Why good prompts still give bad answers
You wrote a careful prompt. Told the model what you wanted, gave it a role, set a word count, asked for a table. Everything the guides tell you to do.
The answer came back polished and useless.
So you rewrote it. Added a sentence. Moved the constraint to the front. Tried “act as an experienced” instead of “you are an expert.” Fourth attempt, fifth attempt, and each one costs you a few megabytes and a bit of battery you weren’t planning to spend on this.
The rewriting was never going to work. Not because your sentences were bad, but because the problem sat somewhere your sentences couldn’t reach.
The prompt that had everything
Suppose you’re preparing a course outline for an accreditation visit. You write this:
Act as an experienced academic quality assurance consultant. Review a mass communication course outline and identify gaps against national accreditation requirements. Present findings as a table with severity ratings. Be direct and specific.
Look at what’s in there. A task. A role. A format. A constraint on tone. If you have read anything at all about prompting, you have been taught that this is a good prompt, and by the standard those guides use, it is.
It will produce something worthless.
Not wrong, exactly. It will give you a tidy table of generic gaps that any course outline might have, written with total confidence, some of which will apply to your document and some of which will not, and you will have no way to tell which is which without going through your outline line by line, which is the work you were trying to avoid.
The model has never seen your outline. It has never seen the benchmark statement you are being assessed against. You asked it to compare two documents and gave it neither.
No wording fixes that. You could spend an hour polishing that prompt and the ceiling would not move, because the ceiling isn’t made of words.
What actually changes the answer
Attach the outline. Attach the benchmark. Then ask:
Attached are two documents: a mass communication course outline, and the NUC benchmark statement for mass communication. Compare them.
List every requirement in the benchmark that the outline does not clearly address. For each gap, quote the exact line of the benchmark it comes from, and tell me where in the outline it would most naturally fit.
That prompt is plainer than the first one. No role. No expertise claim. It reads less like prompt engineering and more like a note you’d leave a colleague.
It works because the model now has the two things it was being asked to reason about. The instruction to quote the source line matters too: it makes the answer checkable, so you can verify a claim in seconds rather than trusting it.
Why the myth survives
Because rewording occasionally works, and when it does you remember it.
Sometimes a request genuinely is too vague, and sharpening it helps. That’s real. It’s also the minority of cases, and it’s the only category of failure the popular advice knows how to talk about. So people learn one repair technique and apply it to every breakdown, including the ones where it can’t possibly help.
The other reason is that supplying context feels like less of a skill. Pasting a document isn’t clever. There’s no formula to memorise, nothing to screenshot and share. It just works better, which is a much harder thing to build a viral post around.
And the model never tells you. It doesn’t say “I have no idea what’s in the document you’re describing.” It answers anyway, in the same confident register it uses when it does know, which is the single most expensive habit these tools have.
What to do when an answer disappoints
Before rewriting anything, ask one question: could this have been answered correctly by someone who has never seen my situation?
If no, then the fix is information, not phrasing. Attach the file. Paste the email thread. Give it three examples of what good looks like. Tell it the budget, the deadline, the audience, the thing that makes your case different from the default case it assumed.
If yes, if the question really was general and the answer is still poor, then sharpen the instruction. That’s when rewording earns its place.
Most of the time it’s the first one.
The part that isn’t just efficiency
Everything above sounds like productivity advice, and where data is cheap, that’s all it is. Nobody in London counts the cost of a fifth attempt.
I don’t write from there. On a metered connection, every regenerated answer is paid for twice: once in data and once in the battery you’re spending during an outage you didn’t schedule. A habit that wastes four attempts to reach the answer a first attempt could have produced isn’t merely inefficient. It’s a bill.
That’s why this particular misconception is worth taking apart properly rather than filing it under tips. Getting it right is not about elegance. It’s about not paying for the same answer five times.
This is drawn from Chapter 4 of Stop Prompting Like It’s 2023, a practical guide to getting usable answers from ChatGPT, Claude and Gemini in fewer attempts.