Why a vague prompt gets a vague answer
A language model continues text. Given an ambiguous instruction it does not ask what you meant; it picks the most statistically ordinary reading and answers that. So write about remote work returns the most average possible article about remote work, and the model has done nothing wrong.
Prompting well is therefore mostly the work of removing choices. Every decision you leave open is one the model makes for you, and it will make it blandly. The [msm_tool slug="prompt-optimizer"] rewrites a rough instruction into one that closes those gaps.
The parts of a prompt that actually change the output
- Task. One verb, unambiguous. Summarise, compare, rewrite, classify. Two verbs in one prompt usually gets one of them done properly.
- Audience. Who reads this and what do they already know. The single highest-leverage line, and the one most often missing.
- Format. Prose, bullets, table, JSON, a specific number of items. Say it, or you get whatever shape is most common.
- Length. A word count or an item count. Brief means nothing.
- Constraints. What to avoid, what to keep verbatim, what must not be invented.
- Context. The source material, if the answer should come from your material rather than the model's training.
A prompt with all six is usually a short paragraph, not a sentence, and that paragraph is the difference between a usable answer and a paragraph of received wisdom.
Show it what good looks like
Describing a style rarely works as well as demonstrating it. Write in a plain, direct voice is interpretation; two example outputs are specification.
One example fixes the format. Two or three fix the tone and the level of detail as well. If you have an existing piece that got it right, pasting a paragraph of it as a sample is faster and more precise than any adjective you could choose.
This also solves the most common frustration: getting the shape right but the register wrong, over and over. That is almost always a missing example rather than a missing instruction.
Constraints that prevent the usual failures
- Ground it in your text. Answer only from the passage below; if it does not say, write "not stated". This is the most effective single line against invented detail.
- Forbid the padding. No introduction, no summary of what you are about to do, no closing paragraph. Saves an editing pass every time.
- Protect the specifics. Keep all figures, names and dates exactly as written. Necessary for any rewriting task.
- Ask for the uncertainty. Mark anything you are inferring rather than reading. Models comply with this more reliably than people expect.
- Set the failure mode. Tell it what to do when it cannot comply, or it will comply anyway with something invented.
Iterating without starting over
When an answer is nearly right, the instinct is to rewrite the whole prompt. Change one thing instead — otherwise you learn nothing about which change helped.
Diagnose before editing. Wrong shape means the format line is missing. Wrong depth means the audience line is. Right content in the wrong voice means you need an example, not more adjectives. Confident invented detail means you need the grounding constraint.
Keep prompts that work. A prompt that reliably produces what you want is a small piece of tooling and worth saving in a file, not retyped from memory each time with slight variations.
What better prompting cannot fix
No prompt gives a model information it does not have. It cannot read a link, see your database or know what happened last week, and asking more precisely will not change that — it will produce a more precisely worded guess.
Nor can a prompt make output trustworthy on facts. Only use accurate information reads like a constraint and is not one, because the model has no internal flag separating what it knows from what it is assembling. Verification stays your job regardless of how good the prompt was.
And prompts do not transfer perfectly between providers. The same instruction to different models gives different results, so a prompt that works beautifully on one is a starting point, not a finished artefact, on another.
The prompt you type goes from your browser to the provider you chose, under their terms, and never through this site.