A couple ideas here:
- Keep coding your own stuff on the side.
- Don't let the AI give you code that you don't understand.
- Understand the greater architecture of what it's building.
- Be specific in your instructions.
Yesterday I was using Copilot and Claude to write a clone of the old trn Usenet newsreader in Rust. Got pretty far in about 7 hours. 8300 lines of code and 8300 words of design docs in Markdown. The architecture is pretty good, and the thing works. (Still missing is the thread graph display and posting, but I know those can be added without breaking anything.)
Don't get me wrong: if there were a button that destroyed LLMs forever, I'd have already pressed it. I have other projects that I write by hand even though LLMs could definitely do it because I like the challenge and I want to stay sharp. Reading code doesn't do it.