How I Built My Own Tutor, Textbook, and Roadmap to Learn a New Stack in a Week

Interactive learning with AI at its finest. Instead of buying a course, I had AI build me one — around a real project I actually wanted to build.

A few simple agent rules I had to setup to make this project possible.

Between new jobs and new projects, I have to pick up new things pretty often. The usual move is to go buy a course on the topic, and that’s what I did for most of my career. But courses are generic, they go stale, and most of them just have you build yet another boring to-do app you’ll never open again. Like a lot of you, I learn way better when I’m actually interested in what I’m working on — and I just can’t get excited about weather apps and to-do lists anymore, you know?

There’s a concept I really love called passion-based learning: the idea that we learn best when we’re driven by genuine curiosity about something that matters to us, not by a generic syllabus. It’s true for all of us. But there is more: we all have our different learning styles and I’m a definitely a visual tinkerer kind of person, so most of the boring textbooks/tutorial just don’t work for me.

So, I tried something else: I sat down with AI and had it build me a complete, personalized learning system — a textbook, a tutor, a roadmap I check off, a quiz deck, even architecture diagrams — all wrapped around a real project I’d genuinely use. Using the stack I’m actually trying to master. This is how that went, and how you can copy it for whatever you’re trying to learn.

One of the interactive lessons in my “handbook,” which I review before I write any code. Writing it all by hand again? Honestly, exciting.

Rule #1: learn by building something you actually want

The fastest way to kill motivation is to learn in the abstract. So before any of the study material, I picked a fun project I’d want even if I never got the job: a local, always-on onboarding copilot with a knowledge graph.

The idea: a tool that’s always running, listens to my calls, ingests the docs, and during a call quietly explains every acronym, name, and project as it comes up. Afterward it saves the call into a growing “company brain” so the next answer is smarter. When it’s not sure, it hands off to a human. It used the exact stack I was trying to learn.

Claude can be a great teacher.

The architecture became the curriculum. Every step — a thing I have to learn. Learning stopped being a reading list and became “make this work.”

What “the AI built me a course” actually means

Before I started, I had it generate a whole set of artifacts up front, like a course author would:

• A short textbook — numbered sections (§1–§15) covering the all topics I had to learn. Numbered so my AI tutor can say “read §8 before you write this part.”
• An interactive version of that textbook — same sections in a web page, but with quizzes and little live demos.
• A roadmap I check off — seven phases, each with concrete steps, a quiz, an interview question it prepares me for, and a commit point.
• A quiz deck for spaced repetition, and a learning log the tutor keeps updated.
• Agent rules — the most important file. More on that next.

The trick: write rules that make the AI a teacher, not an autocomplete

🐌

This is the difference between “AI did my homework” and “AI taught me.” I wrote a rules file that the AI reads every session, and the core rule is blunt: I write the code. You coach.

So when I hit a new concept, the AI doesn’t hand me a finished file. It explains the idea in a small chunk, points me to the right textbook section, gives me a tiny example, and then makes me write the real thing. After I write it, it reviews what I did, tells me honestly what’s not idiomatic, and quizzes me before we move on. Plumbing I’ve already learned, it can write for me with an explanation. The new stuff is always mine to type.

It feels slower in the moment. That’s the point. The friction is where the learning happens, and the quizzes plus the “explain it back to me in three sentences before you advance” gate mean I can’t fool myself into thinking I get something when I don’t.

The honest caveats

This isn’t magic, and I want to be straight about it:

• You still have to do the work. If you let the AI write the code, you learn nothing. The whole system is built to stop that, but you have to mean it.
• The AI is confidently wrong sometimes. Treat it like a smart, fast, occasionally-wrong study partner, not an oracle. Verify the things that matter.
• Curated > generated, sometimes. For deep theory I still reach for the official docs. Or ask Claude to read them. The AI textbook is a fantastic on-ramp and reference, not the final word.

The recipe, if you want to steal it

1. Pick a real project in the stack you’re learning. Something you’d want to exist anyway.
2. Have the AI turn the architecture into a curriculum — what do you need to learn to make each piece work?
3. Generate the materials: a numbered textbook, a checkable roadmap, a quiz deck.
4. Write tutor rules that keep your hands on the keyboard. “I write the code, you coach, quiz me, point me to sections.” This one file changes everything.
5. Work the loop: read a section → write the code → get quizzed → log what to revisit → commit. Repeat.

I’m not done yet but I already have something a paid course never gave me: a tutor that knows exactly what I’m building, a textbook written for my project, and a roadmap I can physically check off. For free.

If you’ve been putting off learning something because the courses look generic or expensive, try this instead. Build the thing you want, and let the AI teach you on the way there.