How to Write AI Prompts That Get Better Answers

How to write AI prompts that get better answers

There is a lot of mystique around prompt writing and almost none of it is deserved. A prompt is a brief. If you have ever handed work to a freelancer, you already know how to do this.

The four things every prompt needs

Not every prompt needs all four. Most bad prompts are missing at least two.

ElementWhat it answersExample
TaskWhat am I doing?Rewrite this email
ContextWhat do you know that I don’t?The client is annoyed about a missed deadline
ConstraintsWhat are the limits?Under 120 words, no apology in the first line
FormatWhat should it look like?Plain paragraphs, no bullet points

Start with the outcome, not the instruction

Most people write prompts describing an activity: “write”, “summarise”, “explain”. Better prompts describe a finished thing: “a 200-word summary a busy manager could read in 30 seconds and know whether to act.”

The second version contains the standard the output will be judged against. That is the part the model cannot invent.

Context is the part people skip

The model has no idea who you are, who is reading, or what happened last week. Every fact you withhold becomes an assumption it makes for you.

A useful test: if you handed this prompt to a competent stranger with no other information, could they do the job? If not, the missing pieces are exactly what you need to add.

Constraints do more work than adjectives

“Professional”, “engaging” and “high quality” mean nothing operationally. Numbers and prohibitions mean everything.

  • Not “keep it short” — “maximum 150 words”
  • Not “make it engaging” — “open with the specific problem, not background”
  • Not “be professional” — “no exclamation marks, no rhetorical questions”

How to fix a bad answer

The instinct is to start a new chat. Resist it. The context you have built is an asset.

  1. Name what is wrong. “Too generic” is not actionable. “It describes the feature but never says who it is for” is.
  2. Give the missing information. Nine times out of ten a wrong answer is a context problem, not a model problem.
  3. Ask for a revision, not a rewrite. “Keep the structure, replace the second section” preserves what worked.

What does not help

  • Threats and bribes. “You will be penalised” and “I will tip you $200” have no reliable effect.
  • Stacking roles. “You are an expert marketer, copywriter, psychologist and SEO specialist” is worse than one clear role.
  • Copying prompts blind. A prompt tuned for someone else’s context will not fit yours. Take the structure, replace the content.

A template that works for most tasks

Task: [what you want]
Audience: [who reads it]
Context: [what the model cannot know]
Constraints: [length, tone, things to avoid]
Format: [structure]
Before answering, ask me anything that is unclear.

That last line is doing more work than the rest combined.

Common questions

How long should a prompt be?

As long as it takes to remove the guesswork, and no longer. For a quick rewrite, one line. For a document that matters, a paragraph of context is normal and worth the typing.

Should I use one long chat or many short ones?

One chat per task. Context helps within a task and hurts across unrelated ones — a long thread about your holiday plans will bleed into a work question asked in the same window.

Do I need to learn “prompt engineering”?

No. You need to learn to write a clear brief, which is a more useful skill anyway and transfers to working with people.