Prompt

What it means

A prompt is an instruction or question for a language model that sets out what it should do, in what context and in what form it should answer. It can be a simple question or a detailed brief with a role, goal, source material, examples and the required output format.

The quality of a large language model's output depends heavily on the precision of the prompt and the material supplied. A vague request produces a generic answer; a clear one with specific data produces something usable. Writing good prompts is sometimes called prompt engineering.

Prompts are how a company works with generative AI when drafting text, summarising documents, analysing data or writing code. With AI agents, the prompt becomes part of the system instructions that govern how the agent acts and which tools it may use.

How it works in practice

A good prompt includes context, such as who the reader is and what the text is for, a specific task, constraints like length or tone, and the output format. Examples of the desired result help, as does telling the model to admit when it is unsure.

A company benefits from a library of tested prompts for recurring tasks, such as drafting a reply to a client, summarising a contract or checking a text. It ensures consistent quality and makes onboarding easier.

Model output must always be checked, especially on specialist subjects like tax, accounting or law. A model can cite a non-existent section or an outdated rate, in other words hallucinate, and the person who uses the result is responsible for its accuracy.

What to watch out for

Clients' personal data, trade secrets and confidential documents must not be pasted into prompts without thought. Public tools may use the content for training or store it outside the EU; a company should have rules on which tools may be used for which data.

Prompts can also be attacked. In systems that process external text, such as emails or web pages, a hidden instruction in the text can change the model's behaviour. Automation must therefore limit what the model can do without human review.