Teaching Prompt Engineering to Students: A 5-Lesson Unit That Works

Most pupils already type things into an AI chatbot. Very few of them know why one request produces a useful answer and the next produces something bland, wrong or oddly confident. That gap is teachable, and it is one of the most transferable skills you can hand a teenager right now: the ability to describe a task precisely enough that someone - or something - can complete it well.

This unit has been run with mixed-ability classes in Years 8 to 11 and adapts easily upwards or downwards. Each lesson is designed for a single 50-minute period. You need one device per pair, not per pupil; pairs argue about wording, and the arguing is where the learning happens.

What pupils should be able to do by the end

• Break a vague request into a task, an audience, a format and a constraint.
• Predict how changing one element of a prompt changes the output.
• Spot a confidently wrong answer and describe how they checked it.
• Explain, in their own words, when using AI is appropriate for a piece of schoolwork and when it is not.

Notice that only the first two are about AI. The rest are comprehension, verification and academic honesty - skills your curriculum already asks for.

Lesson 1: The four ingredients

Open with a deliberately bad prompt on the board: write about the Romans. Ask the class what an AI cannot possibly know from that sentence. They will list things fast: how long, for whom, which Romans, what for.

Introduce the four ingredients:

1. Task - the verb. Explain, compare, summarise, generate, critique.
2. Audience - who reads it and what they already know.
3. Format - length, structure, bullet points, table, paragraph.
4. Constraint - what must or must not appear.

Then rewrite the Roman prompt together:

Explain why the Roman army was hard to defeat.
Audience: a Year 8 class who have studied Roman roads but not military tactics.
Format: five short paragraphs, each with a bold topic sentence.
Constraint: use no Latin terms without explaining them in the same sentence.

Run both prompts side by side on your own screen. The difference is obvious and it lands without you having to argue for it. Pupils then rewrite three vague prompts of their own using a simple four-box template.

Lesson 2: Change one thing

This is a controlled experiment, and I introduce it with that language deliberately because it borrows credibility from science lessons.

Pairs take one working prompt from Lesson 1 and produce three variants, each changing exactly one ingredient. They record what changed in the output. A typical results table looks like this:

Change made	Effect on output
Audience changed to primary pupils	Shorter sentences, lost the technical detail we wanted
Format changed to a table	Easier to revise from, but the reasoning disappeared
Constraint removed	More jargon, more confident, less checkable

The plenary question is the point of the whole lesson: which version would you actually hand to your revision partner, and why? Pupils start judging output against a purpose rather than against how impressive it sounds.

Lesson 3: Catching the confident mistake

Give every pair an AI answer you have prepared in advance that contains two factual errors and one invented source. Do not tell them how many. Their task is to mark it like a teacher, with a red pen, and to write one sentence next to each correction explaining how they verified it.

This is the lesson that changes behaviour. Pupils who have personally caught an AI inventing a book title stop treating output as fact. Keep a running wall display of errors the class has caught during the term - it becomes the most effective anti-plagiarism poster in the room, because they wrote it.

A useful checking routine to teach here:

• Can I find this claim in two places that are not AI-generated?
• Does the source actually exist, and does it say what the answer claims?
• Is a number given without a date or a population? Treat it as unverified.

Lesson 4: Prompts that help you learn, not prompts that do the work

Pupils sort a deck of twenty prompt cards into two piles: helps me learn and does the work for me. Expect a genuinely heated discussion, because the boundary is not obvious.

Examples that belong in the first pile:

• Ask me five questions about photosynthesis, one at a time, and tell me what my answers are missing.
• Here is my paragraph. Do not rewrite it. Point out the two weakest sentences and explain why.
• Explain this concept at three difficulty levels so I can find the one I understand.

Examples that belong in the second pile are the ones they are already using, and they know it. Finish by having the class draft its own AI usage rule for your subject, in three bullet points, and display it. Rules pupils write are rules pupils police.

Lesson 5: Assessed task

Pupils receive a real problem from your subject and submit three things: their final prompt, the output, and a 150-word commentary on what they changed and why, plus one thing they verified independently.

Mark the process, not the polish:

Criterion	Descriptor
Prompt precision	All four ingredients present and appropriate to the task
Iteration	At least two improvements, each justified
Verification	One claim checked against a named non-AI source
Judgement	Explains where the output was still not good enough

A pupil who submits a mediocre output with an excellent commentary should outscore one who submits a slick output with nothing to say about it. That signal matters more than the marks.

Three things that went wrong the first time I taught this

Devices one-to-one. Pupils stopped talking and started racing. Pairs fixed it immediately.

No prepared bad answer. Asking a current model to produce errors on demand is unreliable. Save a flawed answer to a document in advance and reuse it for years.

Letting the unit run without a subject anchor. Prompt engineering taught in the abstract becomes a technology lesson pupils forget. Anchor every task in content you are already teaching that half-term.

FAQ

Do I need paid AI accounts for this? No. Every lesson works with a free tier, and Lessons 1, 3 and 4 work with no pupil accounts at all if you drive a single screen at the front.

What if my school blocks AI tools for pupils? Run it as a teacher-demonstration unit. Pupils still write and critique prompts on paper; you execute them. Roughly 80 per cent of the learning survives.

Is this not just teaching them to cheat more efficiently? The opposite, in practice. Pupils who understand how these systems produce text become much harder to impress with it - and Lessons 3 and 4 give you shared vocabulary for honesty conversations you would otherwise have one-to-one after the fact.

How do I fit five lessons into a crowded scheme of work? Compress to three: merge Lessons 1 and 2, keep Lesson 3 intact because it carries the most weight, and run Lesson 5 as homework. Do not cut Lesson 3.