How to Write Good Prompts

What a prompt really is

When you type into an AI chatbot, your words are called a prompt. The AI reads your prompt and predicts the most fitting reply, one piece at a time. It does not read your mind, so the wording you choose is the steering wheel. Change the prompt and you change the answer.

This lesson is about the craft of writing prompts. If you want the friendlier kids' version first, read Asking AI Good Questions. Here we go a level deeper.

The four parts of a strong prompt

Most good prompts contain some mix of four ingredients:

1. Task — what you want done. "Summarise this article."
2. Context — the background the AI needs. "I am 12 and new to this topic."
3. Format — the shape of the answer. "In 5 bullet points."
4. Example — a sample of the output you want (optional but powerful).

You rarely need all four, but naming the task and the format alone fixes most weak prompts.

Before and after

A vague prompt:

Tell me about volcanoes.

A sharpened version:

Explain how a volcano erupts to a 12-year-old. Use 4 short steps, plain words, and one real example.

The second prompt gives the AI a task (explain eruption), a context (age 12), a format (4 steps), and a constraint (one example). The reply will be far more useful.

Show, don't just tell: few-shot prompting

Sometimes describing what you want is hard, so you show an example instead. Including one or two examples is called few-shot prompting (zero examples is "zero-shot"). For instance, if you want short rhyming definitions:

Define words like this: Cat: a furry pet that loves to nap. Sun: a burning star that brings the day. Now define: Moon

The AI copies the rhythm and style of your examples. Few-shot prompting is one of the most reliable tricks, because patterns are exactly what these models are built to follow.

Ask for reasoning, step by step

On problems with several steps — a math word problem, a logic puzzle, planning a trip — adding "think step by step and show your working" often gives better results. Walking through the steps lets the model catch its own slips instead of blurting a fast guess. This does not make the AI truly reason like a person, but on multi-step tasks it measurably reduces careless errors.

Iterate: the first answer is a draft

Prompting is a conversation, not a single shot. If the answer misses, do not start over. Refine:

• "Make it shorter and simpler."
• "You missed the cost; please include it."
• "Give me three options instead of one."

Each follow-up is a new prompt that builds on the last. Skilled users treat the first reply as a rough draft to push and shape.

The honest limits of prompting

Prompting is powerful but it is not magic. Be clear-eyed about what it cannot do:

• It cannot teach the model facts it never learned. If the information is missing or out of date, no wording invents the truth — the AI may hallucinate a confident-sounding but false answer.
• It cannot make a small model as capable as a large one.
• It cannot remove bias baked into the training data, only nudge around it.
• It does not give the AI real understanding. It still predicts text.

So the final, unskippable step is always you: check important answers against a trusted source. A great prompt improves your odds; it never replaces your judgement. To understand why the AI sometimes makes things up, see How Chatbots Work.

Try it yourself

Take a lazy prompt like "help with my essay" and rewrite it with a task, context and format: "I'm writing a 200-word essay on recycling for school. Suggest 3 strong opening sentences, each one line." Notice how much sharper the reply becomes. That gap is the whole skill.