There's a scene in The Office where Michael Scott asks his accountants to explain the budget "like I'm a second grader."
Everyone in the room thinks he's being ridiculous.
He's not.
The most common mistake operators make when working with AI tools is quietly accepting outputs they don't fully understand. They nod along, copy the answer, move on — because admitting you don't understand feels like weakness.
It isn't. It's the fastest path to worse outputs and compounding errors.
Here's what I actually do:
When an AI assistant explains something in language I don't recognize — a technical architecture, an unfamiliar framework, a process I've never seen before — I stop it and say exactly this:
"I don't know what that means. Explain it in the most basic terms possible, like you're teaching someone with zero background. Then continue."
Not "please simplify." Not "can you clarify?" Full stop, full honesty, Michael Scott mode.
Two things happen, both valuable:
The explanation gets better. AI systems match your stated level. When you perform expertise you don't have, you get expert-level explanations that build on assumptions you can't verify. When you drop the performance, the explanation drops to where you actually are.
You start catching errors. Simple explanations surface assumptions. When you demand basic language, you often discover the model was assuming something that doesn't apply to your situation. That assumption, left unchallenged, would have sent you down the wrong path. The second-grader question is a probe.
The corollary: you can instruct the AI to educate as it executes. Tell it upfront: "As we work through this, explain your reasoning and flag anything I should understand before we proceed. Don't assume I know the terminology."
This is how operators who aren't engineers get production-grade outputs. Not by pretending to know everything. By explicitly instructing the system to meet them where they are.
Michael Scott was onto something. He just didn't have the follow-through.
If this is the framework you're operating with: Opus plans, Sonnet executes, and you ask every question you actually have. The discipline is in the loop, not the pretending.