When output is wrong, the reflex is to reword the prompt and try again. That works occasionally and teaches you nothing. A method is faster.
First, classify the failure
Wrong task. It answered a nearby question. Usually an ambiguous instruction, or an instruction buried at the end of a long block of context where it competes with everything above it. Move the instruction to the top, and state the goal in one sentence.
Right task, missing knowledge. It answered well but got facts wrong, or invented specifics. It did not have what it needed. No amount of rewording adds knowledge — supply the material.
Right content, wrong shape. Too long, wrong structure, wrong register. A format and constraints problem, not an instruction problem.
Inconsistent between runs. Same prompt, meaningfully different answers. Either the task is genuinely underspecified, or randomness is doing more than you want.
Ignored a rule. You said "at most three" and got five. Usually the rule was one clause inside a long paragraph. Rules that matter belong on their own line, phrased positively.
Positive rules beat negative ones
"Do not write an introduction" is weaker than "Start with the first finding." Negative instructions require the model to represent the forbidden thing in order to avoid it, and leave the space where it would have gone undefined. Say what should be there instead.
The same applies to the most common non-fix in the field: adding "do not hallucinate". It names a behaviour without changing what the model knows. Replace it with something actionable — "only use the text above" — which is checkable.
Randomness, briefly
Temperature controls how much the model varies its choices. Lower is more repeatable, higher more varied. If you need the same answer every time, lower it — but do not expect it to fix a wrong answer. A consistently wrong answer is still wrong. Temperature affects variety, not accuracy.
Failures no prompt will fix
Be willing to conclude the prompt is not the problem:
- Knowledge the model does not have. Anything after its training data, or private to you. Supply it or accept it is unavailable.
- Exact character-level work. Counting letters, precise reversal. It reads tokens, not characters.
- Arithmetic at length. Predicting plausible digits is not calculating. Use a calculator, or give the model one as a tool.
- Genuine ambiguity in your own request. If two people would read your instruction differently, no phrasing rescues it.
The habit worth keeping
Keep the prompts that work. Most people rewrite the same prompt dozens of times because the good version was in a chat window they closed. A saved prompt with a note on what it is for is worth more than another hour of rewording — which is what the Prompt Playground in this platform is for.