[theshrike79]

I'm thinking that the calculator analogy is better and better

An LLM will produce a solution for you, but in most cases you need to know enough about the problem to ask it to apply the right tools. Or even have enough "known unknowns" that you can direct it to search for a better solution.

Like a trivial example: an LLM will happily write a CLI argument parser from scratch. It WILL work. It'll also be complete crap UX wise.

But if you know that a specific library (Cobra, Kong, urfave/cli) fits this specific problem better, you can ask it to use it specifically.

Same with calculators - it'll always tell you the correct answer to whatever calculation you give it. But YOU need to know that this specific problem was solved by a really smart person 100 years ago and there's a formula you can apply to solve it faster - still using the calculator, but now you're using refined algebra instead of brute-forcing a solution