How do I fix a prompt that gives irrelevant answers?
Find the ambiguity first
Irrelevant answers usually mean the model filled a gap you left open. Read your prompt as a stranger would and underline every word with more than one meaning, like 'short,' 'professional,' or 'recent.'
Add the missing constraints: who the answer is for, what it should include, what it should leave out, and how long it should be. A prompt like 'Write about dogs' invites anything; 'Write a 150-word guide for new dog owners on crate training' does not.
If the topic is broad, give two or three examples of the kind of answer you want. Models copy patterns well, so a sample output often does more than another sentence of instruction.
Restructure and retest
Put the task at the top, then context, then constraints, then the output format. Long prompts get better results when the actual instruction is not buried in the middle.
Change one thing per attempt so you can tell what worked. If the answer drifts again, ask the model to restate your request in its own words before answering; that surfaces what it thinks you asked for.
- Name the audience: 'for a beginner,' 'for a hiring manager.'
- Name the format: bullet list, table, email, 3 paragraphs.
- Name the length: word count or sentence count.
- Name what to exclude: no jargon, no pricing, no speculation.
- Add one example of a good answer if the task is unusual.
Common mistakes
- Assuming the model knows your context; it only knows what is in the prompt or conversation.
- Rewriting the whole prompt after every bad answer, which hides which change actually helped.
- Using vague praise words like 'engaging' or 'high quality' instead of concrete requirements.
