Latest / In The Loop / Claude Cowork: How To Use It & Why It Matters

Transcript

Claude Cowork: How To Use It & Why It Matters · In The Loop

0:00
Claude Code has gone viral, one of the fastest growing apps of
0:04
all time, reaching over a billion dollars in run rate
0:07
revenue within just six months of its launch.
0:10
It's so popular even non developers are now using it for
0:13
things that Anthropic never ever anticipated.
0:17
I'm definitely one of those. I've started using Claude Code
0:20
in Visual Studio, but just a programming interface for tasks
0:24
I never expected, including even creating documents.
0:29
But for most people, Claude Code still feels very technical.
0:33
Going into a terminal, running commands, staring at a dark
0:36
screen, it's a it's a big barrier.
0:39
And so onthropics team who built Claude Code has now built Claude
0:44
Co work. So it's the same agent wrapped
0:48
in an interface that everybody can use, your head of marketing,
0:51
your operations team, probably even your dad could use it.
0:55
Today I want to breakdown exactly what Claude Co work
0:58
actually is, explain to you how it works and why I think this
1:02
might be one of the biggest product launches in AI for many
1:06
years and actually what it signals about where the entire
1:10
industry is heading. This is in the loop with Jack
1:13
on. I hope you enjoy the show.
1:27
So let's start with what Quad Co work actually is.
1:30
Quad Code has been named one of the worst name products in AI
1:33
because the Co part just actually just distracts from
1:38
what the tool actually does and represents.
1:41
It was never just about coding. It's a, it's a general purpose
1:45
agent that just happens to be really good at coding, but it
1:48
can do a lot, lot more Co work isn't just this dumbed down
1:52
version of clawed code. It's literally the same agent,
1:54
just with a completely new interface.
1:56
Mindset itself has taken a huge amount of inspiration from
2:00
clawed code, the way it does its reasoning, the way it structures
2:05
its tools and systems that it can access.
2:07
So being able to provide this type of technology just with a
2:10
new interface for everyone is going to be incredibly powerful
2:13
because ultimately you download the desktop app, there's going
2:17
to be a few different tabs when you access clod, there's chat,
2:21
there's Co work, and then there's code.
2:23
Obviously the code is clock code and Co work is Clod Co work.
2:27
Now on the face of it, it's like, OK, cool, what does it
2:30
actually do? Well, it can do a lot of things
2:33
that regular clod cannot do. And Clod code can do so it can
2:38
interact with files on your computer, it can control your
2:41
browser. It can connect to external
2:43
systems, as you know clod does, but doing all of that in one
2:46
place is where it becomes very powerful.
2:49
In particular, I'll talk about why interacting with files is so
2:53
damn important. So for example, instead of just
2:55
chatting back and forth and then creating a artifact.
2:59
So if anyone's used clod, it's an artifact.
3:01
The equivalent on ChatGPT is a canvas where you know, text
3:05
appears or whatever you're creating appears.
3:07
So it's where the chat just moves to the left and creates
3:10
what looks like a kind of document on the right hand side
3:13
that you can interact with. And a lot of companies have
3:16
spent a lot of time trying to define what this is going to be.
3:18
And I think it's really interesting here, and I'll come
3:20
on to this later, that actually perhaps just the winner of the
3:24
new artifact or asset that AI creates and then you share with
3:28
others is not a artifact or a canvas or a whatever thing that
3:33
people are naming it. It could just be a local file
3:35
that's downloaded straight onto your desktop.
3:37
So yeah, called Co work and then open those files on your
3:40
desktop, rename them, create spreadsheets, navigate websites.
3:44
It could send an e-mail and you might go, OK, cool.
3:47
That's what a lot of these tools seem to do.
3:49
But there are big differences. Now I'll come onto exactly what
3:52
those are and why it matters later.
3:54
So to help understand why this is also important, I'm going to
3:57
first start with how it actually works, like how you could use
4:00
this, what you'd use it for. Because really one of the things
4:04
I'd love for people to do is go and test this and really use it.
4:07
It's not, you know, like a new browser that's kind of
4:09
interesting, but probably unlikely to make you change.
4:12
This is actually something really, really powerful if you
4:16
can systematize it and get it across all your company.
4:36
So when you open Co work, you're starting I guess with a blank
4:39
slate. You know, it can't see any
4:41
files, it can't access the browser, it can't do anything.
4:44
You have to grant access explicitly.
4:46
So for files, you can give it access to specific folders, just
4:49
the ones that you want to work with.
4:51
For the browser, you can obviously you can install a
4:53
Chrome extension that lets you see and interrupt with any web
4:56
app for external services like Slack or you know, whatever
5:00
system that you're using, you can just connect it through
5:03
MCPS. So it's just like the cloud
5:05
interface today. And anytime it wants to do
5:07
something, so change a file, interact with something, it will
5:10
come and ask you first and you can always say no.
5:14
So you can stay in a lot of control.
5:16
So I'll give you a, the most simple example, which doesn't
5:19
really make it seem like super exciting, but then I'm going to
5:21
build on that a little bit. So you might have a shed load of
5:25
receipts in your folder because you know, you've been travelling
5:27
with work, you've been doing stuff and you've taken
5:31
screenshots, you've got it on your PC or your phone or
5:33
whatever it is. And essentially it's just a a
5:35
bunch of files with very unhelpful names like receipt,
5:37
one dot, PDF, March, GPG, you know, whatever it might be.
5:42
You can then give Claude access to that folder, giving Co work
5:47
access just to that specific location where everything's
5:50
stored. And now you can ask, I have a
5:53
receipts folder. Can you just rename all those
5:55
files to say match the dates on the receipts?
5:58
And it's super simple, but actually quite helpful for
6:01
anyone that's up to, to do invoicing and, and get paid by
6:04
their company for the things that they've already bought at
6:06
work. But anyway, so, so the agent can
6:10
just open that folder, look at each receipt, read all the
6:12
content and information inside it to find the correct date, and
6:15
it just then renames the files on your local device on your
6:19
laptop. If one of the dates might be
6:20
missing a receipt, instead of guessing or doing something kind
6:22
of weird, it stops and asks, you know, this receipt doesn't have
6:26
a date. What should I do?
6:27
And it's because Anthropic, they've spent a lot of time on
6:29
trying to get the model to ask for clarification or this agent
6:33
to ask for clarification when it's uncertain rather than
6:36
making assumptions that are just often crap.
6:40
And it sounds like a small thing, but it makes a massive
6:42
difference in practice. Now, you could then go further
6:45
with this receipt example. You could then say, can you put
6:47
this into a spreadsheet? And it will be able to actually
6:49
do that. It creates an Excel file with
6:51
all the receipt data, dates, amounts, whatever was on there
6:54
organize into columns just like that.
6:56
And then maybe you don't want this local file just sitting on
6:58
your computer. Maybe you want it in Google
7:01
Sheets so you can share it with your team and then everyone can
7:04
access it from anywhere. So you say actually go make this
7:06
a Google Sheets instead, and the agent opens the browser and can
7:10
navigate to Google Sheets, create a new spreadsheet and
7:13
start populating it with the information.
7:16
You can watch it type around, control the interface, and now
7:19
it's not quick, it's not instant, but you can watch it
7:22
working through every single step.
7:23
And at one point it might notice it didn't do something correct,
7:27
so it has to go back and it fixes it.
7:29
So it's checking its own work as it goes.
7:31
There's been a lot of progress made on computer use in the last
7:34
kind of 6 to 12 months. And then you say now, OK, go and
7:37
take this and send it to this person, let's say Amy, and the
7:42
agent will be able to open Gmail, Click to compose or
7:46
create new e-mail, type in Amy in the receipts field and
7:49
actually pull from contacts. Now an objection many people say
7:53
is that this is just very slow. It's cool, but very slow.
7:56
I could do all of it myself. And this is where cloud cowork
7:58
gets a massive amount of insiration from cloud code.
8:02
Because yes, you could do something like that for a single
8:04
task faster. And you're definitely faster
8:06
than most agents right now for doing tasks on the Internet,
8:10
let's say. But the real shift is doing
8:11
things in parallel because while cowork is doing one thing, you
8:15
could be just doing something else.
8:17
So this is something that called code got really, really right.
8:20
You can create many agents that are doing many different things
8:22
all at the same time. For example, you could have one
8:26
creating a plan, 1 executing part of the plan that you're
8:29
happy with, one writing tests to assess the quality of the either
8:34
plan or the things that have been written, the code all on
8:37
one project, whilst another agent is then planning your next
8:41
task. So that's what you can do now.
8:43
You can do that on Claude today with many, many different tabs,
8:47
but cloud Co work is trying to bring the ability the cloud code
8:51
got so right into the main application.
8:53
And what's really important here, and this is what we'll
8:56
come onto and spend a bit of time on, is everything it
8:58
creates, say every file it creates, it persists because it
9:03
can be saved as a local file on your device and you can just
9:07
point it at that local file. You don't need to make AGPT or a
9:09
cloud project, or you might create loads of, let's say,
9:13
artifacts or canvases that you might put in a Google Doc and is
9:16
then lost in many ways for contexts that another future use
9:20
of AI could have benefited from. Instead, you can literally every
9:23
file, save it, put it in the folder, organize it in a
9:27
specific way, and suddenly the agent will then access it
9:30
anytime. And this overcomes many, many,
9:32
many big problems, which I'll come onto in a minute.
9:35
Now developers have been doing this for years, creating and
9:38
storing files on their local device, committing it to the
9:42
entire project. This is a very normal thing.
9:44
And This is why I think non developers are about to go
9:48
through what developers have been doing for many years and in
9:50
particular with AI over the last 12 months.
9:52
And they're going to do a break net speed.
9:56
So we're going to see a huge amount of change in how people
9:58
generally work because I really think the difference between
10:01
say, chatting with normal cloud and working with coworkers, the
10:04
difference between having conversations and continually
10:07
building a system. And it's a real mental shift
10:09
that everybody must really take on board because the ability to
10:14
create files and save them straight on your device
10:16
overcomes, oh, many big problems.
10:18
So for anyone that's been using normal cloud, you might end up
10:21
hitting context window problems. And you do the same on GP2.
10:24
So you've put too many files in, you've had too many
10:27
conversations, you've had too many iterations.
10:29
It says I've hit the file context in it, which is a pain
10:32
because you might be on a really good track and suddenly you have
10:34
to go make a new chart. All of that can go away because
10:39
you can suddenly just create a chart and all the files you've
10:43
saved and assets. Not if I've put it into a
10:45
folder, point cowork at that folder quickly and it's done so
10:49
super, super quick. Another huge benefit outside of
10:52
context windows here is the way you can create a giant system.
10:58
So a good example here with how we work.
11:00
So essentially if you create the right folder structure and give
11:04
Claude access to certain folders to run instructions, and you
11:08
tell it to always follow this folder full of the instructions,
11:13
it will always do that. So it's like creating a giant
11:15
project, but the agent won't just be stuck on one set of
11:19
instructions. So we've got a folder which
11:22
essentially is an agent because when people go, what is an agent
11:26
or why the what's the value of an agent?
11:28
It's literally just a series of prompts that fire at the perfect
11:33
moment to provide an LLM with the perfect amount of context to
11:36
complete that task. So what you're doing by creating
11:40
a folder structure is essentially pointing an LLM at
11:42
folders full of information. You might have a folder called
11:46
planner, you might have a folder called researcher, you might
11:48
have a planner called Presentation Maker, whatever it
11:51
might be. And inside that there is a set
11:54
of instructions, which is essentially just a document
11:57
saved as an MD file. An MD file is just the best
12:01
format for an LLM to read. But you know, imagine it's
12:04
equivalent as a Doc X. You can download a file as an MD
12:07
file. So you create a folder on your
12:10
local device with a set of instructions, which are
12:12
essentially your agents. Because when you say to an LLM
12:16
essentially go read that folder, it's going to use that context.
12:20
You might then create another one called skills.
12:22
Skills are processes that this agent can follow.
12:25
So you might have one called a product requirement document
12:28
creator. You might have another one
12:31
called Python Writer, you might have another one called e-mail
12:37
writer. So these are skills, which is
12:39
essentially a document which lays out in a specific format
12:42
how an LLM should follow a set of instructions.
12:46
So as soon as it then understands that this person is
12:48
asking for a certain task, it will be able to essentially
12:52
access that folder with that skill and read it and then go,
12:56
OK, I know I need to use this skill in this particular way.
12:58
And it will fold the process that you set out.
13:00
And all you can start to do here is continually build this
13:02
system. And every time you create a new
13:04
file, you can commit it to the right folder and an agent can
13:07
then read it further and further over and over again.
13:09
So you're creating a giant system here.
13:12
And this is what we've done it from a engineering perspective.
13:16
It's what we're doing across the company.
13:17
It is literally going to transform how people work.
13:21
Another major mental shift with this here is just this idea of
13:24
having many, many, many different claws or AIS all the
13:27
time. See, it is like tending to your
13:30
claws or tending to your AI because instead of going super
13:34
deep in one thing, you have many things going on at once.
13:38
You're going to check in on each one of them because it might be
13:40
going for three minutes, 5 minutes, 12 minutes, thirty
13:43
minutes on a specific task and you check in to make sure it
13:46
doesn't need to be redirected for some reason.
13:49
And you're just doing this across all your different
13:51
clauses. And that's how programmers have
13:53
been coding for the last 12 months with all these new AI
13:56
tools. It's how I've been using
13:57
different systems and now Co workers bringing that to
14:01
everybody. One thing again, I mentioned it
14:03
earlier about skills. I think this is going to make
14:05
things increasingly better for people over time.
14:08
Essentially, skills is when the agent creates, let's say a
14:12
spreadsheet, it can load up something called an Excel skill
14:16
behind the scenes, which is essentially a set of
14:18
instructions that teach claw the best way to create a Excel
14:22
spreadsheet. Now Anthropic provides
14:24
prepackaged skills. We as mindset, we build skills
14:28
all the time. But what you can also do is
14:30
build your own set of skills with MD files so you can start
14:33
to build these skills that are very particular to your team and
14:35
process. Now that said, if you're just
14:37
getting started, of course, don't over customize, don't go
14:39
too deep. Start simple.
14:42
Go online and search for Co work file structure, for example, and
14:47
look online for the best way to create and set up the system.
14:51
You know, Anthropic have actually run competitions for
14:53
programmers on the best system design and that's where we've
14:56
got a lot of inspiration from. So yeah, start very simple.
14:59
Install the Chrome extensions, install the app so you can
15:03
actually interact, you know, with web as and do things and
15:06
justice. Start using it and force
15:07
yourself into this new way of working because I promise you
15:10
it's going to be infinitely better.
15:13
There's a reason Cloud Code has been so successful.
15:33
So yeah, I think let's move on to a few interesting use cases
15:36
and what I think is practical tips here outside of how this
15:38
all works and why it's a bit of a mental shift.
15:40
I think some of the really interesting use cases here is
15:42
large processing of content. So there was someone called
15:48
Lenin Rychisky. So he's a big podcaster in the
15:51
space, in the product management space.
15:53
He basically gave access to 320 podcast scripts on his laptop,
15:59
which is about 450 to 600 hours of content, so a huge amount
16:04
that would take forever in let's say if you were uploading that
16:08
to ChatGPT or clot because you can't drag 320 documents into a
16:11
chat window badly. With Co work you can just
16:14
literally point it at a folder done as another use case.
16:17
One of the fantastic things about accessing your actual
16:21
local files is the ability to produce lots of local files and
16:25
then get AI to organise them better.
16:28
So you might constantly be creating folders and systems.
16:31
And This is why looking online for a really good initial system
16:34
to structure this around is good.
16:35
But as you create more files, you download it, you save it,
16:38
add it to new folders, you can then just ask Claude to go and
16:42
reorganize it, tidy up and make it efficient for whatever system
16:46
you're trying to design it. So again, using AI to prune the
16:49
AI system that it's going to be following.
16:50
Another big use case here, I think is actually starting to
16:53
like unlock things like APIs without being someone that's
16:56
technical because most vendors, most technology platforms offer
17:01
APIs that allow you to do really interesting stuff, pull
17:03
interesting data straight into your conversation.
17:08
It allows you to just do a lot of different things.
17:10
And for most people, they cannot use those unless they're a
17:13
developer. So suddenly a big use case, it's
17:16
emerging and it started because of non-technical people using
17:19
cloud codes or any coding tool for that matter.
17:22
But suddenly with Co work, because it's the same underlying
17:25
agent, it's going to be very good at doing that as well.
17:27
Another big use case area is browser based tasks.
17:30
So it's really the team has really tried to be good at using
17:33
the browser. I think there's going to be an
17:34
emerging space to try and create general purpose agents.
17:38
So an agent that will sometimes use skills, sometimes use
17:40
browsers to essentially be able to go and complete specific
17:46
tasks, e-mail things to people and do things for people.
17:49
And again, it can be slow and cumbersome.
17:51
But if you don't see it as a function of speed all the time,
17:54
but having many different tasks all happening at once and you're
17:58
just checking in on your clouds or your AIS, then I think this
18:02
becomes and is more powerful over time.
18:04
And finally, when it comes to research, I think it's really
18:06
interesting because you know, if you do research on different,
18:09
you know, AI systems, you download the file, you might use
18:13
that somewhere else. You have to upload it now.
18:15
You could just get it to research many different things
18:18
but smaller tasks and just download it all into a project
18:21
called Research on xtopic and then start a new chart with 0
18:25
context. So the context window is big,
18:28
which means you can do lots of work on that research and point
18:31
it straight at the folder. So again, I hope you're starting
18:34
to see the compounding impact of doing task saving files,
18:39
organizing the files with AI, and then starting new chats and
18:43
doing it again. But pointing at that folder that
18:46
you've created just then. It's really powerful.
19:05
So finally, I guess you can see that I'm relatively excited by
19:07
this, but it's going to why I think this overall is going to
19:10
be a really big deal. So I think #1 it's a platform
19:14
moment, and it really showcases the power of, I guess, nobody
19:18
really knowing where things are going to go.
19:20
You know, when the App Store launched, the initial apps were
19:23
beer drinking simulators. You know, where you got the iPod
19:25
and you pretend you're drinking a beer.
19:27
Nobody predicted Uber or Tiktok, you know, nobody had an idea
19:31
what that killer use case was. And I think we're at the exact
19:34
moment with agents Cloud code was created for programmers, a
19:36
very specific thing and suddenly everybody lived it.
19:40
And so they suddenly had to diverse and move out of just
19:45
coding. See, I think it's going to spawn
19:46
a load of new potential ways of doing things.
19:49
And I think the other big area of the reason I think it's
19:51
really exciting is the exponential impact of this is
19:54
real. You know, a year ago people at
19:56
Anthropic were saying by the end of 2025, most people won't be
20:00
writing code. And I think it sounded very
20:02
aggressive. I actually commented that it
20:03
sound quite aggressive, but many engineers are having most of
20:06
their code written by Claude. It's certainly the case that
20:10
most of the code that is produced at Mindset is now with
20:13
AI, and I think that's the trajectory for coding.
20:15
But the same thing is going to happen across all forms of work.
20:19
If you listen to last week's episode on what trends are going
20:21
to be there for 2026? A couple of them were everybody
20:25
at knowledge workers becoming an AI manager and knowledge work
20:28
having a transformation akin to what programmers has.
20:32
And this is a big reason why. See, I think often we think in a
20:37
they think progress is going to happen in a linear way, but
20:40
sometimes it can happen exponentially.
20:41
I think this is going to be one of those instances.
20:44
I think another big reason this is going to be so impactful is
20:46
that everybody is going to follow.
20:49
Open Eye is obviously going to copy this.
20:51
Google is going to do something, Microsoft is going to do
20:53
something, probably built into Windows and distributed to
20:55
billions of people. Apple are going to incorporate
20:58
it in some way over the next few months.
21:01
It's going to be in added into things like Adobe and other
21:04
forms of systems like that. The entire industry is just
21:07
going to converge on this pattern.
21:08
And what I think is really interesting it actually that
21:11
there's a lot of investment into things like the idea of a canvas
21:15
or an artefact or a new, I guess, file type that happens
21:19
within the AI system that you can create things with
21:23
collaboratively with AI. And what's funny is that I think
21:27
maybe what we've just seen is that simply downloading a local
21:30
file onto your laptop, the oldest thing in the book, I
21:33
think that's been around for the longest, might just be that
21:36
winner. I think this is going to be
21:37
massive. I think that's going to reshape
21:39
much of the industry. I think this is going to lead to
21:41
a big transformation of how products are going to be built
21:43
and designed. In short, I think the genie is
21:46
out of the bottle. So yeah, in conclusion, I think
21:50
Co work is clawed code without the terminal, the same agent.
21:53
It's accessible to everyone, it works on your files, it controls
21:56
your browser, run tasks in parallel.
21:59
So you can have many different Claws or AIS going at once.
22:02
And I think this shift is going to be massive.
22:05
And in particular, I'm really excited to see people start to
22:07
think about files that you save in folders on your laptop being
22:11
something that you spend a lot of time thinking about the
22:13
system and structure of that, and whether that is going to
22:17
just persist for those that live.
22:19
I guess it might feel for some people somewhat technical, even
22:23
though it's really not, because it's slightly different to the
22:26
way they work today. Or whether this becomes the de
22:29
facto standard across the industry.
22:31
Or whether some people try to replicate the idea of files
22:34
being saved in that format but on an application that feels
22:37
more like AG drive. But anyway, I think the industry
22:41
is going to converge on this pattern and the people who build
22:44
intuition and new processes around this tool now will just
22:48
have a huge advantage to how they work.
22:50
So yeah, download the app, connect your folders, start
22:54
simple, but please do start. This is in the Libra Jackal and
22:57
I hope you enjoyed this week's episode.
23:00
I had a lot of fun going through it, getting really stuck into
23:03
Cloud Co work and I look forward to speaking to you next week.