It's a trip back to a guesswork stage because most of the training data looks like the process you described your beginnings have been like.

Now you know how to do things first shot. When you write the code, your inner voice reminds you ahead of time to do it like this and that. And that "this and that" is what your prompt needs to have to avoid the trip to guesswork. Working with agents is like working with juniors. If you don't give them the direction and explain how, they'll be as lost as the LLM. The difference is, your brain finds guardrails as you solve the problem, the LLM doesn't have this context, you have to give it upfront. Like, you know, a good manager would.