That is the whole problem imho. I've found that I can use LLMs to do programming only if I fully understand the problem and solution. Because if I don't, it will just pretend that I'm right and happily spend hours trying to implement a broken idea.
The problem is that it's very hard to known whether my understanding of something is sufficient to have claude propose a solution and for me to know if it is going to work. If my understanding of the problem is incorrect or incomplete, the plan will look fine too me, but it will be wrong.
If I start working on something from poor understanding, I will notice and improve my understanding. A LLM will just deceive and try to do the impossible anyway.
Also, it overcooks everything, atleast 50-60% of the code it generates are pointlessly verbose abstractions. agian: imho, ymmv, ianal, not financial advice ;)