Myths & anti-patterns

The research does not support a courtesy bonus. A cross-lingual study found that rude prompts often do hurt performance, but that extra politeness gives no guaranteed gain, and that the optimum differs by language. A 2025 follow-up on multiple-choice accuracy found the opposite of the folk wisdom: very polite prompts scored 80.8% against 84.8% for very rude ones. Neither finding is a reason to be rude — the honest conclusion is that politeness is a small, inconsistent effect, and clarity is the variable that actually moves results.

Roles work, and every major vendor documents them. Superlatives are the part that does not. "World-class" and "award-winning" contain no information a model can act on — they do not name a discipline, a seniority, an audience or a constraint. Anthropic frames a useful role as telling the model what it is accountable for. Compare "a world-class copywriter" with "a copywriter who writes onboarding email for people who have already paid and are now confused": only one of those changes a single word of the output.

This is the clearest reversal in current vendor guidance. OpenAI documents that removing repeated instructions and duplicated examples improves both performance and efficiency, and gives a concrete failure mode: repeating rules such as "ask first", "do not mutate" or "wait for approval" causes the model to issue unnecessary approval requests. Repetition also crowds out the instructions that were only stated once. Say it clearly, in one place.

Genuinely excellent advice in 2022; largely spent today. Current frontier models reason internally by default. Independent testing found the accuracy gain from explicit step-by-step prompting on reasoning models to be small at best while adding 20–80% to response time, and medical-domain work found that chain-of-thought changed a large number of answers and hurt more of them than it helped. Google advises simplifying the prompt and raising thinking_level instead of engineering the reasoning yourself.

Incentive and pressure prompts circulated widely on the back of anecdotal 2023 testing and have never appeared in the prompting documentation of any provider on this site. Google goes further and advises against persuasive language outright, telling you to state the goal plainly. If you want more effort, there is a documented way to ask: Anthropic recommends stating it directly — request the thorough version rather than hoping the model infers urgency from a fictional stake.

Length is not the variable; specificity is. OpenAI states that shorter, clearer instructions often perform better on its current models and that they need less scaffolding than older ones. Google asks for concise input prompts. Anthropic asks for clear and direct ones. A long prompt fails in two ways at once: it buries the instructions that matter among ones that do not, and it usually contains contradictions you cannot see because you wrote them ten lines apart.

A prompt with six deliverables gets a budget-split answer: each item receives a sixth of the attention and none of them is finished. It also removes your ability to diagnose failure, because a single weak section forces you to re-run everything. Anthropic documents chaining complex prompts as the alternative, and it is one of the few techniques that improves quality and debuggability at the same time.

Tone adjectives are the single least reliable part of most prompts, because they mean different things to the model, to you, and to your colleague reviewing the draft. OpenAI’s current guidance is explicit: replace vague descriptors with concrete writing choices — whether to state answers directly, when to acknowledge a problem, what to do with hedging.

Fluency and accuracy are produced by the same mechanism and are not correlated in the way a human reader assumes. A model states a fabricated statute, a wrong figure or a non-existent citation in exactly the register it uses for a correct one. This is why CLEAR ends on Reflective, why GCSE has a Source slot, and why every vendor documents grounding techniques. Treat confident output about facts, law, medicine, money or people as a draft to verify, never as an answer.

A model’s self-report about its own version, cutoff date, context window or capabilities is generated text like any other, and it is frequently wrong — a model may name a cutoff earlier or later than its actual one, or describe features it does not have. Anthropic addresses this directly under the heading of model self-knowledge. The vendor documentation is the authority; the model is not a source about itself.