Prompt

DefinitionExample
On this pageDefinition
Definition
The text an AI model reads before it answers: the task, the background and the rules, which together decide how useful the answer is.

Why it matters

A model does not know your business. It only has the words in front of it. A vague prompt such as "write a follow-up email" produces a generic email. A prompt that names the customer, what was agreed on the last call, the tone the company uses and the one action wanted produces something close to sendable. Most disappointing AI output comes from missing information, not from a weak model.

How to apply it

Build a prompt from four parts and check each one is present.

  • The task, stated in one clear sentence.
  • The context the model cannot guess, such as who the reader is and what happened before.
  • The limits, such as length, tone and things to avoid.
  • The shape of the answer, such as a table, three bullet points or a structured format another system can read.

Add one or two examples of good output when style matters. Then run the prompt on several real cases, not just one, and fix the prompt rather than editing each answer by hand.

What it is

A prompt is everything the model reads before it answers. It can be one line, such as "summarise this email", or several pages holding a role, background documents, examples and rules. In a chat app the prompt is what you type. In a product built on AI, a builder usually writes the prompt once and fills in fresh data each time it runs.

Common mistakes

  • Writing instructions that contradict each other, so the model has to guess which one wins.
  • Assuming the model remembers earlier chats. Each new conversation starts empty unless the context is supplied again.
  • Pasting a whole document and asking a vague question about it.
  • Treating a prompt as finished after one good result. A prompt that works once may fail on the next customer.
Worked example

Suppose a recruitment firm wants a model to draft follow-up emails after candidate calls. A prompt that says only "write a follow-up email" returns a generic message that a recruiter must rewrite. The team rebuilds the prompt in Claude with four parts: the task in one sentence; the candidate's name, role and what was agreed on the call; the limits of 120 words, a warm tone and no salary figures; and the shape of the answer, a subject line followed by the body. They test it on ten real call notes, not one. Seven drafts need only a light edit. The three weak ones share a fault: the notes never recorded a next step, so the prompt now asks for that to be flagged when it is missing.

Tools in the example

Some links are affiliate links: we may earn a commission at no cost to you. It never decides a ranking. How we work with partners