China’s college entrance exam — the Gaokao — is taken by 12 million students every year. One exam. One day. Every major life decision that follows.
My son is one of those 12 million.
Last month I found him at his desk at 11pm, notebook open, writing the letter C in a column of 20 identical rows. He was checking answers from a 1998 past paper. By hand. One by one.
I counted the Word files on his laptop: 58 documents. 30 megabytes. Exam papers going back to 1990.
He had every resource he needed. He could not use any of it.
The old way of “studying” #
Here is what actually happened when he sat down to practice with those files:
- Open a Word file
- Read through the passage — eyes scrolling, losing his place constantly
- Write answers A, B, C, D in a separate notebook
- Manually scroll to the answer key buried at the bottom — sometimes a separate file entirely
- Check each answer one by one
- Hit an unfamiliar word. Open Google Translate in another tab. Copy. Paste. Switch back.
- Lose focus. Repeat.
One full cloze test this way: 30+ minutes, most of it spent on logistics rather than actual learning. The friction was the problem. Not laziness. Friction.
My first attempt failed #
I decided to build something better. My first instinct was GPT-4.
I fed it the first few exam papers and asked it to generate a quiz interface. It gave me React code — which requires Node.js, npm, a build step, a terminal. For a 16-year-old who just needs to answer multiple choice questions on his phone before school. Wrong tool entirely.
So I switched to Claude and tried a different approach: feed in all 58 Word documents and extract every question, answer, and explanation automatically.
That is when I discovered the real problem. The 2016 papers worked fine. The 1991 papers used a completely different format for answer keys. The 1995 paper had an answer key line that looked exactly like an option line — and the parser kept misreading it. The 2005 paper used a fullwidth hyphen (－) instead of a regular dash in the answer ranges, which broke the extractor entirely.
Three sessions. Dozens of debugging rounds. Edge case after edge case.
That is what the token bill actually paid for.
$74. And I would do it again.
If this tool saves my son 15 minutes per session and he uses it 100 times before the exam, that is 25 hours recovered — hours spent actually learning instead of fighting a file manager.
What we ended up with #
One HTML file. 395 kilobytes. No installation. No login. No internet required once you have it. Open in any browser on any device.
What that 395KB contains:
63 complete cloze tests, 1990–2023 —** 1,276 questionstotal 36 reading comprehension sets**—** 441 questions**total- Full answer explanations for every question
- A built-in vocabulary system covering nearly 2,000 words
The feature I did not plan — but changed everything #
Click an answer. The passage updates instantly. Green for correct, automatic reveal for wrong — explanation right there, no scrolling required.
But the thing that ended up mattering most was something I added almost as an afterthought: every word in every passage is now clickable.
Tap any unfamiliar word and a small popup appears with phonetic pronunciation and Chinese definition. No tab switching. No copy-pasting. No breaking concentration to reach for a dictionary.
For a student doing exam practice, the bottleneck is rarely whether they can read English at all. It is the specific academic vocabulary they do not know yet — the words that appear again and again across 33 years of papers. Those words used to require a separate tool. Now they do not.
30MB to 400KB — what actually changed? #
The 30MB was not wasted. It was raw material. Every answer, every explanation, every passage was already in those Word files. AI helped compress 33 years of content into structured data, strip the formatting noise, and rebuild it as something interactive.
The size reduction is not magic. It is the difference between a format built for printing and a format built for learning.
One file. Offline. Instant feedback. Click-to-translate.
And somewhere in building this, a question started bothering me that I have not been able to shake:
Will our children even need to learn English — or will AI just translate everything for them in real time?
I do not have an answer yet. But I think the process of struggling with a language, looking words up, getting answers wrong and understanding why — that might matter even if the destination changes. My son is not just preparing for an exam. He is learning how to learn.
That part, I do not think AI replaces.
Try the tools — free, no account needed #
📚 Gaokao English Past Papers — Free Interactive Practice
→ Complete Cloze Tests — 63 sets · 1,276 questions · 1990–2023
→ Reading Comprehension — 36 sets · 441 questions · 2002–2023
→ Full Gaokao English Toolkit — vocabulary, grammar, writing & more (19 tools total)