the bulk of the skill lies in shoring up the fundamental limitations of llms

they cannot be held accountable - therefore, you need the skill and expertise to verify outputs, or set up a verifier

they have a limit to their context window - therefore, you need the skill of knowing what context is useful and exposing that

they cost money and time to run - therefore, you need the skill of learning what parts can be handed off to code, when you can use a cheaper model, when you tradeoff accuracy vs latency

they are nondeterministic - therefore, you need the skill to set up evals, guardrails, knowing when to reach for the llm vs. code depending on the problem

they are not omniscient - therefore, you need the skill of setting up the processes and infra to supply legible context to the llms

etc etc

We still have yet to find what is possible without a career's worth of experience doing it all by-hand.