I stopped asking AI for answers. I asked it to quiz me instead.

I use Docker, but not every day. I know the concepts, but the syntax fades. Multi-stage builds? I could explain the idea, but write the actual Dockerfile from scratch? Not so sure.

Sound familiar?

This time I tried something different. Instead of re-reading the docs (again), I asked the AI to teach me — with challenges, not explanations.

It gave me a scenario: “You inherited a 940MB Node image. Your tech lead says get it under 100MB. Rewrite the Dockerfile.”

I stared at it. I was rusty. I didn’t fully remember what npm run build produces. So the AI broke it down — not by giving me the answer, but by giving me just enough context to take the next step myself. “Think of it like a construction site. Stage 1 is the workshop. Stage 2 is the delivery truck. You only put the finished product on the truck.”

With that, I wrote my attempt. It wasn’t perfect — I missed --from=builder, forgot a RUN keyword, got the copy order wrong. But the structure was mine. Two stages, alpine for runtime, production-only deps. I’d reasoned my way there.

That’s what made it stick. Not the AI’s corrections — the fact that I had to think it through before seeing the answer. The AI coached me into taking the lead, and only stepped in to tighten what I’d built.

20 minutes. One challenge. I walked away with a Dockerfile I actually understand, not one I just copy-pasted from a chat.

I tend to struggle with syntax and concepts I don’t practice often enough. Now I can get a structured refresh in a breeze — and it feels less like studying and more like pair programming with a patient colleague.

The trick? Don’t ask AI to explain things. Ask it to challenge you.