Ask HN: How do you learn with LLMs?

I've had a bit of success building little games. Not exactly what you'd consider "advanced technical stuff", but I used AI to help me get better at reading the water for fly fishing.

For more technical stuff, like deep database expertise, I used it probably like you do, just tons of follow-ups, ELI5 commands, etc.

I love the question. I haven't been too intentional in LLM driven learning but for code I do like to get the LLM to make the code and then I review it closely making sure I understand it. Question it etc.

Socratic method seems like a good idea maybe there are skills out there for it.

I don’t typically use AI when programming.

When putting my perf tests behind a setTimeout I found I could push my WebSocket message send speed from 3,250,000 messages per second to about 3,250,000 messages in about 0.002 seconds. I did not believe the numbers so I asked about this on IRC. People said my perf must be broken and I should use AI to find the error. After a few factual corrections about the code Claude confirmed the numbers as valid.

Learn by doing. Even more true nowadays, I think the best way to learn and understand how LLM and advanced topic like agent harness works is by, get your hands dirty and start building one. Trust me, it will pay off, one day.

I generally outline the problem first that I would like to solve, then ask for best practice approaches and industry standards that apply to situations. Next, I ask to summarize the pros and cons of the various possible solutions and how they apply to my particular situation.

I remember using LLMs to learn SEO last year. I knew some SEO basics before, not a total beginner, but I'd never done it hands-on. When I decided to learn it systematically, my approach was: first, I used deep research to compile a complete knowledge base. For anything I didn't understand, I went back and forth asking questions for several rounds until I had a solid grasp. Then I had Codex design a practical project for me — it suggested starting with long-tail keyword research. Following Codex's advice, I step by step built a website about Black Friday deals (didn't get much traffic so I shut it down later). From the first prompt discussing requirements to the site getting real organic traffic took about a month. Whenever GSC or GA had new data, Codex would pull it and notify me. Now, I wouldn't call myself an expert, but I'm at least competent.

I have done a variety of projects with LLMs. This incldues looking up articles on AI and libraries, then checking their quality. It also includes learning vibe coding, by doing it. I learned to write my prompts first in the free engines like Claude and Gemini, then put the prompts into the paid engine I am using. I also learned how the audit tools work in Claude, Gemini, and Replit to make the application I am working on more stable. One engine can be used to check another. I also have learned how to use Claude and Gemini to fix Windows 11 problems, they have pretty solid basic PC support information. You can use the phone if your computer is not working to troubleshoot with AI.

Clarification: I meant interesting applications of LLMs to improve the “learning process” itself