What is a prompt in an LLM?

A prompt is simply the input you give an LLM — the question, instruction, or text you type in to get a response. It's how you tell the model what you want. That sounds obvious, but the prompt does more than ask a question; it sets the entire context the model uses to generate its answer.

A vague prompt tends to get a vague answer, while a clear, specific prompt with the right details and examples usually gets a far better one.

What actually goes into a prompt?

The text you type is only part of it. Most apps quietly wrap your words in extra context before the model ever sees them, and all of it together forms the prompt the model reads:

• Your input — the actual question or instruction you type.
• A system message — hidden rules the app sets, like "you are a helpful support agent, stay polite."
• The recent conversation — earlier turns, so the model can follow up in context.
• Retrieved or attached material — documents, files, or search results the app pulls in.

The model has no memory of you beyond this. Everything it knows about your request comes from the prompt in front of it.

Why does prompt wording matter so much?

Because the LLM has no idea what's in your head. It can't guess the format you want, the audience you're writing for, or the level of detail you need unless you say so. That's why prompt engineering became a real skill: small changes in how you phrase a request can noticeably change the output.

The clearer and more specific your prompt, the less the model has to guess — and the closer the result lands to what you actually wanted.

How do you write a better prompt?

A few reliable techniques go a long way. Give the model a role ("act as a financial analyst") to set its tone and perspective. Show a couple of examples of what a good answer looks like, so it can copy the pattern.

Spell out the format you expect — a table, bullet points, three short paragraphs. And for anything complex, break the task into explicit steps instead of asking for everything at once. None of this is magic; it's just removing ambiguity so the model spends its effort answering rather than guessing what you meant.