BlueAIBack to chat

BlueAI hub

How to write a prompt that gets a useful answer

Most disappointing AI answers come from the same handful of missing details. Here are six habits that fix them, each with the weak version and the better version written out.

· BlueAI team · about 6 minutes

1. Say who the answer is for

A model with no audience in mind writes for a general reader, which means it writes for nobody. Naming the reader changes the vocabulary, the length and the amount of assumed knowledge in one move.

Explaining a technical subject

Weak

Explain how vaccines work.

Better

Explain how vaccines work to a worried 70-year-old with no science background. Two short paragraphs, no jargon, and address the fear that it “overloads” the immune system.

The audience sets the register, the length caps the answer, and naming the actual worry means the reply addresses the question behind the question.

2. Ask for the shape you need

If you do not specify a format you will get prose, because prose is the default. Asking for a table, a checklist, an email or five bullets costs four words and saves you reformatting the answer by hand.

Comparing options

Weak

What are the differences between these three suppliers?

Better

Compare these three suppliers in a table with columns for price, lead time, minimum order and one risk each. Then one sentence recommending which to pick and why.

The table forces the same dimensions for each option instead of three paragraphs that each praise something different, and the closing sentence makes it commit.

3. Give it the material

This is the single highest-value habit on this page. A model asked about your situation in the abstract has to guess; a model given the actual document, error message, draft or numbers is working from evidence, and the failure rate drops sharply.

Debugging

Weak

Why is my Python script not working?

Better

Here is the function and the full traceback. It works on files under 10MB and fails above that. What is the cause, and what is the smallest change that fixes it?

The traceback removes the guesswork, the working-versus-failing boundary points straight at the cause, and asking for the smallest change stops it rewriting everything.

4. State the constraints up front

Length, tone, what to avoid, what must be included, who must not be offended. Constraints stated afterwards mean a second round trip; stated first, they cost nothing.

Writing something difficult

Weak

Write an email telling my team about the delay.

Better

Write an email to my team about a two-week delay caused by a supplier. Under 150 words, direct but not grim, take responsibility without naming individuals, and end with the new date and one concrete next step.

Every constraint removes a way the draft could come back wrong. The result usually needs one edit rather than three attempts.

5. Treat it as a conversation, not a slot machine

The most common mistake is throwing away a mediocre answer and rewriting the question from scratch. The second question is almost always where the good answer comes from, because the model now has your reaction to work with.

Point at the part that is wrong
“The second paragraph is too defensive — rewrite just that.” Far better than regenerating the whole thing.
Ask for alternatives
“Give me three openings with different tones.” Choosing is easier than specifying.
Make it argue with itself
“What is the strongest case against this recommendation?” This surfaces weaknesses better than asking whether it is sure.
Ask what it needs
“What would you need to know to answer this properly?” Useful when your own question is still fuzzy.

6. Ask for sources on anything factual

An answer that cites its pages can be checked in thirty seconds. An answer that does not has to be trusted, and trust is the wrong relationship to have with a system that produces fluent text whether or not it has grounds for it.

Anything with a number in it

Weak

What are the rules on statutory sick pay?

Better

What are the current statutory sick pay rules in the UK? Search the web, cite the official pages, and flag anything that changed in the last year.

It forces a live lookup instead of recall, the citation makes the claim checkable, and asking about recent changes catches the case where the training data is simply out of date.

None of this is prompt engineering in the sense people market courses about. It is being specific with a system that cannot see your desk — which is the same thing that works when briefing a competent stranger.

Frequently asked questions

Keep reading: how to check an answer, what an AI chatbot is, getting started, AI chat for students.