How to Write a Good Prompt: A Plain Guide for Working People

Step one happens before you type anything. Decide what the finished result should look like: an email you can send, a one-page summary your manager will read, a list of questions for a supplier call. If you cannot describe the result in one sentence, the assistant cannot guess it either, and you will spend the next few exchanges discovering what you wanted.

Step two is context. Tell the assistant who you are in this situation, who the output is for, and what that person already knows. 'I coordinate deliveries for a furniture company' and 'the reader is a purchasing manager who is annoyed' change the answer more than any clever phrasing. Context is the part of a prompt people skip most often, because it feels obvious to them. It is never obvious to a tool that has only your message to work from.

Step three is the task, stated as one clear action. Use a verb that names the work: summarize, rewrite, compare, draft, list, classify, explain. Avoid stacking five jobs into one sentence. If you need a summary, then a reply, then a slide outline, ask for them in order or number them, so each one gets proper attention.

A useful test for steps two and three together: imagine handing your prompt to a capable colleague who has never worked with you, and ask whether they could start right away. If they would come back with three questions, put the answers to those questions into the prompt. The test works for any AI assistant, because the missing information is the same for a person and for a tool.

Here is a common case: you coordinate deliveries and a shipment is going to arrive late. A weak prompt reads like this: write me an email to a client about the delay, make it good and professional, not too long. The assistant must now guess who the client is, how long the delay is, what caused it, how upset the reader is, what you can offer, and what 'good' means. It fills every gap with a safe, generic guess, and you receive a polite, vague email that could have come from any company.

The rewrite adds, in this order: your role and the reader's role, the facts (the new date and the reason), the reader's situation (her team moves into a new office on Monday), the tone (apologetic but calm, no excuses), the format and length (under 150 words, with a subject line), and two constraints (offer the two delivery options and end with a question). Each addition answers a question the assistant would otherwise have to guess. The rewritten prompt is longer, and the result is shorter, more specific, and much closer to something you can send.

Notice what the rewrite does not do. It does not use magic phrases, and it does not ask the assistant to act as the world's best copywriter. It supplies information. Both versions appear in the prompt examples on this page, so you can paste them one after the other into whichever assistant you use and compare the results yourself. Doing this once with your own tool teaches more than any description.

Then comes the follow-up. Suppose the first draft is good but the second paragraph sounds defensive. Instead of rewriting the whole prompt, you reply with a precise correction, and the assistant changes only what you named. A decent first prompt plus one or two sharp follow-ups is what working with an AI assistant usually looks like in practice.

Swap the details for your own situation and the structure still holds. A late monthly report for your manager, a supplier invoice that needs approval, a meeting you must reschedule with five busy people: each one has a role, a reader, a few facts, a tone, a length, and a constraint. Once you see that pattern, you can write better prompts for almost any task without memorizing templates.

A reliable prompt structure is a short sequence: context, task, format and length, then example or constraint. You can write it as one flowing paragraph or as four short lines with labels. Labels such as 'Context:', 'Task:', 'Format:', and 'Rules:' help when the prompt is long or when you plan to reuse it, because you can edit one part without rereading everything.

How long should a prompt be? As long as it takes to remove the guesses, and no longer. A request for a synonym needs one line. An email to an upset client or a summary of a long document usually needs a short paragraph or two. A prompt rarely fails because it is long and every sentence adds a fact, a rule, or an example. It fails when it is padded with repetition, apologies, or instructions that contradict each other.

Keep your instructions consistent. If you say 'be brief' in one place and 'cover everything in detail' in another, the assistant has to pick one, and it may not pick the one you meant. When you paste in source material such as meeting notes, a policy, or a customer message, mark clearly where the material starts and ends, and say what you want done with it.

Do prompts work the same in ChatGPT, Copilot, Gemini, and Claude? The habits do: context, a clear task, a format, and iteration help in all of them. The details differ, because each tool has its own interface, its own ways of attaching files, and settings that change from time to time. Results also vary between tools and even between runs of the same prompt. Learn the method once, then test it in the tool you actually use at work and adjust to what you see.

Begin with the Strategic Prompting lesson 'Context Is Everything', which is 5 minutes long. It covers the ingredient that changes results the most, and it matches step two of the method above. The same week, take one task from your job and rewrite your last prompt for it so that it says who you are, who the output is for, and what the reader already knows.

Next, in the same course, take 'Claude Prompting Fundamentals for Everyday Work', which is also 5 minutes long. It uses Anthropic's Claude as the example assistant and focuses on prompting for ordinary working tasks, and the habits carry over to other tools. Afterwards, choose one recurring task, such as a weekly status update or a customer reply, and write a reusable version of the prompt with labelled parts: context, task, format, and rules.

Then move to Master the Essentials and take 'The Iteration Loop'. It speaks to step six: improving an answer through targeted follow-ups instead of restarting. Put it into practice by finding an answer you were not happy with, and improve it through a few specific follow-up messages. Notice how much faster that is than writing a new prompt from scratch.

After that, the rest of the Master the Essentials course has free lessons on Document Analysis and Research and Thinking, which apply the same habits to longer material. If your workplace uses a different assistant, Nightschool AI also has first-week courses for ChatGPT, Microsoft 365 Copilot, and Google Gemini, each with free lessons. The ChatGPT course, for example, opens with 'Getting an answer you can actually use'. Pick whichever matches the tool you open every day.

AI output can be fluent and still wrong. Before anything reaches a colleague or client, check the names, dates, numbers, quotations, and any claim about a policy, a contract, or a regulation against the original source. Then read the text as the recipient would: does the tone fit, and does any sentence promise something you have not confirmed? For summaries, compare a few key points with the source material rather than trusting the whole.

You can also ask the assistant to list the claims in its own draft that deserve verification, as in the checking prompt above. That is a useful way to find what to look at, but it is not a substitute for looking. The assistant can miss its own errors, so the final check belongs to you.

A few weeks of steady practice will give you a method, noticeably better first drafts, faster iteration, and a sense of which tasks suit an assistant and which do not. It will not make you an expert in a field the assistant writes about, it will not replace your judgment, and it will not produce identical results every time you send the same prompt. Treat it as a working skill you build gradually, not a guarantee about any outcome.

Be careful about what you paste in. Do not share confidential documents, personal data about customers or colleagues, or anything your organization does not allow you to put into an AI tool. Check your workplace's rules, remove names and identifying details where you can, and remember that you remain responsible for whatever you send under your own name.