So What? Setting up Claude Cowork

Katie Robbert – 00:35

Happy Thursday. Welcome to So What, the Marketing Analytics and Insights live show. I am Katie, joined by Chris and John. Hello.

Christopher Penn – 00:46

It’s a mess.

Katie Robbert – 00:47

It’s us. This week we are talking about setting up Claude Cowork. Chris and I have been talking a lot about how we get ourselves out of the weeds of running the business day-to-day. Because as we’ve mentioned many times over, we’re a very small company, and that means we wear a lot of hats—all of us do. We’re all just trying to keep moving forward, sometimes just not as quickly as we would like to.

The solution is that we could either hire more people and add more headcount—which, quite frankly, budgets are tight all around—or we could find ways to automate more and figure out how do we use what we have and make it more efficient. One of the things Chris has really been talking about a lot is Claude.

As a large language model, Claude has a lot of different things it can do, and Chris is going to get into it. One of the things he’s talked about a lot is Claude code. We’re not talking about Claude code today; we’re talking about Claude Cowork. I think it’s worth acknowledging there’s Claude the LLM chat, there’s Claude code—which Chris will talk about—and then Claude Cowork, which is what we’re going to focus on today. Do I have all of that straight so far?

Christopher Penn – 02:07

So far, so good.

Katie Robbert – 02:09

All right, that’s right. I’m going to tap out now because I don’t want to get it more confused. So Chris, over to you.

Christopher Penn – 02:15

All right, so a couple of things upfront. Claude Cowork is a brand new product. It is about two weeks old at this point. It is iterating rapidly and it is in very limited testing. Only people on Pro or Max plans can use it; the free plans cannot use it. It is inside the Mac desktop version of the Claude app, so it’s not available on the web or on Windows. According to the Claude team, they are working very quickly on a Windows-based version and they expect to have that out soon—with a capital S.

Today we’re going to go over how it is right now. It is going to change. For example, in Cowork, skills and agents are not formally supported yet. The team—Boris Cherney over at Anthropic—has been saying they’re working on it as fast as they can. It’s very funny; they actually used Claude code to write Cowork, and they’re trying to add as many things as quickly as they can.

Every other vendor is looking at the Claude ecosystem now and getting the photocopiers ready because they’ve got all the buzz. Valley investors are all paying very close attention to what Anthropic is doing. Claude Cowork actually came out of Claude code because Anthropic was looking at how people were using Claude code and noticed people weren’t just coding—they were planning vacations. They wondered why people were planning a vacation in Claude code, and it was because it has so many tools to do automated work, like finding flights, where you could just come back later and it’s done. They decided to make a version that didn’t require you to be a gearhead just to get the thing installed.

Katie Robbert – 04:18

Yes, please.

Christopher Penn – 04:20

And that is Cowork. Let’s start digging in. The first thing you’re going to need to do is download it from Anthropic’s website. If you go there, you are looking for Claude on the desktop. When you install that app, there are three tabs in the upper left: Claude chat, which is the regular plain vanilla version; Cowork, if you’re on the Pro or Max plans; and then Claude code, which is their cloud-based version of Claude code.

Katie Robbert – 05:01

If I’m looking off-screen, it’s because I am following along with Chris. I just installed Claude for the desktop for myself to learn how to do this. I’m literally using this live stream to learn how to do this from Chris live.

Christopher Penn – 05:20

It’s a company training. To orient you, there are a few different things. Obviously, there is the existing chat list. There are some starter prompts. On the right-hand side is the status panel that tells you what Claude is working on. Then there’s the ability to connect to a local folder on your desktop and use connectors.

Some of the connectors available are things like Claude in Chrome, which I’ve installed on my computer with the big asterisk that it can control Chrome. That means if you are not thoughtful in how you prompt it, you might be telling it to go post some deepfakes on LinkedIn for you. If you don’t think that through, it will do those things.

The reason why we want to use something like Cowork is twofold. One, it has access to our local files, which is such an important thing. I’ll show you in a minute how to connect it to something because we want to put guardrails around it and say this is the only little sandbox you’re allowed in—you’re not allowed to go other places. Second, it can control a browser. Those two things give it the ability to do autonomously a lot more stuff than the regular Claude chat can do.

Katie Robbert – 06:46

We talked about the guardrails in more depth on the podcast this week, which you can get at TrustInsights.ai/tipodcast. That was a big question I had for you: if you’re installing this stuff on your local machine and you likely keep sensitive data there to keep it off the web, how do you make sure it’s protected? If you want to get more of a deep dive, check out the podcast or join our free Slack community, Analytics for Marketers, where we can go into more depth and learn what everyone else is doing to protect their data.

Christopher Penn – 07:30

Exactly. So we’re now ready to start doing some of the configuration. First things first, it’s a good idea to give Claude its own folder per project or per task. Don’t just give it access to your desktop; that’s a terrible idea. Instead, give it a place on your desktop to work and maybe even give it a folder inside that place.

On my desktop today, we wanted to talk about the Trust Insights newsletter, Inbox Insights. I’ve created a folder called “Inbox Insights Strategy.” I’m going to tap “Work in a folder,” find my way to that folder, and it will ask if I’m sure I want it to be able to read and write files in this folder. I’m going to say yes because I’ve given it this specific folder to work in. Please do not just give it your entire hard drive.

I have also enabled the Claude in Chrome connector to allow it to control my browser because I’m going to show an example of how that might be useful. Again, the caveat is you want to be careful how you use it so it’s doing things intelligently and not just running amok.

Katie Robbert – 08:57

It sounds like for those just starting out, maybe wait for Phase Two before giving Chrome permission. Just get used to this first before you let it take over your Chrome.

John Wall – 09:18

Just let it take your browser and go everywhere it can and start buying random stuff. That sounds like a complete recipe for disaster.

Katie Robbert – 09:26

“Where did all this stuff come from? I was just researching million-dollar homes and now I’m suddenly the proud owner of five different estates.” Whoopsies.

Christopher Penn – 09:37

Exactly. And a moat.

John Wall – 09:40

That log cabin I’ve been looking for.

Christopher Penn – 09:44

The next thing we want to do has nothing to do with Claude, but this is the part that’s going to make Katie’s soul happy. We want to think about the governance inside our Claude folder and how to organize it. You don’t want to just have empty folders or pile stuff in there; that is like how my kids clean their room by shoveling everything into the closet.

I already threw some stuff in here and it’s kind of a mess. I would suggest having at least three or four folders for every Claude Cowork project. We get this from Claude code. I’m going to create a folder called “Data” for data sets, a folder called “Input” for things like transcripts, a folder called “Output” where I want Claude to put things as it’s working, and a folder called “Docs” for rules-based documents—like where my code standards live.

Because Cowork does not formally support agents and skills yet, we’re going to create two folders for those as well. An agent and a skill are nothing more than prompts, so we’re just going to borrow existing prompts for those things. Now I’ve got six folders. Katie, why don’t you describe what you noticed in our subscriber counts for Inbox Insights this week and what our goal is?

Katie Robbert – 11:34

If you’re brand new, we have a weekly newsletter called Inbox Insights which you can find at TrustInsights.ai/newsletter. We have a brand new landing page as of last Friday. What we’ve noticed is that in 2025, our subscriber numbers stalled out and plateaued. Then we did a deep clean and data hygiene of the list itself, and unsurprisingly, the number went down.

Moving into 2026, the total number of subscribers is lower than where we were at the beginning of 2025. We know from our own data that our email newsletter is one of the largest drivers of engagement and eventually conversion for our services. It’s in our best interest to increase that subscribership, which means promoting it more, writing about topics people care about, and getting more awareness for it.

There are a lot of different things we need to do to hit a target number of subscribers in 2026. If we were to follow the 5P Framework, the purpose is to increase the total number of subscribers to the Inbox Insights newsletter so it continues as a strong channel for engagement and conversion.

Christopher Penn – 13:13

Right, which is a fantastic use of Claude Cowork. One of the things that makes it different than a regular chat is that it will build its own to-do lists and try to do tasks if you write them out well. First, we need to figure out how to put this together. We’ll start with some basics, like our Ideal Customer Profile (ICP). I’ve chosen our Mid-Market Growth ICP from our sales playbook. It is not the entire 135-page document—just this one particular segment.

One thing we’ve started doing is removing anything involving age because it’s discriminatory and because what we really care about is someone’s ability to buy from us. I don’t care how old you are or where you live; all money is green.

Katie Robbert – 14:15

Well, also not a true statement, but that’s for a different episode.

Christopher Penn – 14:19

That is for a different episode. So we have our ICP. Second, we need to know the landscape. I’ve been in email marketing now for 30 years. What’s changed? What’s new? The easiest way to get that information is to commission a deep research report.

I kicked off a deep research report following the Trust Insights Casino Framework. I wanted to know who the audience is and what’s happening in the email marketing world. I focused on 2023, 2024, and 2025. Specifically, I want to know the best practices for retaining your audience and acquiring new subscribers, as well as the mistakes people make that cause them to lose subscribers.

I went ahead and ran this. The nice thing about Claude Cowork is that it can read Google Docs directly. In my case, I exported it to a Google Doc and then dumped it as a markdown file, which is the lingua franca of AI. That deep research report is now in plain text so the AI will have an easier time reading it.

Katie Robbert – 16:11

When you said that you were old and liked to have things, I was fully expecting you to say you printed it out and used a binder clip and a highlighter. Making things markdown so they are easier for the AI to read doesn’t count as being “old.”

Christopher Penn – 16:32

That’s fair. Now, another question we would want some research on is how we do our newsletter. Before the show, I spent 15 minutes just talking about how we do our newsletter—the process, who writes what, and the major sections. I recorded it and transcribed it into text. We also need to know the existing metrics.

Katie Robbert – 17:22

Probably the baseline metrics. Here’s our current number of subscribers and here’s the goal we want to hit. We need the engagement rate, the open rate, and the rough makeup of the current subscriber list. We need the baseline because Claude might look at our ICP and our current list and say, “Those are two different audiences; good luck with that pivot.”

Christopher Penn – 18:01

Exactly. The easiest way for me to do that was to take screenshots of our marketing automation system for the last full year of issues. Screenshots are just fine because Claude can read images. The other thing that would be helpful is to see how much of our website traffic comes from our newsletter. The place to get this would be Google Analytics.

Because I have the Chrome connector turned on, I can say, “I need some help building an exploration in Google Analytics to find out how much website traffic I’m getting from my Inbox Insights newsletter.” I can give it the session source and the host name and ask it to help me build this exploration for every day from January 1, 2025, onwards.

As an instruction, Claude will fire up Chrome and try to figure out where our Google Analytics is.

Katie Robbert – 19:54

Claude is doing this? You’re not?

Christopher Penn – 19:56

Here are my hands; I’m not doing anything at all. It’s now going to the explorations. This takes it about 20 minutes to go through the whole thing, so I’m going to pause it. But you can see the window says “Claude MCP,” which tells you this window is being controlled by Claude.

This is what it came up with, and it is correct. This is the correct way to get this data. It correctly set the host name, the session source, the values, and the sessions by date.

If you have a piece of software and you don’t know how to use it, but you know what to ask for, you can tell Claude Cowork to take control of your browser and start doing the thing. You can hit the emergency stop at any point, like I just did. But if you have Salesforce or HubSpot or whatever and you’re not sure how to do something, Claude Cowork can walk you through it.

Katie Robbert – 22:06

I think that’s the key. Kelsey has a question: what about In-Ear Insights, which is our podcast? Would you also want the podcast involved for this particular task?

Christopher Penn – 22:44

No, because we’re focusing specifically on growing the newsletter subscriber list. We will repeat this process offline to figure out how to grow the podcast downloads. I did mention in my prep recording that we mention the newsletter on the podcast so Claude knows what all the pieces are.

I have my Google Analytics data. I’m going to rename it so Claude can understand the context. Another thing Kelsey asked about was XCO—AI search optimization. Using our own homegrown XCO tool, I extracted search terms and queries from Gemini. That’s going to go in as an ingredient because if we’re talking about growing our subscriber base, we need to know what topics people are looking for and what Google is recommending.

The final ingredient is Kelsey’s 5P analysis on how to make our marketing more fresh and relevant to a broader audience. We’re going to put all these things together into Claude Cowork.

My subscriber data and Google Analytics data will go in the “Data” folder. The research data goes into the “Input” folder. Our current processes go in the “Docs” folder. I’m also going to put my screenshots in a folder called “Mautic” so Claude understands they are from the marketing automation system.

Katie Robbert – 26:06

We are more than halfway into the show and we haven’t touched a large language model at all yet, other than the Google Analytics pull. We’re still just getting set up and doing requirements gathering. In order to have any system work, you have to do this work. This is the stuff you can’t skip over; this is what sets you apart from everyone else generating mediocre things.

Christopher Penn – 26:57

Our recipe now is what we put in as the master prompt in Claude Cowork. We’re going to start by saying: “You are an email marketing expert. Your task today is to help us build and refine a marketing strategy to grow the Inbox Insights newsletter from its current position of 33,000 subscribers to 40,000 subscribers by the end of calendar year 2026.”

That is our Purpose. For People, we have an ICP already provided in the input folder. For Process, it’s going to review the “Docs” folder to understand how we currently do our newsletter and then review the remainder of the files in the “Input” folder.

In Platform, we’ll tell it that it has two different agents it can use to review its work. One is called “Co-CEO Katie,” which is an avatar of our CEO and can provide business sense. The other is called “VOC,” which stands for Voice of the Customer. We also have a skill called “Fact Check.”

Finally, for Performance, we define the output. Once it has gone through the preparatory process, it will do a full QA assessment of our strategy—what’s good, what’s bad, what’s missing, and what’s unnecessary. Then it will build a refined strategy, tactics, execution, and measurement plan.

Katie Robbert – 32:58

One of the things we do is offer the newsletter in a few different formats—the email, YouTube, and the podcast. We also put it up on LinkedIn. There is a dissemination strategy once the newsletter is completed to expand our social footprint. I’m assuming all of that is included.

Christopher Penn – 33:57

You are correct; that is in the 15-minute recording where I talked about UTM tracking codes and promotional tools. We want to give the LLM as much information as possible. We could just copy and paste this and hit go, or we can create this as a recipe file that stays in the folder. I prefer the recipe file because I forget things. Next quarter, when we want to re-evaluate, I have the exact prompt ready.

Katie Robbert – 34:58

It’s a best practice for scalability. LLMs are really good at consistent, repeated tasks. You could eventually hook this up to a tool like n8n and have it kick off once a month to evaluate how we’re doing. You’d suddenly get a file delivered with new ideas and an assessment of what’s not working.

Christopher Penn – 35:39

Exactly. I’ve dragged in the recipe file and Claude says, “I noticed you used the 5P Framework. This is a substantial project; how would you like me to approach it?” I’ll tell it to do a full deep dive.

You can see it’s already putting together its to-do list of 12 different items. It has a scratchpad where it takes notes as it goes through the process. It’s launching multiple agents to start reading and assembling notes on each of the files.

Katie Robbert – 36:37

Is that all stuff that we as humans can then access to see the notes it’s taking to learn how it’s thinking about it?

Christopher Penn – 36:46

Yes, you’ll be able to get that once it’s done. It’s now looking at the Mautic screenshots and checking the fact-check skill. It got to a point where it was like, “I don’t have high confidence in the information provided; I need to go and do some fact-checking.” It’s going to fact-check the research, which is pretty cool.

Katie Robbert – 37:39

It’s helpful because we’re humans and we make mistakes and have biases. Using tools like this is a good way to pull yourself out of it so you aren’t just following your own agenda. I let the data drive a lot of those decisions.

Christopher Penn – 38:20

As each of these agents run, they have their own task list. It kicked off six different web searches to figure out if what was in that deep research document is actually true. I had Claude use the fact-check skill on a different project once, and it came back and said 27% of the document was fabricated. It checked the sources and found the errors.

Katie Robbert – 39:06

I like how Claude is throwing shade at Gemini. When the LLMs start fighting, you know you’ve lost control.

John Wall – 39:18

This is virtual war. They’re actually punching each other behind the screen.

Christopher Penn – 39:24

So Claude has finished. It highlighted some fact-checking results and spit out a strategy document. Strategic thesis: “Retention is the new acquisition in a zero-budget environment.” It says the most efficient path to 7,000 new subscribers is activating the existing 33,000 as a distribution channel while fixing discovery gaps.

It recommended we create more content about agentic AI because that’s what people are asking about. It suggested we have more giveaway content—like prompts people can take action on immediately. It also recommended a 48-hour delay for LinkedIn so that email subscribers have an incentive to stay on the list.

Katie Robbert – 49:09

I’m a little underwhelmed because I’m looking for where we get those new subscribers from. It’s focused on keeping the existing base, but I want to know if we should be on other social channels letting new audiences know this thing exists.

Christopher Penn – 50:16

You would absolutely include that as guidance and revise the recipe. That is the beauty of Cowork—I can change the recipe, rerun it, and come back when it’s done. At no point did I have to babysit this thing while it was running.

Imagine taking this and saying, “I agree with the evergreen scheduling; go ahead and log into my Agora Pulse account and schedule these posts for me.” Then you walk away and it’s done. You should check its work the first few times, but it takes away execution tasks that don’t require executive function or problem-solving.

Katie Robbert – 52:00

Why would someone use this desktop version instead of just uploading files to the web version of Claude?

Christopher Penn – 52:35

Two reasons. Number one, you want the autonomous agent-driven version. The regular chat doesn’t have that. It kicked off multiple sub-steps and ran the “Co-CEO” by itself because it was in the recipe. In the web version, you have to manually do each step.

Second, every time one of these sub-processes run, it runs in its own short-term memory. Claude’s context window is 200,000 tokens. In this folder, we have so much content that if you did this in one chat, Claude would run out of memory. When these agents run, they come back with a short summary, which allows the orchestration side to keep things on the rails. It doesn’t have to try and remember everything all at once.

Katie Robbert – 54:06

It’s really just a matter of your working style. If you want to give it piece-by-piece and refine as you go, the web version is fine. But if you have all the pieces and a recipe, the desktop version is a better option. Can I see this data in the web version?

Christopher Penn – 54:44

No. You’d be able to see it in your desktop version, but not on the web, because Cowork doesn’t exist on the web. The other thing is repeatability. You just put new data in those folders next quarter, run the recipe again, and come back after lunch.

Katie Robbert – 56:07

Our next step internally is to take the output from this instance and refine it. It’s not ready for us to use yet, but we’d recommend people try this if they have access to Claude and a Mac.

Christopher Penn – 56:19

Hopefully the Windows version will be out soon. If you have someone technical, they should run Claude code. Claude Cowork is for people who don’t want to touch coding stuff; it gives them the capabilities with none of the cognitive burden.

John Wall – 57:18

It gave us 10 possible things to try. You just need to look at that list and say, “Okay, we’re going to take a shot at three of these.” Coming up with the list of which way to go next is often the hardest part. If you have some direction based on research, you’re way better off than just doing what the intern came up with last week.

Christopher Penn – 57:57

On that note, we’ll see you all next time. Be sure to subscribe to our show and check out the Trust Insights podcast and our weekly email newsletter. If you have questions, join our free Analytics for Marketers Slack group. See you next time.