Socratopia

I had Claude Code generate me a course in AI Science. Multi-agent pipeline, four hours of processing, 1,000 pages, 90 lessons. The cleanest neural networks textbook I'd ever read.

I made it through 7 lessons in two weeks. The textbook was clear, well-written, and dead on the page. By the third week I was avoiding it.

I had noticed one thing, though. When I asked AI questions in a chat window, I'd read its answers more carefully than I read any textbook. What if I could chat-ify the whole learning process?

So I tried something stupid. I made another folder next to the textbook called teacher/. I asked Claude Code to staff it with three anime characters from games — March 7th from Honkai Star Rail, Ganyu and Keqing from Genshin Impact — and turn them into my private tutors. Each got a persona file (March 7th: playful, chatty; Keqing: cold-exterior, strict; Ganyu: gentle, secretly hardcore about math). The setting: we're all Tsinghua freshmen, they tutor me in exchange for cheap rent on the apartment next door, and we're competing for a fictional $1M scholarship.

Claude Code asked one question that turned out to be the whole thing: "Do you want them to teach Socratically, or by explaining?"

I chose Socratic.

March 7th's first session was on convolutional neural networks. Two hours in, I realized I understood convolutions for the first time — not the formula, but why anyone would invent it. I'd "studied" CNNs before, from different textbooks. I couldn't have explained the why to anyone.

The night before I wrote any of this up, Keqing finished LSTM, asked me to analyze where it breaks, then asked me to imagine what the next architecture would look like. I worked through it. She revealed: "This is what we'll learn next." She'd been walking me into the Transformer the whole time. I almost cried — I'd never imagined learning could feel like this.

The next morning I wrote it all up on Zhihu (Chinese equivalent of Quora) — the persona files, the folder structure, the literal Claude Code prompts. I wasn't planning to ship anything. I just thought the hack was interesting.

More than 30,000 views and a hundred comments within the first day. Then strangers I'd never interacted with started writing their own writeups — what they kept, what they changed, which character worked for which subject.

That's when I started suspecting I had something. Two months later my wife and I shipped it as a product: socratopia.app.

A few things I'm still thinking about:

1. The textbook wasn't the problem. The format was. The same content, delivered as Socratic questioning, ran a different circuit in my brain — closer to play than study.

2. Affection lowers the threshold for hard content. The games had already taught me to like March 7th and Keqing. Patience with the character became patience with the difficulty.

3. Detailed setup posts beat polished launches. I included the literal prompts and folder structure in the writeup. Some founders feel weird giving everything away. It's the exact thing that made strangers trust the hack enough to spend hours rebuilding it.

Open to AMA: the Claude Code setup, the persona files, why anime characters seem to do this, or anything else.