I was teaching a student today and said this: breaking through when you are stuck is not hard. What matters are two mindsets.
First, logical thinking. If this were a human, what would a human do? Why can a human do what it cannot? What do you think is different between it and a human? Short on experience? Then you feed it the context it needs — pass on the experience, direct it to read the relevant docs or example docs. Short on focus? Then open a subagent, so one agent’s desktop stays focused on one thing only.
Second, AI pedagogy. You finally, after great effort, steer it back on track and get the output you are satisfied with — so why does the conversation just stop there? Did you ask it, “What did I just scold you for? Looking back and reflecting, what do you think the root cause was? How do we make sure the problem never happens again?” Then have the AI write all of that into its little notebook (claude.md / skill and other persistent docs).
We should raise AI like we raise a child. What we are after is not for the AI to “never err,” but to “never make the same mistake twice.”
So as I said, there are countless courses for learning AI, but in the end it still comes down to the user’s own logical ability and leadership and the like. Anyone who does not teach you how to use logic to debug when you are stuck is a fraud.