I introduced Gemini, NotebookLM, and Synthesia into a team's daily workflow this year, expecting to spend most of my time on tooling questions. I spent almost none. The hard part was never which model or which button. It was the same hard part it's always been: can you describe what you actually want clearly enough that another party — human or model — doesn't have to guess.
The briefs that failed before AI ever entered the room
Long before prompt engineering was a job title, I was writing briefs for designers and freelance writers, and the bad ones had a pattern: vague adjectives ("make it pop," "more premium feel"), missing constraints (no audience, no format, no length), and no example of what "good" looked like. Those briefs failed with humans for the same reason they fail with a model — the executor is left inventing the parts you didn't specify, and then you're disappointed by choices you never actually made.
The fix was never a clever phrase. It was structure: who is this for, what does it need to do, what does the ceiling look like, what's explicitly out of scope. That structure ports directly onto a prompt with almost no translation.
Where the analogy breaks
It's not a perfect mapping. A model won't push back on a bad brief the way a good designer will — it will confidently produce the wrong thing rather than ask a clarifying question, unless you've explicitly told it to. So the one genuinely new skill is building the clarifying question into the brief itself: asking the tool to state its assumptions before it produces the output, so you can catch the wrong guess before it's baked into four paragraphs of copy.
The tool changed. The actual skill — saying precisely what you mean — didn't.
If your team is rolling out AI tools and the friction feels like a tooling problem, it's worth checking whether it's actually a briefing problem wearing a tooling costume. In my experience, it usually is.