Most disappointing model outputs trace back to one of three missing pieces: the model wasn't told who to be, wasn't given the context it needed to answer well, or wasn't told what shape the answer should take. Fix those three and a huge share of "the AI just isn't that good" problems disappear - not because the model got smarter, but because the prompt stopped leaving so much to guesswork.
Role: telling the model who to be
A role instruction - "you are a senior tax accountant," "you are a blunt, no-fluff code reviewer" - narrows the model's behavior toward the tone, vocabulary, and judgment calls appropriate to that role, instead of the generic, hedge-everything voice it defaults to. Role prompting works because it activates a narrower, more consistent slice of the model's training: a tax-accountant persona is less likely to explain basic terms you already know and more likely to flag the specific thing an accountant would actually flag.
Context: giving the model what it needs to actually know
Context is the background information the model can't infer on its own - your company's specific policy, the previous three messages in a conversation, the actual document you want summarized, the constraints of your specific situation. The single most common cause of a wrong or generic answer isn't a bad model, it's a model correctly answering a question it was never given enough information to answer well. Pasting in the relevant document, policy, or prior context is frequently worth more than any clever phrasing.
Three inputs, one model, one shot at a good answer - each missing piece is a guess the model has to make on your behalf.
Format: telling the model what shape the answer should take
Even with the right role and context, a model left to choose its own output shape will often default to a hedge-y paragraph with a summary and a disclaimer. An explicit format instruction - a table with these columns, a JSON object with these keys, three bullet points and nothing else, a subject line under 60 characters - removes that ambiguity and, as a side effect, makes the output far easier to parse or drop directly into a downstream tool or document.
Putting the three together
A well-structured prompt usually reads in that order: who you are (role), what you need to know (context), what you're being asked to do, and what shape the answer should take (format). You don't need every piece for every prompt - a quick factual question doesn't need a role - but for anything you'll run more than once, writing all three explicitly is what separates a prompt that works today from one that works reliably next week too.