How to ask AI for what you actually need

When an AI answer is no use, it is easy to assume you need to learn a new technique or memorise some special word. Often the problem is simpler: the tool does not know what result you are after, who is going to read it, or what an acceptable answer means to you.

You do not have to write like a programmer. A good request looks more like explaining a job to another person: you say what you want, give the context that is needed, state the format, and make clear what it should not do.

Short answer

Before writing, complete these four sentences:

1. “I want to achieve…”
2. “The situation is…”
3. “I need the result as…”
4. “It will be right if…”

That alone will get you a far more useful first answer than “do it better”.

1. Start from the result, not from the tool

Do not say “use your full potential”. Say what you want to have at the end:

• an email ready to review;
• a shopping list with criteria;
• an outline for a meeting;
• an explanation for someone non-technical;
• a plan with a clear first step.

A concrete result stops the conversation turning into a demonstration of the tool. If you do not know what result you want, ask it first to help you clarify the problem — but do not present that first answer as the final solution.

2. Give context that changes the answer

Context does not mean telling it your life story. It means giving the facts that change the decision:

• who will read the result;
• what experience they have;
• what time or budget there is;
• what has already been tried;
• which limits cannot be crossed.

Compare these two requests:

“Write a guide about backups.”

“Write a guide for someone who keeps their photos on their phone and has never made a backup. They have little time, want an understandable solution, and need to know what they could lose if they never check the backup.”

The second is not more technical. It is more specific about the person and the need.

3. Ask for a format you can review

An answer in flowing paragraphs can hide gaps. Ask for a structure that lets you check the work:

• a three-sentence summary first;
• then numbered steps;
• risks and open questions at the end;
• a table only where there really are comparable options;
• a list of claims that need a source.

You can also give an approximate length. “Between 700 and 900 words” is more useful than “not too long”, though you will still have to read it and decide whether the length makes sense.

4. Define what doing it well means

An AI does not know your standard unless you explain it. You can ask for:

• adult, clear language;
• no invented facts;
• examples kept separate from real facts;
• a warning wherever information is missing;
• an answer you can start using today, not an endless list.

If you are reviewing your own text, add: “Do not change my opinion and do not add experiences I have not had.” If you are preparing a recommendation, add: “Distinguish what was tested, what was researched, and what I cannot yet claim.”

A complete request, with no jargon

Here is an example you can adapt:

“I want to turn these notes into a practical guide for an adult who is not technical. The goal is that they can decide the first step without feeling lost. Structure it with a summary, five steps, common mistakes and a checklist. Do not add facts that are not in the notes; mark as pending anything that needs a source. Use a clear, direct and respectful tone.”

The request contains the goal, the audience, the structure, the tone and the standard of confidence. It does not need to say “act as an expert”.

Work in versions, not by magic

The first answer is a draft. A simple way to improve it is to review one dimension at a time:

1. Sense: does it answer the real problem?
2. Structure: is it clear what to do first?
3. Accuracy: are there facts that need checking?
4. Tone: does it sound like how you want to talk to that person?
5. Usefulness: could someone apply something today?

After each review, ask for one specific change. “Make it more human” helps less than “remove the grandiose sentences and keep the practical examples”.

OpenAI’s own guidance recommends writing clear, specific requests and refining them iteratively. The important idea is the process: there is no magic phrase that replaces reviewing.

When the request contains rules

Rules help if there are few of them and they do not contradict each other. Put your priorities in order:

1. Do not invent facts.
2. Solve the main need.
3. Keep the requested format.
4. Use a specific tone.

If you ask it at the same time to be exhaustive, extremely brief, informal, academic and to leave nothing out, the tool will have to decide which rule to break. Better to choose what matters most.

Ask it to ask you questions when something is missing

You do not always have to fill the gap yourself. You can say:

“Before drafting, ask me at most three questions that would change the answer. If a piece of information is not essential, carry on with a clearly marked assumption.”

That avoids two extremes: answering blind, or getting into an endless interview. If you do not want questions, state the limits and accept that some parts will stay open.

Check the answer without becoming a specialist

Reviewing does not mean knowing everything. Start by checking:

• whether it answered what you asked;
• whether it respected your limits;
• whether it added names, figures or dates you never gave it;
• whether the links exist and say what it claims;
• whether the advice could do harm if it is wrong.

When you do not know a fact, ask it to flag that and go and find a primary source. For important decisions, do not use the AI itself as your only means of verification.

Common mistakes

Adding instructions until the goal disappears

A request with twenty rules can bury the real need. Start with four facts and add only the rule that is missing.

Saying “more context” without defining it

Useful context is the context that changes the answer. Details that change nothing only make reviewing harder.

Giving an example that is too perfect

If the example does not resemble the result you need, the tool will learn the wrong direction. Use a realistic example, but with no private data.

Asking it to sound like you

The tool does not know your voice by default. Describe what you want kept: short sentences, a direct tone, your own examples, and no personal claim you have not lived.

What I would do

I would write the request in four sentences and ask for a first version. Then I would correct one single thing at a time. On my own website project, working this way showed me that “automate publishing” was far too broad: first I had to decide what content was real, what a person had to review, and which part the machinery could actually carry out.

Template to copy

“I want [result]. I need it for [person or situation]. The essential context is [facts]. Give it back to me as [format and length]. It will be useful if [standard of quality]. Do not invent [limits] and flag whatever is still to be checked.”

Security and privacy

Test the template with public or invented information. Do not put in passwords, keys, banking details, medical records, identifiable contracts or other people’s information. If the result feeds into a decision that matters, verify the facts separately and keep the original source.

Conclusion

The best answers do not depend on talking like an engineer. They depend on the tool receiving a request it can understand, and on you checking whether it delivered. Goal, context, format and quality: four simple things that give you more control.

Sources checked

• Best practices for prompt engineering in ChatGPT, OpenAI Help Center, checked on 2026-09-02.

Related

• Getting started with ChatGPT without learning prompts
• Information you should never give an AI

Review

Responsible for publication: Ibrahin. Date: 2026-09-02. Last checked: 2026-09-02.