There is a genre of content that presents prompting as a secret art: magic phrases, formulas to memorise, “the one prompt that changes everything”. Most of it is noise. But something real sits underneath, and the reason prompting works is simple enough to explain in one paragraph — after which most of the folklore either makes obvious sense or obviously doesn’t.
A language model produces a probability distribution over the next token, given everything before it. The prompt is everything before it. So the prompt does not instruct the model in the way a function argument instructs a program. It selects a region of the space of possible texts. Write like a support ticket and you get the continuation of a support ticket. Write like a peer-reviewed methods section and you get the continuation of one.
A prompt is a conditioning statement, and generation is sampling from what remains. Every prompting technique that works, works because it narrows that region toward the text you wanted.
Specificity works because vagueness has many continuations
“Write about churn” is compatible with a blog post, a slide, a lament, a definition, a sales email. The model has to pick, and it picks whatever is most common in text following that kind of request — which is the blandest option, because bland is what such requests are usually followed by.
“Write 200 words for a data team’s weekly note explaining why churn rose in February, given these three numbers, ending with what we plan to check next” is compatible with far fewer texts. All of them are closer to what was wanted. The improvement did not come from persuading the model to try harder. It came from deleting most of the alternatives.
This gives a useful test for any prompt: how many quite different documents would satisfy it? If the answer is “lots”, the output will disappoint, and the fix is not a magic phrase but another constraint — audience, length, format, what to leave out.
Examples work better than descriptions
Showing two or three input/output pairs before the real input usually beats a paragraph describing the format. This surprises people, and it shouldn’t.
A described format has to be translated into behaviour, and the translation is itself a guess. A demonstrated format is already the pattern. When the prompt contains three items that look exactly alike, the likeliest continuation is a fourth item that looks exactly alike. Nothing is being taught; the model’s weights do not change. The examples simply make one kind of continuation overwhelmingly more probable than the rest.
Two practical consequences. Examples should cover the edges you care about — including the awkward ones, since a model shown three clean cases will handle a messy fourth as though it were clean. And examples are expensive in tokens, so they earn their place on structured, repeated tasks and rarely on one-off prose.
Why role framing shifts output
“You are a senior epidemiologist” is not a switch that loads expertise. The model has no expertise to load beyond what its weights already contain.
What it does is select a register. Text written by epidemiologists uses different vocabulary, hedges differently, cites different things, and worries about different failure modes than text written by generalists. Conditioning on that framing shifts the distribution toward that corpus. The model does not become more accurate about epidemiology; it becomes more likely to produce text with the characteristics of expert epidemiology writing.
Sometimes those two things line up — expert-styled text about a well-covered subject is often more correct, because careful writing and correct writing travel together in the training data. Sometimes they come apart badly, and you get fluent, technical, wrong. Role framing is a style control that correlates with quality. It is not a knowledge control. Be clear about that difference before you trust an answer because it sounded qualified.
Most “prompt tricks” are just clearer instructions
Run the popular techniques through the conditioning lens and they sort themselves quickly.
| Technique | What it really is |
|---|---|
| ”Think step by step” | Conditions toward worked reasoning, where intermediate steps constrain the final one. Real effect on multi-step problems, mostly noise on one-step ones. |
| Assigning a role | Register selection. Style, not knowledge. |
| Few-shot examples | Pattern demonstration. The strongest lever for structured output. |
| Output format spec | Removes a whole axis of ambiguity for near-zero cost. |
| Offering a tip, threatening, pleading | Folklore. Effects are small, unstable across models, and vanish on re-test. |
| ”You are the world’s best expert” | Superlatives condition toward marketing copy as often as toward quality. |
The pattern is hard to miss. The techniques that survive are the ones that give the model more information about what is wanted. The ones that don’t are attempts to motivate something that has no motivation.
Step-by-step reasoning deserves its own note, since it is the one technique that does more than clarify. Each generated token is conditioned on the tokens before it, so writing out an intermediate result makes that result available as context for the next step. The model gets to use its own output as working memory. That is a genuine computational change, not just a stylistic one — and it is why the gains show up on arithmetic and multi-hop questions rather than on “summarise this”.
The most useful of the practical prompt-formula videos is Master the Perfect ChatGPT Prompt Formula by Jeff Su, which is essentially a checklist for supplying the constraints described above.
None of this makes prompting a skill worth putting on a CV. It makes it a special case of something older and more valuable: writing a brief that a competent stranger could execute without asking a follow-up question. Anyone who can do that for a colleague can do it for a model — the same discipline that separates a useful analysis request from a vague one, as covered in communicating results to stakeholders. The model is just a stranger who will never ask the follow-up question, and will confidently guess instead.