The most powerful AI workflows I've found aren't about prompt engineering. They're about conversation.
One reason LLMs are so good at solving homework problems is that the problem statements are designed to collapse the possibility space in which students work. We provide the geometry, operating conditions, assumptions, constraints, and objectives. The student works within that scaffold. That's exactly what we're doing when we chat with an LLM.
So where does this leave us with genuine knowledge work? The big, messy, Capital-P Problems?
Think about a time when you were part of a group that solved something difficult. Was there a single person with a complete vision assigning tasks to everyone else? Or was it a more ambiguous process?
In my experience, it's almost always the latter.
People sit around and talk. Ideas emerge. Most of them aren't very good. A few gain traction. There's a lot of "yes, and..." and "no, but..." and gradually the problem becomes more defined and the path forward starts to emerge.
This is where I've found the highest value from AI.
Not in producing one-shot answers or replacing expertise or executing a perfectly engineered prompt. Instead, helping me sort out half-formed ideas. Helping me brainstorm solutions for problems that I know exist but don't yet fully understand. Helping me identify blind spots, explore unfamiliar territory, and find structure in ambiguity.
The LLMs were trained on natural human language: books, articles, chat logs, forum discussions, and countless examples of people working through ideas together. The interactions that collapse possibility space for deep knowledge work often look surprisingly similar to the interactions that do the same thing in human collaboration.
They look like conversations.
They look like telling a colleague what's frustrating you about a problem. They look like sharing information that may not seem relevant at first. They look like discussing constraints, concerns, goals, and dead ends. Sometimes they even look like talking about how you feel about the problem, because those feelings often reveal constraints that are every bit as important as a physics parameter or a process objective.
So if you find yourself frustrated with AI output and wondering where all these high-impact workflows are hiding, consider talking to it for real.