Where a good prompt starts

On Wednesday break, a Primary 4 Chinese Language teacher opens AI in the staff room to draft three oral-assessment scenarios for next week. She types "Give me Primary 4 Chinese oral topics", waits a minute, and gets items that look scraped from the internet: "Describe your favourite festival", "Talk about protecting the environment" — nothing aligned to her unit My community, nothing for the six pupils who need Cantonese oral and literacy support. She closes the tab and tells the colleague next to her: "Same as always — I still have to fix it myself."

A good prompt is your professional judgement written down, not a politer question. This is lesson 1 of the prompt engineering course and the opening of Module 1, Prompting fundamentals — establish that clarity beats length before lesson 2 shows you how to split a prompt into a structure you can check every time.

---

One need, two ways of asking

The second attempt puts everything that had been in her head into the text:

You are a planning assistant for a Hong Kong primary Chinese Language teacher working on oral assessment.
I need three oral-assessment scenarios for Primary 4 unit "My community", each with five guiding questions.
The class has 28 pupils; six need Cantonese oral support and two have literacy difficulties.
Students already know community-facility vocabulary (market, park, library, bus stop). The learning focus is describing how local facilities affect residents' daily lives, with concrete examples.
Output in Traditional Chinese: three scenarios, each with a scenario description (50 characters max), five questions from easier to harder, and a simplified opening line for language-support pupils.
Where a place name, figure or policy is involved and you are not certain, mark it "teacher to verify" rather than supplying one.

The second output can go straight into the assessment sheet: all three scenarios stay on community life, the questions tier sensibly, and the language-support lines are ready. The model did not get smarter between attempts. The teacher filled the blanks.

The model fills your blanks

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 — the blandest, safest, most generic phrasing on the internet.

Like ordering at a cha chaan teng without a menu choice

Walk into a cha chaan teng you have never visited and say only "One set meal, please." The waiter serves what most customers order — milk tea with instant noodles, or fried beef ho fun. Neither is wrong; neither may be what you wanted today. AI is in a worse position: it cannot ask "What would you like?" It simply serves its best guess.

When you state year level, learning focus, class profile and return format, you are ordering with specifics: less chilli, no spring onion, rice on the side, smaller portion. Same kitchen, much closer plate.

A one-line prompt returns the average of everything the model has ever read; a prompt that states audience, purpose, class profile and return format returns material you can take straight into class.

For 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.

Eighteen lessons in four modules

The full prompt engineering course has eighteen lessons in four modules:

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

Two prompts you can use today

Skeleton version — replace the bracketed parts; lesson 2 shows how to turn this into a more stable structure:

You are a [planning / marking / admin] assistant for a Hong Kong [primary / secondary] [subject] teacher.
I need [one sentence on the deliverable, e.g. "a lesson starter activity"].
Audience: [year level], unit "[name]", learning focus: [list]. Class profile: [size, ability spread, SEN needs].
Return [list / table / prose], [how many items], [what each item must contain].
Use Hong Kong English and local curriculum terminology; mark uncertain facts, figures or dates "teacher to verify"; give me one version first and I will ask for revisions.

Upgrader — if you already have a weak prompt, do not start from zero:

You are a prompt engineering coach working with Hong Kong classroom teachers.
Review the prompt below, list three missing elements (one sentence each), rewrite it, then one sentence on what difference I should expect.
My real use is [year level, subject, scenario].
The rewrite must state who this is for, what to do, class conditions and return format; leave square brackets where I must supply detail — do not invent class facts.
Mark uncertain facts "teacher to verify".

My original prompt: "[paste]"

Where people drift off course

• Writing the prompt as keywords. "P5 General Studies environment worksheet" produces stitched-together generic content. Write full sentences and say which lesson this material is for. Self-check: read your prompt aloud — if there is no verb (design, mark, summarise), it is probably still too thin.
• Asking for four deliverables at once. "Write the lesson plan, slides, worksheet and letter home" makes all four shallow. Four focused prompts take less total time. Self-check: count the word "and" in your task line; more than one deliverable means split the request.
• 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. Self-check: if removing the year level leaves a prompt that still makes sense, it is probably too generic.
• Treating the first output as final. Treat it as a draft and reply "item three is too advanced — use an everyday example instead". Revision is far faster than regenerating, and the model still has the earlier context. Self-check: if the first version is entirely unusable, you usually omitted class profile or format, not a broken model.

Try it once

Time: 10–15 minutes.

Do this: Pick a unit you are actually teaching next week. Write a one-line prompt (like the teacher's first attempt), then rewrite it using the skeleton above. Run both through the same AI tool.

Hand in: Keep both first outputs. In three sentences, note three kinds of information the second prompt added (for example year level, class profile, return format) and which output you would rather take into class.

You got it right when: at least one item from the second output can be pasted into a plan or worksheet without a full rewrite, and you can name one sentence in the prompt whose removal would immediately flatten the answer — which means you are starting to see which words actually do the work.