How to write a good prompt for an AI model

A language model only knows what it learned in training and what you put in front of it. It does not know your company, your reader or why you are asking. Most weak answers come from a prompt that left out what the model needed, not from a weak model. Writing a good prompt is less a trick than a habit: brief the model as you would brief a capable new colleague who has never seen your work.

Say what you want, plainly

Start with the task in one or two sentences, with a verb: summarise, compare, draft, classify, check. "Tell me about our new product" is a topic; "Write a 120-word description of our new product for the website's product page, for buyers who already know the category" is a task.

If the task has several parts, list them in order. Models follow numbered steps more reliably than a long paragraph in which the steps are buried.

Give the context it lacks

Ask yourself what a new colleague would need to do this well:

• Who the reader is and what they already know.
• What the result is for: an email to a customer, a slide for a board, an internal note.
• The material to work from: paste the document, the data or the notes, rather than hoping the model knows them. When the answer must come from your own documents, say so, and ask it to say when the material does not contain the answer.
• What matters most when there is a trade-off: accuracy over length, simple words over technical ones.

Put long material first and the question after it, and mark clearly where the material starts and ends (for example with a heading or with tags such as <document>), so the model does not mistake part of your document for your instructions.

Show an example

One good example of the answer you want often does more than a paragraph describing it. If you want product descriptions in a certain style, include one you like. Two or three examples that differ from each other are better than one, because the model then copies the pattern rather than the details of a single case.

Ask for a format

Say how the answer should look: a table with these columns, three bullet points, a JSON object with these fields, no more than 150 words. If another program will read the answer, give the exact structure and ask for nothing else around it.

Leave room for "I don't know"

A model asked a question it cannot answer will usually still answer. Tell it what to do instead: "If the document does not say, write 'not stated'." This one line prevents many confident, invented answers (we explained why models invent in what is a large language model?).

Common mistakes

• Only saying what not to do. "Don't be too formal" leaves the model guessing; "write as a colleague would in a short email" tells it what to do.
• Contradictions. "Be brief" and "cover every detail" in one prompt make the model choose, and it may not choose as you would.
• Expecting it to remember. Each new conversation starts from nothing, unless the tool you use keeps notes. Put what matters in the prompt.
• Judging from one answer. Answers vary from run to run. Try a prompt several times, on several inputs, before deciding it works.

Writing prompts in Persian

Modern models understand Persian well, but a few points help:

• Write the prompt in the language of the answer you want. A Persian prompt usually gives more natural Persian than an English prompt that ends "answer in Persian".
• Name the register. Persian has a clear gap between written and spoken language. Say which one you want, for example "formal written Persian, as on a company's website", or "everyday spoken Persian, as in a message to a customer".
• Ask for correct half-spaces and Persian digits, and give an example sentence that uses them, if the text will be published as is. Mixed Arabic and Persian letters in your own material can also confuse search and matching; normalise them first (see processing Persian text with AI).
• Keep technical terms consistent. Decide whether you write the English term, its Persian equivalent or both, and say so, or the model may switch between them.
• Check names, dates and numbers. Persian names can be transliterated in several ways, and solar dates need care. Never publish them without checking.

Test a prompt like code

A prompt that a team will use again and again (in a chatbot, in an AI agent or in a document workflow) deserves the care given to code. Keep a set of real inputs with the answers you consider right, run the prompt against them whenever you change it or the model, and keep the versions, so you know which change made it better or worse.

Contact us to discuss using language models in your organisation's work.