Prompting Techniques: Role-Play, Examples and Iteration

The basics get you an answer. Techniques get you a good one. The three below are the ones that repay the effort fastest, and each is a matter of English rather than technology — of showing the model what you mean instead of only telling it.

Give an example

The single most powerful move in prompting is to show the model the pattern you want. English has a word for this that has crossed over from teaching: few-shot — a prompt that includes one or two examples.

Without an example, you describe:

Rewrite this sentence in a formal style.

With an example, you show:

Rewrite this sentence in a formal style, like this: Informal: "We need to fix this fast." → Formal: "This requires prompt attention." Now rewrite: "Can you send me the file?"

The example removes all guessing about what formal means to you. The model matches the pattern instead of interpreting an abstract word.

An example is a constraint written concretely. Saying formal leaves the register open; showing "This requires prompt attention" pins it down. Wherever a word like professional, concise or engaging feels too vague, stop describing and give a sample.

Ask it to reason step by step

For a task with several steps — a decision, a diagnosis, a comparison — add a phrase that forces the model to show its work:

• Think step by step.
• Explain your reasoning before the answer.
• List the steps, then give the conclusion.

This changes what the model produces from a bare conclusion to a chain you can check. The technique has a name, chain-of-thought, but you do not need the name to use it.

• Weak: Which supplier should we choose?
• Strong: Compare the two suppliers on price, reliability and lead time, think through the trade-offs, then recommend one with your reasoning.

The strong version does two things at once: it forces reasoning, and it fixes the shape of the answer so the model does not skip a criterion.

Iterate

A prompt is rarely finished after one try. The real skill is the loop: run it, read the output with a critical eye, and rewrite the prompt to close the gap. Each round adds the thing the last answer was missing.

Round 1: Write an email asking for a deadline extension. Output: a correct but flat email. Round 2: Write an email asking for a deadline extension. It is going to a strict manager, so be respectful, give a concrete reason, and propose a new date.

Iteration is where the earlier skills meet — the four parts of a prompt get added one at a time, as the output reveals which one you left out. And reading the output critically is the same precision you would bring to describing a technical problem.

Asking the model to criticise itself

A useful closing move: ask the model to look at its own answer for weakness.

• What is the weakest point in this argument?
• What could be wrong with this answer?
• Rewrite this, then list the assumptions you made.

This is a hedge against the model's habit of sounding certain — including about hallucinations. You are not asking it to be right; you are asking it to show where it might be wrong, which is information you can use.

The techniques together

A single prompt can stack them:

Act as a careful editor. Rewrite this paragraph to be clearer. Here is an example of the clarity I want: "The meeting was delayed" → "We moved the meeting to Thursday." Think about what makes the example clearer, then apply that to the paragraph. List the changes you made.

Role, an example, reasoning, and a request to show the work — the whole toolkit in four sentences.

Test yourself

The quiz below checks the techniques and the phrasing that makes them work.