I can't code. Not in the classic sense — no computer science degree, no years in a dev team. I'm a dad and a product person.

And yet SchulHeld now runs as a working demo: a GDPR-first school app with an AI tutor, a secure class channel, and direct parent–teacher communication. Trilingual (German, English, Greek), built for the DACH region and Greece.

I built it in dialogue with Claude (Anthropic) — step by step, in conversation. People now call this "vibe coding." Instead of typing every line myself, I describe, decide, review, correct — and the AI implements.

After roughly 150 hours of work, I want to give an honest account. Not the hype version. The real one.

How it started — at the kitchen table

Two things wouldn't let me go.

First, the daily battle at home: homework versus screen time. The phone as leverage, as punishment. I wanted to flip that — turn the device my kids love into something that motivates them to learn, instead of dividing us.

Second, the lazy shortcut in every debate: "AI in school = letting students cheat." I found that both wrong and unambitious. A good AI should do the opposite: not solve the homework, but ask questions, explain, push for original thinking. Socrates, not a cheat sheet.

Out of those two frustrations came an idea. And out of the idea — over many evenings and weekends — came a product.

How it evolved

What started as "a small app" grew into a full system:

Four roles, each with its own interface: student, teacher, parent, school admin.
A secure class channel instead of private WhatsApp groups — account and communication data stay in the EU.
A Socratic AI tutor that explains and asks follow-up questions but never hands over the answer — switchable off by the school or parents at any time.
An "AI driver's licence" — a module that teaches kids, age-appropriately, to use AI critically: spotting hallucinations, checking sources, understanding bias.
A spaced-repetition flashcard system (FSRS), gamification (XP, streaks, quests), homework and exam management, timetable.
Privacy and the EU AI Act as a design principle from day one — human oversight (Art. 14), transparency (Art. 50), an audit log.

There were moments I could barely believe it worked — when the first AI tutor actually asked a demo student a sharp follow-up question instead of spitting out the answer. And there were days I spent three hours hunting a single stubborn bug.

The numbers — real, from the Git history

No estimated marketing figures. This is what's actually in the code:

~150 hours of work
~107,000 lines of code across 351 files
182 React components, 7 backend modules (Cloud Functions)
407 automated tests
3 languages (DE/EN/EL), throughout
Stack: React/TypeScript, Firebase/Firestore, Cloud Functions, Claude via Google Vertex AI, Vercel

To be honest about it: lines of code is a weak metric. AI-generated code tends to be verbose, and the trilingual copy and legal text alone inflate the count. The number shows scope, not genius. But it shows scope honestly.

The comparison that either stings or inspires

Imagine a traditional team had built the same app. What would it have taken?

My estimate — based on scope of functionality, not line count, with assumptions stated openly:

Team: ~3–4 people — a senior full-stack engineer as lead, one or two mid-level devs, plus part-time design and project management/QA.
Duration: ~6–9 months to reach this demo-ready state.
Cost (EU/Germany): roughly €200,000–450,000 — toward the lower end with employees fully loaded, toward the higher end via an agency.

In person-hours, a team like that quickly reaches 3,500–4,000 hours. I needed about 150.

This isn't exact science — but the order of magnitude is clear: roughly 20–25× less time, and costs in the range of a few hundred euros (AI tokens + hosting) instead of six figures.

Where vibe coding clearly wins

Speed & cost: Weeks instead of months. Labour cost essentially zero.
Breadth: One person touches everything — frontend, backend, copy, design, privacy. No "that's not my ticket."
Zero coordination overhead: no sprints, no handovers, no meetings. Thought → implementation in minutes.
Democratisation: And this is the real point. People with product vision but no CS degree can suddenly reach what used to strictly require a team. The barrier to entry has collapsed.

Where a team wins — and this is the honest part

Vibe coding gets sold as "10x." The truth is more nuanced. The exact mistakes that happened to me are the ones an experienced team with a review culture would have caught earlier:

Truthfulness of claims: At one point my app made privacy promises that were too strong ("runs exclusively in the EU"), even though the AI was still temporarily routing through a US provider. The AI writes what you tell it — whether it's true remains your responsibility.
Cleanliness & security: There were secrets in the Git history and branch chaos. A senior would have avoided that by default.
Deep architecture: Decisions that only bite you two years later — an experienced team sees them coming. An AI optimises for what you ask right now.
Domain depth: Real Greek pedagogy, solid legal review — AI can prepare these, but not replace them.

In other words: vibe coding gets you to 90% astonishingly fast. The last 10% — truth, compliance, maintainability — still costs judgment and review.

What I learned

My most valuable contribution wasn't a single line of code. It was product instinct, the persistent "no, that's not right," and simply staying with it when something didn't work.

The AI wasn't a tool I operated — more like a tireless, extremely fast junior-to-senior hybrid that implements anything, but needs direction, taste, and accountability from me. Anyone who thinks they can just say "build me a school app" will fail. Anyone who knows what they want and why gets a lever that didn't exist before.

In summary

Vibe coding doesn't replace a development team — it shifts who's even allowed to start.

A single person with a clear vision can now reach, in ~150 hours, a state that used to take half a team and half a year. That's a quiet revolution.

But: responsibility for truth, privacy, and quality does not automatically scale with the speed. That's exactly why I deliberately call SchulHeld an honest demo today — not a finished product.

And that's exactly why I'm writing this. I'm looking for:

EdTech experts to tell me honestly whether there's real potential here.
Pilot schools willing to test it.
Partners in crime who'd like to build SchulHeld with me.

👉 Try the demo (link in the comments) — 2 minutes, no sign-up. And tell me: where do you see the limit of vibe coding?