A 47-minute compilation that pulls together six previously separate Claude Cowork lessons, what an agent actually is, how to start delegating, the mental model for Cowork, the best setup, how to stop generic output, and how to make Claude truly know your work, restitched into the order they should have been watched in the first place.
Six months of separate lessons compressed into one deliberately sequenced path. The gap now isn't understanding Cowork, it's picking the first real thing to build with it.
If you've watched every Claude Cowork video you can find and still feel like there's one piece nobody's handed you, you're not alone. Five months ago I made a video called "I Wish Someone Had Explained Cowork Like This Earlier." Since then I've made five more, chasing the same feeling every time, because there was always one more thing I wished someone had handed me before I started.
In this video, I'm doing something different. This is a compilation, all six stitched together in the order I actually wish I'd watched them in.
It has taken me 20 years of teaching and hundreds of hours testing AI to realize that the important stuff isn't always explained in the right order.
So, by the end of this video, you'll have everything: what an agent actually is, how to stop prompting and start delegating, the mental model for what Cowork is for, exactly how to set it up, the three shifts that fix generic output, and the one file that makes Claude truly know you and your work.
Starting with the word everyone's been nodding along to without asking what it means.
You've heard the word "agent" so many times by now that you've stopped asking what it actually means, and somewhere along the way, asking started to feel like admitting you're behind.
You're not behind. In this video, you'll learn the one plain test that separates chatting with AI from running an agent, simplified. And by the end of this video, you'll understand exactly what an agent is.
Starting with the word itself.
But this isn't a beginner problem. It's a vocabulary problem the whole AI industry created by using one word to mean about five different things. If you've nodded along in a conversation about AI when someone started talking about agents, that's not a gap in you. Nobody ever drew the line for you between using AI and running one.
Almost every explanation of "agent" skips the one distinction that actually matters.
For a long time, I used AI the exact same way you probably do right now: a string of random conversations. I'd get a great result on a Tuesday, come back a few days later, ask for something similar, and get something different. Not wrong, just not the same. It wasn't how AI worked. It was how I was using it. And once I saw the difference, I stopped explaining myself from scratch every single time.
There's a three-second test that tells you whether something is actually an agent or just a longer chat. I'm going to call it the Three-Second Agent Test, and it's three questions.
Can it keep going without you? A chat waits for you to tell it what to do next. An agent can take a goal, figure out the next steps, use the tools it needs, and keep working without you prompting it through every step.
Does it remember the job, not just the last thing you said? A chat mostly works from the conversation in front of it. An agent has a goal, instructions, and context it can keep working from, so you don't have to explain the job again at every step.
Does it hand you a finished thing, not just a reply to react to? A chat gives you an answer you now have to do something with. An agent works toward a finished thing: a draft that's ready, a brief that's done, a file that's saved.
Think of it this way. A chat is you in the kitchen, looking up how to make each part of the meal yourself. An agent is calling the chef, telling them what you want, and sitting down when it's ready.
You might start it yourself. A schedule might start it. Another event might start it. Doesn't matter. What matters is what happens after it starts. If you're still doing the cooking, step by step, you're chatting. If someone else takes the order and doesn't come back until the plate hits the table, that's an agent.
I have 15 real agents I use that are all scheduled to run at specific times and intervals. Let me show you two of those, running in my own account right now.
This one runs itself, Tuesday and Friday, 3 a.m. while I sleep. It scans the news across four categories I care about, builds a five-minute brief, drops a draft email in my inbox, and files it away, clearing out anything older than a month. I read it on my phone when I wake up. That's the chef, not the recipe app.
This one runs every Thursday morning. It checks my sponsor tracker, picks up to three companies I haven't pitched yet, in the right order, flags anyone who's gone quiet for a week, and hands me a ready-to-paste email brief.
These agents carry their jobs forward. That's the difference, right there on screen.
Here's the mistake: once people learn what an agent is, they try to make everything one. But not everything should be. A one-off question is still just a one-off question.
The real test isn't "can I automate this," it's "can I hand this off with clear boundaries and let it run." That judgment call, knowing what's actually worth delegating, is what separates one dependable agent from five abandoned ones. The agent just does the doing; the judgment is still yours.
If it's one dish you enjoy making, you don't call a chef. Same with agents.
That was the newest one I made. This next one is the oldest, from back in April, and it's still the video that changed how I use Cowork more than any other. Because knowing what an agent is doesn't help much if you're still opening Cowork and typing a prompt like you're talking to a search bar.
Want to know the best way to start using Claude Cowork? Don't start with prompts. Start by telling it what is frustrating you.
In this video, I'll show you the first three things to do in Claude Cowork so it can start doing real work for you.
Before we go any further, let's make it clear what Claude Cowork is. It's not just normal Claude chat. Regular Claude chat is like regular ChatGPT and Gemini. But Claude Cowork is much more than that. It's a desktop-based version designed to help you carry out multi-step work. It's meant for situations where you want Claude to help organize, prepare, build, or carry something through instead of just giving you a text response.
That difference is huge.
If regular Claude feels like a smart assistant you can talk to, Claude Cowork is the 160 IQ teammate you can actually hand things to. That's why this matters so much for business professionals. We're moving away from AI as a tool you occasionally ask for help, and toward AI as something that can take real work off your plate.
Getting Claude Cowork working on your computer is pretty simple. Just take these simple steps.
First, install the Claude Desktop app on your Mac or Windows computer. Just go to claude.com/download.
Then sign in with a paid Claude plan. The Pro plan gets you both regular Claude chat and Claude Cowork for about $20/month.
After that, inside the desktop app, switch over to the Cowork tab. That's where Claude can begin helping with more active, multi-step tasks.
And one quick note on privacy: you stay in control of what Claude can access and do. Cowork asks for permission before accessing applications, and Anthropic says you control which local files Claude can access. My advice is simple: start small, be selective, and only grant access to the files or apps you actually want it to use.
Now here's where most people have been going wrong. They think the way to win with AI is to become some kind of prompt expert.
They think they need a perfect three-part instruction with the right persona, the right task explained, the right magic words. But we're moving past that quickly.
The big shift is this: stop trying to talk to AI like a programmer. Start talking to it like a manager. Stop giving it tiny isolated tasks and start giving it workflows that need to be regularly completed.
Interestingly, this is where things get surprisingly simple because it is so easy to use effectively. The best first move on your part is to start complaining.
Your daily frustrations are where the best automations begin.
Start complaining to Cowork about the problems you are having difficulty with. This should be easy because everybody already knows how to do that. Just type or speak into your computer's built-in microphone, and tell Claude what's annoying you. You might say, "I can't figure out how to get this thing done," "I feel stressed before meetings because I never feel fully prepared," or "Every morning I waste time trying to figure out what matters most first."
Claude will start to take over and will start putting together real solutions that it can take care of for you, and even perform regular work for you on a schedule while you are away or even sleeping.
That's the mindset shift I want you to take from this video. You don't need clever prompts here. Just talk to Cowork and tell it what's making your work harder than it needs to be.
The second thing I want you to do is this: ask Claude to ask you questions.
This is one of the smartest beginner moves because it removes the pressure of trying to explain everything perfectly on your own.
If you're not sure what to say or type, that's fine. Let Claude help you figure it out. Simply say, "Claude, ask me questions so you can help me build the right workflow for this problem."
That changes the experience completely. Now Claude is helping you discover the workflow instead of waiting for you to describe it perfectly from the start. You just need to know what is bothering you, and then let Claude interview you to uncover the right solution.
Now let's create something useful. For a lot of people, one of the smartest places to start is with what I'd call a high-agency daily intelligence brief. This is not just a simple morning summary. It is a focused, personalized briefing designed to keep you informed about what has changed in the last 48 hours, what trends are emerging, and why those developments matter specifically to you.
Instead of waking up and manually checking five different sources, you let Claude help gather the signals that actually matter, filter out the noise, and give you a sharper sense of what is happening in your world and where things may be heading.
Here is an example of what you could tell Claude Cowork to get started. You could ask it to track the most relevant developments in your industry and related fields, identify the biggest trends or shifts from the previous 48 hours, explain why those developments matter based on who you are and the work you do, and give you a short summary of the most important things you should be thinking about today. Give it a try.
Morning brief prompt: link in description.
You are moving from random prompting to delegated awareness. You are teaching Claude to monitor your world, spot what matters, and help you think ahead. Again, get this one working for you today.
A month after that one went up, I made a video that ended up being the most watched on this channel. Not because it showed a new trick. Because it finally gave people the mental model for what Cowork is actually for, and once that clicked, everything else got easier.
You've got Claude Cowork, you've watched the tutorials. And right around "MCP servers and sub-agents," you checked out, not because it's too hard, but because nobody's explaining it to you.
Most tutorials are built by tech people, for tech people. They teach the tool from the inside out. That's backwards.
Most tutorials teach tools, not systems, and those are completely different things.
So, in this video, I'm going to walk you through all four, with real examples you can try today. What Cowork is for, how to set it up, how to build workflows that work, and how to make those workflows run automatically. Everything I'm about to show you, you can do too.
Let's start with the thing nobody says up front.
Cowork is not ChatGPT with extra features. It's not "Claude but better." It's a completely different kind of tool. ChatGPT is built for conversation. You ask it something, it answers. Cowork is built for work.
You give it a task, and it goes and does it, sometimes across multiple tools, sometimes without you watching and even while you're sleeping.
That's the distinction most tutorials skip. And it's the reason people download Cowork, open it, and then just keep using regular chat mode because they don't know what to do with the rest of it.
And yes, you're going to hear words like MCP servers and sub-agents. Don't let those stop you. That's the plumbing. You don't need to understand the plumbing to turn on a faucet.
Here's what actually helped me understand it. When you use regular Claude chat, you ask a question, get an answer, and the conversation ends. And the next time, you start over again. Cowork flips that. With Cowork, you describe a job once, and it does that job on its own, on a schedule, without you asking again. That's a huge difference.
That shift is what makes Cowork so useful. Without it, Cowork is just Claude with a bunch of permissions you haven't used. With it, you've got something that can actually take recurring work off your desk.
After I finish a script for this YouTube channel, I usually spot two or three ideas worth posting on LinkedIn. But turning those ideas into actual posts, every single week, adds up. So I automated it.
I even had Cowork build me a simple web interface to run the whole thing. I open it in my browser, drop in the script. Cowork pulls out the key ideas and writes three posts: one practical, one personal, one proof-of-results. It knows what each post needs to do. I just have to decide which ones to publish.
That's what Cowork is for. Doing tasks that have multiple steps. Tasks you need done the same way every time.
The real question is this. What work are you doing every day or every week that's repetitive, structured, and annoying? That's what Cowork is for. Not the creative thinking. Not the judgment calls. The repetitive structured work that has to get done the same way every time and takes longer than it should.
One of the most useful things you can do with Cowork is also the simplest: just tell it what's bothering you. Describe what's repetitive and frustrating about your week. Then ask what it can do about it.
Before you build anything, there's one step you can't skip, but most people do.
When you open Cowork for the first time, it asks you to choose or create a workspace folder. Everything Cowork does starts with that folder.
It's where Cowork reads files from and writes files to. If you skip that step or just pick a random folder, nothing else will make sense.
So do this first: create a dedicated folder for Cowork. Not buried in your documents somewhere. On your desktop or somewhere you will always know where to find it easily. Name it something obvious. Claude Cowork Folder. That's it. Now everything Cowork creates will live there. Two minutes of setup at the beginning saves you a lot of frustration later.
The second part of setup is connectors. Connectors let Cowork talk to your other applications. If you want Cowork to read your email, you connect Gmail. If you want it to check your calendar, you connect Google Calendar. If you want it to post to Slack, you connect Slack. Most of the apps you already use are in there.
Don't connect things just because you can. Connect the things your workflow requires. For example, my morning brief only works because Gmail and Google Calendar are connected. Without those two, Cowork has nothing to read. Once your folder is set and your connectors are in place, you're ready to build.
This is where most people stop paying attention to the tutorials, but it's also where Cowork actually becomes useful.
Cowork comes with some built-in capabilities, things like web search, file editing, running code. You don't have to configure those. But the powerful stuff, the workflows that actually save you time, you have to build.
That's where skills come in. A skill is just a set of instructions you get Cowork to write once and then trigger with a single command. Think of it like teaching someone a procedure. You walk them through it step by step the first time. After that, you just say "run the morning brief" and they know what to do.
I have a text file in my Cowork folder called goals.md. It has my three priorities for the month. When I trigger the skill, Cowork reads that file, looks at what I've actually worked on that week, and gives me a one-page check-in, what moved, what's stalling, and one thing to do tomorrow. Takes about thirty seconds to read.
You probably have goals you haven't looked at since you wrote them down. A skill like this fixes that. You tell it what you're working toward once. After that, every time you ask, it already knows the context.
I didn't write any code. I just told Cowork what I wanted, answered a few questions, and it built it. That's a skill. And once it's built, you never have to explain it again.
Think about the thing you explain to people most often, a process, a format, a standard you hold yourself to. That's a skill waiting to be built.
Scheduled tasks don't take you out of the loop. You still built the workflow. You're still checking the output. The only thing you're no longer doing is remembering to run it. And that's actually where the time comes back.
Once you've built a workflow that works, you can schedule it. You no longer have to be at your computer. Cowork will run it at 6 a.m. every Tuesday, or every time a new file appears in a folder, or whenever you want it to.
That's what scheduled tasks do. They take a skill you've already built and make it automatic.
I have one that runs every weekday morning before I wake up. It's called the morning brief. Cowork reads my inbox from the last eighteen hours, checks today's and tomorrow's calendar, and sorts everything by urgency, what needs a reply, what needs a decision, what just needs my awareness. It adds one line at the bottom: the single most important thing to do in the first thirty minutes.
I wake up, make coffee, and read it on my phone. The day is already mapped before I've opened my laptop.
And I didn't have to build it from scratch. I just described what I wanted once, and now it runs on its own.
That's the mental model. Purpose, Setup, Skills, Scheduling.
The mental model is only useful once it's actually wired into your real tools. So this next one, from June, is the one where I stop explaining Cowork and walk you through my own setup, connector by connector.
First you want to connect it using something called connectors, and once you understand what they do, the rest of this makes sense.
Claude can connect directly to your email, your calendar, your files, but only if you tell it to. Think of it like plugging a printer into your computer. Connectors do the same thing for Claude.
At the top, just click "customize" then "Connectors."
Here's the one that saves me the most time every single week: the Gmail connector.
Most of us spend at least an hour a day in email. I used to do all of that manually. Now Claude does most of it.
Here's how it works. I go into Claude and click customize at the top left, then I click on Connectors. I click the plus button and search for Gmail. I click it and log in to Gmail. Now Claude has access to my inbox.
That means I can ask things like, "Find the last email about the contract renewal and tell me what's still unresolved." Claude searches my inbox, finds it, and gives me a summary. Or I can say, "Draft replies for anything that needs a human response." Claude writes the reply. I read it, make a small adjustment if I need to, and send it.
The reason this works is because Claude isn't just guessing what my emails say, it's actually reading them.
Now, I know some people hear that and immediately think, "Wait, is that safe? Is Claude reading all my emails?" Fair question. Here's the answer. Claude's security is handled by Anthropic, the company that built it.
They don't store your emails on their servers. Claude reads what you ask it to read, and that's it. If you're worried about it, don't connect it. I've had this connected for months and it's never given me a reason to second-guess it. You stay in control of what it reads and what it doesn't.
The second connector I use is Google Calendar. Same setup process: go to Connectors, choose Google Calendar, log in. Now Claude can see my schedule.
So instead of opening my calendar, scrolling through the week, and trying to figure out where I have breathing room, I just ask, "Find me an hour this week to work on my project but not first thing in the morning." Claude looks at my schedule and finds it. That's it.
But here's the part most people don't think about. Claude can also analyze how you're spending your time.
I told Claude what actually matters to me right now. Then I asked it to look at my calendar and tell me if I was spending time on those things. The honest answer was no. I was filling my week with stuff that felt productive but wasn't connected to anything I'd said was a priority. Claude pointed it out.
Now, those two connectors, Gmail and Google Calendar, are the ones I use most often. But there are others. Claude can connect to Slack if your team uses that for communication. It can connect to Google Drive so it can read and reference documents you've stored there. It can connect to Notion if that's where you keep your notes and project plans. The process is the same every time. Go to Connectors, choose the service, log in.
The second thing is to load it, and that's where Claude Projects come in. And I'll be straight with you, this isn't something I use day to day anymore, because Claude Cowork handles most of what Projects is designed for. But if you're spending most of your time in Claude Chat, this matters.
A Project is basically a folder with a memory. You create one, give it a name, and load it with everything Claude would need to help you with that area of your work. Your documents, your standards, how you want Claude to respond. All of it lives in one place.
Setting it up takes about two minutes. In the left sidebar, click New Project. Give it a name, something specific like "Client Reports" or "Weekly Newsletter." Then click Project Knowledge and upload whatever Claude needs to do that job. A style guide, a template, a document with your preferences. Finally, add a short instruction at the top telling Claude how you want it to behave inside this project. That's it. Next time you open it, all of that is already loaded.
The problem it solves is this. Every time you open a regular Claude chat, you're starting from zero. Re-explaining who you are, re-pasting the background, re-describing what you need. A Project means you do that once. After that, Claude knows and remembers that area of your work. This is perfect if you only are on the free plan.
But I moved past this stage when I saw what Cowork can do. If you're not there yet, Projects is the bridge. It's the first time Claude starts feeling like it actually knows you.
Third is the hand off. That's Claude Cowork. And honestly, this is the one that made the biggest difference to how much I can get done and how much more is possible.
Before I get into it, I want to give you a quick picture of how Claude works across its three different modes, because most people don't know all three exist, and it changes how you think about what's possible.
Claude Chat is what you've been using, the web interface or the mobile app where you type something and Claude responds. Everything I've shown you so far, connectors, Projects, that all lives here. Most people spend all their time in this mode, and there's a lot of leverage still untapped in it.
Claude Code is the most advanced. This is more for developers, people who write software and want AI working directly inside their codebase. If that's not you, you don't need it. I'm not a developer and I don't use it. But it exists, and it's worth knowing you have it.
Claude Cowork is the one I want to spend time on, because it's the most relevant for the professionals watching this. Let me make sure that landed. Chat is for questions. Cowork is for jobs. That distinction is everything.
Cowork lives inside the Claude desktop app, that's a separate program you download and install on your computer, but they're all connected. Once you have it, you'll see Cowork in the top left corner. When you open it, it gives Claude the ability to control your computer and do tasks for you while you step away. It can even do work for you while you watch it in real time.
Let me be clear about what that means, because it sounds more complicated than it is. You give Claude a task that involves clicking through different programs, maybe pulling a report from one app, putting the numbers into a spreadsheet, creating a chart, and sending it to your team. Normally, you'd do all of that yourself. With Cowork, you describe the task and Claude does the clicking for you.
Here's a real example. Every week, I want to know what's happening in AI, but there's so much going on. I used to try to piece it all together but it took at least 30 minutes and I still felt I was missing something.
Now I have Cowork scheduled to do it for me. I wrote out the task once, search for AI developments across the topics that matter to my work, filter out the noise, and compile everything into a short brief. Now that task runs every Tuesday and Friday at 3 a.m. When I wake up, the brief is already waiting in my inbox.
That's the shift. I'm not asking Claude a question and waiting for an answer. I've given it a job, and it's doing that job on a schedule without me touching it.
Cowork isn't perfect. Sometimes it gets stuck. When it does, I just tell it what went wrong the same way I'd talk to someone sitting next to me. It adjusts and keeps going.
None of this is hidden. Connectors, Projects, Cowork, it's all right there. But rarely do people explain what it's for. Now you have the setup. And that changes what's possible.
Don't try to set up everything at once. Pick the one that connects most obviously to your biggest daily frustration and start there. One thing. See if it saves you time. If it does, add another.
Here's the thing nobody told me until I'd already done all of that. You can have the perfect setup and still get generic answers, because setup and communication are two different problems. This one's about the second half.
You set Claude up. You connected your tools. You followed every tutorial. And you're still typing questions into it like it's Google. That's not a setup problem. That's a communication problem. And nobody talks about it.
In this video, three shifts. Each one with a demo.
I've lived in South Korea for a while. I can follow a conversation in Korean, pick up the mood, and read the room. But sometimes my wife's Korean friends are all talking, they all laugh at something, and I almost get the joke. Almost.
I know what it feels like to be smart enough to follow but still miss the one thing that makes it click. That's exactly where most people are with Claude right now. And here's the thing I've noticed after watching a lot of people use these tools: the ones getting real results aren't necessarily smarter or more technical. They just approach it differently. Most people open Claude and start with the tool. What can it do? What should I ask it? But the people getting the best results start with the problem. What do I actually need done? Then you can use Claude to do it.
That one shift changes everything about how you communicate with it. And what's missing isn't a better prompt. It's those three things I mentioned earlier: context, outcome, and role.
Confusion almost always comes from the same place. People come looking for an answer rather than looking to understand. I see the same pattern in how most people talk to Claude. They ask it questions when what it actually needs is a proper handoff.
Let me show you where I learned that. But it wasn't with Claude.
I teach a university writing course. To prevent AI cheating, students handwrite major assignments. But I built an AI system to grade the handwriting. The irony was not lost on me.
I hit several walls building that system. It was a struggle to get right because I was correcting the AI more than it was helping me.
But things changed when I stopped asking questions and started briefing it. Context, constraints, what I'd already tried. Real students were being run through this system. And I'd promised results soon. It ran at 85 to 90 percent accuracy on the first run.
That experience taught me the three shifts I want to show you now. Each one changes what Claude does with your request. And together they change how the whole process feels.
When you open a new Claude session and type a request straight away, Claude is guessing. It's like calling a taxi and just saying "drive." It doesn't know your role, history, or what success looks like. It's working partially blind.
Before you write what you want, ask yourself this: if I handed this task to a capable new colleague who has never worked with me before, what would they need to know first? That's your context. It goes in before the request.
The before: "Write a follow-up email to a client." The after: "I'm a business consultant. I had a call yesterday with a long-term client about renewing their contract. The call went well but they asked for two weeks to confirm the budget. I need a follow-up that's warm, gives them space, but keeps momentum going." Same task. Completely different foundation.
Most people describe what they want Claude to produce. Write me a paragraph. Summarize this document. That's describing the format, not the goal.
The better version describes what the output needs to do. "Write me a paragraph that a non-technical senior manager can read in 30 seconds and understand what action they need to take." Now Claude knows who reads it, how long they'll spend, and what success looks like. So, before any request, ask: what does a good version of this actually do? Answer that, and Claude stops guessing.
If you ask Claude to "look at this draft" then it reads it as a general reader. But if you ask it to "read this as a skeptical manager who has seen a hundred proposals and is deciding whether this one deserves ten more minutes" then it reads it completely differently. Same draft. Different lens.
The role is not a character you invent. It's the perspective you actually need. Choose it deliberately.
Applying all three together didn't just improve individual outputs. The entire dynamic changed. Sessions stopped feeling like a back-and-forth and started feeling like a handoff. That's a meaningful difference when you're doing this every day.
Let me show you what this looks like when it actually matters. Not a writing task. A real decision.
Imagine you're a senior manager. You have two final candidates for a project lead position. One has more experience but came across as guarded in the interview. The other has less experience but impressed you. You need to make a recommendation to leadership by Friday. And you can't shake the feeling you're missing something.
Here's what most people type into Claude Cowork: "I have two candidates for a project lead position and I need to make a decision by Friday. One has more experience but I'm not sure about the culture fit. The other is less experienced but really impressed me in the interview. What should I consider?"
And Claude gives you a reasonable answer. It talks about culture fit, interview performance, the role requirements. It's not wrong. But it's generic. It could apply to anyone hiring anyone anywhere. It ends by asking you a question back, because it doesn't have enough to work with.
Now watch what happens when we add context: "I'm a senior manager at a mid-sized marketing agency. I'm hiring a project lead who will manage a team of five and report directly to me. We're in a growth phase and the team is under pressure. I have two final candidates. Sarah has eight years of experience managing large accounts but came across as guarded in the interview and asked very few questions. Marcus has four years of experience but was engaged, asked sharp questions about the team dynamic, and has managed smaller projects successfully. I need to make a decision by Friday. What should I consider?"
The output is immediately more specific. Claude understands the pressure the team is under. It takes the culture fit concern seriously rather than treating it as a gut feeling. It starts giving you things to actually do before Friday. Better. But we're not done.
Now we add the outcome. Same context, but now Claude knows you have to walk into a room and defend this recommendation to leadership if challenged. It knows what success looks like: not just a decision, but a defensible one.
Now Claude tells you exactly what argument to make for each candidate. It anticipates the pushback you'll get from leadership. It tells you what information you still need before Friday and how to get it fast. This is something you could actually walk into a room with.
Then we add the role. Same prompt as earlier but we start it with this: "You are a senior talent advisor who has helped executives make hundreds of high-stakes hiring decisions. You know that the most costly mistakes usually come from what people overlook, not what they consider."
Same problem. But a much different perspective in the room. Claude stops analyzing the candidates and starts analyzing your blind spots. It tells you that the most common mistake in a decision like this is letting interview performance stand in for job performance. It tells you what "guarded" actually might mean and why you need to find out before Friday. It tells you what to ask on the reference calls, not just whether to make them.
And it ends with this: "What you're looking for isn't who impressed you more in a room. It's who is most likely to still be functioning well six months from now when the growth phase is still on and the team has had two hard quarters."
That's not a search engine answer. That's what a trusted advisor tells you.
Here's the template. "You are a [specific perspective]. I need this to [what it needs to do]. Here's the situation: [the specific context that matters]."
But imagine how time consuming and frustrating it can be to rebuild that context every time. I have a way around it.
It's a single file Claude reads automatically at the start of every session. Before you type a word, Claude already knows your role, your priorities, and how you think. You don't explain yourself. You just start.
There's a kit that walks you through the whole thing, and the next one is exactly that.
Now, before the last piece, if any of this has been useful so far, a like or a comment genuinely helps this video get in front of more people who are stuck exactly where you were five months ago. And if you want the rest of what I've learned since, subscribe.
Here's the last piece. Context, outcome, and role fix a lot. But you're starting each session without Claude intimately understanding you and your work. Because if Claude doesn't know you, it can't tell the difference between the right answer and the right-sounding answer. In professional work, that gap has consequences.
Have you ever asked Claude for advice on a real work decision? It probably sounded solid. But there's a good chance it was generic, because most people never give Claude the right stuff to work with. I know because I used to be one of them.
If Claude does not know you, it can't tell the difference between the right answer and the right-sounding answer. And in professional work, that gap matters.
Anthropic actually studied this. They looked at 400,000 Claude sessions to find out what separates people getting strong, useful output from people getting generic output they have to fix themselves.
The answer was not prompting skill. It was domain expertise. That means professionals who knew their field deeply got better results than skilled prompters working outside their area of expertise.
That should be good news. The years you have spent in your field are your advantage. But here is the part that does not add up: if the tool works and your expertise is real, why does Claude still fall short of understanding the work you do and producing the output you actually expect?
By the end of this video you will know exactly why Claude gives you generic advice, and you'll see the exact fix that turns Claude into the AI you've been hoping to finally have.
I ran into that same gap repeatedly before I found the answer. It was not better prompting. It was the questions Claude never thinks to ask: about your work, your decisions, the things you would never think to mention.
So I built what I call the Professional AI Interview Kit. I run my own business through it. I stay on track of my goals with it. I build agents and systems with it that work the way they're supposed to. And now, I'm about to show you what it is and what it fixes.
The entire thing is a single file: a set of instructions that turns Claude into a professional interviewer. If you use Cowork, setup is simple: take the file, rename it CLAUDE.md, and drag it into whichever folder you use for Cowork.
Cowork reads this file automatically every time you open a session in that folder. No project to create. Nothing to paste. Nothing else to configure.
If you use claude.ai on the web instead, you paste the same file into a Claude Project's instructions. Same result.
Open a new Cowork session pointed at that folder and type "I'm ready."
Claude says it needs a moment to check your files, then asks once if you would like to add anything else, a resume, a work sample, before it starts. Say "let's go."
Then it introduces itself, explains what is about to happen, and asks its first question. Notice what it does not ask. It does not say "describe your job in one sentence." Instead, it asks how you got to this point in your career, what path actually got you here. That is a question Claude never thinks to ask on its own: about your work, your decisions, the things you would never think to mention.
It's not "what is your job title." It is asking about the decisions you make repeatedly, the things a new colleague would take months to learn.
If your answer is a summary, it follows up. It is not satisfied with the surface. That is what makes the profile it produces usable instead of generic.
At the end, 25 to 35 minutes in, Claude produces a file called the Personal Operating Document. This is what you keep: a structured profile of you as a professional, written in a format Claude can work from directly.
That is not a summary of the conversation. That is working context: how you communicate, what a good answer looks like in your work, the reality Claude needs to operate inside instead of a generic version of it.
Claude then asks if it should save the document to your workspace automatically. Say yes. One step, and it is permanent. From this point on, every session you open in this folder starts with Claude already knowing your context intimately.
This is where it proves itself. You have felt this before: an answer from Claude that read fine, that you used, and that only later turned out to be built for someone who was not you. Here is exactly why that happens, and exactly what changes when it does not.
Here is Claude when it does not know you, asked the same question with no background loaded: "What are the three most important things I should focus on in my business over the next six months?" The output is reasonable. It is also built on defaults, the most generic useful thing Claude can construct for no one in particular.
Now ask that same question with the complete Personal Operating Document loaded. Same question. Different output. Not because the question changed, but because what Claude was working from changed.
This is the cold start problem. Every new Claude session starts the same way: it does not know your role, your clients, your communication style, or what actually clears the bar in your work, so it fills the gap with defaults, reasonable sounding, built for no one in particular. The AI Interview Kit closes that gap, once, permanently. That is what the research was pointing at. The expertise was always there. Claude just did not have access to it until now.
The research did not find that Claude works better for some professionals. It found that some professionals give Claude more to work from. This interview is how you give it that, and it sticks.
That is the Professional AI Interview Kit, and that is exactly how I use it. If you want to try it yourself, it is in the description.
So those were the major things I wish I knew before starting Claude. They all matter. But the file at the end, the one that turns Claude into something that already knows you, that's the single thing I'd protect if I could only keep one piece of all of this.
If you want the exact prompt that builds that file, the link's right below this video.
Mine continues to evolve with me, and so will yours.