Prompting Fundamentals That Actually Move Output

Most beginners fix a bad prompt by adding to it: a role, three more examples, a paragraph of rules, a magic phrase someone swore by. The prompt swells, the output gets worse rather than better, and the reflex is to add even more. This is the single most common self-inflicted wound in prompting, and it runs backwards from what a practiced operator does.

Two forces drive results, and neither one is volume. The first is specificity. A model cannot read your mind. When you ask for "a short professional email" it has to guess what short means, what professional means to you, and what the email is supposed to accomplish, and it fills every gap with the statistical average of everything it has seen. The fix is not more words. It is more specific words. "Three sentences, plain, no exclamation marks, asking the client to confirm the Friday deadline" removes the guesses. Notice that is shorter than a vague version padded with adjectives, not longer.

The second force is subtraction, and it is the part nobody teaches a beginner. Everything you put in a prompt competes for the model's attention. Every instruction you add can quietly conflict with one you wrote earlier and forgot. Past a point, more material does not sharpen the output. It dilutes it. This is not a motivational claim, it is measured. On a benchmark that packs hundreds of instructions into a single prompt, even the strongest current models top out around 68 percent instruction-following at high density, and they bias toward the instructions near the start of the prompt while the later ones quietly fall away. Sophistication does not save you either. More elaborate prompts sometimes score lower than plain ones on the same task. The prompt that finally works is often shorter than the draft you started with.

Examples are the sharpest version of this trap, because they feel so obviously helpful. Showing the model one good example of the output you want usually beats a paragraph describing it. So beginners reason that if one example helps, eight must help more. They do not. Example count has a saturation curve, and past the peak it does not plateau, it collapses.

That curve is real. A team testing a routing task scored 93 percent with zero examples, climbed a little as they added a few, then watched accuracy collapse to 30 percent by the eighth example. On a separate code-fixing task another model fell from 56 percent all the way to 0 percent as examples piled up. Order matters as much as count. Rearranging the same set of examples, changing nothing but the sequence, has swung accuracy on a classification task from near-chance to near-best, because the model over-weights whatever label it saw last and whatever label it saw most. Examples work like a precision instrument rather than a volume knob: a few that span the boundary you care about, including the hard or ambiguous case, beat a long list of easy ones.

So the operator habit is the inverse of the beginner reflex. When a prompt underperforms, get more specific about what you actually want, and strip out everything that is not pulling weight, before you add a single new thing.

Subtraction is not a universal law, and treating it as one fails just as hard as over-prompting. If the model is wrong because it genuinely does not know something, no amount of trimming or rephrasing fixes a knowledge gap. That is a job for giving it the actual information, which is a different lesson. Cut too far and you hit the opposite failure: starve the prompt of a constraint it really needed, and the model will not stop to ask. It will guess, confidently, and the guess looks exactly as polished as a correct answer, so you will not catch it by glancing at the output. There is also no gauge for any of this. Nothing tells you the moment a prompt tips from lean to starved or from rich to bloated. You read it from the results, which means you only know after you run it. And one more honest caveat. A model provider can quietly update the model underneath you, so a prompt you tuned last month can drift even though you never touched a word of it.