Big Idea

The argument in one line.

Claude Cowork Projects replicates the full AI operating system architecture in a GUI, but its isolated VM sandbox means local MCP tools and custom integrations still require moving to an IDE.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A solopreneur or consultant who wants to automate lead research, content, or operations without touching a terminal.
Someone who has heard of Claude Projects but hasn't figured out how to structure it into logical business departments.
A business owner who wants scheduled AI agents running daily playbooks — GTM research at 7 AM, ContentOS on autopilot — without writing code.
A consultant building AI systems for clients who needs a transferable four-department framework.

SKIP IF…

You already work in Claude Code or a full IDE-based agentic setup — the GUI layer adds nothing you don't already have.
You want a deep dive into any specific skill (GTM OS, ContentOS, Research Lead) — this video surfaces each one briefly and points to separate deep-dives.

TL;DR

The full version, fast.

Claude Cowork Projects is a GUI version of the AI OS architecture: four department workspaces (GTM, Content, Delivery, Operations) each containing skills (plain-English SOPs), plugins (skill bundles), scheduled tasks, a context folder, and isolated memory. The GTM workspace automates daily lead research through Apollo and LinkedIn enrichment. The ContentOS runs an AI SEO-to-YouTube pipeline on a scheduled task. The real structural win is logical isolation — each department workspace knows only what it needs. The real ceiling is MCP: Cowork runs in an isolated VM, blocking local MCP server creation, and its domain allowlist feature is currently bugged. For power users the IDE is still superior; Cowork is the accessible entry point.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:17

01 · Intro

Hook on Cowork Projects dropping; promises overview of AI OS for content, sales, and operations.

00:17 – 02:01

02 · What are Projects?

Projects as isolated department workspaces. Four departments: Delivery OS, GTM OS, Content OS, Operations. Business logic: acquire, deliver, support.

02:01 – 03:14

03 · How Skills & Plugins Work

Skills are SOPs bundled into plugins. AtomicOps GTM plugin demo shows all skills that form the GTM department. Plugins are the distribution unit.

03:14 – 04:44

04 · Scheduling Automated Tasks

Creating a scheduled task: 'sales research' runs daily at 7 AM on Opus 4. Logical separation helps humans organize across departments.

04:44 – 06:14

05 · What is a Skill?

Skills are plain-English operating manuals. Research Lead skill walkthrough: Apollo leads, LinkedIn+Perplexity enrichment, relevance-first outreach test.

06:14 – 07:47

06 · Content, Delivery & Ops

ContentOS: AI SEO, competitor analysis, automated YouTube pipeline on scheduled task. Operations: daily briefs, email triage, Telegram dispatch.

07:47 – 09:25

07 · Context & Memory

Context folder in Cowork maps to VS Code context folder. Project-level memory persists key info across sessions; replaces basic RAG plugin need.

09:25 – 10:19

08 · Project Instructions

Project instructions are a lightweight claude.md for each workspace. Keep them rules-only; tone and formatting live inside the skill SOPs.

10:19 – 11:08

09 · Creating Projects

Three ways to create: from scratch, import from existing chat, or use existing local folder. Plan logically before building.

11:08 – 13:28

10 · MCP Connectors & Limitations

Remote MCP connectors work (HeyReach, FireCrawl). Local MCP blocked by Cowork's isolated VM. Domain allowlist whitelist currently bugged. IDE is still superior for custom integrations.

13:28 – 15:41

11 · The AI Operating System Play

Four-department blueprint fits most businesses. Scheduled tasks handle automation; conversations handle what can't be automated. Agent and human work in parallel.

15:41 – 15:57

12 · Outro

CTA to description deep-dives for each skill. Community and consulting links.

Atomic Insights

Lines worth screenshotting.

Claude Cowork Projects is a GUI wrapper around the same AI OS architecture power users build in an IDE — the ceiling is just lower.
A skill is a plain-English SOP: if a human employee could follow it step-by-step, Claude can execute it as an automated playbook.
Plugins are skill bundles — the distribution unit that lets you push a whole department OS into a workspace or sell it to a client.
Project-level memory persists key decisions across sessions so you no longer need a RAG plugin for basic cross-session awareness.
Cowork runs in an isolated VM that blocks local MCP server creation — a hard limit the IDE version does not have.
The domain allowlist inside Cowork Capabilities is supposed to enable local MCP proxies but is currently bugged for all users.
Personalization in outreach is dead; the Research Lead skill is built around one test: could this message only have been sent to this one person?
Scheduling tasks by department — GTM at 7 AM daily, ContentOS on autopilot — is the operational core of the AI OS, not the conversations.
The four-department blueprint (GTM, Content, Delivery, Operations) covers the surface area of most solo and small-team businesses without custom architecture.
Overly complex project instructions confuse Claude — keep them rules-only since tone and formatting belong inside the skill's SOPs.
Context folder in Cowork maps directly to the VS Code context folder; isolating it per project keeps each department focused without polluting others.
The IDE will always have more freedom, but Cowork removes the intimidation barrier for non-technical operators — which is the whole point.

Takeaway

The four-department structure that runs a one-person business on autopilot

WHAT TO LEARN

Splitting your business into GTM, Content, Delivery, and Operations before building any AI skill is the decision that makes everything else composable.

Skills are plain-English operating procedures: write them as if handing a playbook to a new employee, and Claude treats them identically.
Project-level memory eliminates the need to re-establish context every session — commit key decisions once and they persist across conversations.
Scheduling is where the AI OS becomes real: daily lead research, content pipelines, and ops briefs running at fixed times free you for high-leverage work.
Keep project instructions lean and rules-only; tone, formatting, and task logic belong inside each skill's SOP, not the top-level workspace config.
Custom MCP connectors in Cowork require a remote MCP endpoint — tools without one cannot be integrated until you move to an IDE-based environment.
The four-department blueprint (GTM, Content, Delivery, Operations) covers the operational surface area of most solo businesses without needing bespoke architecture.

Glossary

Terms worth knowing.

AI OS (AI Operating System): A structured set of AI workspaces, skills, and scheduled tasks that automate the recurring functions of a business — sales, content, delivery, operations — so human attention stays on high-leverage decisions.
Skill: A plain-English standard operating procedure loaded into Claude that tells it exactly what to do for a specific task — equivalent to a written playbook you would hand a new employee.
Plugin: A bundle of related skills packaged for distribution into a Cowork project or for sharing with clients; the distribution unit of the AI OS.
GTM (Go to Market): The department responsible for finding potential customers, making an offer, and closing them — covers outbound lead research, outreach campaigns, and sales calls.
MCP (Model Context Protocol): A standard that lets Claude connect to external tools and services; in Cowork, only remote MCP endpoints are supported — local MCP servers cannot run inside the isolated VM.
Cowork: Anthropic's GUI interface for Claude that adds Projects, scheduled tasks, context folders, memory, plugins, and MCP connectors in a sandboxed environment.
ContentOS: The content department workspace: runs AI SEO tracking, competitor analysis, and an automated YouTube pipeline (scripts, thumbnails) on a scheduled task.
Dispatch: A Claude feature that links your phone to Cowork so scheduled task outputs and alerts are delivered directly to mobile.

Resources

Things they pointed at.

03:27toolHeyReach

05:15toolApollo

05:15toolPerplexity

07:21toolTelegram

12:05toolFireCrawl

Quotables

Lines you could clip.

04:54

“A skill is literally just like a standard operating procedure. Imagine you handed a playbook to one of your employees.”

Clean analogy, no context needed, lands the concept in one sentence→ TikTok hook↗ Tweet quote

05:15

“I think personalization is dead, and relevance is the only thing that ever mattered anyway.”

Contrarian claim, standalone, no setup required→ IG reel cold open↗ Tweet quote

14:05

“The goal of our AI operating system is to do all of the work that slows us down growing in a business.”

Crisp mission statement for the whole AI OS concept→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogystory

00:00Cowork projects just dropped and it changes everything for nontechnical users. In this video, I'm gonna give you an overview of how to use it as an AI operating system from running your content journey to automating your sales and all those boring business operations in between. So let's get into it.

00:14Okay. So here we are in my Claw desktop app, and you'll notice on the left we have these shiny new projects folder over here. So the way that I like to think about projects are just isolated workspaces.

00:25If you look at any business out there, they're going to have specific departments and functions in them obviously, But at a higher level, you can think of a business as having the need to acquire new customers. They need to have something that they deliver to these customers, either a service or a product, and then they need to be able to support these customers.

00:42And they usually do that with some kind of back end operations. So with projects, we can split that out by function. So we can have our delivery operating system here, which has all of our skills and plug ins specifically related to delivering our product or service to their client.

00:57Same thing for GTM. Now GTM is just a term that basically says we're gonna make a really good looking offer, find a whole bunch of people to put it in front of, and then hopefully close them. That's really the entire process of going to market.

01:09And then obviously for support and operations, we would have a workspace as well down here that deals with all of the back end stuff that people don't really focus on, but also perhaps some of the documentation and things like that that we need to give to these people as part of delivering a service or doing our GTM such as proposal statements of work, things like that.

01:26And then content generally is one of the ways that we would serve all of these people. The point is we are using these folder spaces to logically isolate specific parts of our business.

01:35And as a consultant, this is extremely handy for me because if you've seen this channel, I've been talking about the AI operating system for a few months now. Generally, I work in my IDE and this tends to scare a lot of non technical people because they just see this box over here and all of this stuff and they think it's complicated.

01:50So Anthropic obviously acknowledged this. So they've built co work as a user friendly way of doing most of the things that we can do in the back end, but realistically, I still prefer working in an IDE. The point is though, with this logical isolation, it helps us understand things a lot better, but it also has a lot of added functionality that we're gonna look into now.

02:09So let's just start with the GTM operating system because that's probably the easiest concept to understand. So when I talk about an operating system here, what I'm talking about is a list of skills that are bundled together to serve this specific business need or department. In this case, it's going to market, which is taking that offer, putting it in front of the right people, and closing those people on a call.

02:29And as part of that, I have several skills. Skills, generally, if we're trying to distribute them, we're going to push them as plug ins because it's a really good way to bundle everything up and then push it into a system like this. So you can see over here I've got my AtomicOps GTM.

02:43I've split up my plugins just like I have my project folders and you can see over here all of the skills that would form the going to market plugins.

02:51So everything that I'm ever gonna need to go to market is written down in a skill that AI can take care of for me. And just as a side note, any plugins that I do talk about in this video, I have deep dives on my channel into how I built them and I generally give them away for free. So I'm not gonna be focusing too much on what all of these things actually are and how they work.

03:07You can watch the deep dive if you're interested in that. I wanna stick to what project is and how we would be using it for this specific video. But as a part of that, I obviously need to give you an example.

03:16So as part of going to market, I want to go and research all of my leads and this specific skill that we have over here called research lead, it's going to go out there and it's going to do hyper relevant outreach for us. So it starts with a LinkedIn profile, and it researches all the way through using several systems until it gets to a point where it's in a database for me to approve.

03:33Once it does that, it will go to HeyReach and it will automatically form part of a campaign that brings me leads on demand. So as a part of that, I wouldn't wanna have to go and do all of this stuff manually because coming in here and going research lead, why would I wanna do that? I would much rather schedule tasks.

03:48We can do that over here. Now for the past few weeks, it has been available in Cowork as a whole. But again, this logical separation just helps us as humans organize our working life.

03:57Me as a consultant or a solopreneur, I understand that, okay, this folder space over here is exactly for everything that I need for going to market. So I'm gonna schedule my tasks for researching my lead over here, and I would just simply come down here and I would say sales research, and I could set a frequency.

04:16I want this to run every day. I want it to run at seven in the morning, and I can run it on a different model if I want to.

04:22So for this, I would probably leave Opus because it's a really long research lead workflow. But for smaller tasks, you might want to use one of these models down here. And then all I would have to do is hit save and we can now see that it's a part of my GTMOS that every morning at 07:00 this thing is gonna run and research all of my leads for me and put them in front of me for my approval.

04:39And at this point, I'm sure some of you might be wondering what the hell is a skill or a plugin, so we're gonna have a brief look at that over here. I've gone over to the customize tab and you can see we have my personal plugins loaded. We clicked on the GTM plugin over here, and you can see all of my skills.

04:52So a skill is literally just like a standard operating procedure. Imagine you handed a playbook to one of your employees, and this thing told them exactly what to do step by step in order to do the job that they were paid to do this morning.

05:06That's exactly what a skill is. We are just telling Claude what to do for this specific thing. So if we come on over to the research lead thing that we just scheduled, you'll see over here, the goal is take leads from Apollo, enrich with LinkedIn and Perplexity Research, and generate signal built outreach that passes one test.

05:24Could this message only have been sent by this one person? Because for this whole thing, I'm focusing on relevance. I think personalization is dead, and relevance is the only thing that ever mattered anyway.

05:33So this skill kind of focuses on that. More importantly though, you can see how simple this is. It is in plain English, and like I said, if you had to hand this to a human and you were paying them money, they would understand exactly what to do.

05:45It just outlines the system flow. It tells you what tools the system is going to be using in order to achieve its goal. In this case, we're using several MCP servers.

05:53We're using a bunch of scripts down here. We have a bunch of references of what my writing might look like, what a good post looks like when I'm reaching out to people. All of those things live inside our skill because it helps AI be a lot smarter at doing whatever it needs to do.

06:08And as you can see in your business, you're gonna have tons of different skills to help you do different things within your business. So as a human and perhaps even as a system, makes logical sense for us to group these things, which is why we have this concept of an operating system. Because if I want a GTM workspace, I want to make sure that all of my stuff lives in that workspace.

06:27And the easiest way to group all of these skills is to bundle them into something called a plug in so that I can distribute them into my own GTM operating system. I share it with my community that way. You can even use this to go and sell it to businesses as part of a tailor made AI operating system for them, which is part of the model that I talk about a lot on this channel if you've seen any of my other videos.

06:46Then obviously, it's not just about GTM. We need to fill up our other project spaces or departments within our business. So we have skills for content.

06:53Here we have AI news monitors, AI SEO that tracks what YouTube is currently trending, competitor analysis to see what people in the space are talking about so we can find contrarian approaches. Then as part of this, it turns it all into an automated YouTube pipeline that just writes me scripts, generates me thumbnails, all of that stuff on demand that all lives inside our ContentOS, runs on autopilot on a scheduled task.

07:15We do the same thing for our back end ops. We can build things like daily briefs. So instead of checking our emails, reading this place, reading that place, can get all our news delivered to Telegram or even now with dispatch.

07:26This is literally your phone linking to co work. We can get it all sent to us so that we have to do far less manual work. That is the whole point of all of this.

07:35And this is how I am currently building my entire AI operating system. So that at a very high level is just how these things work and I'm gonna flip on back into our GTMOS over here. And we're gonna talk about the next most important thing which is context.

07:47Now context is something that has existed in Claude for a while. Specifically with this, we're literally giving it a folder workspace over here on my computer much like I would when I come on over to Versus Code over here, you'll notice that I have this context folder, And inside this context folder is all the context that I need for my business.

08:04The way that I write, the way that I speak, everything to do about my personal journey over the last twelve years, it lives in here. So every time I either talk to Claude or run any of the skills that you've now seen, it has all of this information that it needs, and it can just pull it out of there whenever it needs it.

08:19So we now have that living right here inside Cowork for us, but more specifically, it's inside our project folder, which which just helps this specific project achieve its goals a lot better because context makes everything when we are working with AI. Then next up, and this is really handy, we now have memory and co work.

08:35So specifically for each of these project folders, it has isolated memory for this and it remembers things across sessions. So if I talk to it one day today and I say a bunch of things, anything it thinks that it's important, it will commit to memory or you can specifically ask it to do that. But more importantly, next time you have a conversation with it, it will remember that important thing and bring it into the context window that you are now working with.

08:57So previously, you had to use a plug in or a skill to help you do this. And to some extent, you might still want to use one of those because we've built some pretty crazy memory systems on this channel where we stash things in rag and we have a whole bunch of temporal decay and other fancy things that help us have really intelligent memory systems.

09:13But for most businesses, they really don't need it. They just needed something that helps make Claude a little bit more context aware compared to starting a new session every single time they open up Claude. But then last but not least, we have instructions and these again are for project level instructions over here.

09:28So this is kind of like your claude.md for this specific workspace and it just gives it a few rules that we would might want to use for this workspace. So instead of doing things like adding tone or formatting, which to me I would stash in my context over here or within the skill when I give it an example of what good writing or good tone looks like, I would focus on rules, specifically how to guide the way that this thing works.

09:50So I might just give it some very basic context saying you're a GTM operating system and you're helping this business achieve this thing. Something simple like that. This doesn't need to be complex.

10:00Oftentimes having a massive claw dot m d file or giving it a lot of instructions when it doesn't really need them can just confuse the system. You need to think about that before you're adding instructions where you don't need them especially now that we have context in here and most of our instructions live inside our skill where we have that exact operating manual that we need.

10:18So creating projects is extremely simple. All you have to do is click on this plus sign over here. You can either start one from scratch, you can import a project from a previous chat that you had, or you can use an existing folder that you might have lurking around here somewhere.

10:32But for me, like I showed you over here, this is how I planned it. I didn't just come in here and kind of y'all on my way through it. I've logically broken down aspects of my business and I've been doing this for a few months now working in the back end version of Claude because I think it makes the most sense to do it this way.

10:48Even if you're in a larger organization and you have a sales marketing department, they would have a project folder, but inside there, they would have separate things that help them achieve their goal as a sales marketing department. It would be up to you as a business or the consultant coming in and installing this for a business to understand how they would want to break things down logically.

11:05But for the most part, I think what I've got here does a pretty good job at that. So I think it definitely comes down to proper planning because you wanna make sure that you get this thing right. Once you have logically isolated it, building the specific skills that you need is easier than it's ever been because you can just talk to Claude and say, I'm trying to build this aspect of my business.

11:21What do I need? There are some limitations that we need to talk about though because while this is getting to where it needs to go, the back end version is still far superior. I have a lot more control, functionality, and I have a lot more freedom despite some people thinking that it's a little bit intimidating.

11:36Some of the things that you need to think about here are when we have all of our skills, they need to be able to go out there into the real world and do a lot of actions on our behalf. That's how this whole automation thing works. And right now, there are a few limitations with the MCP connectors inside Cowork.

11:51So we have a whole bunch of the main applications that people use every day and that's great. But then we can also add custom ones. So if we don't have something over here when we're browsing our connectors, you do have the opportunity to come over here and just add a custom connector.

12:05But that custom connector that you want to add, they need to have a remote MCP service set up for you so that you can just put in the URL and connect. Now thankfully for things like HeyReach and FireCrawl, I've been able to do that. You can see here these are two custom ones that I set up earlier.

12:19But we're limited here in what we can actually do if this functionality doesn't exist. In my IDE, if I don't have an MCP server, I can literally build one like that and run it on the spot. With this, I can't do that because the way that co work actually works, it creates an isolated virtual machine on your computer and it locks you behind that so you don't have full access to do whatever you want.

12:38They do that to protect you from prompt injection and a few other things that don't really matter to you right now. But the point here is that you're supposed to be able to override this to some degree in the capabilities over here. If you look down here, you can see we have a domain allow list.

12:52This would give us a little bit of network access to go and maybe set something up locally and then use that to kind of proxy our way through to a specific tool that we need to use. But right now there's a massive bug where no one is actually able to do this because this white listing isn't actually working. But I digress.

13:08The whole point of me saying that is that if a tool does not exist in MCP, you're not going to be able to run a specific workflow that you might have built in your IDE version back here. It's probably not going to work in here.

13:20So that's one of the biggest issues at the moment. I imagine they're gonna fix that very soon So I think the play here is really simple for building an AI operating system. You need to look at your business and to be honest it's mostly going to look like this because most businesses have these functions.

13:32You need to sell something that's why you're in business. You need in front of those people which is why you have go to market. You're probably gonna get that done with content at the moment.

13:40That's a very popular way of doing this as a part of direct lead gen. So you're gonna have your inbound and your outbound. And then you obviously have all the operations and the boring stuff on the back end.

13:49You're gonna automate as much of this stuff as you can. So you're going to set up scheduled tasks for your skills or your playbooks that go through these things and the agent does all of the work for you, and then you're gonna work alongside it by talking in here and this thing will pop up in some chats for the stuff that you cannot entirely automate.

14:06And you can do that. So while this thing is running, can pop over to my GTMOS, and I can work in here for a little bit to come back when this chat is done and carry on working in here while these things are also autonomously running on their scheduled tasks.

14:19So the goal of our AI operating system is to do all of the work that slows us down growing in a business. So you as a solopreneur or a consultant, if your goal is to make money, you obviously need to be focusing on areas where your time personally has the most amount of leverage on making that money, and the machine takes care of everything else.

14:37That is the point of this operating system. And so understanding this structure is the most important because once we've nailed this part, all we need to do are build a specific list of skills to get each one of these things done. For delivery, I need this specific list.

14:50For GTM, I need this specific list and you can see we've broken that down here. If I come and have a look at my delivery OS, we just go to plug ins.

14:58Say you're a consultant and part of what you do is either audits or you build prototypes. You can see here we have a build app skill. So this is something where I would need to manually come in here.

15:08So while all of my other stuff is working in the background, I can come and build my prototypes inside Cowork using this specific skill. Then I would flip over to my GTM and I would use other skills while some of the skills are running on their native schedules. And that to me is how we work in the system and it's only going to get better specifically in co work.

15:26For now, I'm definitely still sticking to my IDE because like I said, there is far more functionality inside here. I can get away with doing a lot more a lot easier because I have complete control and complete freedom. So this was just a very quick video to give you an overview of projects as a whole.

15:40I have a ton of deep dives into all of these things that I just spoke about here, so you can check those in the description below. Otherwise, leave some comments, and let's talk about this more. I will get back to you as soon as possible.

15:49If you need some help or want this type of thing set up for you, you can also get in touch with me or check out my community. Thanks very much for watching. I'll see you guys in the next

The Hook

The bait, then the rug-pull.

When Anthropic shipped Cowork Projects, the pitch was simple: everything IDE power users build behind a terminal, now in a GUI. Mansel Scheffel took that claim seriously, rebuilt his entire AI operating system inside four department workspaces, and then told you exactly where the ceiling is.

Frameworks

Named ideas worth stealing.

00:45model

Four-Department AI OS

GTM OS
Content OS
Delivery OS
Operations

Most businesses need to acquire customers (GTM), generate awareness (Content), fulfill their offer (Delivery), and handle back-end processes (Operations). Each maps to a Cowork project workspace.

Steal forstructuring any solopreneur or small-team AI agent setup

02:01model

Skills > Plugins > Projects (Distribution Ladder)

Write a skill (plain-English SOP)
Bundle into a plugin (shareable/sellable)
Deploy into a Cowork project (department workspace)

The three-layer packaging model for building, distributing, and deploying AI operating system components.

Steal forproductizing AI workflows for clients or communities

05:15concept

Relevance-Over-Personalization Test

For outreach, filter every message through one question: 'Could this message only have been sent by this one person?' If yes, it is relevant. Personalization (name, company) is table stakes; relevance (problem fit) is the differentiator.

Steal forcold outreach, lead nurturing, sales messaging

CTA Breakdown

How they asked for the click.

VERBAL ASK

15:41link

“Check those in the description below. If you need some help or want this type of thing set up for you, you can also get in touch with me or check out my community.”

Soft, non-pushy. Points to deep-dive video playlist in description and community link. No hard sell.

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.