Prompt engineering was the skill of the last couple years. The way you asked the model mattered, and the people who got good at asking pulled ahead. That's mostly over now. Models are good enough that a sharp prompt and a sloppy one land close to the same place, and the gap between users isn't really in the phrasing anymore.

What's actually separating people now is the harness around the model — everything it's given before, during, and after the ask. The files it can see. The tools it can call. What it's been told about prior decisions, what it's been told to ignore, what it's allowed to trust as the source of truth. The memory it carries versus the memory it throws away. A model with a thin prompt and a strong harness beats a model with a perfect prompt and no harness almost every time.

Different skill than prompting, different muscles. Less clever language, more systems thinking. What does the model need to know to do this well. What would I, as a human, be carrying into the room without even thinking about it. What's persistent versus task-specific. What tools does this job actually need.