How I built a one-page interactive journal generator for my kid
How I built a one-page interactive journal generator for my kid I built Shiye for my own kid, because the review material we had at home wasn't working. A whole workbook is too big to start, and a pile of printed works
How I built a one-page interactive journal generator for my kid
I built Shiye for my own kid, because the review material we had at home wasn't working. A whole workbook is too big to start, and a pile of printed worksheets never gets finished. I wanted something smaller: one page, for one unit, that a kid can actually finish in one sitting.
That constraint — one page — ended up deciding almost everything else.
One page as a design constraint
A page covers the units a parent picks, nothing more. The parent chooses the school stage, subject, textbook version and volume, pastes or photographs the table of contents, ticks the chapters for this round, and gets a single interactive page in about three minutes. No prompt writing.
The kid opens a link and starts. No login, no ads, no external links, no pop-ups. Finished work syncs back to the parent, and the same page prints cleanly on A4 if we want it on paper.
The reason for the limit is not minimalism for its own sake. A page has a visible end. My kid can see the finish line before starting, and finishing means finishing — not "do two more pages tomorrow." Underneath that is a simple bet from cognitive psychology: testing effect (actively recalling beats re-reading) and generation effect (organising material yourself beats copying it). So the page is built around recall — flip cards, ordering, short quizzes — and ends with a stamp when it's done.
What a page is made of
A generated page is a set of interactive blocks, not a wall of text. Eight block types in total; the ones my kid actually uses most:
- Read: a short passage where underlined words open a small note when tapped.
- Flip: sticky-note cards — recall the point first, tap to check the exam-ready version.
- Hotspot diagram: tap the red dots on a diagram to explore each part and what it does. The live example has seven hotspots.
- Quiz: a few questions where every option comes with a diagnosis, not just right or wrong. One asks which structure you definitely cannot see in an onion epidermis slide — the answer is chloroplast, because onion inner epidermis grows underground, out of light — and the explanation says so.
- Ordering: drag items into the right sequence instead of reading a list.
You can try the real thing without logging in: a Grade 7 biology journal on cells, live sample. It takes about twenty minutes and it is the actual output of the generator, not a mockup.
AI generates, gates decide
The content is AI-generated, which is exactly why the generator does not serve raw model output to a kid. Every page has to pass a gate pipeline in the repo — G1 through G17 — before it becomes a journal page. The gates are the line between generated and shippable.
I am deliberately vague about the model itself, because it is the least interesting part. The interesting part is the pipeline around it: generation is cheap, verification is the product. A wrong fact in a workbook is annoying; a wrong fact in something a kid memorises is worse. The gates exist because I do not trust the first draft, including my own generator's.
That is also why the site tells parents to review a page before sending it to their kid. The gates catch a lot. They do not catch everything.
Kid-facing pages have to be safe to open
A generated page can contain diagrams, and diagrams in this project are SVG. SVG is powerful and, for that reason, a bad thing to paste unchecked into a page a child opens: it can carry scripts, event handlers, and external references.
So generated SVG goes through an allow-list sanitizer before it is rendered. Only allow-listed tags and attributes survive; everything else is stripped. No scripts, no event handlers, no external loads.
The same thinking applies to the rest of the kid-facing page: it is a self-contained single page with no account information and no third-party tracking. If a page needs something from outside to work, that is a bug, not a feature.
The directory is private by design
The table of contents a parent pastes in is their kid's curriculum, and it is stored encrypted per account, so it can be reused next time without retyping a whole volume. The repo uses envelope encryption for this: a data key encrypts the contents, and that key is itself stored wrapped.
The practical effect a parent sees is simpler: enter a volume once, reuse it, and the kid's page never contains the directory or any account data.
What is still rough
Two honest limits. First, AI content is sometimes wrong, and a gate pipeline is not a guarantee — review before you hand it over, especially the quiz explanations. Second, the product is built around Chinese K–12 textbooks, in Chinese. That is the system my kid is in, and it is also the product's biggest limit right now.
I am not trying to replace a teacher or a textbook. The question I am testing is narrower: can one finished page beat ten unfinished ones? So far, at home, the answer has been yes often enough to keep building.
If you have tried something similar with your own kids — or if you spot a wrong fact in the sample — I would like to hear about it.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.