Role, goal, constraints, format

Most bad output comes from a vague request

Not a weak model. A vague request.

A prompt that works reliably has four parts. You do not need all four every time, but when output disappoints, the missing part is almost always one of these.

1. Role

Who should the model act as? This sets vocabulary, assumptions and depth.

You are a senior financial analyst reviewing a small business P&L.

2. Goal

What outcome do you want? Not what topic — what outcome.

Identify the three line items with the largest month-over-month variance and explain what likely caused each.

3. Constraints

What should it avoid, assume or respect?

Do not speculate about causes you cannot support from the numbers. Assume the reader is not an accountant. Keep it under 300 words.

4. Format

How should the answer come back?

Return a markdown table with columns: Line item, Variance (AED), Variance (%), Likely cause. Then one paragraph of commentary.

Put together

Those four sections stacked produce output you can use without rewriting. Compare it to "analyse my P&L" and the difference is not subtle.

The iteration loop

You will still not get it right first time. The loop:

1. Send the prompt
2. Identify specifically what is wrong — not "it's bad", but "it invented a cause for row three"
3. Add the constraint that prevents exactly that
4. Resend

Three rounds of that produces a prompt you can reuse forever. Save it. That is how a prompt library gets built — not by collecting other people's prompts, but by keeping the ones that solved your own problems.