When we started building LUMI Chinese Institute, an AI-powered Chinese tutor for Hong Kong students aged 4–16, we made the classic edtech mistake: we generated a lot of reading passages and a lot of questions, and assumed quantity was the product.
Then we spent a week auditing every single passage in our library against Hong Kong's actual assessment frameworks — TSA (Territory-wide System Assessment), HKAT (the Secondary One placement exam), and the DSE Chinese curriculum. This post is what we learned, because almost none of it was obvious.
1. Hong Kong doesn't test what you think it tests
The TSA reading papers follow a strict word-count ladder by grade level:
| Level | Passage length | Benchmark |
|---|---|---|
| P1–P2 | 30–100 characters | Pre-TSA, picture-sentence drills |
| P3–P4 | 120–200 characters | TSA P3 → P6 bridge |
| P5–P6 | 230–320 characters | TSA P6 → HKAT S1 placement |
Our original passages drifted up to 25% below or above these bands. Sounds trivial? It isn't: a P4 student practicing on 160-character passages is training on P3 material and will hit the real paper unprepared. Word count is not formatting — it IS the difficulty spec.
2. Every answer must be findable in the text
The single most important rule in Hong Kong elite-school paper design: every question's answer must be directly locatable in the passage. No inference questions where the answer lives in the teacher's head. Our audit found "hard questions" that were actually just ambiguous questions — unanswerable from the text, which punishes careful readers instead of weak ones. All deleted.
3. Distractors come from inside the passage, not from a vocabulary list
This was counterintuitive. A good cloze (fill-in-the-blank) question in the Hong Kong style draws its wrong options from words that appear elsewhere in the same passage — testing whether the student can discriminate between similar words in context, not whether they've memorized more vocabulary.
Generic AI question generators get this wrong constantly. They pull distractors from a frequency dictionary, producing questions where the wrong options are actually more plausible than the right one. We ended up writing an audit script that flags every distractor not present in the source text for human review.
4. Question order follows text order
Question 1's answer should sit in the early paragraphs; question 5's, near the end. This "sequential principle" reduces cognitive load and mirrors how the real papers are structured. Randomized question order — the default in most quiz apps — actively trains students against the exam format.
5. The ability ladder is: recall → explain → reorganize → extend → evaluate
DSE defines five reading levels. A well-formed question set moves up this ladder within a single passage: a vocabulary question, a retrieval question, a "why does the author..." question, then an open response. Our first question bank was 80% recall. The rebuilt bank enforces a mix.
What we built from this
Three things, all open to steal:
- A written spec. Grade-by-grade word counts, paragraph structure rules (P4+ requires ≥3 meaningful paragraphs), question-type coverage matrix (we required all six core question types per grade: punctuation, sentence reconstruction, expansion, picture-to-sentence, MCQ, short answer).
- A regression audit script. Runs on every content change: word-count compliance, cloze-answer-in-text, MCQ index validity, quotation consistency, paragraph counts. Fail = block the deploy. Our content CI, basically.
- The discipline to delete. We removed or rewrote 56 passages that didn't meet the ladder. Losing a third of the library to ship the right two-thirds was the best decision we made.
If you're building anything for the Hong Kong education market: the parents here are extremely sophisticated consumers of assessment design. They will read your sample papers. The product surface that looks like "content" is really your spec compliance.
We're at lumi-chinese.com — there's a 24-hour free trial if you want to see the rebuilt bank in action.
Top comments (0)