AI questions that actually look like the syllabus

Feature A · 6 min read

The first time I asked a general AI to "give me a Primary 5 math question", it spat out a perfectly fine algebra problem. It also could have been a Primary 5 question in California, or London, or anywhere. Nothing in the wording tied it to the Singapore syllabus. That bothered me, because the whole point of practising is to practise the thing you'll be tested on.

So the question engine in Professor Turtle isn't just "ask an LLM for a math problem". There's a stack of decisions underneath that decide what gets generated.

It starts with the curriculum, not the AI

When you pick your level and subject, the app pulls the Singapore syllabus scope for that band — what's been taught, what hasn't, what command words are used, what the expected difficulty is. The model is told: "you're writing a Primary 5 Science question on Energy Conversion. Use the MOE phrasing style. Use SEAB's mark-scheme conventions. Don't introduce content the student hasn't been taught yet."

That last bit matters. A child shouldn't be marked down on vocabulary or a concept they haven't covered. The whole syllabus structure (which topic falls under which band, which sub-topic, which question types are typical) feeds the prompt.

Inspired by MOE materials, guided by SEAB

I won't pretend the AI is doing what a real marker does — it can't, and you should never let it replace a teacher. But the model was tuned against two specific anchors:

• MOE teaching materials. These shape the style of the questions — the phrasing, the context (Singapore-relevant scenarios where possible), the way command words are used.
• SEAB marking rubrics. These shape how the AI marks the answers — what counts as a method mark, what's an answer mark, what partial credit looks like.

That second anchor is the bit I worked on longest, because marking open-ended answers is a Singapore-specific skill. A Primary 6 science explanation has a recognisable shape, and the AI has been coached to recognise it.

What this means for a child sitting down to practise

In practice, here's what shows up:

• Vocabulary matches the textbook. "Mass" not "weight" where it matters; "energy conversion" not "energy change" if that's what the syllabus uses.
• Command words are correct. "State", "describe", "explain", "suggest" each have a specific meaning in MOE papers, and the model knows the difference.
• Difficulty stays in band. A Sec 1 question doesn't accidentally require Sec 3 knowledge.
• No two sessions are identical. Every time you generate a fresh set, the questions are new — different numbers, different scenarios, different ways of asking the same concept.

The honest bit

Is it perfect? No. The model still occasionally produces a question that's a bit off — too easy, too hard, or phrased in a way that doesn't quite match what the child sees in school. That's why there's a daily token quota, and why I'd always recommend treating the app as practice, not as the source of truth. If a question looks weird, skip it and generate another one.

What it doesn't do

A few things I deliberately left out:

• No past-year papers. I'd rather not scrape official papers — the AI generates fresh questions in the same style, but they aren't copies of real papers.
• No "harder than the syllabus" mode. Some kids want to be challenged past their level; that's a tutor's job, not an app's.
• No drawing or construction questions. If a question needs a protractor or a compass, it isn't generated — students don't have those tools in a phone.