GPT prompts that work: brief it like a colleague

A good prompt is a short brief, not a secret formula. Here is what GPT needs from you, how to turn a vague request into a clear one, and which follow-up lines rescue an answer that missed.

6 min readUpdated

Most disappointing AI answers come from short questions that leave the model guessing. GPT fills every gap you leave with the most average choice: an average reader, an average length, an average tone. The fix is not a clever phrase. It is the same short brief you would give a new colleague who is smart, fast and knows nothing about your situation. This guide shows what goes into that brief, how to rebuild weak prompts, and what to type when an answer misses. Everything here works in our free GPT chat with GPT-6 Luna or GPT-6.1 Terra.

The five-part brief

OpenAI's own guide to the GPT-6 family suggests starting every assignment with the result you want, who it is for, the relevant context and constraints, and what counts as done. In everyday terms that gives five questions to answer before you press send:

  1. Result. What should exist at the end? A list, an email, a plan, a table, a decision?
  2. Reader. Who will read or use it? Your manager, a ten-year-old, a customer who is already annoyed?
  3. Material. What does the model need to see? Paste the text, the data, the error message or your notes.
  4. Limits. Length, tone, words to avoid, budget, deadline, things that must stay unchanged.
  5. Done when. How will you judge the answer? "Fits on one screen", "every claim has a number", "ends with one clear next step".

You rarely need all five in full sentences. A line each is plenty, and for a quick question one or two of them is enough. The point is to notice which gap the model would otherwise fill on its own.

Weak prompts, rebuilt

Here are three requests people send every day, and what they look like once the gaps are closed.

Weak:
Complain to the shop about my kettle.

Strong:
Draft a calm complaint to an online shop. The kettle I bought from
them stopped heating after three weeks, order number [number]. Their
returns page promises a replacement within 30 days. Ask for a free
replacement and a prepaid return label. About 100 words, friendly
but clear, no legal threats.
Weak:
Explain inflation.

Strong:
Explain inflation to a 15-year-old who gets weekly pocket money.
Use one everyday example with prices, then say in two sentences
why central banks raise interest rates. No jargon, under 200 words.
Weak:
Give me some business ideas.

Strong:
I can bake well, have about 500 euros to start and work weekends
only. Suggest five small food businesses I could test within a
month in a mid-sized town. For each: first step, rough cost, and
the biggest risk. Put it in a table.

Notice what changed. The strong versions are longer, but every extra word removes a guess. You can try the kettle example directly: open it in the chat and change the details to match your own case. If you are unsure which model to use, the GPT-6 Luna page explains why the fast default suits tasks like this.

Paste the material, do not describe it

"Make my cover letter better" gives GPT nothing to work with. "Here is my cover letter, make it shorter and more specific" followed by the actual letter gives it everything. The same goes for code, contracts, meeting notes and spreadsheets: the model is far better at working on real text than at imagining what your text might say.

A few habits help when you paste:

  • Put your instructions first and the material after a clear divider, such as a line of three dashes or a label like TEXT:.
  • Say what to leave alone: "keep the numbers and names exactly as they are".
  • Remove private details you do not need to share. Our privacy and safety guide shows quick ways to do that.
  • For a photo or screenshot, use the + button in the composer and ask a specific question about it. The guide to asking about images has examples.

Ask for a shape

Telling GPT how the answer should look is one of the cheapest improvements you can make. Some shapes worth asking for by name:

  • A table with named columns, for comparisons and plans.
  • A numbered list of steps, for instructions you will follow one at a time.
  • Three versions of a headline, subject line or opening sentence, so you can pick.
  • A one-line answer first, then the explanation, when you are in a hurry.
  • Questions back to you: "Before you write anything, ask me up to five questions you need answered." This is useful when you are not sure yourself what the brief should contain.

Length limits work better as concrete targets ("about 150 words", "five bullet points") than as vague wishes like "short" or "detailed".

Follow-ups that fix a miss

The first answer is a draft. Instead of starting over, tell the model exactly what was wrong. Useful follow-up lines:

  • "Too formal. Rewrite it the way I would say it to a colleague I like."
  • "Cut it by half without losing the deadline and the price."
  • "The second point is wrong: we close at 18:00, not 20:00. Fix only that."
  • "Give me two very different alternatives, not small variations."
  • "Which part of your answer are you least sure about?"

Our chat gives you two more ways to steer. Regenerate asks for a fresh answer to the same message and keeps the earlier ones, so you can flip between versions with the arrows and keep the best. Edit on your own message lets you change the prompt and resend it, which is often cleaner than adding a correction at the end of a long chat.

Shortcuts built into gpt.me

You do not have to write every brief from scratch.

  • The prompt library has ready-made prompts by topic, from writing to career. Insert one into the composer and fill in the parts in square brackets.
  • Assistants come with their own instructions for jobs like editing, tutoring or coding, so you can skip the setup lines. You can also create your own assistant with a brief you reuse often.
  • The AI tools are single-purpose pages, such as the email writer or the summarizer, where the brief is already built into the form.
  • For a hard problem with many conditions, press Think in the composer to switch to GPT-6.1 Terra. Our page on reasoning models explains when that is worth it.
Quick test before you send: read your prompt as if you were a stranger. If you would have to ask "for whom?", "how long?" or "based on what?", add that line now. It takes ten seconds and saves a round of corrections.

FAQ

Is there a magic phrase that makes GPT answer better?

No. Phrases such as "you are a world expert" have little effect compared with clear facts about the task. A concrete reader, real material and a defined result matter far more than any wording trick.

Should I write "think step by step"?

For GPT-6.1 Terra there is little point: OpenAI's guidance for reasoning models says they already work through problems internally, so that line is unnecessary. If you want to check the logic, ask for the steps to be shown in the answer instead.

Do long prompts use up more of my free messages?

No. Each message you send counts once, whatever its length. Very long chats are trimmed, though: the chat sends the model the most recent part of the conversation, so restate key facts if a chat has grown long. The FAQ explains the daily allowance.

Can I write prompts in my own language?

Yes. GPT answers in the language you write in, and the same five-part brief works in any of them. If you want the answer in another language, say so in the prompt.

Where can I keep prompts I use often?

Create your own assistant with that brief as its instructions, or keep the prompt in a note and paste it when needed. Assistants you create stay in your browser along with your chat history.

Got a question? Ask it now

No account and no download. Open the chat, type what is on your mind, and the answer starts appearing in a second or two.

Start chatting