Collect no child data at all
The app runs on anonymous device identity, with an age range as the only targeting signal. That sidesteps COPPA risk and earns trust, and it means there is nothing sensitive to leak. A child profile would have improved targeting slightly and cost the whole privacy position.
A plan completes by being viewed
No quizzes, no streaks, no pressure mechanics. The goal is parent learning, not engagement farming — and engagement mechanics would have undercut the calm, supportive tone the product depends on. Positioned deliberately as learning rather than coaching: it builds a parent's judgement instead of prescribing or grading.
A curated library, not open AI generation
Wiring the app to generate guidance on demand would have been faster and would have quietly reintroduced the exact trust problem the product exists to solve. A closed set of vetted sources is the value proposition, not an implementation detail.
A no-shame filter, enforced above source authority
Nothing that blames or shames the parent or child gets in, however reputable the source it came from. Three rules hold the line: every claim paraphrased in plain language and copyright-safe, every claim cited with a working link, and nothing that shames.
Links are audited, not trusted
When one source quietly removed its articles, the dead links were caught by a routine audit and the entire topic was repointed to a live, equally credible source — rather than left pointing at a homepage and silently degrading.
Production-first, even in beta
It would have been easy to wire the app straight to an AI API and ship a demo. Instead the app never calls an external model directly — only its own backend proxy, which keeps API keys off the device and puts every outbound call behind one controllable boundary. Summaries are pre-generated and cached rather than produced per request, so the app stays fast offline and predictable under load. The principle: you are always building for production, even while testing in beta.