Teaching by example, and output consistency

Short answer: when you have explained the format three times and the output is still wrong, stop adding adjectives. Paste two examples and the problem usually disappears on the spot.

A real panel problem

A Chinese Language panel head needed a Primary 6 vocabulary revision paper, with five teachers writing ten items each. She tried AI first:

Write ten Primary 6 vocabulary items, multiple choice, moderate difficulty.

Item one came back as a gap fill, item three as a matching task, item seven as four options printed on one line, item nine with an explanation the others lacked. She added "please keep the format consistent" and got consistency — in the format she did not want.

Second approach, pasting two past items she was happy with:

You are a question-setting assistant for a Hong Kong primary Chinese Language teacher.
Task: continue with ten more Primary 6 vocabulary items in exactly the format of my two examples.
Context: first-term Primary 6 revision paper. The learning focus is distinguishing near-synonyms by context. Keep vocabulary within textbook units 1 to 4.

Example 1:
[stem sentence with a blank]
A. option  B. option  C. option  D. option
Answer: A
Distinction: one sentence explaining why the other three do not fit this context.

Example 2:
[stem sentence with a blank]
A. option  B. option  C. option  D. option
Answer: A
Distinction: one sentence explaining the difference in usage.

Output format: exactly the four-line structure of the examples — stem, four options on one line, answer, distinction. One blank line between items.
Constraints: all four options must be genuine near-synonyms, with no obvious filler; vocabulary must stay within Primary 6 range; if you are unsure whether a word appears in the primary common word list, mark it "teacher to verify" rather than deciding yourself.

This time all ten items matched the examples exactly, and the panel head could merge five teachers' items into one clean paper.

Why this works

A model can learn what shape you want in two ways: you describe it, or you show it. A description has to be translated — you say "moderate difficulty" and the model has to guess what that means. An example needs no translation. Layout, sentence length, punctuation and register all copy across at once.

It is like training a new teacher to write report comments. Tell them "be more specific and more positive" and they will still miss. Show them three comments you wrote yourself and their next one lands. Examples carry the things you cannot articulate but recognise instantly.

One caution: the model imitates content as well as form. Choose example topics deliberately different from your target topic, or you will get ten near-duplicates of the examples.

Template 1: fixing the register of written feedback

A Primary 4 Chinese Language teacher wants a consistent voice across a set of composition comments, rather than something stiff one day and gushing the next.

You are a feedback assistant for a Hong Kong primary Chinese Language teacher.
Task: write one comment per composition, matching the register and structure of the three examples below.
Context: Primary 4 narrative unit, learning focus is conveying feelings through action and dialogue. Students must be able to read the comment themselves and know what to do next.

Example 1: You described the weather on sports day very concretely, so the reader is right there. Next step, try adding one action after "I was so nervous" — sweaty palms, or checking the clock — and the feeling becomes visible.
Example 2: The order of your three paragraphs is clear and the piece holds together. Have another look at the ending. Right now it is "what a happy day"; replacing that with one thing you actually did would leave more resonance.
Example 3: Your dialogue sounds natural and the characters speak like primary pupils. Watch the punctuation — full stops belong inside the closing quotation mark, and there are four places to fix.

Output format: one paragraph per composition, three to four sentences. First sentence names one specific strength and quotes the student's own words. Second names one area to improve. Third gives one concrete action the student can take immediately.
Constraints: no marks or grades, feedback only; do not rewrite the student's sentences or paragraphs, only indicate a direction; any quotation must be words the student actually wrote, never invented.

Template 2: a negative example to block drift

A General Studies teacher keeps getting closed questions that can be answered in a single line.

You are a planning assistant for a Hong Kong primary General Studies teacher.
Task: design six inquiry questions for the Primary 5 unit on public transport in Hong Kong.
Context: students will research in groups and present in two weeks. The learning focus is comparing the effects of different modes of transport from several perspectives.

Good example: if the government cancelled every bus route in our district tomorrow, who would be affected? Name at least three different groups of people.
Good example: why do some people take the MTR and others take a minibus for exactly the same journey?

Not like this: how many MTR lines does Hong Kong have? (a closed question answered by a lookup, not an inquiry)
Not like this: what is your favourite way to travel and why? (personal preference, nothing to compare or argue)

Output format: six questions, each followed by one sentence on what students must gather to answer it, and one sentence giving the teacher a follow-up question that pushes the discussion deeper.
Constraints: not one of the six may be answerable in a single sentence; do not state specific fares or service frequencies — where such data is needed, mark it "teacher to verify" and leave it for the teacher to fill in.

Template 3: turning past papers into examples

The fastest source of examples is material the panel already owns.

You are a question-setting assistant for a Hong Kong secondary [subject] teacher.
Task: first analyse the three past-paper items I paste below and summarise the house style they share, then write eight new items in that same style.
Context: [year level], unit "[name]", a unit test worth 20 marks in 25 minutes.
Output format: part one is five bullet points naming the style features you found (stem length, command verbs, option design, mark allocation); part two is eight new items in exactly the past-paper format, with answers and marks.
Constraints: no new item may reuse a scenario from the past paper; cover only this unit's learning focus, nothing cross-unit; where a figure, regulation or date is involved and you are unsure of it, mark it "teacher to verify".

Past-paper items:
[paste three]

Common mistakes

• Examples that disagree with each other. If the field order or punctuation differs between the three, the model averages them and you get something that matches none. Put them side by side before pasting.
• Example topics too close to the target. Use near-synonyms as your example and then ask for near-synonyms, and all ten items will circle the same two word pairs. Take the example from unit 2 and ask for unit 5.
• Examples that are too long. Paste three full model compositions and the format signal drowns in content. Choose the shortest passage that still demonstrates the shape.
• Forgetting to de-identify. Pasting student work as an example hands over personal data. Strip names and class, or rewrite as a synthetic sample. Lesson 9 and lesson 14 deal with this in detail.

Going further

• A different subject: Mathematics examples must show the full working layout (one step per line, equals signs aligned) or the model compresses the steps into a paragraph. English Language examples should demonstrate the marking code you use, for example T for tense and SV for subject-verb agreement.
• A different year level: the example is itself the year-level signal. Primary 4 examples produce Primary 4 output far more reliably than "make it simpler" in the constraints line.
• SEN adaptation: supply one standard example and one adapted example side by side, and the model will produce both versions for every subsequent item. An adapted example might split the stem into two sentences, one instruction each, with key words in bold.
• Even more consistency: if the output has to import into an assessment system or spreadsheet, examples alone are not enough — specify a machine-readable structure instead. See lesson 15, structured output.

Examples solve "what should it look like". The next lesson handles a different failure: questions that need reasoning, where the model answers quickly and wrongly. See lesson 4, chain of thought and step-by-step instructions.