How to Write a Prompt That Works in ChatGPT

You asked for “write an email chasing a client for payment” and got something correct, polite and completely useless: too generic to send, too formal for your client, and missing every detail that would have made the chase work. So you rewrote it by hand and decided AI is not good for this.

It is. What was missing was information, not capability. This is lesson 2 of the course, and by the end of it you will know how to build a request that returns something usable on the first try, and how to steer it in the second message instead of starting over. About 10 minutes of reading, with a ready prompt to test at the end of each section.

The golden rule: it answers what you wrote

In lesson 1 we settled that AI does not know, it estimates. The practical consequence for prompting is this: it answers what you wrote, not what you meant.

When you write “write an email chasing a client for payment”, the AI has to fill about fifteen gaps on its own: who you are, who the client is, how overdue the payment is, whether this is the first chase or the fourth, whether you want to keep the relationship or have already given up, whether the tone is partnership or final notice. It fills those with the average of what it has seen. Average is exactly what sounds generic.

A good prompt is not a long prompt or a pretty prompt. It is a prompt that closes the gaps that matter and leaves open only what you genuinely do not know.

The five pieces of a prompt that works

Almost every good prompt has the same five pieces. You do not always need all five, but when the result comes out bad, it is nearly always one of them that was missing:

Piece	What it answers	Short example
Role	Where the AI speaks from	“You are an employment lawyer”
Context	What it has no way of knowing	“I am a sole trader billing $8k a month”
Task	What to do, with a verb	“Summarise”, “Compare”, “Rewrite”
Format	The shape of the answer	“As a table, with 3 columns”
Constraint	What to avoid	“No jargon, 200 words maximum”

The next sections take one piece at a time, each with a prompt from the catalogue that uses it well so you can see it working.

Role: why “act as” changes the answer so much

Giving a role is not theatre. It is the cheapest way to pick the vocabulary, the level of detail and what the AI will treat as relevant. “Explain compound interest” and “you are a high school maths teacher, explain compound interest” produce different texts because the second sentence selects a repertoire.

A role works better when it describes how that person works, not just the job title. Compare “you are a teacher” with “you are a teacher who always starts from an everyday example before showing the formula”. The second version gives an instruction about method, and that is what changes the output.

Two ready examples that live on this: the private tutor for any subject, which sets level and pace before teaching anything, and the kind proofreader who explains your mistakes, where the role carries an attitude (correct without humiliating) that changes the whole text.

Context: what it cannot possibly guess

This is the piece that goes missing most often. The AI does not know your job title, the size of your company, what you earn, who you are talking to or what has already gone wrong before. All of that changes the answer, and it has none of it.

A quick test: read your prompt pretending to be a freelancer who joined the project this morning and has never spoken to you. If they would need to ask questions before starting, those questions are the context you left out.

Notice how the CV from scratch with no experience starts by digging up what you have already done outside formal employment, instead of writing straight away. And dealing with a difficult boss or coworker only works because it asks for the concrete situation before suggesting a sentence, since the same advice helps or backfires depending on who is on the other side.

Format: ask for the shape along with the content

Most frustration with AI is not wrong content: it is right content in a shape you cannot use. You wanted three lines and got seven paragraphs. You wanted a table and got prose.

Saying the shape is cheap and fixes it immediately. “In bullet points”, “as a table with columns X, Y and Z”, “5 lines maximum”, “in the tone of a WhatsApp message”, “give me three options to choose from”. Asking for options, by the way, usually pays better than asking for the answer: you compare instead of accepting.

The meeting summary is a good model for this piece, because it splits decisions, open items and owners into fixed blocks, and that is what makes the summary usable. In the same vein, the caption with hook, body and CTA and the LinkedIn post with authority impose a structure before a single word is written.

Constraint: saying what you do not want

Constraint is the piece almost nobody writes and the one that saves the most time. AI tends toward the middle: average text, average length, average vocabulary. Constraint is what pulls it out of the middle.

The ones that work best day to day are length (“200 words maximum”), vocabulary (“no technical terms, no English loanwords”), attitude (“do not invent data I did not give you”) and scope (“do not suggest anything that costs money”). That third one is especially useful, because it is an explicit request for the AI to say “I do not know” instead of filling the gap with a plausible invention.

Look at explain it like I am 12, which is a vocabulary constraint turned into a whole prompt, and the medicine leaflet translator, which restricts scope on purpose so it does not turn into medical advice.

The second message is where the prompt gets good

Here is the habit that most separates people who find AI useful from people who find it disappointing: the first answer is a draft, not a delivery.

People who give up rewrite the prompt from scratch. People who get value reply in the same conversation: “too formal, redo it in a conversational tone”, “cut it in half”, “the second option is the best one, develop only that”, “you invented the number in the third line, take it out”. The AI keeps everything already said and adjusts only what you pointed at, which costs a sentence instead of a new prompt.

The gift consultant was written for exactly this rhythm, because the first suggestions exist so that you can say what does not fit. The chef of whatever is in the fridge follows the same logic: you remove what you do not like and the list improves each round.

Text prompts and image prompts do not follow the same rule

Worth spelling out, because it confuses a lot of people. Everything above is about text prompts, the kind that becomes an email, a summary, a plan or an explanation.

Image prompts run on another logic: instead of role and context, what matters is framing, light, lens, style and what you want preserved from the original photo. A conversational sentence tends to produce a poor image, and a stacked visual description tends to produce a good one. If that is what you are after, the path is the guide to image prompts, which deals only with that and with examples.

When a careful prompt does not fix it

Prompting does not fix everything, and insisting costs more time than doing it by hand. Three situations where the missing piece is not the wording:

• When the data is not in the conversation. If you did not paste the contract, no prompt makes the AI know what it says. Paste the text or attach the file, and then be careful with the request.
• When the answer has to be right and you cannot check it. A medicine dose, a legal deadline, a severance calculation. The AI answers with the same confidence whether or not it has a basis, and no prompt changes that.
• When it is arithmetic, not text. A language model handles words. For numbers, ask for the formula and run it in a spreadsheet, or use a tool that actually calculates.

There is also the excess in the other direction: a two-page prompt with fifteen constraints usually performs worse than six well written lines, because too many instructions compete with each other. If your prompt is turning into a document, there are probably two tasks in there that should be two conversations.

Build yours in five minutes

Take a real task of yours and follow the order. The idea is to leave this lesson with a prompt of your own, not with theory:

1. Write the task in one sentence, starting with a verb. “Rewrite”, “compare”, “summarise”, “list”. If you cannot pick the verb, the task is still fuzzy in your head, not in the AI’s.
2. Add the context only you have. Apply the freelancer test: what would they need to ask before starting?
3. Set the role, with method. Not just the job title, but how that person works.
4. Say the shape. Length, structure, tone, and ask for options rather than a single answer.
5. Add at least one constraint. If in doubt, use “do not invent information I did not provide”.
6. Run it and correct in the second message. Point at what went wrong instead of rewriting everything.

If you prefer to start from something ready and modify it, you get to see the structure before having to invent it. The used item listing that sells fast and the impartial decision advisor are good ones to dissect, because the five pieces are visible in the prompt text itself.

Quick questions

How long should a ChatGPT prompt be?

There is no magic number. The criterion is different: the prompt has to close the gaps that change the answer and no more than that. In practice six to twelve lines handle almost everything day to day.

Do I need to be polite to the AI?

It makes no difference to the result. “Please” and “thank you” neither improve nor worsen the answer. Write in whichever way is more comfortable for you.

Does the same prompt work in ChatGPT, Gemini and Claude?

Most of the time yes, with small differences in tone and answer length. The five pieces hold for all three. When the result comes out unexpected, it is usually a matter of adjusting the format, not rewriting everything.

Why does the same question give different answers?

Because the answer is generated by probability, not looked up in a table. Two runs of the same request follow slightly different paths. If you need consistency, ask for a fixed format and constrain the length.

Is it worth asking the AI to improve my prompt?

It is, and it is underrated. Paste your request and ask: “what information is missing here for you to give me a better answer?”. It usually lists exactly the gaps you had not noticed.

How do I stop ChatGPT from making things up?

You cannot eliminate it, you can reduce it. Provide the material instead of expecting it to know, ask explicitly for it to say “I do not know” when it has no basis, and check any number, date or quotation before using it.

Do I have to write prompts in English?

No. For text prompts the result is equivalent in your own language, and writing in it helps you be specific, which is what actually matters.

What comes next

You already have the five pieces and the second-message habit, which together fix most bad results. What is left is repetition, and the fastest way to get practice is to start from a ready prompt and modify it until it is yours.

Start with the text prompts, which is where everything in this lesson applies directly, or see the other lessons in Courses. In lesson 3 we compare ChatGPT, Gemini and Claude task by task, so you know which one to open in each situation instead of guessing.