Modern Creator
Why Modern Creator?
Mansel Scheffel · YouTube

Claude Just Changed the Rules: Rebuild Your AIOS From Scratch

A 56-minute walkthrough of rebuilding a Claude Cowork setup around the new cloud-first model, in five layers from context files to scheduled routines.

Posted
2 days ago
Duration
Format
Tutorial
educational
Views
1.6K
64 likes
Big Idea

The argument in one line.

An AI operating system should be built bottom-up in five layers, starting with portable context files so the whole system survives any change to the tool running it.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You built a Claude Cowork or Claude Projects setup before the cloud migration and want to know what moved, what broke, and how to rebuild it.
  • You are a consultant or operator who sets up AI systems for clients and needs a repeatable framework for where context, data, skills and schedules live.
  • You run a small business through Claude and want always-on scheduled tasks without keeping a laptop open or renting a VPS.
  • You keep collecting skills from random repos and want a grounded way to decide which skills your business actually needs.
SKIP IF…
  • You only use Claude as a chatbot for one-off questions and have no recurring workflows to systematize.
  • You are a Claude Code power user looking for CLI or agent SDK specifics. This is the desktop and web Cowork product, not the coding tool.
  • You need a data governance or compliance walkthrough. The video explicitly defers that to other videos.
TL;DR

The full version, fast.

Claude Cowork now runs tasks in Anthropic's cloud, so any file a task needs gets copied up and goes stale the moment you edit it locally, and there is no local-only mode anymore. The video rebuilds a working setup in five layers. Layer one is context: connect your SaaS tools as MCP connectors, lock down write permissions, then run an onboarding skill that pulls your business and voice into plain markdown files. Context splits into curated knowledge, changing state that belongs in a database, and memory Claude learns on its own. Layer two is projects, organized by four business pods: acquisition, delivery, support and operations, each with a coordinator that dispatches isolated worker threads. Layer three is skills built from a workflow audit rather than downloaded repos, bundled into plugins and distributed through a private marketplace. Layers four and five are the synced interfaces and scheduled routines, with monitoring still missing on consumer plans.

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 – 03:00

01 · Why your Cowork setup needs a rethink

Cowork tasks now run in Anthropic's cloud containers. Files are copied up, go stale once edited locally, and there is no local-only mode. The five-layer ladder is introduced.

03:00 – 16:15

02 · Build your business brain: context, data and memory

Connect MCP connectors first, lock tool permissions, upload the onboard and pod-mapper skills, then run onboarding to produce business and voice files. Context is split into knowledge, state and memory, and data is separated from context.

16:15 – 28:02

03 · Set up the new Projects and connect your files

Projects are reorganized around four pods. A new Acquisition project is created with a goal, the coordinator-and-threads model is explained, and the library shows the difference between uploaded and linked folders. Settings cover models, memory and environment.

28:02 – 33:34

04 · Run skills with Claude's worker threads

A YouTube content pipeline skill set is run inside the project. Threads spin up in isolated context windows, send results back to the coordinator, and the conversation compacts over time. Cloud-only folders cannot reach local databases.

33:34 – 45:55

05 · Map your workflows and build skills and plugins

The pod-mapper audit interviews you, maps a workflow trigger by trigger, lists needed context files and produces an automation map. Three ways to build skills, the Customize tab, the vetted marketplace, plugins and private marketplace repos.

45:55 – 48:30

06 · Turn business data into useful dashboards

A business rescue skill audits all four pods across connected systems and presents the findings as an artifact: a verdict, money lost, and actions, shown with mock data.

48:30 – 50:31

07 · Use your system on desktop, browser and mobile

The same project, threads and library appear in the browser and on the phone. Linked local folders only show while the computer is on and do not yet appear on mobile.

50:31 – 56:04

08 · Scheduled tasks, routines and monitoring

Out-of-the-box scheduled tasks, manual task setup, in-project routines, the Claude Code routines screen with GitHub and API triggers, a morning brief test run, and the lack of monitoring on consumer plans.

Atomic Insights

Lines worth screenshotting.

  • Cowork tasks now run in Anthropic's cloud by default, and there is no local-only mode for them anymore.
  • Files a cloud task pulls up are snapshots: edits you make locally after the task starts never sync, so you start a new task to see them.
  • Context is three different things living in three different places: curated knowledge in markdown, changing state in a database, and memory Claude learns on its own.
  • Memory feeds context but does not equal context; it lives in a separate system from the files you wrote.
  • Connect your SaaS tools before anything else because they already hold your voice, your clients and how you work.
  • Set tool permissions so Claude can never send a client-facing email or change a live system without approval.
  • Build context files first because they are portable: the same markdown plugs into Cowork, Claude Code or any other model.
  • Every business breaks into four pods: acquisition, delivery, support and operations, and each pod gets its own project.
  • In the new Projects, a coordinator manages the job and worker threads do the work in isolated context windows that only send results back.
  • Thread isolation is a token strategy: the coordinator reads a result, not the whole trail of work each thread produced.
  • A linked local folder keeps data on your computer while Claude still works in the cloud, but only while that computer is on.
  • A skill is a handbook for a new employee: structured instructions beat telling the agent to go figure it out.
  • Skills come from pain, not from repos: audit a real workflow step by step before writing one.
  • Plain-English instructions are probabilistic; scripts inside a skill are deterministic and anchor the output so it comes out the same every run.
  • Give a skill a reference of what good output looks like and it stops inventing a new format each time.
  • If a task runs once, let the agent improvise; if it repeats, turn it into a skill.
  • A plugin is a bundle of skills, and a private marketplace repo lets you push vetted skills to clients and keep them in sync.
  • Schedule recurring work the moment it is built inside the project, rather than at the end from a separate screen.
  • Claude still has no native webhooks, so event-driven triggers need an outside tool or an always-on computer.
  • Consumer plans get a usage page and nothing else, so serious users have to build their own monitoring dashboard.
Takeaway

Build the context layer first and everything else survives.

WHAT TO LEARN

Cowork moving to the cloud exposed every setup that depended on the tool instead of on portable context, and the fix is to rebuild in five layers from the bottom up.

01Why your Cowork setup needs a rethink
  • Cowork tasks now run in Anthropic's cloud containers; files are copied up at task start and do not sync afterward, so a changed file means a new task.
  • There is no local-only mode anymore, but anything that depends on a local script still runs on your machine as long as it is on.
02Build your business brain: context, data and memory
  • Connect MCP connectors first because they already hold your voice, your clients and your working patterns, then restrict any tool that can send or change something.
  • Run an onboarding interview that pulls your business and voice into plain markdown, since those files port to any model and any future version of the tool.
  • Treat context as three things with three homes: curated knowledge in markdown, changing state in a database, and memory Claude learns on its own.
  • Data is everything available; context is the smallest useful slice for the job at hand, and loading more does not make the output better.
03Set up the new Projects and connect your files
  • Split a business into four pods, acquisition, delivery, support and operations, and give each its own project with its own goal and context.
  • The coordinator is a project manager, not a worker; threads do the work in isolated context windows and send back only results, which keeps token use down.
  • Uploading a folder copies it to the cloud; linking a folder keeps it on your computer, and only the linked option reaches local databases.
  • Set the coordinator effort higher than the default low for real work, and use the usage split to see what threads versus the coordinator consume.
04Run skills with Claude's worker threads
  • A long-running project conversation compacts its own context over time, so it will not remember everything forever even though it scrolls forever.
  • A cloud-only folder cannot reach a local database; move that data to a hosted database or link the folder on an always-on machine.
05Map your workflows and build skills and plugins
  • Audit a painful workflow trigger by trigger before writing a skill, listing tools, data sources, required context files and steps that can be cut.
  • Skills from random repos bring prompt injection and clutter; every skill should trace back to a real recurring pain.
  • Pair plain-English instructions with scripts and reference examples so the output is deterministic and matches what good looks like every run.
  • Bundle related skills into a plugin and distribute through a private marketplace repo so clients get vetted, synced skills without touching your work.
06Turn business data into useful dashboards
  • A dedicated audit skill can read every connected system across the four pods and present losses and actions as an editable artifact a non-technical reader understands.
07Use your system on desktop, browser and mobile
  • Projects, threads and the library sync across desktop, browser and phone, but linked local folders only appear while that computer is on and not yet on mobile.
08Scheduled tasks, routines and monitoring
  • Schedule recurring work from inside the project the moment it is built, and write unattended instructions that forbid questions and connector suggestions.
  • Routines in Claude Code add GitHub and API triggers, but native webhooks still do not exist, so event-driven flows need an outside tool.
  • Consumer plans only expose a usage page, so anyone serious about running a system should build their own monitoring dashboard until Anthropic ships one.
Glossary

Terms worth knowing.

AIOS
AI operating system: a structured setup of context, connections, skills, interfaces and scheduling that lets an AI assistant run repeatable business work rather than answer one-off chats.
Cowork
The task-running mode of the Claude app, as opposed to chat. It runs skills and multi-step jobs, now inside Anthropic's cloud containers rather than only on your machine.
MCP connector
A Model Context Protocol connection that gives Claude read or write access to a SaaS product such as Gmail, Slack or a calendar, with per-tool permissions you can restrict.
Coordinator and threads
The new Projects structure: a coordinator acts as project manager and dispatches isolated worker threads, each with its own context window, that return only results.
Skill
A folder containing a plain-language instruction file plus optional scripts, references and assets that tells Claude exactly how to perform one repeatable task.
Plugin
A bundle of related skills packaged together so a whole workflow can be shared or installed as one unit.
Marketplace
A source, such as a GitHub repository, from which plugins can be installed and kept in sync, either Anthropic's public directory or a private one you run for clients.
Pod mapping
An interview-style audit that walks through one business function step by step, lists the tools and data involved, and decides what AI should own, assist with, or leave to humans.
Routine
A scheduled or event-triggered run of Claude in the cloud, the Claude Code name for what Cowork calls a scheduled task, with extra triggers such as GitHub events and API calls.
Artifact
An HTML page, document or deck that Claude creates and can edit inside its own app, used here to present audit results in a readable layout.
Context compaction
The process by which a long-running conversation summarizes what it thinks matters into checkpoints so it can keep working after it runs out of room.
Resources

Things they pointed at.

05:38linkOnboard skill (free download) ↗
05:38linkPod Mapper skill (free download) ↗
21:50toolSupabase ↗
21:50toolSQLite
21:50toolNeon ↗
18:12toolFathom / Fireflies call recorders
36:40toolCalendly, LinkedIn, HeyReach, Google Docs (tools found in the mock workflow audit)
42:40productClaude for Small Business plugin
45:20productAnthropic plugin directory (marketplace source)
54:00tooln8n (webhook workaround) ↗
55:40productAI Native community (Skool) ↗
Quotables

Lines you could clip.

00:00
“Claude Cowork has just moved to the cloud and every Cowork setup built before this week was built for the old version.”
clean statement of the stakes in one line→ TikTok hook↗ Tweet quote
02:30
“There is no local only mode for cowork tasks in the way that they used to be.”
the single change that breaks old setups→ IG reel cold open↗ Tweet quote
04:22
“Once we nail this stuff, you can literally plug this into cowork or any AI model and it's completely portable, meaning you will never be locked into a system.”
the portability argument for context files→ newsletter pull-quote↗ Tweet quote
04:50
“I will never let this thing write an email.”
short, firm, memorable permissions rule→ TikTok hook↗ Tweet quote
05:00
“You don't want to be focusing on that stupid flashy shit that you see on YouTube where they zoom in and out of a bubble. You want to focus on the boring things that are repeatable.”
contrarian, sweary, cuts against hype→ IG reel cold open↗ Tweet quote
21:05
“While memory feeds context, memory as a term alone does not summarize the entirety of context.”
clears up a common confusion in one line→ newsletter pull-quote↗ Tweet quote
28:38
“Would it be better to just say, hey, go and organize all of that stuff over there for me? Or would it be better to give them a handbook?”
the new-employee analogy for skills→ IG reel cold open↗ Tweet quote
33:50
“You don't just go out there into some random repo and pull down a bunch of shit that somebody told you your business absolutely must have.”
contrarian take on skill hoarding→ TikTok hook↗ Tweet quote
44:10
“Imagine it was probabilistic. One day you click on the start menu and it opens up my computer. The other day it opens up one of your favorite games.”
vivid analogy for why skills need scripts→ IG reel cold open↗ Tweet quote
46:45
“If you're just gonna be running something once, probably you don't need to build a skill for it. But if you know something is going to be repeatable, it absolutely makes sense to turn it into a skill.”
the decision rule in two sentences→ newsletter pull-quote↗ Tweet quote
53:50
“Considering Grokbot and ChatGPT can do both of these things natively, I think it's kind of silly that we can't do that yet.”
honest criticism of the missing webhooks→ TikTok hook↗ 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
Claude Cowork has just moved to the cloud and every Cowork setup built before this week was built for the old version. The way you used to work with it and your projects has changed entirely. So today, I'm going to show you how to set up an AIOS in Cowork that actually works now, so your context stays current, your work keeps running even when your laptop's closed, and you're not burning through your tokens unnecessarily.
I'll break it down into five simple layers, so by the end of it, you'll know exactly where everything goes. For those of you who don't know me, my name is Mansel, and I've spent the last 12 years consulting for Fortune 100 and 500 companies. I built two consultancies during that time, and now I teach people how to do the same.
Let's get into it. Okay, so when we're trying to build an AI operating system for ourselves or for our clients, the absolute last thing that we want to do is YOLO our way through it. So we have this little ladder here that's going to help us build this thing step by step.
making sure that we cater for everything inside our specific little ladder over here. But before we start climbing this ladder, I do need to talk about what's actually changed in the new version of Cowork because it's important to set the scene about how you might be working with all of these different layers for you or your clients.
So if you log into Claw today, you might have had an update. It is still rolling out to some people, so it depends on your region and a few other things, but it will eventually come to you. You'll notice here that my Cowork tab is gone and we now have this unified interface.
That doesn't mean that coworker is dead. It just means that Claude itself will make the decisions for you. If you just ask a simple question, it's just going to answer you like a chatbot.
If you have something that involves tasks or skills and things like that, it will run a coworker task. The difference comes in where that task runs. Now, if you don't select anything over here, of course, this thing is just going to run, but it will be running natively inside Anthropix Cloud.
what they do is they create a little container which is like a virtual machine in their cloud it runs the task and it brings you back the results that's how they're getting this always on thing between the desktop app the cloud app and your cell phone as well so for those people who did want some kind of always on functionality without having to leave their computer on or have a VPS or something like that.
It's really going to cater for you. There are some differences though. There is no local only mode for cowork tasks in the way that they used to be.
So what happens is when I click add a folder to that little cowork tab now, it's going to push up whatever it needs to run inside Anthropix Cloud. So the task itself will run there and any of the AI work that needs to get done. But if there are any specific files that this actual task needs, it will pull them up into its cloud to complete the work.
Those copies will in its cloud until that session is deleted and any changes that you made after you have already started the task, it does not sync automatically up to the cloud. meaning that any files that you do change during this session, the version that is in the cloud will probably be outdated, in which case you need to start a new task.
Anything that you do run locally will still run locally. So for instance, if you have some scripts that rely on something on your local folder on your computer, they will still run as they need to inside your environment as long as your computer is on. Cool, so we can get started with the first leg in our ladder.
There are lots of other big changes for 2026, but we'll get through those as soon as we get to the specific layer in the ladder. Right now, I wanna focus on data and context because this is the foundation of everything that you're gonna be building upon. Once we nail this stuff, you can literally plug this into cowork or any AI model and it's completely portable, meaning you will never be locked into a system.
It also has all of the information that you're going to need for anything that you do with AI. So the more time we spend doing this stuff, the better any system that you build will be.
Now there are several ways that you can get context into a system. But the first thing that we want to do is head on to customize over here. And then we're going to go to this little connectors tab over here.
Connectors are just MCP connections to your SAS products and the tools that you use every single day. The reason why we want to connect these first is because a lot of them already have context about who you are or how you work, how you write, things like that. So if we get these things loaded up front, we can actually run an onboarding skill that will go and pull out some of this context for us as a first step.
So for your first step, make sure that you come down here and just connect whatever it is that you use. At a minimum, I would say at least connect Gmail or whatever email app that you're using so that you can get your voice from there and perhaps some other context around your business. Once you've connected these things, I just want you to click on them and then go to your tool permissions over here.
What this thing is going to allow you to do is configure any permissions that you might have for your tools out there. For things where it could do something client facing like send an email or change something inside a system. you want to make sure that you cater for the permissions accordingly.
For instance, I will never let this thing write an email. If you have a system change that you wouldn't want it to do, you could either get it to ask for approval or you could just deny it. But that's definitely something that you want to take into account when you're setting these things up.
The second thing that I want you guys to do is grab the onboard skill and the pod mapping skill from down below, and you'll be able to then add it. to your cloud environment. Those are going to be vital for building some of the workflows that we're going to be doing, especially for onboarding.
So after you've downloaded them, all you would need to do is head on over to the skills tab and then hit add an upload skill. run through the process of uploading them literally by dragging and dropping them hitting upload and you will then have them available inside your cloud environment something to note if you do upload skills this way they will be synced to your cloud account and therefore can be used anywhere if you're doing any of the stuff through cloud code which is this thing's little cousin that we're not covering in this video it is a separate environment so that's just something to take into account if you're a little bit technical and you're going to be using both of these systems but now that we've done that we need to look at what context actually is before we can onboard ourselves so whenever you're trying to build a business brain you don't want to be focusing on that stupid flashy shit that you see on YouTube where they zoom in and out of a bubble.
You want to focus on the boring things that are repeatable and live inside systems that actually make sense because a lot of people don't cater for that stuff and if you build something like that for your clients you're going to end up with more hassle than something that just works straight out of the box. But when we're looking at what context really is, it's made up of so many different moving parts.
If we try and compartmentalize them into three things, we have knowledge. Knowledge are the curated facts that you might know up front. So it could be your business voice, your clients, what type of people you're serving, what your offer might be, things like that.
These are facts that live inside systems that either the system itself or the skills that it's going to be running need. So the information that those systems need. Then you have something called state, and this is always changing.
It's never a constant thing unless, for instance, a lead is closed. But at some point, that lead in the system would have been in progress, maybe it turns to one, maybe it ends up being lost, whatever. That is a state change, and this...
does not live in a normal little file. This would ideally live inside a database of some kind because it's a lot better at managing state. And then finally over here we have memory and these are the experiences and things that the AI learns as a result of working with you.
Whether that's by questioning you on something or even if that is something that it just picks up from working inside the console. And we're going to look at where each of these things live inside our AI operating system as we build up through the layers.
So now that we're back in Claude, the first thing that I want you to do is make sure that you do create a local folder. We are going to be using that just to set things up and then ultimately we will branch out into the cloud. For those of you who do want to have some kind of always -on solution, you can obviously always keep using this local folder for your specific hybrid runs.
but the point is i want you to add one you can call it whatever you want it really doesn't matter i'm just going to call mine ai os test and i've already got a test folder in there so i'm going to create a second one and then i'm going to hit open then it will ask you if you can have permissions you're just going to hit actually always allow because you always want claude to be able to access this folder and after you've done that we're going to get on board so we're going to take that skill that you grabbed from down below and set up just a second ago and you're going to do forward slash onboard.
If something pops up, it's just asking you whether Covert can run this thing, just click allow. And you'll see here it automatically opens up a session. Before we get onboarded, there's actually one more thing that I want to tell you guys about, and that is obviously making sure that you keep your settings private.
Now, depending on whether you are running either the Pro or Max plan, or if you've got a Teams or Enterprise plan, there are different things going on here. So for most of us, I imagine you're on the Consumer plan, meaning the Pro or Max plans. in which case you want to head on over to privacy and just make sure that you untick these toggles over here so that you're not helping improving any air models and that you don't share your location metadata of course this doesn't mean that you're part of any regulatory compliance or adhering to any data governance laws you need to cater for that separately if you are on the teams or enterprise plans you have a lot more control over your data things are turned off by default and so on and so forth you obviously still need to check with whatever workflows that you're building on how to deal with the data governance behind it and how to take care of that for your clients i have separate videos on that that i will link down below for those of you wondering so if we're coming back to our onboard skill we can just get a brief look at what this thing does what it is going to do is it's going to see what's already here if you have any files already that this thing can use for context it will use those but of course we probably don't if we're starting from scratch
So what it will do is it will check if you have a website. It will ask you what that is. Go to your website and pull down any context that it can from there.
Of course, that will tell you who you serve, who your clients might be, some information about your business. This is all things that you're going to be needing for your systems. Then it will ask you for any of the stuff that might be missing.
So it will interview you to pull out any of the knowledge from your head that it couldn't find from your website. Then what it's going to do is it's going to create however many files it needs to based on the things that we've done. For most of you guys, it will just be my business and my voice.
Those are the two things that we're setting up first. And also one of the reasons why we connect to Gmail, because Gmail contains our voice that this skill can then go and run through to get that information based on how you send emails. Finally, this thing will end with helping you do your first task, whether that's creating a morning brief or something related to this work that you're doing.
The idea here is just to get you a really simple start before we start climbing into the more complex stuff. Okay, run through the skill and also check my Gmail. You can grab my voice from there for the file.
Any questions you have, feel free to interview me. so for your thing this will obviously go and run through your live systems and check real things i've just asked you to use mock data for this video test so that i don't have to redact anything that it might pull out that i don't want the world to see not that i'm keeping people in my basement but you can see here that it starts to run through everything in here and whatever it does find will ultimately get stashed in that folder that we've created we're not always going to be using that folder forever unless we're working in this kind of hybrid approach, but when we get to projects, things are about to change quite a lot.
This thing on the side over here is actually really handy for you guys because it can tell you not only what task it's currently working on, but also what it is currently doing for every single part of the session. So any connectors that it would have gone into. For instance, if you connected multiple MCP connections to this and you told it to scour several different systems out there for separate reasons.
you'd be able to see what it did and what it did inside there as well just by clicking on it and then it will give you some more information over here. It will also tell you any skills that ran and also where this thing ran. Currently, like I said, we are still using most of this on the desktop because the skill is forcing it to do everything locally in this instance.
So the next thing that the skill will ask you as part of onboarding is what do you actually want to do as your first task? This is important not just for you but also for your clients. If you were doing this live demo for someone, you would want them to start getting into this as fast as possible because the more that somebody gets something solved really fast, the more likely they are to adopt a product.
So doing something like this when you're training a team or perhaps going out there and showing someone something in a live workshop, it's very important to get wins as fast as possible and as easily as possible and slowly build up in complexity. just like we're doing in this video. I'm just gonna choose a client follow -up email for this.
And then, your notes to self are raw and sweary, and your emails to people are short and plain. Which should the voice file be built on? This is a very important thing because, again, depending on which email you have connected to, it might get a different voice.
If you connected this thing to your personal Gmail, you would get one style of voice. But if it was just to your client emails, that would be entirely separate. Now, if you're using the system solely for business, it's obviously better to connect the Gmail for your business.
Annoyingly, at the time of me recording this, you still cannot connect more than one Gmail account, which is absolutely ridiculous because pretty much everything else can. I'm going to tell this thing to split by context because it's obviously aware of who it will be speaking to. To some degree, it can choose the voice based on whatever it thinks is best.
Then what it's going to do is it's going to create the two draft files around your business and what it thinks that it is and also around your voice and what it thinks your voice sounds like. And these will pretty much be the backbone for a lot of the things moving forward. You can, of course, create as many context files as you want, and you absolutely should.
make sure that you cater for everything that might not have been catered for this i don't want to run through every single onboarding skill that i have i have a separate video on that that walks through them in depth all of the different interviews and things like that i will link that down below and it will include all the skills and everything that you need to actually do it for now i'm going to save this as it is and i want to show you some different things here so what we just built is context in the sense of harvesting that from someone's brain but like i said there are different parts there are databases which live in state and then there is also memory and you'll notice here for any eagle -eyed viewers that this thing suddenly pulled in memory over here and that's obviously because i've already been working with claude for the last year and a half so it's learned a little bit about me and i absolutely didn't tell this thing anything but if i click on this little memory tab over here you can see it has a specific memory for my business atomic ops and then it's got a summary that it wrote by itself and all the details and this is why i was saying that it is different to knowledge as in the stuff that we know up front or we curate from systems this is something that it
learnt by working with me over time. So while memory feeds context, memory as a term alone does not summarize the entirety of context. And it's very important that you understand that because this lives in a different system compared to the context files that we just built that live in plain markdown.
And both of those live in different systems compared to a database. which would obviously live in something like SQLite or Superbase or Neon, whatever database it is that you choose to use. Anyway, we've now got our outputs from running this onboard skill.
You can see that it's saved two files inside our local folder. We've got our voice and our business, and then we've also got three other folders in here. We've got deliverables, we've got data, and we've got a temp file.
And these are going to be handy later on when we head on over to projects. And then finally, as an example, it will obviously show you what a demo would look like for a response to one of the emails that it found inside your Gmail. And this would test your voice.
to how you're replying to somebody's email. And from there, you pretty much have the absolute basics set up. So coming back to level one, you'll see that we have both context, but then we also had data and I didn't really talk anything about data over here.
It's important that we separate these two because while data does form part of context, there is actually something quite different with data and how we would use it. Data is super important though, because data is everywhere, it is everything. If you're running content engine, let's say you have Instagram and you have YouTube, all of those comments on there.
Those are data points for your business. For instance, when you guys watch one of my videos, I monitor to see how people respond.
I see what they say. I see what they have problems with and what they want to hear so that I know for my next video, I should incorporate those into this. And I did that a lot for this video in itself.
So it's really important to have a system that can harvest all of that data from those platforms and then store it inside our AI operating system that forms a part of the content pipeline that I build. So it is most certainly context because this data will shape the context that I use for my videos. But the format of this is entirely different.
It usually belongs in a database, not those simple little markdown files that we've been working with. It's also important to note that it's not limited to something like YouTube. You have data in every single system.
Like I keep saying, you have discovery call logs that you have perhaps from Fathom or Fireflies, whatever it is that you use. You have offers that live in separate systems. You might have several different databases with several different things inside them.
And one thing that we can do is take all of this data, put it through a little sieve, and put it in a central database so that we have the central location of our most important usable data that shapes our systems later on down the line. much like i do with my content pipeline and youtube so this is certainly something that you might want to consider it's definitely not something that every business needs but in the modern age because we just have mcp connected over here with all of our sas tools like we did in the beginning of the video we can easily go and map these systems and then we can pull out that data into a central database if we want to but i'm not going to be covering that in this video because we would be here forever i have of course already recorded that it will be down in the description below i'm just going to put a whole playlist at this point what i want to do now for most of you guys is head on over to the projects tab this has been completely overhauled and we're going to set up some different projects inside here aligned to a framework that i like to use logically to break down my business so next up we have projects now projects are super useful for those of you new to them you've joined at a really interesting time because there's just been a massive change
For those of you who already have some projects, you're probably going to have to migrate them. A little box would have magically popped up and said, hey, we're going to be doing a bunch of stuff that you may or may not want, but you don't really have a choice for much longer. You will be able to defer any kind of upgrade.
I don't know how long you can defer for, but I was able to do it for a day and then eventually they were just like, nope, we're moving you across. so you can see i've got a bunch of stuff in my archive folder over here these are pretty much useless they just live here there for you to go and grab some stuff out of if you want to they will create copies of these and migrate them across to the new project format but those are mostly just trash anyway so i deleted them we're going to take a look at this project in just a second it's from a live demo that i did last week it's got some important information that i want to go over with you but for those of you starting to build a project for yourself for a business We need to have some kind of framework that helps us design our project properly.
You can do whatever the hell you want. I'm just giving you this framework so that you have some kind of reference to understand how to break things down for a business. So whenever we're working with a business, generally, they're going to have some form of acquisition.
How do we find the right people? How do we find our clients? What are the systems that live inside acquisition?
You would then also have delivery. And this is how do we serve our clients? So how do we give them the product that we built or the information that they need?
Things like that. We then have support. And this is how do we help our customers?
What are the support channels that we have inside our business that help our clients? Then finally, we have operations, and this is for your own business. This is generally the backend stuff, connecting things between teams, connecting things between these different teams in your business, the backend ops, basically.
So by having this structure of grow, serve, help, and run, we have these four pods that we can align our business to. For me, this gives my brain a logical framework of what kind of systems I would want to build in each one. For instance, the systems and processes that I have inside acquisition.
they're going to be vastly different to the things that I might have inside my backend operations. So by breaking this down, I know that creating specific project spaces for them, the systems that live inside there are only for that pod or lane. So we can do that quite easily.
Let's do it for one of them now. We'll start with the acquisition one and I'm going to be focusing on my content pipeline because that's mostly what I do for clients nowadays. The rest is just referrals from my 12 years in consulting.
But moving on, let's just go and make a new project right now. So we're going to call this thing acquisition, if I can spell properly. They have this concept of a goal.
This is entirely optional, but it's pretty good to have a goal because it's going to help the project understand what it is that you're trying to achieve in here. You manage all my client acquisition from sales and lead gen to content creation. Our goal is to attract people into our business to then later close them as clients.
So again, this is totally optional. You can still do everything that you would need to do if you didn't fill this in, but it can be useful. The next thing over here, you'll see this magical word context pop up again, the stuff that we harvested at the beginning of this conversation.
which is why I say it's so important to use that as the foundation because everything you do afterwards gets built up from it. Our options when we hit this little add button over here, we can either add a GitHub repo, we can add a file, we can add a folder, or we can add Google Drive. So anything that we add here will be uploaded to their cloud except for a GitHub repo because that's already in the cloud and obviously the same for Google Drive.
But if you're adding files or folders, what it's going to do here, let's just say we add a folder and say we wanted to add that folder that we built earlier, AIOS test. If I hit upload, that's going to do exactly what it sounds like it is uploading it this is not some local link like we had when we were talking in our normal chat over there it's very important to understand that whatever was in this folder will be uploaded and that's perfectly fine if it's just markdown files telling claude what your business is about who cares but this would be a point where you would maybe want to take a pause and think Hmm, some of the stuff that I'm about to add to this do.
I want Claude to actually keep this in their cloud. For most of you, I absolutely think there's no reason why not, especially if you're running literally everything through this LLM anyway. It already has that information.
For those of you who are in a more regulated environment and things like that, of course, you're on a Teams or Enterprise plan, in which case things might be a little bit different. But for now, I'm not actually going to add anything here because I want to show you how to do this once we get inside a project because there is a new concept that we need to look at.
So I'm just going to create this project. And you'll see that we're now met with this shiny new menu over here. Cloud will immediately start working.
What it's doing is acquainting itself with this project. So if you did upload any context beforehand, it would read it to get an understanding. what they've done here with projects now is split it out into more of an agentic flow you have this concept of a coordinator and threads so the coordinator you can think of it like a project manager it's not really going to be doing any of the work all its role is is to coordinate tasks between the different workers so you the humans set the job this is what we want to do we spoke about the goal when i put it in that project we add some context which are the sources and we give it a definition of done It then hands the task off to this coordinator.
The coordinator says, okay, cool guys, I need a bunch of you to go and work on X, Y, Z. It will then create a thread and it will go and do some research. It would create another thread for any review that it needed to do.
It would create another thread to go and draft a bunch of stuff. These are all separate threads inside their own isolated context window. And all they do once they have completed their work is push that work back up to our project manager.
The project manager then looks at it, sends it back to the human and says, cool brah i think this is done we can carry on living can you please review this at which point you the human need to use your own intelligence and say hmm this looks pretty good or no actually we need to make some changes in which case it would go back to the little worker and get the work done so coming back to this screen over here i want to get our context connected now there are multiple ways that we can do this one of the easiest ways is to come on up to the top over here and you have this library you can see i've already added this folder ai os test this is what we made earlier and inside here we have our context and we have my business and my voice so claude can automatically now use these things always on anywhere in the cloud on any device so that's one of the reasons that they're pushing for this cloud first approach so if you wanted to go and add this you can then hit add on the side here you can see we can upload from a device we could upload our files and folders we can add a google drive folder so anything that you have in google drive you can link automatically here and it would form part of the context or you can link a local folder now it's super important to understand what is happening here firstly none of this is in the docs so i'm just going to go ahead and i'll link aios test one
this linking setting that i'm currently doing it's actually not in the docs at the moment it wasn't included when they originally released projects it was only cloud only and i think that pissed off enough people to the point where they're like hey we actually want to run things locally so you can see here immediately there is a separation it still lives in the library but this one over here is totally cloud native and this one blatantly says that it's running on this computer and it is also the default setting so for those of you who did want computer only projects you can have that by linking a folder over here in terms of claude doing the work obviously it's still going to be in the cloud but at least your local data can live solely on your computer again your computer needs to be on for that for most of you guys this will probably be everything that you need to do for projects if you're just a general user and you just want to get this thing kicked off
Once you've got your context in here and any other files and folders that you need as a part of your project, you can then just start chatting to this thing and say, hey, I'm trying to achieve X, Y, Z. Can you go and do it? It will then coordinate in here and will send out a worker thread.
We'll take a look at how that works in just a second. But I do need to cover the settings because this is very important in understanding how projects work under the hood. So we have this tab called project settings over here.
You can see that at the top, we have the goal that I typed in at the beginning over here. But then further down, we have models. We have the coordinator model.
Again, the coordinator being the project manager, this main chat screen that was behind there. The default is currently Opus 5 .5 and the default ever that they have is set to low. This is obviously going to be task dependent.
I would probably set this to medium or high, but that obviously depends on the type of work that you're doing. Remember, the coordinator itself is not doing the actual work. It is just coordinating.
It is the project manager. The threads are what actually do the complex work behind there. So they have a default of Opus 5 .5.
You can change it if you want to, but to be honest, I would just leave it there. Same for the effort level of medium. I might bring this up to high for more hardcore work.
But realistically, the coordinator will also take care of that. Then coming back to memory. So those project settings that we made earlier, this might be somewhere where you want to put in some project settings.
Like it already knows inside its library that some context exists, but sometimes it's good to give it a little bit of context about the space that it's working in. So you could copy some of the settings from your claw .md inside here for this project specifically. We've already given it a goal, telling it what it's trying to achieve, but we could perhaps say this is for AtomicOps and our consulting business that serves XYZ.
Just something brief over here. As you scroll further down, you'll see auto memory is enabled for this one and you have your memory .md file. Now, again, memory is what Claude learns as a result of working from you or directly asking you itself.
So you can go into here and for this specific project, you can see here four minutes ago, it created this little memory over here and it's just some kind of internal notification for it. where it's setting up remote control for the environment. That's because we added that local folder.
If you wanted to, you could also get it to change or remove specific memories. So if you clicked on here and had a ton of memories in here, if there were things that you wanted to remove, you could just remove those from here. I do have a separate skill that actually caters for this because it's an entire process within itself.
Then we have the ever -important environment tab. This is definitely going to be for more power users out there. The everyday person probably won't need to touch the screen.
By default, you would just be able to use projects like I said before we dived into the settings menu. All of this would have already been taken care of for you.
But for those of you who do want to configure your own environment, you can add GitHub repositories. So specific repos that you have, you can link them so that every single thread that does the work has access to a specific project repo. The other thing that you might want to do is create a cloud environment.
So I've already created a bunch. You'll be familiar with this if you're using the managed agents. But you can always add one over here.
You can call it what you want. You can define the network access that this thing gets. You can give it specific API credentials if it needs to reach out to any systems that don't have MCP or you just prefer taking an API route.
These setup script and environment variables, they're mostly going to be tied to managed agents. I don't see much use for them inside a project. but it's just a part of configuring the cloud environment.
You can then see that we have our connectors. So connectors give Claude access to your apps like we already spoke about and we connected. And it's also why we connected them first.
Again, because if we have this, Claude will then automatically be able to go out into those systems. But if you wanted to, you could click manage and add some more now. You then also have the ability to let Claude use your device.
So again, this is obvious. We want this if we want to use our folders locally, you can tick this box over here. And then pre -approved folders is pretty much doing exactly.
So when I hit the add folder button over here and I said link a local folder, What it was doing is just adding that setting over here. So that's the exact same folder that we added.
And it's currently the default linked pre -approved folder. You can also use get work trees if you want to. You can enable that.
I haven't enabled it because I don't really use them inside my projects. And then folders and files, this just opens up that library that we've already got open over here. so most of you guys you're probably just going to be using this with the default settings and then maybe changing the model it is important to note that you can of course just change the models down here like normal you can also change the effort level and then we have this little usage toggle down at the bottom and it will be split by coordinator and threads so that you can track who is doing what and how much usage they're using so that is an in -depth tour of projects what we need to do now is actually get some skills running and some stuff working inside our project so that we can see how these two things work together but also so that we can ultimately get this work scheduled and all of the other layers inside our AI operating system.
So we've now climbed from the first ladder. We did the second ladder as well. That was all of our connectors.
And we're now at level three where we have our skills and agents that actually do the work for us. So to show off what a skill is, it's probably better to just give you a practical example. Now, we set our goal and we did exactly what we needed to do for our project.
And I could literally come on over here and just say, hey, go and research some content ideas for me. And it would go and it would do that. It's got some context about my business.
So it has a rough idea of the things that I might want. All it would do is then go and dish off a task to one of the workers over here and it would go and do the work that I just asked it to do. What a skill enables us to do though is to give it a more structured approach to doing the work that meets the needs that we want directly.
Think of it like this. If you had a new employee join your business, Would it be better to just say, hey, go and organize all of that stuff over there for me?
Or would it be better to give them a handbook on exactly how you want that stuff ordered and organized and what you want them to present to you at the end of their working day? The second option would obviously get a much better piece of work from them. So that's all we're doing over here.
Instead of just YOLOing our way through Claude going and doing things for us, we're giving it a very specific set of instructions. about how we wanted to go and grab content ideas for us and then ultimately turn that into a content piece or a video idea. So I'm just going to tell this thing.
I want to make a new YouTube video, go and run through my entire YouTube content pipeline skill set. And I'm just going to hit enter on that. And this is now going to run through a whole bunch of skills that I've got specifically related to my YouTube stuff.
It's going to decide how the work gets done over here. You can see automatically that our threads begin. It will give you a list of things that are waiting on the user.
So that's stuff that requires human in the loop. It will tell you anything that is currently working and then anything that is finished. this list of tasks over here it will exist for a very very long time each of these little workers you can see that we can come in here we can control them we can change the model we can steer the thread any which way that we want to and talk to it ourselves if we want to take over the coordination layer here for this specific thread these threads in themselves they do run in their own isolated context window meaning that the work they're doing here is not by default visible to say another working thread.
You can go into a thread and say hey one of the other threads is working on xyz can you go and look at that and grab some of its context. So while they do by default work in isolation you can view things from them. The reason that it's done that way is because obviously context isolation means that one we don't muddy up the waters of some of the work that the separate threads are working on but also it's a lot more token efficient because all of the work that they do in here they only send back what the coordinator needs to hear in order to present us with information instead of everyone sending back everything into this single window over here that not only would we have to sift through but also this thing would have to go through tons and tons of tokens and all the work that these little threads have done instead all it gets is a result and that's all that really matters in terms of how long this conversation lasts This conversation lasts pretty much forever.
You will always be able to scroll through here. At least that's what the docs are currently pointing to. But that doesn't mean that this thing will remember everything forever.
Just like we have with other chats and different parts of using AI, it needs to compact its context over time. So it will carry on working and working and working, but eventually it will compact its context, meaning it's going to take specific things that it thinks is very important based on how it's been working in here and create little checkpoints of knowledge that it has so that it can keep working later on in this conversation.
Something that's very important for you to understand at this point is that if you opted for a cloud only context, let's say you only added a folder here and you uploaded it to cloud, you didn't link it. That means that anything that was reliant on your local computer inside these little threads that we have working, it wouldn't have access to that.
So practically speaking, if I had a database with my competitors, and some of my old news and perhaps seo signals if they lived in a local database in one of my folders on my computer if i was using a cloud only folder this thing would probably break and say hey i don't have access to your local database so i can't actually complete this task so that's something that you need to factor in for your planning with yourself and your clients how are they going to be using this if they need some kind of database that currently you're using locally then you're probably going to be wanting to push that across to something like superbase because obviously that is always on in the cloud alternatively if your device was always going to be on then you could do what we did earlier and just link that local folder as well inside your library over here and then it would be able to use both whenever it needed to you can see so far it's built as a youtube folder because now it's starting to complete tasks and it needs to deliver the files that we need in order to do the tasks so you can see here it started to put in markdown files which is which is just the type of file that agents like working with and you can see here moments later this thing is done so it's created what it thinks is going to be my next best video
with a little bit of instruction behind it, and then you can see the assets that it created, all the markdown files that we saw living inside the library. It then went back and it told the coordinator, hey, I'm done. The coordinator comes and tells us.
But as you can see, we have control over both of these environments. You might be asking yourself at this point, okay, Mansell, but how the hell do I know what skills my business needs? And again, that ties in with this four -part framework that I come with.
If I look at acquisition over here, I know what systems run inside sales. You know what you're aware of, what your business is using.
Maybe you're doing outbound with cold email. Maybe you're doing LinkedIn. Maybe you're doing YouTube.
Whatever it is, inside there you have manual processes. You have stuff that you are doing every single day by yourself or some human is doing with their hands on. So that is exactly what gets turned into a skill.
Instead of doing all of that stuff manually, you're just outsourcing all of those clickety clicks into something that Claude is going to repeat for you. In terms of how you build those skills, There are multiple ways that we can do that.
Okay, so when we're looking at trying to get skills in our business, you don't just go out there into some random repo and pull down a bunch of shit that somebody told you your business absolutely must have. That's one of the worst things to do because a lot of those repos are just filled with prompt injection and people trying to take your stuff.
More importantly though, you don't want to fill everything with clutter and things that you're never going to use. It's a really big problem and it can just screw with your entire interface and your system that you've built. So what we need to do is, again, start with a grounded approach.
We're going to look at a skill that I told you to grab as well called PodMapper. So I've already run this thing in the background. We're going to take a look at the output that it does.
But essentially what this thing is going to do is it's going to audit you. You can see that over here, interactive business workflow audit, map any business function into a pod, which is the four pods we already spoke about, with clear workflows, tools, data sources, automation opportunities, and technical translation. I'm not going to go through this entire thing, but essentially it's going to sit down with you.
talk about a specific lane whatever you think your most painful is in this example that we've been running with throughout our entire process we've been talking about content because for me that is my absolute pain so in doing so i pick my engine for this case it would be acquisition because that's where content lives and then it will ask me some questions then inside phase two it will map my workflow and walk me through every single step that i take inside my content creation journey it needs to do this so that it can understand everything that is involved in that process and the more information and context that i give it the better the output of the skill is that it's ultimately going to create for me.
So it's very important not to rush through this kind of thing. It can seem really appealing to do that because AI push one, two buttons, job done. But realistically, like I keep saying, you're going to get a much better result if you focus on doing the groundwork, the foundations upfront really well.
All the work you do from here is a lot simpler and more accurate. So if we come back to the screen, you can see this little mock journey that I put together here. It started by interviewing the person.
It asks them some questions. We get a business snapshot delivered to us. pretty much what's going on with this specific lane.
It's chosen acquisition for this mock data that it ran through because that's where the user identified the most pain. When we scroll down to phase two, the workflow over here, you can see that it's literally what is the trigger, what happens next, what happens after that, what happens after that. And it does that until there are literally no more steps in the part of the process.
In doing that, it identifies all of the systems that the user is currently linking with. You can see they've got Calendly, they've got LinkedIn, Heyreach, Google Docs, all of these other things. So if anything wasn't connected via MCP already, this thing would be able to say, hey, you know, it's looking like you've got most of these things connected, but there are a few that you still need to connect in order to turn this into a working system that we can run inside Cowork.
It's also very easy for you, the user, to understand what your journey looks like on paper because really you might be walking through this and think to yourself, well, I don't actually need a bunch of this stuff anymore so I can cut this and then we won't have to build such a long skill. Then it's going to tell you some of the context files that you're going to need as a part of this system.
So it identifies here, You need to have your ICP, your voice, your offers, and your GTM profile. It would help you build all of this, of course.
That's the whole point of this. And again, it would stash it in that context folder where we put all of our other things. Then what it's going to do is it's going to analyze your workflow for you and say, well, you know, it seems like you're doing a lot of stuff that isn't actually providing a lot of value.
We could cut a bunch of stuff over here. And then you could work with the AI to cut the parts that don't matter. Until ultimately, cut a long story short, you end up with a perfectly clear automation map.
of what the AI can own in its entirety, what it can assist you with, and what should be kept as a human -only part of this process. So that's pod mapping in a nutshell, but it is the easiest way for you to understand the best ways to write your skills. But not only that, as soon as you agree with whatever the plan is that you guys worked on, you would then just say to this thing, cool.
go and create the skills that we need for this and then it would just use the skill creator to go and do it so for most of you guys this is going to be the most accurate way to start but now there are actually other ways that we can build skills one of them you could just come over here and you'll see we have record a skill what this thing is going to do is it's going to record your screen every action that you take will be part of the process that it outlines for the skill that it's going to build for you so if you were doing something let's just say doing lead gen research on LinkedIn.
You could show your daily process of whatever it is that you do. It would take notes of that and build a skill from that. For me, I think it's mostly unnecessary.
I know the interview process can take a bit longer, but it's far more accurate and it can help you build step by step. It's the same way as having a really smart consultant who knows the business inside out. helping you build that skill as opposed to a few clickety clicks.
The third thing that you could do if we had a really long conversation going on inside a project or even a previous chat, something you could do is literally come down here and say, cool, this thing that we just worked on for the last 30 minutes, go and turn it into a skill for me. Because if you know that you're going to have to do it again and again.
There's no point in just keeping it in a conversation and coming back to that conversation every time just to run it all over again. Next up, we need to look at this little customized tab that I keep dipping in and out of. There is a lot going on over here.
You can see we've got our skills, we've got our connectors and we've got our plugins. Then there are these two little toggles over here. We have yours and we have discover.
yours pretty self -explanatory this is everything that either you created or you have uploaded to your environment like i mentioned earlier anything that you add by uploading a skill if it's on the desktop app or the cloud it will be synced between the two environments so you'll have those things that you need you can then also create a skill over here which then just takes you to a menu where you can write this manually but we're absolutely not going to be taking that option you can also create with claude where it just opens you straight into a chat and says let's use the skill creator and work together tell me what what skill I should do first.
This is a really mediocre way of doing my pod mapping version. And then they also have record a screen over here. Something else you can do, they have this marketplace of skills under the discover tab.
And you can see other people that have made skills. You can go and try them. You can click on them and see what's inside them.
All of the skills inside here, they have been security vetted. That's one of the benefits of having these in here. It's a lot better than going into someone's dodgy repo out there in the wild and pulling things in.
Because every time somebody wants to upload a skill, it goes through a security. security scan, which you can see just by clicking upload over here. After you hit upload, it will run that scan just to check for any prompt injection or things like that.
But like I said, make sure any skill that you do add, it comes from a position of pain, somewhere where you realize you have a problem and you think, I wish I had a skill for that. Don't just do it because it sounds cool to have an agent force ADLC inside your business if it has absolutely no business value. And then next up over here we have plugins.
And plugins are really just a group of skills packaged together to solve a specific business problem. For instance, The YouTube content thing that we ran through earlier in our projects demo, that entire bundle has about seven or nine skills in there that walks through my entire YouTube content pipeline.
So instead of having seven zip files that I need to share with someone, I can bundle all of that up into a simple plugin zip, which I can then distribute to whoever I want, whenever I want in several different ways. And again, inside the plugin marketplace, we can find tons of things like that. We can have plugins that I have uploaded, just the same as skills.
And we have ones that I can discover from the marketplace, usually for finance, small businesses, things like that. There are a lot of plugins that are really valuable for people who do have a small business and don't care about setting any of this up for themselves. The Claude for Small Business plugin is actually pretty awesome.
If we click on one of these plugins, you can see what's inside them before you do anything with them. You get an overview, shows you the contents of what's inside this folder and everything in there, all of the skills that are going to be inside there, and then you can dip down into what's inside those skills. So if you absolutely wanted to, you could come and read all of this.
You can separate the skills over here and grab specific things from them. And while we're here, I just want to give you an overview of what actually lives inside a skill's contents. So this data map skill that we're going to take a look at in just a second, I ran the output secretly while I was recording this because it takes quite a while.
But the point is over here, we have our skill .md and this is just that standard operating procedure i was speaking about earlier it is that instruction manual it tells the ai the exact instructions that it should take in order to do a very specific task all in plain english written with a few clauses and things like that you can see here this one does call some local stuff that actually lives on my computer so in this case for this skill i would need it linked to my local computer but obviously when you're building this out you would understand that yourself the important part here though is that we have these different folders that all make this thing actually work we have assets we have references and we have scripts now this information over here is probabilistic meaning that we are feeding this to ai and if all we did was give it these plain english instructions it would go and do this task every time but every time there would be subtle differences with that task even if it was right
Now, in business, you generally don't want that kind of thing. You want the exact same thing done in the exact same way every single time. So in order to anchor that into something, we need to make sure that we give it something programmatic because scripts or code is deterministic, meaning every single time it runs.
It's going to do the exact same thing. That's how you use your actual computer's operating system like Windows or Mac. Imagine it was probabilistic.
One day you click on the start menu and it opens up my computer. The other day it opens up one of your favorite games. That would be absolutely a terrible user experience.
And in the same way, we can think of that with AI. We want to make sure that whenever I run this data map that scours through all of my connected systems looking for problems in my business. that it gives me the same result.
It finds those problems and presents it to me in a way that is meaningful. I don't want it to go through there one day, find two things and then automatically build me some really shitty looking data brief that makes no sense to a human. So that is what we are trying to achieve with skills over here.
That we understand what operational excellence looks like for this specific task and that every time we run it, we're going to get the perfect example. But these two things aren't the only things that we need for our skill to do that. One thing that really helps AI is giving it some form of references.
Yes, we want this thing to be able to make its own decisions and do what it needs to do. But that doesn't mean that we can't anchor it by showing it what good looks like. So for instance, here, as a part of this data mapping skill, what I get it to do is go out into these systems and find those business problems.
But one thing that I definitely want is a beautiful looking draft that a human can read easily. So what do I do? I give it a reference of what that looks like, a summary table.
pattern and I give it exact constraints about how this information should be presented to the human in a way that is really meaningful for them so that they understand it clearly. So things like that that sit alongside the skill that really just make it that one step better at giving the human exactly what they want. Now in terms of adding plugins it's pretty much the same as skills.
You can upload the plugin via a zip file if you want to or you can create a plugin from a bunch of skills that you have already put together. But when it comes to plugins, they have a very different concept called a marketplace. Now for you guys, if you just switched on Cowork, you might not already have Anthropix marketplaces connected.
So the screen behind here might look a bit bare. One of the first things I recommend you do is head on over to browse Anthropix sources and you can just add ones that are relevant to whatever the type of work is that you're doing. The Anthropix directory is probably one of the best ones to add.
Something else you can do, though, is you can add a repository. So they have this functionality for those of you who might want to distribute a whole set of plugins to any of your clients. So for instance, in the same sense that Anthropic has this marketplace where everyone who uses Cowork can just grab all of these skills, you could build one specifically for your clients that only they can access, filled with all of the plugins that only their business needs.
And then you would just add it from a repo and connect it to those team members' Cowork sessions. That way you can edit and upload skills and keep them in sync and they can't interfere with any of the work that you're doing and they get trusted skills from you, the person in charge of managing their AI platform. This is the exact same process I take with my community.
Anyone who joins the community, they get access to all the skills that I build in real time that's synced. via our GitHub repository. But at a high level, that pretty much wraps up skills.
So I mean, after you've done the whole pod mapping thing and added your skills, of course, there's gonna be some level of testing. You can't just run this thing and expect on the first try that everything's gonna be absolutely perfect. The whole point of this is iteration and failure until you get to a point where this is extremely reliable, more so than just having an agent YOLO its way through whatever the task is that it needs to do.
It's all about efficiency over time here. If you're just gonna be running something once, Probably you don't need to build a skill for it.
Just get the agent to go and figure it out. But if you know something is going to be repeatable, it absolutely makes sense to turn it into a skill. And this neatly segues us into two separate things.
One where I showcase what a skill looks like, that it's been grounded, but also where I slowly lead us into artifacts. So on the left over here, this is a skill that I started at the beginning of this video is the one that I just mentioned and walk through with you. And its whole point here is to go out into all of our systems, gather all the information that it can from everywhere to figure out where we're going wrong inside those systems.
So it reads emails and sees how long it takes us to reply to someone. That could be a potential gap in why we might be losing clients because we don't get back to them fast enough. And it does that for each of the four pods in the business across acquisition, support, ops, and delivery, so that we can understand all of the different areas of our business that are failing, why they are failing, and what we can do about it.
Then what I get it to do, just like I showed you in that little summary table, is present it in a meaningful way that the human understands. And for me, that is telling it, go and make us an artifact inside Claude. And this is one part of what an artifact really is.
It's just an HTML file that Claude can edit within its own app. and it can create pretty things for us. It can do a bunch of other stuff as well with Word documents now and slides and all sorts of things.
But just know that it's really handy because you can present information about your own business in a really meaningful way like we're doing over here. So you can see up at the top, we have the verdict of our skill run and everything that it found. This is all mock data, by the way.
Of course, I'm not going to reveal my own business problems. But all it's doing is telling the user exactly how much money they're losing and why they are losing it. And it goes through the entire audit of the business, all the systems that it went into, what it found in each system, and then delivers it in a way that people can understand.
It then gives them actions that they can take so that they can act responsibly responsibly based on actual data points instead of a bunch of things that they don't really understand about their business. So could an agent do this without you providing it a skill?
Absolutely. You could just tell it once you've connected to those connectors. Say, hey, go and look at all my systems, figure out what's wrong with my business, and then tell me how I can make it better.
That would totally work. Would it give you something as great as this in the exact way that you wanted to see it with the level of detail that's already been tested multiple times? Probably not.
That's another reason why skills are great is because you run them once. Did it give you the output that was good or not? Could have been better by doing X, Y, Z.
That's feedback. Run again. Could have been better.
Go do ABC. Iteration through failure is how systems get better, which is why this whole building skills thing has like five or six videos on my channel because it is such an intricate process to get to this point. The reason why I don't harp on about it in this video is not just because we'd be here all day, but also because some of those concepts are quite technical and really for most business users, you can create a skill as easily as talking to this thing.
But I digress. We need to move on to our next layer. We have now covered skills and agents and we're moving on to interface.
Interface is pretty simple. This thing that you're using right over here is the interface. But it is not the only interface.
This is using the desktop app. We can easily head on over to... our little browser over here and we could just type in claude and this is another interface but this is a cloud -based interface if you don't have the desktop app installed you will see here immediately that we have been synced so our acquisition project that we created locally on the desktop app is here along with our worker thread over here what you will see as well is our library has synced so everything that it built the youtube stuff that it output is all in here the global folder that we set up And then you can see over here, because my computer is still switched on, obviously, we can see these two folders that we gave this thing access to.
If I had to close this computer and say I was working on another one or working on my cell phone, we wouldn't be able to access any of the stuff on here. So only the rest of these things would remain. In terms of what's on my phone, you can see that we've got the acquisition.
We've got the work that ran in here. We have the video outputs and everything that was packaged up. One thing that we do not have when we hit those three little bars.
is under threads we have library and you can see that those two linked folders that were on my desktop are not there i don't know at the time of me recording this if that functionality for those linked folders is coming to the phone so that you can use it there i'm not sure if they're going to do that again most of these features literally just came out and aren't even documented yet But I would suspect that because they're trying to make this accessible for every single device that we would ultimately get there.
Otherwise, that's the whole point of these interfaces. This always on is meant to be there so that when you're on the go, you can do whatever you need to as long as you put it in the always available space, which is our library. If I wanted to use my phone at this point, I could absolutely go and kick off a new task and it would be able to do whatever it needed to do inside this project as well.
So for me, that's a really handy feature. I don't have to deal with VPS. I don't have to set up VPNs with tunnels.
to something that I've got stashed in my own computer. But obviously that depends on how your business works, what your data privacy, data governance, and how all those sorts of things tie into your business. Then we get onto our final layer, which is the runtime and the ops.
And really what we're saying here is scheduled tasks and some kind of monitoring to see what people are doing in our environment. Now, things are still being shifted around in this little landscape over here. But if we head on over to scheduled tasks, you'll see they give you some things automatically out the box.
These are some of the most important things that most people do whenever they are running a business for instance you have daily briefing inbox triage basic stuff like that so for some quick wins you could definitely just click on that claude will run through it and say okay cool your mcp is connected all of this stuff is done if it has any questions it will ask you otherwise it will literally just go figure out your local time zone and set up a scheduled task for you that will run every morning at i'm guessing 8 a .m yeah 8 a .m so that's one way to do it the other way that you can do it is that you can come on over to new task and you can set one up manually so this again you just type in the name of this thing give it some instructions you can call out skills if you need to set a frequency how often this thing runs give it permissions so does it need approval or does it have auto where it runs without asking but with some certain limits attached or does it need to ask you things in
i think setting it manually is pretty silly because then it's no longer a scheduled task you can also set up notifications for when this thing has finished so you can get push email or slack and you can also require this computer to be on for the specific task that it is running something to note though the time of me recording this it is entirely different to the way that you would work with projects for instance if we come over here it says work in a project or a folder but you'll notice it does not have my projects i think they're still merging this over into here because if we come back to our projects over here you can see that there was one more little tab that we never clicked on inside here that was our clock and that blatantly just says ask claude in a thread to put recurring work on a schedule like a morning digest or a weekly report so that would be another way to do it and honestly it's probably going to be one of the best ways to do it because while you're working on specific things when you know that the thing is built
just get it scheduled then and there don't bother going to do this right at the end get things done while you're doing them if you know that you're already going to be using them later on something else to note though is that these are called routines and routines are actually something a little bit different to just a standard scheduled task i suspect that everything is going to be merging over to this routines thing but they actually belong in claude code if we look over here routines are pretty much the exact same thing they just called it this because they can do other things So for instance, in Cloud Code, you can still run local and cloud.
I'm going to hit cloud because that's clearly where they are heading inside the co -work tab we were working on. But for the most part, it's exactly the same thing. We just have more options.
Apart from scheduling a task, we can also respond to a GitHub event. So if someone does something in GitHub repo, we can respond to that. But we can also trigger from our own code or sending some kind of request via API.
Annoyingly, again, at this time, we still don't have native webhooks, which is ridiculous. A webhook is essentially when one system creates an action and then sends that event trigger to Claude. Claude would then be able to respond and go do a bunch of stuff.
We don't have that natively. So one of the ways to do that right now is with NNN or to create something locally and then always leave your computer on. But considering Grokbot and ChatGPT can do both of these things natively, I think it's kind of silly that we can't do that yet.
But I'm sure it will be coming soon. So pay attention for that. Heading back to the scheduled tasks, you can see here that it created that one automatically with Claude.
If we come on down here, you can see it filled in the instructions. So if we check the instructions, you can see it will run my morning brief using the skill that I have. for my morning brief.
It will then write it in English and my time zone is New York, America. It will look at my calendar and specific things in there, pull some things out of there. This is an unattended scheduled run.
Do not ask questions, don't suggest connectors, just gather information. It runs at 7 .55 so that everything is delivered to me at 8 a .m. So I hit off a test run over here by clicking run now.
You can see here that you get your history that will be on the left -hand side. When you click here, it goes through to the skill. You can see the output that it gives you over here, your morning brief.
Super simple and easy to understand what your day is going to be looking like. As mentioned, by default, this will be running in the cloud. It doesn't need your computer for anything unless you've ticked that little box inside the new task that you created.
So that's it for scheduled tasks. And then really all we're left with is monitoring, which unfortunately, if you're just on the pro consumer plan, you don't really have much options. If we head on over to our usage.
This is pretty much all we get. So building a custom dashboard, if you're really interested in understanding your metrics and the things that are running and how often they run, until Anthropic give us some kind of proper monitoring dashboard, it's vital that you actually build one for yourself if you're serious about how you use Cowork and how your environment runs.
But I do think that it will be coming. It's already on the teams and the enterprise plans. It gives them a little bit more functionality and understanding who is running what, how often they run and things like that.
ChatGPT obviously has that baked in as well. So in terms of who's trying to win on which platform, it's very interesting at the moment because they're both making moves that the other one could have done better and they're all doing it in their own way. But I think ultimately everyone is going to get to the same place a year from now.
It really just absolutely will not matter what you are using. But other than that, thanks for sticking around so long, guys. If you do have any questions, leave them down in the comments below.
You will have all the resources that you need. The things that I spoke about in this video, they'll all be stashed down there as well. If you do need more help, if you're a business owner or a consultant, I have a community that caters exactly for you guys to help you get from wherever you are now to a system that actually works and landing your first client.
So if that's something you wanted to explore, the link is down below as well. Thanks very much for watching, guys. I'll see you in the next one.
The Hook

The bait, then the rug-pull.

The opening claim is blunt: Cowork moved to the cloud this week, and every setup built before that is built for a product that no longer exists. What follows is less a feature tour than a rebuild manual, climbing a five-rung ladder from markdown context files to scheduled routines, with the host stopping at each rung to show what changed and where it bites.

Frameworks

Named ideas worth stealing.

14:03model

The Five-Layer AIOS

  1. Data + Context
  2. Connections
  3. Skills + Agents
  4. Interface
  5. Runtime + Ops

Build bottom-up: own your data and context first, get connections approved second, then give skills and agents methods and decisions, keep humans directing and approving at the interface, and let the runtime run and recover. Loose coupling between layers keeps each swappable, including the model.

Steal forany AI systems setup, client onboarding doc, or course outline
05:19model

Knowledge, State, Memory

  1. Knowledge: curated, stable facts in markdown
  2. State: live, changing records in a database
  3. Memory: lessons the AI learns from working with you

Three kinds of context that need three different homes. Knowledge rarely changes and gets a human check first. State changes constantly and needs real-time updates. Memory is captured from experience and lives in Claude's own memory system.

Steal fordeciding what goes in CLAUDE.md versus a database versus auto-memory
15:04concept

Data Is Not Context

Data is everything available. Context is the smallest useful slice for this job, right now, that the agent is allowed to see. More is not better; the right slice is.

Steal forscoping what to load into any agent prompt
17:52list

Grow, Serve, Help, Run

  1. Acquisition (grow)
  2. Delivery (serve)
  3. Support (help)
  4. Operations (run)

Four pods every business maps onto. Each pod becomes its own project, and the systems inside each pod are different enough to warrant separate skills and context.

Steal fororganizing projects, skills and automation audits for any business
34:41model

Pod Mapper audit

  1. Pick your engine (highest pain pod)
  2. Map the workflow trigger by trigger
  3. Identify tools, data sources and missing connectors
  4. List required context files
  5. Cut steps that add no value
  6. Produce an automation map: AI owns, AI assists, human only

An interview-driven workflow audit that produces the plan a skill creator can build from, instead of pulling random skills from repos.

Steal forany process-automation discovery call
39:45list

Three ways to build a skill

  1. Interview via pod mapping, then skill creator
  2. Record a screen session of the manual process
  3. Convert a long chat you will repeat into a skill

The interview route is slower but most accurate. Screen recording is quick but shallow. Converting a conversation is the right move when you notice you keep coming back to it.

Steal fordeciding when a workflow is worth packaging
CTA Breakdown

How they asked for the click.

VERBAL ASK
55:36product
“If you're a business owner or a consultant, I have a community that caters exactly for you guys to help you get from wherever you are now to a system that actually works and landing your first client.”

Soft, placed in the last 30 seconds after the value is delivered. Free skills are given away in the description first, and the paid community is framed as the next step for consultants rather than pushed mid-video.

MENTIONED ON CAMERA
05:38linkOnboard skill (free download) ↗
05:38linkPod Mapper skill (free download) ↗
21:50toolSupabase ↗
21:50toolNeon ↗
54:00tooln8n (webhook workaround) ↗
55:40productAI Native community (Skool) ↗
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Storyboard

Visual structure at a glance.

open
hookopen00:01
what goes up, what stays
promisewhat goes up, what stays01:56
knowledge, state, memory
valueknowledge, state, memory05:19
the five-layer AIOS
valuethe five-layer AIOS14:03
data is not context
valuedata is not context15:04
new project
valuenew project19:58
coordinator and threads
valuecoordinator and threads28:02
pod mapper
valuepod mapper34:41
business rescue artifact
valuebusiness rescue artifact45:55
scheduled task
valuescheduled task51:30
morning brief
valuemorning brief54:24
community CTA
ctacommunity CTA55:36
Frame Gallery

Visual moments.

open00:01Frame at 01:01 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch01:01what goes up, what stays01:56Frame at 02:27 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch02:27Frame at 03:22 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch03:22Frame at 03:51 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch03:51
open00:01Frame at 01:01 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch01:01what goes up, what stays01:56Frame at 02:27 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch02:27Frame at 03:22 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch03:22Frame at 03:51 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch03:51Frame at 04:33 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch04:33knowledge, state, memory05:19Frame at 05:57 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch05:57Frame at 06:50 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch06:50Frame at 07:21 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch07:21Frame at 08:03 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch08:03Frame at 08:45 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch08:45Frame at 09:27 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch09:27Frame at 10:09 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch10:09Frame at 10:51 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch10:51Frame at 11:33 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch11:33Frame at 12:15 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch12:15Frame at 12:57 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch12:57Frame at 13:39 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch13:39the five-layer AIOS14:03data is not context15:04Frame at 15:46 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch15:46Frame at 16:28 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch16:28Frame at 17:10 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch17:10Frame at 17:52 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch17:52Frame at 18:22 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch18:22Frame at 19:16 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch19:16new project19:58Frame at 20:43 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch20:43Frame at 21:22 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch21:22Frame at 21:45 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch21:45Frame at 22:46 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch22:46Frame at 23:28 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch23:28Frame at 24:10 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch24:10Frame at 24:52 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch24:52Frame at 25:34 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch25:34Frame at 26:16 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch26:16Frame at 26:58 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch26:58Frame at 27:54 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch27:54coordinator and threads28:02Frame at 29:05 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch29:05Frame at 29:47 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch29:47Frame at 30:29 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch30:29Frame at 31:11 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch31:11Frame at 31:53 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch31:53Frame at 32:35 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch32:35Frame at 33:17 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch33:17Frame at 33:59 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch33:59pod mapper34:41Frame at 35:23 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch35:23Frame at 36:05 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch36:05Frame at 36:47 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch36:47Frame at 37:29 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch37:29Frame at 38:11 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch38:11Frame at 38:53 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch38:53Frame at 39:35 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch39:35Frame at 40:17 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch40:17Frame at 40:59 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch40:59Frame at 41:41 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch41:41Frame at 42:24 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch42:24Frame at 43:06 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch43:06Frame at 43:48 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch43:48Frame at 44:30 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch44:30Frame at 45:12 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch45:12business rescue artifact45:54Frame at 46:36 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch46:36Frame at 47:18 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch47:18Frame at 48:00 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch48:00Frame at 48:28 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch48:28Frame at 49:24 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch49:24Frame at 50:06 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch50:06Frame at 50:31 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch50:31scheduled task51:30Frame at 52:12 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch52:12Frame at 52:54 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch52:54Frame at 53:36 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch53:36morning brief54:24Frame at 55:00 from Claude Just Changed the Rules: Rebuild Your AIOS From Scratch55:00community CTA55:36
One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.