Prompt engineering in one lesson: treating AI like a print shop that needs a job ticket

After lunch on Wednesday, two Primary 6 maths teachers queue at the same staff-room machine. Teacher A types "make fraction word problems" and gets a Primary 4-level sheet with a wrong answer key. Teacher B pastes half an A4 job ticket and, three minutes later, has six questions plus a marking-points table. She changes one character and prints. Teacher A says: "But we used the same AI."

Prompt engineering is not a programming language. It is the skill of writing a job ticket. In the LLM applications module, this lesson follows your understanding of hallucination and bias, and shows how to state what you want and what you forbid before handing work to the model. The next lesson connects your school's own documents so the AI reads before it answers.

---

A print shop only reads the fields on the ticket

Imagine handing a worksheet to the school print room. The colleague at the desk will not guess whether you want A3 or A4, colour or black and white, a teacher version or a pupil version. You fill in subject, year group, copies, paper stock and deadline. Anything left blank is filled from habit, and habit may not match your panel's conventions.

AI is that print room: fast, widely read, never asks a follow-up, and knows nothing about your scheme of work, your phrasing habits or your marking culture. Six words produce whatever sample was most common in training data. A complete ticket stands a much better chance of producing your version.

Five fields: context, audience, deliverable, format, off-limits

Fill all five and most "wrong answer" problems shrink.

Context. Who you are and where this will be used. "Hong Kong primary maths teacher preparing in-class practice before a Primary 6 unit test." One sentence fixes register, difficulty and cultural setting.

Audience. Pupils or staff? Does an SEN group need simpler wording? "Audience: Primary 6 mainstream class, secure on fractions, ratios not yet taught."

Deliverable. Do not stop at "set questions". State count and type. "Six fraction word problems: two with mixed numbers, one requiring two steps, each with a model answer and a common-error note."

Format. "Output as a table with columns for number, stem, model answer, common error and marking point." Named columns are trim lines for the print shop; the model omits less when it has boxes to fill.

Off-limits. The field most often skipped and most often worth money. "No ratio questions beyond the scheme. No Taiwanese or mainland usages. Write teacher to verify for any local example you are unsure of. No answers on the pupil version."

Off-limits rules are not negativity. They switch off the model's default urge to write long, rewrite everything and finish incomplete thoughts, which is exactly where hallucination and drift begin.

Off-limits beat adjectives

A model's default is to say as much as it can, as elegantly as it can, as completely as it can, inventing plausible text when it is unsure. "Please be specific, professional and constructive" rarely moves it. "Do not rewrite the essay; list three improvements only, each under 30 words" usually does.

The line "if unsure, write teacher to verify" drags hallucination into the open. It does not remove invention, for reasons set out in where hallucination and bias come from, but it shows you where to check.

One example beats five adjectives

If you want comments in a particular style, do not describe "warm but precise". Paste a comment you already like:

"Write comments at this length and structure. Example: you set up step two correctly, but the unit conversion is missing; after fixing it, check once with the original numbers."

Models imitate formats they have seen. They guess your internal standard badly. One real example carries more information than a pile of adjectives.

A science panel A/B test

At one secondary school the science panel ran a blind comparison during co-planning. Eight teachers each wrote a prompt asking AI for feedback on a Secondary 1 lab-report draft, not a rewrite. When the eight outputs went on the wall, the pattern was clear. Everyone who had written off-limits rules such as "do not rewrite paragraphs, do not supply a full model answer, list three improvements only" had usable output. Among those who had not, four times out of eight the AI simply handed back a "corrected" report, which helps a pupil not at all because they cannot see their own mistake.

The panel head saved the two best tickets plus a blank five-field template in the shared folder. Supply teachers next term only fill in topic, class and word count.

From personal skill to panel asset

The endgame of prompt engineering in a school is not every teacher becoming a prompting expert. It is turning proven tickets into shared panel templates. A prompt that works is worth as much as a validated lesson framework: archive it, name it, note the year group, and stop retyping from scratch.

Teaching pupils to prompt is teaching in its own right. Saying clearly what you want and clearly what you forbid transfers into writing and speaking, not just tool training. Managing a template library is covered in the Prompt Engineering Classroom.

Where people go wrong

• The drift: writing "you are a professional teacher, answer carefully" and assuming role-play is enough. Self-check: is output still missing year group, format or off-limits rules? If you edit heavily every time, a field on the ticket is still blank.

• The drift: stacking all requirements at the top, then pasting a long pupil essay or textbook extract below. Self-check: are your most important off-limits rules buried in the middle of a long input? Move them to the final line and compare again.

• The drift: closing the window when output looks good and starting from zero next week. Self-check: does the panel shared folder contain verified prompts? If not, you are repeating mistakes someone else already made.

Try it once

Time: about 12 minutes. Materials: any AI tool; optionally one real comment or question you already use.

What to do: first paste the minimal version, "write three comments for a Secondary 2 pupil". Note whether output is too long, rewrites the whole piece or uses the wrong regional Chinese.

Then paste the ticket version:

Context: Hong Kong secondary Chinese teacher giving feedback on pupil work.
Audience: Secondary 2, already writing in paragraphs.
Deliverable: three comments, each naming one strength and one improvement.
Format: table with columns "quoted sentence", "comment", "one-line suggestion".
Off-limits: do not rewrite the essay; do not supply a full model answer; each comment under 40 words; write teacher to verify for any fact you are unsure of.
Example comment: "The example in paragraph two is on topic, but the closing sentence does not tie back to the argument; next time lead with therefore."
[paste anonymised pupil work here]

What to hand in: both outputs plus three lines comparing them: which is usable as-is, and which off-limits rule made the difference.

How you know it worked: you can explain to a colleague that the gap is not the AI version but whether all five fields were filled, and you can point to what the off-limits section and the example each fixed.

Next lesson: the school knowledge base and retrieval-augmented generation. Previous lesson: where hallucination and bias come from.