Prompt engineering in one lesson, or why the same question gets such different answers

On the same morning, two Primary 5 Chinese teachers each set out to use AI to prepare a reading comprehension worksheet.

The first typed six words: "make me a reading comprehension". She got a short passage about pandas at roughly Primary 2 level, five multiple choice questions, and one wrong answer key.

The second typed about eighty words. She got a passage at the right difficulty, three comprehension questions, two inference questions, one open-ended question and a marking points table, in Hong Kong Traditional Chinese with no Taiwanese usages. She changed two characters and printed it.

They used the same AI. The difference was not luck.

Picture yourself briefing a supply teacher

Imagine a supply teacher arriving at your school for the first time. She is extremely clever, has read enormously, and types fast. What she does not know is which textbook series you use, the ability range of this class, the worksheet format you prefer, or which expressions your school avoids.

If you say "make a reading comprehension" and walk away, she will invent one from her own assumptions. What she produces will not be wrong. It just will not be what you wanted.

AI is that supply teacher, except she will never ask a follow-up question and never say she is unsure. She fills in every blank herself. All of prompt engineering amounts to not leaving her blanks to fill.

The four parts of a sound prompt

Include all four and nine tenths of the problem disappears.

One, role and audience. "You are a Hong Kong primary Chinese teacher writing for Primary 5 pupils." That single sentence fixes vocabulary conventions, difficulty and cultural context at once. Without it you will often get mainland or Taiwanese phrasing.

Two, the task. The more specific the better. Not "make a comprehension" but "a 350 to 400 character expository passage on a topic connected to daily life in Hong Kong, with three comprehension questions, two inference questions and one open-ended question".

Three, the output format. "Present the questions as a table with columns for number, question type, question, model answer and marking points." Specifying format buys you two things: no reformatting afterwards, and a model that omits less because it has columns to fill.

Four, what must not be done. This is the part most often forgotten and the most valuable. For example: no Taiwanese usages, no answers on the pupil version, write "teacher to verify" rather than inventing any local fact, keep each marking point under 25 words.

Why constraints work better than requests

Because a model's default behaviour is to say as much as it can. It tends to write long, write elegantly and write completely, even when it is not sure.

So the most effective prompting technique is often a negative instruction: tell it what it may not do. The line "if you are unsure, write teacher to verify" converts hallucination from something hard to spot into something clearly flagged. It does not remove hallucination, for reasons set out in where hallucination and bias come from, but it shows you where to look.

One example beats five adjectives

If you want comments in a particular style, do not describe the style. Give an example of it:

"Write comments in this format and at this length. Example: the structure is clear and each of the three paragraphs has a topic sentence, but the example in paragraph two does not obviously support the argument, so next time ask yourself what this example proves."

Models imitate extremely well and guess your internal standards extremely badly. One real example carries far more information than "please be more specific, more professional and more constructive".

An example from the classroom

A secondary English panel ran an experiment at a co-planning meeting. Eight teachers each wrote a prompt asking AI to give feedback on a Secondary 3 argumentative essay, and the eight sets of output went up on the wall anonymously for comparison.

The pattern was unmistakable. Every teacher who had written "do not rewrite the essay, give feedback only" produced usable output. Among those who had not, the AI simply handed back a corrected essay about half the time, which helps a pupil not at all, because the pupil cannot see what to change.

So the panel head did something practical. She took the three best prompts from that day, added a checklist of the four parts, and turned them into a shared panel template in the common folder. Three months later new teachers did not need to learn prompting technique. They filled in blanks.

That is what prompt engineering really looks like in a school: not every teacher becoming an expert, but expert prompts becoming a panel asset. Managing a template library is handled fully in the Prompt Engineering Classroom.

What this means for your classroom

1. When AI gives you a poor answer, check what you left out first. Four times out of five the model is fine and the prompt is missing the year group, the format or the constraints.
2. Save the prompts that work. A prompt that proves itself is worth as much as a good lesson plan and should be archived and shared across the panel rather than retyped.
3. Always add the verification instruction. "Write teacher to verify where you are unsure" should be the standard closing line for staff prompts, especially anything involving dates, regulations or curriculum detail.
4. Teaching pupils to prompt is teaching, not just tool training. Saying clearly what you want and clearly what you do not want transfers directly into writing and speaking.

A complete prompt you can copy

This one contains all four parts and can go straight into any AI tool:

You are a Hong Kong primary Chinese teacher writing for Primary 5 pupils.
Task: design a reading comprehension worksheet. First write a 350 to 400 character expository passage on a topic connected to daily life in Hong Kong. Then set six questions: three comprehension, two inference, one open-ended.
Format: give the passage first, then a table of questions with columns for number, question type, question, model answer and marking points.
Constraints: write in Hong Kong Traditional Chinese and avoid Taiwanese or mainland usages; give no model answer for the open-ended question, only marking points; write "teacher to verify" for any local fact you are unsure of rather than inventing it; keep each marking point under 25 words.

Once you have tried it, delete the whole constraints paragraph and try again. You will see immediately why this lesson spends so long on constraints.

What comes next

However well a prompt is written, the AI will still not know your school's teaching schedule or homework policy, because it has never read them. The fix is to have it consult the school's own documents before answering, which is the next lesson, the school knowledge base and retrieval-augmented generation. To revisit why Chinese prompts cost more than English ones, see tokenisation and why Chinese costs more. For a large library of ready-made classroom prompts, see what makes a good prompt.