A good prompt is a good brief. Tell ChatGPT what you need, who it is for and why, give it the material to work from, describe what a good result looks like, and say what to avoid. That is five parts, and it works the same way in Claude, Gemini and Copilot. There is no secret wording. OpenAI's own guidance comes down to the same thing: be clear and specific, give enough context, then refine based on what you get back.
Most bad answers come from short requests. "Write an email to my landlord" leaves the AI to guess everything that matters, so it guesses the most average version. Below is the template, real before-and-after examples, and what to do when the answer is still wrong.
The 5-part template
| Part | What to write | Example |
|---|---|---|
| 1. Task | what you want, as one clear action | "Write a reply to this customer complaint." |
| 2. Context | who it is for, why, and what you want to happen | "She is a regular customer. I want to keep her and offer a fix, not a refund." |
| 3. Material | what it should work from, pasted in full | the customer's email, your refund policy, the order details |
| 4. Good result | format, length, tone, an example if you have one | "Under 120 words, warm but direct, sign off as Anna." |
| 5. Limits | what to avoid or never do | "Don't promise delivery dates. Don't mention the supplier." |
You do not need all five every time. A quick question needs only the task. But the moment a result matters, the parts you leave out are exactly where the answer goes wrong.
A copy-ready version:
Task: [what you need, in one sentence]
Context: [who it is for, why, what should happen as a result]
Material: [paste the text, data or notes it should use]
Good result: [length, format, tone; paste an example you like if you have one]
Avoid: [things it must not do, say or invent]
Before you start, tell me in three sentences how you understand the task.
That last line is the most useful one in the whole template. If its summary is wrong, you fix the misunderstanding before it writes a single paragraph.
Before and after: four real tasks
1. An email
- Weak: "Write an email to my landlord about the heating."
- Strong: "Write a short email to my landlord. The heating has been broken for 6 days, I reported it on the 3rd and got no reply. I want a repair date this week. Firm but polite, under 150 words. Don't threaten legal action yet."
The weak version gets a generic, overly apologetic letter. The strong one gets something you can send.
2. A summary
- Weak: "Summarise this report."
- Strong: "Summarise this report for my manager, who has two minutes. Three bullet points: what changed, what it costs, what we need to decide. Then one line on the biggest risk. Use only what is in the report."
"Use only what is in the report" matters. It reduces the chance that the AI fills gaps with things that sound plausible but are not there.
3. A decision
- Weak: "Which of these two quotes is better?"
- Strong: "Compare these two supplier quotes in a table: price, delivery time, payment terms, warranty, anything unusual in the small print. My priority is reliability over price. Don't recommend one; just make the differences clear."
Asking it not to recommend keeps the decision with you, where it belongs.
4. Something to learn
- Weak: "Explain VAT."
- Strong: "I'm starting a small online shop in Germany selling to customers in the EU. Explain in plain language what I need to know about VAT in the first year. Assume I know nothing. End with five questions I should ask an accountant."
The context turns a textbook answer into one about your actual situation.
Six habits that matter more than wording
Paste the real material. The single biggest improvement. An AI working from your actual email, contract or numbers beats one working from your description of them.
Show an example. If you have a past email, post or report you liked, paste it and say "in this style". One example explains tone better than ten adjectives.
Name the reader. "For a customer who is annoyed", "for my boss who hates long emails", "for a 12-year-old". The reader decides length, words and tone.
Use plain tone words. OpenAI suggests descriptive words such as formal, friendly, professional or serious. Pick one or two, not five.
Let it ask you questions. For anything bigger, end with: "Before you start, ask me up to five questions that would most improve the result." It often asks what you forgot to say.
One job per request. "Summarise this, then write a reply, then make a social post" gives three mediocre results. Do them one after another in the same chat.
When the answer is wrong: fix the specific thing
Do not start again with a new chat, and do not just say "make it better". Say exactly what is wrong:
| Instead of | Say |
| "Make it better." | "The second paragraph is too apologetic. Cut it to one sentence." |
| "Shorter." | "Keep it under 80 words. Keep the deadline and the offer, drop the rest." |
| "That's wrong." | "The price in line 3 should be $240, from the quote I pasted. Check the other numbers against it too." |
| "Sound more human." | "Use short sentences, no phrases like 'I hope this finds you well', and match this example of my writing: [paste]." |
That last row answers one of the most searched prompting questions: how to make ChatGPT write like a human. The reliable way is not a magic phrase. It is to show a sample of how you actually write and to name the phrases you never use.
Five things a prompt can't fix
- Missing facts. If the information is not in your material or widely known, the AI may invent it. Check numbers, names, dates and quotes.
- A task you can't judge. If you would not recognise a good result, a better prompt will not help you spot a bad one.
- Confidential data. No prompt makes it safe to paste passwords, card numbers or clients' personal data into a personal account. Remove them first.
- The latest information. Unless the tool searches the web and shows sources, treat anything time-sensitive with caution.
- Your judgement. For medical, legal, tax and money decisions, use AI to understand and prepare questions, then decide with a professional.
Save your best prompts
When a prompt works, save it. Keep a simple document with your five to ten most-used requests, written with the blanks in square brackets, as in the template above. After a month you will have your own set built around your work, which is worth more than any list of "100 best prompts" written for everyone.
If you use ChatGPT, you can also store instructions inside a project so you do not paste them every time. And if you are looking for ready prompts for paid work specifically, there are 12 in ChatGPT prompts to make money.
The money math
A clear prompt does not just give better answers. It cuts the back-and-forth. If a five-part brief saves you two rounds of corrections, about five minutes, on ten tasks a week, that is around 3.5 hours a month. At $30 an hour, roughly $100 of your time, from writing one extra paragraph at the start.
If you are still at the stage of finding which tasks to give AI in the first place, start with how to start using AI and the list of 20 tasks worth giving ChatGPT.
Does this work in Claude, Gemini and Copilot?
Yes. The template is about explaining a task clearly, not about one company's model. Anthropic's guidance for Claude makes the same point: be clear and direct, give context and examples, and say what you want. Minor differences exist, for example some tools handle very long documents better than others, but a good brief works everywhere. A comparison of the tools themselves is in ChatGPT vs Claude vs Gemini.
FAQ
How do I write a good prompt for ChatGPT?
Include five parts: the task, the context (who it is for and why), the material to work from, what a good result looks like (length, format, tone, an example) and what to avoid. Then ask it to restate the task before it starts.
Do I need to learn prompt engineering?
Not for everyday use. Clear explanation matters far more than special phrases. Prompt engineering as a job is a different, more technical field.
What is the best prompt template for ChatGPT?
Task, context, material, good result, avoid. Add "Before you start, tell me in three sentences how you understand the task" at the end. It works for emails, summaries, comparisons and learning.
How do I get better answers from ChatGPT?
Paste the real material, name the reader, show an example of the style you want, and correct specific problems instead of asking it to "make it better".
How do I make ChatGPT write like a human?
Paste a sample of your own writing, ask for short sentences, and list the phrases you never use. Specific instructions work better than asking it to "sound human".
Does the same prompt work in Claude and Gemini?
Yes. A clear brief with context, material and a description of a good result works in every major AI assistant.
Sources
- Prompt engineering best practices for ChatGPT (OpenAI Help Center)
- Prompt engineering overview (Anthropic documentation)
