How to Write a Good Prompt

The four ingredients of a good prompt

The AI does not know your situation. Information a person would pick up from your face or the earlier conversation does not get through unless it is written down. That is why a short question gets the most average, flattest answer. Writing a sentence or two for each of the four items below makes a visible difference. You do not have to fill them all, but when an answer misses the mark, start by checking which one was missing.

• Context: who you are and what situation you are in
• Goal: what you will do with the result
• Format: length, tone, number of items, table or not
• Constraints: what must be included and what must be left out

Describe the situation, not just a role

A well-known tip is to assign a role such as "you are a world-class expert", but on its own it often changes little. What works better is telling it who will read the result and what it is for. Compare "write a meeting notice" with "a meeting notice for ten new hires attending for the first time; include the room and what to bring; five lines at most". The second gives it grounds to decide what to stress and what to drop. Knowing the reader sets the tone and the level.

Examples speed things up

If the shape you want is hard to describe, attach one or two examples. A short sample marked "like this" aligns tone, length and structure in one go: show one product blurb you like when asking for several, or fill in the first row yourself when asking for a table. It may copy the example too closely, though, and drift into similar content, so add "use this for format only and write new content".

Break big jobs into steps

Ask for a whole report in one go and you tend to get something polished but hollow. Have it produce an outline first, edit the outline yourself, then have it write one section at a time; if the direction goes wrong, you can correct it midway. Complex problems work the same way: start with "list what needs to be checked first" and proceed one step at a time, and mistakes drop. Splitting the work also makes it much easier to find which part went wrong. Another option is to ask it to summarise how it understood your request before writing anything. If it misunderstood, you can correct it before wading through a long answer.

How to ask for revisions

Disliking the first answer is normal. Repeating "try again" just produces more of the same, so point at the problem: "too long, make it three sentences", "drop the jargon so a thirteen-year-old could follow", "change the example in the second paragraph to one from our industry". When a conversation runs very long, conditions set early on can fade, so it is often better to take a clean list of requirements and start a fresh conversation.

Give it the material and keep it inside

When summarising a long document or asking about what it says, paste the original and add "answer only from this material; if it is not in there, say so". That reduces the risk of invented content creeping in. Asking it to quote the sentences each answer rests on makes checking against the source easy. Marking the beginning and end of the material keeps your question and the text from blurring together. First, though, make sure the material contains no personal data or company secrets.

Rewriting a flat question

A few extra words change the answer. Here are short requests on the left with context and format added.

• "Give me a diet plan" → "a week of dinners, as a table, for an office worker who eats lunch in the company canteen and only cooks in the evening"
• "Write an email" → "an apology to a client contact that delivery will be three days late, offering two alternative dates, polite but short"
• "Explain Python" → "five things to learn first for an office worker who only knows spreadsheets and wants to automate repetitive tasks"

Common mistakes

However well you write the prompt, you still need to check whether the answer is true. Prompting is the skill of fitting the answer to your situation, not a guarantee of accuracy. These mistakes come up often.

• Packing several requests into one sentence and not noticing one was skipped
• Setting a word count and never counting the result
• Believing that writing "don't make mistakes" makes it accurate
• Assuming conditions from earlier in the chat still apply