GUIDE

How to write a good AI prompt

Most disappointing AI answers start with a short request that leaves the important parts unsaid. The AI fills every gap with a guess, and guesses drift toward the generic. A good prompt closes those gaps. It does not need magic words; it needs the right information, stated plainly.

Here are the six things a good prompt tells the AI, one request rewritten from start to finish, and the mistakes that most often undo an otherwise good prompt.

Create my prompt Describe your goal in a sentence. Mint Prompts writes all six parts for you.

Updated

The six parts of a good prompt

A request like “write an email to a client about a delayed project” states a goal and leaves everything else to chance. These six parts close the gaps. A sentence or two for each is usually enough.

  1. Goal. What should exist, or be different, when the AI is done. Name the outcome, not just the task: “Tell a client their project will arrive late, and keep their confidence in the work.”
  2. Context. Who you are, who the answer is for, and the facts the AI cannot know: names, dates, numbers, what already exists. “I manage the project. The original date was [date]; the new date is [date].”
  3. Must-haves. The points the answer has to include, in the order that matters: “State the delay in the first two sentences, give the new date, and say what is already finished.”
  4. Limits. Tone, length and what to leave out, including what the AI must not invent: “Warm and accountable. No excuses, no jargon, no promises beyond the new date.”
  5. Format. The shape you will actually use, whether that is an email, a table, a checklist or a one-page plan: “A subject line and an email under 150 words, ready to send.”
  6. Self-check. How the AI should check its work before it answers: “Before you finish, confirm the delay, the new date and the next step are each stated once.”

The same six parts work for code, research, plans and emails. What changes is the detail inside each one.

One request, before and after

Before
Write an email to a client about a delayed project.
After
Goal: Write an email telling a client their project will arrive later than agreed, and keep their confidence in the work.

Context: I manage the project. The original date was [date]; the new date is [date]. The cause is [brief reason]. Use only these facts and leave a placeholder for anything I have not supplied.

Must-haves: State the delay in the first two sentences. Give the reason in one plain sentence without blaming anyone. Give the new date, what is already finished, and one step we are taking to protect that date.

Limits: Professional, warm and accountable. No excuses, no jargon, no promises beyond the new date. Do not invent dates, causes or compensation.

Format: A subject line and an email under 150 words, ready to send.

Self-check: Before finishing, confirm the delay, the new date and the next step are each stated once and clearly.

The first version leaves the AI to guess the client, the tone, the dates, the cause and the length. The second answers each of those questions, so the AI can spend its effort on the writing. Both are illustrative examples.

Five habits that make any prompt better

  1. Label what you paste. Keep your instructions apart from the material you paste, and say what the material is (“Here is the client’s brief:”) so the AI knows what it is looking at.
  2. Give facts, and forbid inventions. Supply the details that matter and tell the AI to mark anything missing as a placeholder rather than make it up.
  3. Show, don’t describe. One short example of output you like teaches tone and format faster than a paragraph of adjectives.
  4. Ask for the format you will use. If you need a table, say so and name the columns. If the answer has to fit somewhere, give the length.
  5. Ask it to check itself. A closing instruction to compare the answer with your requirements catches missed points before you see them.

Common prompt mistakes

  • Instructions that pull against each other, such as “be brief” and “cover every detail”. Decide which one wins.
  • Quality words with no definition, such as “professional”, “engaging” or “better”. Say what they mean for this job.
  • Asking for everything at once. A plan, a draft and a review in one prompt usually makes all three thin; ask for them in steps.
  • Leaving out what already exists, so the AI starts from nothing or contradicts earlier work.
  • No stated audience, so the answer is pitched at no one in particular.

Why the wording matters

Researchers keep measuring the same effect: what you tell a model changes what you get back. Adding one line of instruction (“Let’s think step by step”) took one model from 17.7% to 78.7% correct on a benchmark of arithmetic word problems (Kojima et al., NeurIPS 2022). Formatting the same prompt in different ways moved one open model’s accuracy by up to 76 points (Sclar et al., ICLR 2024).

Models have changed since. Today’s reasoning models work through problems on their own, and OpenAI now calls “think step by step” instructions unnecessary for them (OpenAI, reasoning best practices). The lesson that survives is the one this guide is about: the information in the prompt shapes the answer.

These are independent studies of prompting in general. They did not test Mint Prompts, and results vary by model and task.

Or describe it once and let Mint Prompts write it

If writing six parts for every request sounds like work, that is the job Mint Prompts does. Describe your goal in a sentence and Mint Prompts writes the goal, context, must-haves, limits, format and self-check for you, ready to paste into the AI you already use. Advanced asks a few questions about the gaps that would change the result first.

Already have a prompt that is not working? Use the prompt improver.

Questions

How long should a prompt be?

Long enough to state the six parts, and no longer. Most people write too little: Google’s Workspace prompting guide says the most fruitful prompts it sees average around 21 words with relevant context, while the prompts people try are usually under nine (Google, 2024). For code or a plan, a page is normal. Length helps only when every line tells the AI something it needs.

Do I need magic phrases, or to tell the AI it is an expert?

No phrase replaces missing information. Giving the AI a role can help set its expertise and tone, and Anthropic and Google both suggest one, but a role works on top of a clear goal and context, not instead of them.

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

A prompt with the six parts works in any capable assistant. The vendors’ own guides differ in the details, such as how many examples to include and where long material goes: see how to prompt ChatGPT, Claude and Gemini.

What if I do not know all the details yet?

Say so. Ask the AI to list what it needs before it starts, or to mark missing facts as placeholders. In Mint Prompts, Advanced asks about the gaps that would change the result before it writes your prompt.

Sources