens are noise.
The model doesn't care if you "believe in it" or offer $200. It needs clear instructions, not motivation.
Stop prompting AI like it's a person. Every token should help specify what you want. That's it.
full write up: https://keon.kim/writing/prompt-psychology-myth/
Code: https://github.com/keon/prompt-psychology
