I’m of the opinion that most AI coding tutorials out there aren’t that useful. Even worse, I think they might not be the best thing to learn from.

I have nothing against content creators evangelizing better ways to build with AI. At the end of the day, making people work better with AI leads to a net increase in the creation of economic value. What concerns me is that to survive the attention economy and evangelize at all, you have to keep showcasing the next new thing. Today’s AI coding evangelists look a lot like the fullstack web framework evangelists who got so much attention in the pre-AI world. Both advertise things that promise to remove work from people — essentially selling shovels so others can mine gold.

The consequence is that these things also remove critical thinking about how things actually work, as everyone trends toward easy abstractions. In the AI era it’s Claude skills and orchestration frameworks; pre-AI it was things like Next.js templates. Easy abstractions let you not think about what’s happening under the hood. That can be fine in some cases, but in others you’re depending on the abstraction to just work and that’s not guaranteed.

This matters even more now because an LLM is at its core a function that produces non-deterministic outputs: it gives a probability distribution over the next token. Making sure an LLM says the right thing is very hard and still an evolving art. Building good evals is very nontrivial, let alone making an agent like Claude Code do the right thing where agentic evaluation is doubly hard.

I’d think twice before reaching for an AI coding framework. If it genuinely aligns with how you work, use it. But most of my own experience has been vanilla Claude Code + a few homegrown skills to automate repetitive things + my own judgment on how to build things. The last thing you’d want is the AI agents taking the wheel and you blindly trusting its outputs.