We Tested the Prompting Folklore — 474 Answers Later, Nothing Had Changed

Tell the model it is an expert. Offer it a $200 tip. Threaten it. Say please. Add "let's think step by step". We ran all seven of those against a plain instruction, 480 times across two models, and scored every answer with a script rather than an AI judge. All 474 scored answers were correct, in every condition — because the plain instruction was already at 100%, and no trick can improve on that. Getting there took eleven task variants that all hit the ceiling, and a harness bug that briefly showed shouting at the model made it the only thing that worked. Our own prompt builder, meanwhile, was claiming a 30-60% accuracy gain on the authority of internal benchmarks we had never run.