Writing a prompt that works

Why a bare question underperforms

"Write a summary of this" gives the model nothing to aim at: summary for whom, how long, in what tone, leaving out what? It fills every gap with the most average choice, and average is rarely what you wanted. Almost every disappointing AI result is a prompt that left the model guessing.

The four parts of a prompt that works

1. Role. Who should the model be while it answers? "You are an experienced project manager writing for executives." This sets vocabulary, depth, and what counts as obvious.
2. Context. What does it need to know that it can't see? Who the reader is, what happened before, what the output is for. Paste the actual material; don't describe it.
3. Constraints. Length, tone, what to include, what to avoid. "Under 120 words. No jargon. Don't mention pricing."
4. Output format. The shape you want back: a bulleted list, a table, three options, an email with a subject line. If you'll paste it somewhere, say where.

You don't need all four every time. You need to notice which one is missing when the answer is off.

How to iterate

Read the weakest output and name what's wrong in one phrase: too long, wrong audience, missed the point, invented a detail. Then change one thing in the prompt that addresses it and run it again. One change at a time is the only way to learn what the change did. Two changes at once teaches you nothing, even when it works.

The task

Take a real task from your work or life. Prompt it three ways:

1. A bare question.
2. A structured prompt with a role, context, constraints, and an output format.
3. One iteration: look at the weakest output, change one thing, run it again.

What to submit

Three prompts, three outputs, and a short note on what each change fixed.

How it's graded

• The structured prompt contains all four elements.
• The iteration changes one thing and the note explains its effect.
• The improvement is visible in the outputs, not just asserted.