Longer prompts are not automatically better. Once the goal, the key context and the hard requirements are stated, extra content quickly stops helping and starts hurting: conflicting instructions make the model follow whichever is easiest, buried priorities get lost, and mismatched examples pull the answer off course. The test is simple: if deleting a paragraph changes nothing, that paragraph never belonged there.
Where overlong prompts go wrong
First, attention gets spread thin. Pack a dozen requirements into one prompt and each gets executed at a discount; the classic symptom is correct formatting with missing content, or a strong start where the final requirement is quietly forgotten. The more demands there are, the lower the chance any single one is taken seriously. That is not attitude, it is how the mechanism works.
Second, self-contradiction. Long prompts are usually layers of edits: be brief near the top, expand in detail further down; one tone set early, pasted examples in a completely different style later. Faced with a conflict, the model does not ask you; it picks one at random or averages them, which is where unstable output usually comes from.
Third, example backfire. Examples meant to teach format get absorbed as content: facts, figures and tone from the examples leak into the real task. Give three examples and the model may write an example's company name into your final draft, or imitate an example's length and skim the material it was actually asked to process.
Delete these three kinds first
Kind one: repetition. Saying the same thing three ways does not make the model value it more; it only eats space. Keep the most specific version and delete the rest, especially pure emphasis like please make sure, this is very important, never forget, which carries no information at all.
Kind two: background unrelated to this task. Who you are, what your company does, the whole backstory: keep only what genuinely changes the output. Pasting a page of company introduction to get a leave-request email written is the textbook case of dead weight.
Kind three: old requirements that newer ones override. Every time you revise a prompt, subtract first: remove formats, tones and length rules that no longer apply, then add the new. A prompt is like a suitcase; only packing and never removing ends in a burst zipper.
How long is right
There is no universal word count, so judge by subtraction: delete paragraphs from the end backwards, keep deleting any whose removal changes nothing, and stop when the next deletion would lose real information. What remains is the right length for that task. Everyday writing, editing and Q&A usually need a few sentences to two or three paragraphs. Genuinely long prompts belong to genuinely rule-heavy work, like fixed report formats or multi-condition checklists, and even then the requirements should be numbered so each can be checked off. One last self-check before sending: read it once as the executor and ask, could I tell in one pass what to do first and what not to do? If you get tangled reading it yourself, do not expect stable performance from the model.
