How to ask an AI tool for what you want, in plain English

Short answer: tell the tool what you want, who it is for, what it should look like, and what to leave out, in that order, in ordinary sentences. That is the whole method. There is a word for it, prompting, and a great deal of advice attached to it, but the advice reduces to writing the kind of request you would hand to a capable new colleague who has no idea what your business does. This guide gives the short version first, then the steps, then the things that usually go wrong. It is written from the vendors' own documentation, checked October 7, 2026, and applies to any of the chat-style assistants, whichever one you happen to use.

The short answer

A good request has four parts. What you want, in one sentence: a reply to this email, a summary of this report, a list of questions to ask a plumber. Who it is for: a customer, your manager, a nine-year-old, yourself. What it should look like: three bullet points, one paragraph, a table, under 100 words. What to leave out or avoid: no jargon, do not invent prices, do not mention the delay. Anthropic's prompting documentation, read October 7, 2026, puts the same idea in one line: think of the assistant as a brilliant but new employee who lacks context on your norms and workflows, and the more precisely you explain what you want, the better the result. That is the test for every request below: would a new colleague understand it without asking a question back?

Step by step

• 1. Say the task first. Start with the verb: write, summarise, list, rewrite, compare, explain.
• 2. Paste the material. If the task is about a document, paste the text or attach the file; the tool cannot see your screen.
• 3. Name the audience. Who will read the result decides tone, length and vocabulary.
• 4. Describe the shape. Bullet points or prose, how long, which language, any headings you need.
• 5. Add the one thing it must not do. Invent a fact, use a word you dislike, go over a length.
• 6. Read the result and reply with a correction instead of starting again. 'Shorter, and drop the second point' works.

Each step is one sentence, so the whole request is five or six lines. Longer is fine when the job is bigger, but a request that is long because it is vague is worse than a short one that is precise. The order matters a little: the tool reads the task first and interprets everything after it in that light, so the verb goes at the top and the material goes directly under it.

A before and after

Before: 'Can you help me with this email?' followed by a pasted message. The tool will guess whether you want a summary, a reply, a tone check or a translation, and it will guess politely, which is the same as guessing wrong half the time. After: 'Write a reply to this email from a customer asking for a refund. The refund will be paid in full. Keep it to four sentences, friendly but not apologetic, and do not promise a delivery date.' The second request takes thirty seconds longer to write and produces something usable on the first try. The difference is not clever wording. It is that the second request contains the four parts, and the first contains none of them.

Give the reason, not just the rule

One instruction that surprises people: telling the tool why you want something improves the result. The Anthropic documentation states that providing context or motivation behind your instructions, such as explaining why a behaviour is important, helps the assistant deliver more targeted responses. In plain terms, 'keep it short because this is going in a text message' gives better results than 'keep it short', because the tool now knows what short is for and can make the other small decisions that follow from it. The same page suggests a check it calls the golden rule: show your prompt to a colleague with minimal context on the task and ask them to follow it; if they would be confused, the assistant will be too. That check costs nothing and catches most bad requests before they are sent.

What can go wrong

The result is confident and wrong. Assistants will produce a figure, a quotation or a citation that does not exist, in the same tone as a true one. Ask for the source and check it; if the task involves numbers, paste the numbers in rather than asking the tool to remember them. The result ignores a constraint. Usually the constraint was buried in the middle of a long request; move it to its own line at the end, or reply with just the constraint. The result is the wrong length. Length instructions work better as a count ('under 120 words', 'three bullets') than as an adjective ('brief'). The result is in the wrong voice. Say who is speaking: 'write this as the owner of a small bakery', not 'make it friendly'. And the tool asks for a file it cannot see: most assistants cannot open a link on your screen or read a window; paste the text.

When a request is not enough

Some jobs need more than a better sentence. If the same request is made every week, most assistants let you save the standing instructions once so they do not need repeating; the vendors call these projects, Gems, custom instructions or similar, and the names change. If the task depends on a set of documents, tools that let you upload those documents and answer only from them are more reliable than pasting excerpts. And if the result must be exactly right, a contract, a dosage, a tax figure, the assistant is a drafting aid and not the final check; a person who knows the subject reads it before it leaves. None of that replaces the four-part request. It sits on top of it.

Bottom line

Say the task, the audience, the shape and the thing to avoid, in plain sentences, and paste in whatever the tool needs to see. Give the reason when there is one. Read the result as a draft from a capable newcomer, and reply with corrections rather than starting over. That method is the same across the assistants, it is what the vendors' own documentation recommends as of October 2026, and it costs about thirty seconds more than the one-line request that produces a guess.