Prompting that works

Context, examples, constraints and iteration. About 1 h, beginner level.

You'll be able to

• Write a prompt that covers context, task, format, constraints and examples.
• Use examples (zero-, one- and few-shot) and cues to steer the shape of an answer.
• Improve a weak answer by critiquing it and asking for a revision, instead of starting over.

A prompt is a brief

A language model is a very capable collaborator who knows nothing about your situation: not your class, not your reader, not what you already tried. It fills every gap with a guess, and the guess is usually the most generic answer.

So think of a prompt as a brief you'd hand a new teaching assistant. A good brief usually covers five things:

• Context: who you are, who the output is for, and what it's for.
• Task: the exact job, with a clear verb (summarize, compare, draft, critique).
• Format: the shape you want: five bullets, a table, 150 words.
• Constraints: what to stick to or avoid: level, tone, sources.
• Examples: a short sample of what good looks like.

Not every prompt needs all five. Missing context and missing format cause most weak answers.

Instructions and content

Most useful prompts have two parts: an instruction (what to do) and the content it applies to (the text, data or notes). Keep them visibly separate so the model doesn't confuse your notes with your request.

A simple habit: put the instruction first, then the content between clear markers, then repeat the key instruction at the end if the content is long. For example: “Summarize the article below in three sentences for a first-year student.” followed by the article in quotes, and then “Three sentences, no jargon.”

Long inputs push your instruction far away from where the model starts writing. Restating it at the end, which Microsoft's course calls “doubling down”, helps the model keep it in view.

Show, don't just tell: examples and cues

A model copies patterns very well. If you show it one or two examples of the output you want, it usually infers the style, length and structure better than any description.

• Zero-shot: an instruction and no examples. Fine for simple tasks.
• One-shot: one example of input and output, then your new input.
• Few-shot: several examples. Useful when the pattern is subtle, like a grading style or a citation format.

A cue is a lighter version: you start the answer yourself. Ending your prompt with “Top three takeaways: 1.” nudges the model into a numbered list that begins where you want it to.

Examples are powerful, so choose them carefully. The model copies their flaws too.

Iterate like an editor

Your first prompt is a draft. When the answer is off, don't start over: say what's wrong and ask for a revision. “Too formal. Cut it to 120 words and use one example from biology.” is faster and more precise than rewriting the whole prompt.

Three moves that help:

• Ask it to ask you first: “Before you start, ask me up to three questions about anything you need.” This surfaces missing context.
• Give it an out: “If the article doesn't say, write ‘not stated’.” This reduces invented details.
• Save what works: a prompt that worked once becomes a template you can reuse with new content.

Prompting is a trial-and-error process. You bring the judgment about whether the answer is right; the model brings speed.

From a lazy prompt to a good brief

Before: “Summarize this article.”

The model returns a solid but generic paragraph at an unknown level, with the details it happened to find interesting.

After: “I'm a first-year student preparing for a seminar discussion. Summarize the article below for someone who hasn't read it. Use five bullets: the main claim, the evidence, one limitation the authors admit, one question I could raise in discussion, and a one-line takeaway. Plain language, no quotes longer than ten words. If the article doesn't cover one of these, write ‘not stated’.”

What changed: context (who it's for and why), a precise task, a fixed format, constraints on language and quotes, and an out for missing information. The answer is now shorter, checkable against the article, and useful for the seminar.

Try it

Your lazy prompt, your rewritten prompt, both answers side by side, and one sentence on what made the difference.