Why experienced professionals, people who already know their work well, often get the least useful output from Claude Cowork. The more internalized your judgment is, the harder it is to hand off to an AI agent without a deliberate step first.
An AI interview process pulls that hidden knowledge out before you ask Claude Cowork to do anything, which is the same idea behind the Professional AI Interview Kit.
You've been using AI long enough now that you've probably noticed something. It works fine for simple things. But when you try to hand it something that actually matters, something complex, something that requires real judgment, it comes back generic.
It has taken me 20 years of teaching and hundreds of hours building, breaking, and testing AI systems to realize that the people who struggle most with AI aren't the ones who know the least, they're often the ones who know the most.
It's not the prompt. It's what's missing before the prompt.
You've spent years building expertise. You know what good work looks like. But that knowledge is mostly invisible, even to you, it's just instinct at this point. When you sit down with an AI agent, you suddenly realize you're not actually very good at explaining how you do what you do. That's the gap between what's in your head and what you've managed to get out of it.
The more experienced you are, the worse this problem gets. Early in your career, everything was explicit, you followed processes and checked lists. Then you got good, and the process disappears into instinct and pattern recognition. That's what experience actually is: compressed, internalized judgment. Which is exactly why your AI keeps coming back generically disappointing, it's working with whatever partial description you gave it at the start.
One of the best things you can do with Claude Cowork first is ask it to interview you before you ask it to do anything else. Not as a trick, because an interview is how you get the hidden workflow out. When Cowork asks you the right questions, what's the recurring frustration, what does "done" actually look like, you start articulating things you've never had to articulate before, and once it's out, it's usable.
If you're coming over from ChatGPT, this works there too. Run the interview in ChatGPT first, pull out what it already knows about you, and bring that into Claude.
Four areas matter. One, your operating rhythm: what does a real week look like, what takes longer than it should. Two, your recurring decisions: what judgment calls do you make over and over, what information do you need before you can decide. Three, your definition of done: what's the difference between done and done right. Four, your recurring friction: what are the things that are technically your job but feel like they eat time without adding much.
Rhythm, decisions, standards, friction. That's where useful automation starts. Not "help me with work," that's too vague to do anything with.
Instead of trying to write the perfect instruction, say this: "Claude, I want you to interview me. Ask me one question at a time. Help me uncover what I actually do, what slows me down, what decisions I make, what my standards are, and what parts could be handled better. Don't move on until you understand my answer."
From that conversation, you can build something that's actually shaped around how you work, a morning brief, a meeting prep system, a weekly planning routine, an email workflow. But you can't skip the conversation. The conversation is the setup.
Once you've had that conversation, first run the interview, then capture what comes out. The Portable AI Working Identity guide shows you how to turn that conversation into a document you own and can use anywhere, across any tool, any platform, any new chat. The interview gets the knowledge out. The document gives it somewhere to live.
This video points to The Professional AI Interview Kit, the practical next step after watching.