Teaching students prompt literacy

The problem: thirty assignments, ten of them sharing a sentence

The Secondary 4 Citizenship and Social Development project was on Hong Kong's ageing population. Of the thirty submissions, ten used nearly identical sentence patterns, the same three arguments, and the same authoritative-sounding figure that could not be traced to any source.

The teacher asked what prompt they had used. Almost the whole class had typed one line:

Causes and effects of Hong Kong's ageing population, 300 words.

The problem is not only copying. The problem is that the student knows exactly one move — paste the question in unchanged, hand the output back unchanged. It has never occurred to him that he could follow up, ask for two opposing views, or ask the AI to flag which figures need checking.

Give the same student this instead:

Role: you are a discussion partner, not a ghostwriter.
Task: I am writing a project on Hong Kong's ageing population. List five key pieces of data I should verify myself, and two common claims that may not hold up.
Output format: two columns — "data to verify" and "what kind of source to check it against" — then three items under "claims worth doubting".
Constraints: do not write any of my prose. For any figure you are unsure of, write "student to verify" rather than supplying a number.

What he hands in is written in his own sentences, with three checked sources he did not have before. This lesson teaches him to write that kind of prompt.

Why this works

What an AI gives you depends largely on what you cast it as. Treat it as an answer machine and it produces answers; treat it as a discussion partner, an opponent or a proofreader and it produces questions, counter-examples and suggested edits. Same model — the difference is one role line, which is exactly what the role, task, context, format framework is about. Nobody has ever taught students this.

The second thing students do not know: confidence of tone tells you nothing about reliability. A model predicts one token at a time, and fluency and truth are separate properties. You do not need to explain architecture in class — one failed demonstration does the work. Ask the AI for the clause number of a Hong Kong ordinance or a specific statistic, then have the class look it up on the official site together. Two or three times in ten it will not match. The underlying explanation lives in where hallucination and bias come from.

The third thing is integrity. Students cannot tell "AI helped me think" from "AI wrote it" — not because they are devious but because nobody has ever drawn them a line. So the second half of this lesson is not technique. It is rules.

A 40-minute lesson you can run as written

You need two things: a device you can project from, and a printed worksheet. If students have no devices for the verification stage, run the AI output in advance, print it, and have them do only the checking.

Template 1: the student worksheet prompt

Print this on the worksheet for students to copy, leaving the bracket blank for their own topic.

Role: you are my research discussion partner, not a ghostwriter.
Task: my project topic is [                    ]. Break this topic down and tell me what I should be verifying myself.
Output:
1. Three sub-questions this topic splits into.
2. Five key facts or figures I need to verify myself, each with the kind of source to check it against (government statistics, academic report, news coverage).
3. Three common claims about this topic and why each might not hold up.
Constraints: do not write any prose or conclusions for me. For any figure you are unsure of, write "student to verify" rather than supplying a number. Do not mention personal details of any real individual.

Template 2: teaching students to check AI claims

Verification only happens if it is a fixed routine. This prompt lowers the starting cost by having the AI itself list what needs checking.

Role: you are a fact-checking assistant.
Task: below is a passage I am considering citing. List every checkable factual claim it contains.
Output format: a three-column table — "what the passage claims", "type of claim (year, figure, name, legal clause, quotation)", "where to verify it". Below the table, one line naming the claim most likely to be wrong.
Constraints: do not judge whether the content is correct and do not add new information. Your job is only to identify what needs checking. Mark any sentence you cannot classify as "for teacher or student to judge".
[paste the passage to be checked]

Pair this with one simple rule for students: a claim is verified when you find the same sentence or the same figure in the original source. Finding another website that says the same thing does not count. That rule does more work than any amount of explanation.

Template 3: a self-check for the AI-use declaration

Declarations must be specific. This prompt helps a student organise what he actually did before copying it into the box.

Role: you are an academic integrity assistant.
Task: I will describe how I used AI on this assignment. Turn it into a clear use declaration.
Output format: one declaration of no more than four sentences covering three things — at which stage I used AI, what I asked it to do, and what I changed or verified myself.
Constraints: do not polish or soften my description; organise only what I actually wrote. Do not add verification steps I did not describe. If what I describe may already exceed what the assignment permits, say so directly and tell me to ask my teacher first.
My description: [                    ]

A declaration that passes looks like this: "At the planning stage I asked AI to list three possible lines of argument and chose one. The population figures in the essay I looked up myself on the Census and Statistics Department site. I drafted paragraph three myself, then asked AI to point out awkward sentences, and the final wording is my own edit."

An integrity discussion frame: four scenarios

Do not read out the rules. Let students take a position first and they will remember it. Each scenario asks the same question — allowed, declare, or not allowed?

1. Using AI to split my topic into three sub-questions. Most students say allowed, and that is reasonable — this is planning, not content.
2. Using AI to find five figures to verify, which I then check myself. This is the use we want to promote, and it should be declared, because it shaped the direction of the work.
3. Writing a paragraph myself, then asking AI to flag grammar problems which I fix. Usually declare. The line is who wrote the final words, not who read them.
4. Asking AI to write a paragraph and changing a few words. Not allowed. Put one question to the student here: if your teacher asked why you chose this argument, could you answer? If not, the paragraph is not yours.

Afterwards write the class's conclusions as three lines on the wall, and make sure the rest of the panel uses the same three. What students find hardest to accept is different teachers drawing different lines.

Common mistakes

1. Stating the rule without demonstrating it. Students hear "AI makes mistakes" ten times and still believe it, because they have never seen one. Fix: always run a live failure, letting the AI get a checkable fact wrong in front of the class.
2. Vague declarations. "AI was used to assist with this assignment" says nothing. Fix: use template 3's three-part structure, naming the stage, the request and the student's own changes.
3. Teaching verification as "ask the AI again". The student asks "is this figure right", the AI says yes, and he considers it checked. Fix: require verification to leave the AI entirely and land on the original source, with the URL or document name written down.
4. Banning so broadly that use goes underground. In classes with a blanket ban, usage does not drop — students simply stop telling you. Fix: distinguish clearly between assignments where AI is entirely prohibited (assessed compositions, for instance) and those where planning-stage use is fine, and print the line on the assignment rather than saying it once.

Going further

• Chinese Language: turn the verification activity into checking quotations. Students ask the AI for three classical quotations related to the topic, then trace each to the original text and its context. Error rates on classical Chinese quotations are high, which makes the demonstration land hard.
• English Language: rewrite template 1 to ask for counter-arguments, then have students write an argumentative essay in response. The focus shifts from verification to reasoning.
• Information Technology: add a session where students try to make the same model give different answers through different phrasings, then discuss why. It is an introduction to prompt engineering and the best possible proof that AI output is not a single truth.
• Primary 4 General Studies: cut down to two activities — a follow-up drill where one question is pursued to a third level, and checking how many districts the AI says Hong Kong has. Skip declarations; teach only "check it yourself once".
• SEN adaptation: for students with writing difficulties, turn the declaration into tick boxes plus one written line, and run verification in pairs with one student operating and one recording.

Where to go next

The goal is not students who use AI better but students who are still thinking while they use it. Three moves carry a whole year — probe, contrast, verify — plus one honest, specific declaration.

For how students test the boundary in tutor mode, see scaffolded tutor prompts. For what the school has to do at platform level, see the safety net for student AI use and privacy, bias and copyright ethics.