My approach is to write a bit of totally AI-free code every day. So after a day of Claude, I'll spend at least 30 mins wrestling with something. The gnarlier the better, e.g.leet code or Project Euler-type stuff. I think of it as like lifting weights for the mind. The more I struggle at the edge of my knowledge and skill the better.

The other thing is to give your agent a skill not to solve certain key problems unless explicitly prompted. Write the scaffolding sure, but leave the juicy parts alone. And if I get stuck, I have it enter into a dialogue with me, nudging me towards understanding.

I solved a Leetcode Medium on my first try in about 10 minutes the other day and it felt so amazing. Thinking through architecture, project composition, composability, etc when working with agents felt fine for a while but it made me feel disconnected from what I'm actually building.

Actually engaging my brain to solve the lower-level, on-the-ground code allows me to think of better ways to do things while I'm writing them. It's like writing anything. You start with something you want to convey, a thesis, and then it evolves and becomes better as you write it. An agent will just write it with no thought, as in it will reflect one of the LLM 'ghosts' as Andrej Karpathy puts it, doing something in the same way that someone in the training data has done it on a similar or different problem. This is why I get conniptions now when I am sent generated text or am expected to read it on a public forum. It's disrespectful of the time of every person expected to read it.

That said, I wouldn't like to go back to the before times without having the agentic option available. Ideally, businesses should not mandate how LLM's are to be used at their company, and just let the devs find their own flow. That is, if quality is even a factor that any company optimizes for anymore.