Why Wording Changes the Answer
You Are Steering, Not Just Asking
Ask "what are the risks of this plan?" and you get risks. Ask "is this plan good?" and you get a broadly positive assessment with a few caveats. Ask "what would a sceptical finance director say about this plan?" and you get sharp, specific objections about cost and assumptions. Same plan, same model, three genuinely different outputs — and none of them is wrong. Your phrasing does not merely describe the request; it sets the direction the answer travels in. This is why two people can use the same tool on the same problem and come away with completely different impressions of how useful it is. The one getting better answers is usually asking better-aimed questions, not using a better product.
- "Is this good?" invites agreement; "what breaks here?" invites analysis
- Naming a perspective changes vocabulary, depth and what gets prioritised
- Vague questions do not produce neutral answers, they produce generic ones
- If you want criticism, you have to ask for criticism explicitly
Leading Questions Get Leading Answers
Compare two versions of the same request. Weak: "Explain why remote work is more productive." Strong: "Summarise the main arguments and evidence on both sides of whether remote work improves productivity, and say where the evidence is genuinely contested." The first one has the conclusion baked into the question, so you will get a fluent, one-sided case that feels like research and is actually just your own assumption returned to you with better sentences. This is a real trap for anyone using AI to think rather than to write. If you notice the answer agreeing with you suspiciously often, look at how you phrased the question before you congratulate yourself.
- Weak: "Explain why X is better" — assumes the conclusion you wanted
- Strong: "Give the strongest case for and against X, and note where evidence is weak"
- Ask for the counter-argument by name, not as an afterthought
- A useful habit: ask the same question twice, once phrased the opposite way
Some Variation Is Just Built In
Even with an identical prompt, you will not always get an identical answer. Text generation involves an element of sampling, so wording, structure and emphasis shift between runs. Most products expose this indirectly rather than as a setting you control. Two things follow. First, do not read too much into one output — a single bad answer may be a bad roll rather than a bad prompt, and a single great answer is not proof your prompt is reliable. Second, if you need consistency across many uses, that comes from a tighter prompt with an explicit format, not from hoping. Ask for the same structure every time and the variation lands in the wording rather than in the shape of the result.
- The same prompt can produce different text on different runs — this is normal
- Judge a prompt on several attempts, not on the first one you liked
- Specifying an explicit output format is the main lever for consistency
- Regenerating is a legitimate move, but it is not a substitute for a fix
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.