What makes a good prompt

Short answer: a good prompt is not a polite request. It is the act of writing down the four things you normally leave inside your head — who the AI should be, what it should do, the context it is working in, and the shape of the output you want back.

A real gap: five questions for Primary 6 General Studies

A Primary 6 General Studies teacher needs discussion questions for a unit on Hong Kong's water resources. She opens an AI tool and types the most natural thing:

Give me five questions about water resources in Hong Kong.

What comes back is roughly: "Where does Hong Kong's drinking water come from?", "Why should we save water?", "What is Dongjiang water?" Nothing is wrong, but every question reads like it was lifted from a general website. There is no difficulty progression, no link to a learning focus, and not one question she could project for a whole-class discussion. She still has to rewrite them.

The same teacher, second attempt:

You are a planning assistant for a Hong Kong primary General Studies teacher.
Task: design five classroom discussion questions for the Primary 6 unit on Hong Kong's water resources.
Context: students have already studied the water cycle. The learning focus of this lesson is explaining the sources of and limits on Hong Kong's fresh water supply. The class has 28 students of mixed-to-average ability, about five of whom need additional language support.
Output format: a table with four columns — question, cognitive level (recall / understanding / application / analysis), expected answer points, and a simplified wording for the language-support students.
Constraints: the five questions must cover at least three cognitive levels. Where a figure or a date is involved and you are not certain of it, write "teacher to verify" in that cell rather than supplying a number.

The second output can be printed and dropped into the lesson plan. The model did not get smarter between the two attempts. The teacher simply handed over the information that had been sitting in her head.

Why this works

A language model generates text by repeatedly predicting the most plausible next token given the context you supplied. The fewer conditions you set, the closer "most plausible" drifts to the average of everything it has read — which is the blandest, safest, most generic phrasing on the internet.

An analogy: if you ask a relief teacher who has never met your class to "take a lesson for me", all they can do is read from the textbook. Leave a note stating the class, where you got to last time, which five students need extra time and what you want handed back at the end, and the same relief teacher produces something close to what you would have done yourself. AI is that relief teacher — always punctual, endlessly patient, and completely ignorant of your school.

If you want to know why the model behaves this way internally, see the LLM Classroom, particularly why teachers should understand LLMs and where hallucination and bias come from. This course does not repeat that ground; it concentrates on the writing.

What this course covers

Eighteen lessons in four modules:

• Module 1, fundamentals (lessons 1 to 5): a reusable prompt framework, teaching by example, chain of thought and step-by-step instructions, system prompts and school-based personas.
• Module 2, teaching and learning (lessons 6 to 11): planning, tiered worksheets, question setting and rubrics, marking feedback, the scaffolded tutor and student prompt literacy.
• Module 3, administration, RAG and safety (lessons 12 to 14): school administration and parent communication, grounded answers with citations, and prompt injection and student content safety.
• Module 4, advanced technique (lessons 15 to 18): structured output, multimodal prompts, a panel template library and a prompt debugging checklist.

Read lessons 1 to 5 in order, then jump to whatever your workload demands.

The all-purpose template you can use today

Save this in your phone notes or a text file and replace the bracketed parts. It is a simplified version of the framework lesson 2 unpacks properly.

You are a teaching assistant for a Hong Kong [primary / secondary] [subject] teacher.
Task: [one sentence saying what you want, for example "design a lesson starter activity"].
Context: [year level], unit "[unit name]", learning focus: [list them]. Class profile: [size, ability spread, special educational needs].
Output format: [list / table / prose], [how many items], [what each item must contain].
Constraints: use Hong Kong English and local curriculum terminology; where a fact, figure or date is uncertain, mark it "teacher to verify" rather than supplying one; give me one version first and I will ask for revisions.

Fixing a weak prompt: the upgrader template

If you already have a prompt that underperforms, do not start again. Have the model diagnose it:

You are a prompt engineering coach working with Hong Kong classroom teachers.
Task: review the prompt below, identify what information it is missing, and rewrite it.
Context: my real use is [purpose, year level and subject].
Output format: first list three missing elements, one sentence each; then give the rewritten prompt in full; then one sentence on what difference I should expect to see.
Constraints: the rewrite must contain a role, a task, a context and an output format; do not invent class details I never gave you — leave square brackets where I need to fill something in.

My original prompt: "[paste your prompt here]"

Classroom-ready: a five-minute starter

You are a classroom assistant for a Hong Kong secondary [subject] teacher.
Task: write a five-minute starter activity for Secondary 3 [unit name] that reviews the previous lesson.
Context: last lesson covered [content], this lesson introduces [content], and the bridge between them is [concept].
Output format: three questions, each with one sentence on what a wrong answer tells me the student has not grasped, plus one quick show-of-hands question for the whole class.
Constraints: everything must fit inside five minutes; no questions requiring a calculator or the textbook; if a specific fact is uncertain, mark it "teacher to verify".

Common mistakes

• Writing the prompt as keywords. "P4 Chinese expository writing worksheet" produces stitched-together generic content. Write a full sentence and state the purpose, and the quality shifts immediately.
• Asking for four deliverables at once. "Write me the lesson plan, the slides, the worksheet and the letter home" makes all four shallow. Four separate prompts, each focused, take less total time.
• Leaving out the year level. "Explain photosynthesis" is two completely different texts for Primary 4 and Secondary 5. The year level is the cheapest, highest-impact detail you can add.
• Treating the first output as final. Treat it as a draft and reply "item three is too advanced, use an everyday example instead". Revising is far faster than regenerating, and the model still has the earlier context.

Going further

• A different subject: swap the role and the terminology. Visual Arts benefits from "use the vocabulary of visual elements and design principles"; Information Technology from "use Hong Kong secondary curriculum programming terms, not university-level vocabulary".
• A different year level: add to the constraints line "pitch the vocabulary at [year level] and keep average sentence length under [n] words". That works far better than "make it simpler".
• SEN adaptation: ask the output format for an extra column, for example "provide a step-by-step version for students with attention difficulties, one instruction per step". For students with dyslexia, ask for a word bank and sentence frames instead. Lesson 7 develops this fully.

Summary and next step

A good prompt is your professional judgement written down. The more experienced you are, the more unspoken knowledge you carry — and the more prompt engineering gives back.

The next lesson, the role, task, context and format framework, turns the all-purpose template above into a framework you can reuse, and every later template in this course visibly follows those four parts.