Phrases that feel like instructions but aren't
Some sentences look like guidance and don't change behavior — they just spend tokens.
- "Be helpful." The model is already trying to be helpful. This adds nothing, and worse, it pushes toward generic, hedge-y answers.
- "Think carefully." On a non-reasoning model, this prompts a chain-of-thought that you're not capturing. On a reasoning model, the model is already thinking; you've just told it to think more, costing tokens.
- "Take your time." The model has no time. This is anthropomorphism that produces longer prose without higher quality.
- "You are an expert." Sometimes useful for tone, but not for accuracy. See the persona lesson.
- "Do not hallucinate." Hallucination is not a switch. The model can't comply by trying harder; you need boundary conditions and grounding.
What to do instead
Replace each anti-pattern with a positive specification of the behavior you actually want. "Think carefully" → "Show your reasoning in <thinking> tags before answering." "Be helpful" → "Open with the actionable next step." "Do not hallucinate" → "Cite each claim with a source from <docs>; if no source exists, say so."