Latest / AI Fire Daily / #126 Robin: Claude Code 2.0 Can Now Manage Your Chats, Projects, Apps, Docs, Slides, and More

Transcript

#126 Robin: Claude Code 2.0 Can Now Manage Your Chats, Projects, Apps, Docs, Slides, and More · AI Fire Daily

0:00
Imagine building a fully working Mac application
0:03
from scratch. You don't know a single line of
0:06
code. You just ask for it in plain English. Suddenly
0:09
it compiles. It launches right on your screen.
0:12
Right. It sounds like pure science fiction, but
0:15
that is exactly the reality we're unpacking today.
0:19
Welcome to the deep dive. Today we're looking
0:21
at a massive shift. We're exploring the total
0:24
rebuild of Claude code. Yeah, one anthropic engineer
0:28
actually dubbed it Claude 2 .0, which is honestly
0:31
pretty accurate. We're talking about a transition
0:33
that turns a simple text chatbot into a full
0:36
-blown operating system. Now, I have to admit
0:40
something right up front. Managing AI context
0:43
windows usually gets incredibly messy for me.
0:46
I still wrestle with prompt drift myself. Oh,
0:48
it happens to everybody. You start with one goal
0:51
and suddenly the AI is entirely off track. Exactly.
0:54
It's so frustrating. And that drift is exactly
0:57
why this update is so monumental. It finally
1:00
gives you the scaffolding to contain that chaos.
1:03
It completely changes how we interact with software
1:06
So we have a really fascinating journey mapped
1:08
out for you today. We're going to explore how
1:10
the new two -tab workspace actually operates.
1:13
Then we'll uncover how to command master sessions
1:16
to manage massive projects. We'll dive into building
1:19
interactive canvases called artifacts. And finally,
1:22
we'll reveal the hidden slash commands and custom
1:25
mods. Those are the absolute best part. Right.
1:27
They are the power tools that will save you hours
1:29
of work. The jump in capability here is truly
1:32
staggering, but it also requires a totally new
1:35
mental model from you. You're no longer just
1:37
chatting, you're orchestrating. I really want
1:40
to understand the mechanics of this workshop
1:41
first. If we open the desktop app right now,
1:44
the UI looks different. There are just two tabs
1:47
sitting at the very top. Yeah, it's a very clean
1:49
interface now. Those two tabs are simply labeled
1:52
Claude and Code. Okay. Regular chat and co -work
1:55
now share one single prompt box that forms your
1:59
standard Claude tab. So that first tab is basically
2:01
what we're already used to. It's great for quick
2:04
questions, brainstorming, or everyday writing
2:06
tasks. But the second tab, the code tab, that
2:09
seems to be where the magic happens. Exactly.
2:12
The code tab gives the AI direct, secure control
2:15
over your local environment. It can actively
2:18
access your local files. Wow. It can read your
2:21
directories. It can even utilize your computer's
2:23
command line tools. And here is the crucial part.
2:27
You still use plain English for absolutely everything.
2:29
The word code on the tab shouldn't scare you
2:31
away. Let me make sure I'm really grasping the
2:34
difference here. Let's look at a highly specific
2:37
example. Say I ask both of these tabs to build
2:40
a Mac menu bar timer. What actually happens?
2:44
Well, if you ask the regular Colab tab, it simply
2:47
writes out the raw code. It gives you a text
2:49
file full of instructions. Just a text file?
2:51
Right. It leads the actual compiling, building,
2:54
and launching entirely up to you. Right. So I'd
2:56
have to open Terminal, install dependencies,
2:58
and figure it out. But if I use the Claude Code
3:01
tab, it writes the code, builds it, and launches
3:04
the working app directly. Yeah. It finishes the
3:06
entire job without you doing anything else. Let
3:09
me see if this analogy works. Using the regular
3:12
tab is like hiring a consultant who hands you
3:14
a blueprint. But Claude Code is the contractor
3:16
who actually builds the house. That is the perfect
3:19
way to look at it. It's the difference between
3:20
getting advice and getting execution. But, to
3:24
manage this digital contractor effectively, you
3:26
need proper oversight. And that is where the
3:28
Session Rail comes into play. I see this panel
3:31
sitting on the right side of the screen. I assume
3:33
this is the Session Rail. What exactly am I looking
3:36
at here? The Session Rail is your control center.
3:39
It actively shows you everything Claude is currently
3:41
doing behind the scenes. There are four specific
3:44
cards you need to monitor closely. You have environment,
3:47
context, connectors, and outputs. Well, let's
3:50
break those down. What does the environment card
3:52
actually tell me? It shows if your current session
3:55
is working locally on your machine or if it's
3:57
operating remotely in the cloud. It's vital for
4:00
understanding where your data is living at that
4:03
exact moment. OK, that makes sense. And the context
4:06
card, I assume that's tracking how much working
4:09
memory I'm eating up? Spot on. Every file you
4:11
upload, every skill you activate, and every connected
4:14
app takes up space in that memory window. immediately.
4:17
So it fills up fast. Yeah. And if your context
4:19
gets too full, the AI starts to forget earlier
4:21
instructions. Which leads right back to that
4:24
prompt drift that I struggle with. So what do
4:25
I do when I see that context card hitting the
4:27
ceiling? You just click the compact button right
4:30
there on the rail. It automatically summarizes
4:32
the entire chat history. It extracts the essential
4:35
rules and frees up vital space so you can keep
4:37
working smoothly. That is incredibly practical.
4:40
Now, I noticed another feature on this environment
4:43
card. There's a switch labeled remote control.
4:46
What does that actually do? Oh, this is brilliant.
4:48
It lets you continue your local desktop sessions
4:51
right from your phone. Wait, really? Yeah. You
4:54
can be away from your desk, pull out your phone,
4:56
and command your laptop's terminal remotely.
4:59
It keeps the computer awake and working while
5:01
I'm walking, right? Exactly. As long as the desktop
5:04
app is running, your phone acts as a remote command
5:06
center. And speaking of staying organized, the
5:09
new workspace introduces sections. OK, how do
5:13
sections change the workflow? Are we just talking
5:15
about basic folders here? Essentially, yes, but
5:17
they're incredibly fluid. You just click the
5:20
three dots next to any active chat in your sidebar.
5:23
You can move it to a new section and name it
5:25
whatever you want. So I could have a dedicated
5:27
section for content creation and another one
5:30
entirely for client projects. It stops you from
5:32
losing track of important conversations as they
5:34
pile up over the week. Exactly. It keeps your
5:37
mental workspace clean. And then we have side
5:39
chats, which are absolute lifesavers for your
5:42
workflow. Let's dig into that. I type slash BTW
5:45
to open a side chat. Why wouldn't I just ask
5:48
the question in the main prompt box? Because
5:51
every question you ask pollutes the primary context,
5:54
Wenda. If you're building an app and you suddenly
5:57
ask Claude to explain a piece of syntax, that
6:00
explanation becomes part of the project's permanent
6:02
memory. I see. Side chats let you ask those quick
6:06
questions without interrupting the main work.
6:08
So I can pop open a side chat and ask, you know,
6:11
review what is done. Does anything need fixing?
6:13
Exactly. They read the main conversation, but
6:16
they don't inject their own text back into it.
6:19
How do side chats avoid ruining the main AI memory?
6:22
They operate in a parallel memory lane. This
6:25
keeps the main workspace totally uncluttered
6:28
and focused. So they look at the main chat without
6:30
adding their own mess to it. Spot on. And what's
6:33
really cool is you can keep multiple chats open
6:36
simultaneously using Split View. I love a good
6:39
Split View. How do I actually trigger that in
6:41
the app? You just grab a chat from the sidebar
6:43
and drag it over. Or you can hold Alt or Option
6:46
on a Mac and click the chat. It drops it right
6:48
next to your current window. That is absolutely
6:51
perfect for researching in one window while actively
6:53
writing in another. I don't have to constantly
6:55
tab back and forth anymore. Right. Now that we
6:59
have our chats sitting side by side, things get
7:01
really wild. How do we get them to actually collaborate
7:05
with each other? This is where we move from just
7:08
organizing single tasks to essentially running
7:10
a full digital agency. You can now have one main
7:13
chat actively manage multiple other chats. We're
7:16
calling these nested sessions. Yeah, a nested
7:19
session is literally spawned by your main cloud
7:21
session. You don't create it manually. The AI
7:24
creates it to handle a very specific dedicated
7:26
task for you. Walk me through a real world scenario.
7:30
Say I'm launching a massive marketing campaign.
7:32
How does this actually play out? You go to your
7:35
main chat and outline the campaign. You tell
7:37
Claude to create three nested sessions, one dedicated
7:40
solely to writing the sales page, one for drafting
7:43
the email sequence, and one for designing the
7:45
ad copy. So the main chat acts as the project
7:49
manager. It coordinates their work and reports
7:52
back to me when everything is ready. It's f -
7:54
fascinating because each session handles its
7:57
own work independently in its own context window.
8:00
But wait, let me push back for a second. How
8:02
is this actually different from just using subagents?
8:06
That is a great question. Subagents are generally
8:09
just delegated background tools that return a
8:11
simple final result. You don't really interact
8:14
with them. Nested sessions, on the other hand,
8:17
are full active conversations that you can actually
8:19
open up. Let's clarify that term for a second.
8:21
Just to be precise, sub -agents are tiny AI workers
8:24
assigned to complete one specific background
8:27
task. Exactly. They just run and return. But
8:31
with a nested session, you can literally click
8:33
into the email drafting chat, read what it's
8:36
doing, and give it ongoing real -time feedback
8:38
while it works. It's totally transparent. You're
8:41
not just waiting blindly for a result. Now, to
8:44
manage all these sprawling sessions, I understand
8:46
you can create something called a master session.
8:48
Yeah, this is a brilliant workflow hack. You
8:51
take one specific chat and pin it permanently
8:53
to your sidebar. You literally just type into
8:56
it, you are my master session. And what does
8:59
assuming that rule actually allow it to do? It
9:02
automatically gains the ability to track the
9:04
progress of your other active sessions. It can
9:07
scan the metadata of your 20 most recently active
9:09
sessions. It parses what's finished, what's stuck,
9:13
and tells you exactly what needs your immediate
9:15
attention. So it's literally a chief of staff
9:18
for my digital agency. Now everything we just
9:21
talked about happens locally, but there is also
9:25
Cloud Code projects, which is rolling out in
9:27
beta right now. Right, and this is the cloud
9:30
-based coordinator designed specifically for
9:32
pro and max tier users. It takes this orchestration
9:35
to an entirely different level. You give this
9:38
cloud coordinator a massive multi -step task
9:41
to accomplish, something incredibly complex.
9:43
Like, you tell it, research the history of Claude
9:46
code, build a slide deck, and produce a motion
9:49
graphic explaining it. It takes that massive
9:52
prompt and instantly splits it into parallel
9:55
working threads. And because it's a project,
9:57
all those threads share the exact same project
10:00
memory at the same time. Whoa. Imagine scaling
10:05
to a billion queries. The computational power
10:08
required behind the scenes to keep all those
10:10
contexts aligned is truly staggering. And the
10:13
best part? Because it's cloud -based, it runs
10:16
those massive parallel tasks while your laptop
10:19
is completely closed. Are files uploaded to projects
10:22
kept entirely on my local machine? No. Because
10:26
projects run in Anthropix Cloud, those files
10:29
go to their servers. Got it. Projects mean cloud
10:32
servers while nested sessions stay on your local
10:35
drive. Right. It's a very important distinction
10:37
for your privacy and workflow. Use nested sessions
10:40
for local files on your computer. Use projects
10:43
for remote access when you need to step away
10:46
from the keyboard. It's incredibly powerful,
10:49
but running multiple complex sessions simultaneously
10:52
will drain your token usage limits fast. You
10:55
definitely want to start small. Try running just
10:58
two tasks first to see how it feels. Sponsor.
11:02
This deep dive is supported by our sponsors.
11:04
Please take a moment to check them out. OK, we
11:07
are back. So we have our AI digital agency fully
11:10
coordinated. But what are they actually producing?
11:14
I want to look at the tangible outputs this system
11:16
creates. These outputs are robust enough to completely
11:19
replace a lot of traditional software tools.
11:22
The term you're looking for here is artifacts.
11:25
These tangible outputs are called artifacts.
11:28
Basically, they're interactive canvases where
11:30
you create, view, and edit work in real time
11:33
right beside your chat window. Instead of just
11:35
getting a block of text, you get a fully functional
11:38
interface. There are five specific creative tools
11:41
within artifacts that we really need to break
11:43
down. The first tool is Clawed Docs. Yeah, Clawed
11:46
Docs is perfect for long form writing. It creates
11:49
beautiful multi -tab documents. It's great for
11:52
organizing chapters of a book or sections of
11:55
a massive report without losing your place. The
11:57
second tool is Clawed Slides. Now this one fascinates
12:01
me. It builds fully animated presentations directly
12:03
in the window and you can export them directly
12:06
to PowerPoint. That export feature is huge for
12:09
corporate workflows. You're not trapped in the
12:11
Claude ecosystem. You can take the work and drop
12:13
it right into your team's standard format. The
12:16
third tool is Claude Design. This essentially
12:18
gives you a Canva or Figma style design canvas
12:21
to manipulate visuals. I do need to make a quick
12:24
note on that one. If you've been using the standalone
12:26
Claude design website, that actually closes down
12:29
on December 14th. Oh, really? Yep. So how do
12:32
we maintain our brand guidelines if the site
12:33
is closing? They've integrated it seamlessly.
12:36
OK. You can now upload your brand specific design
12:38
system, your colors, fonts, logos directly into
12:41
your primary settings. The AI just pulls from
12:44
there automatically whenever you use Claude design.
12:46
That is actually much more efficient. The fourth
12:49
tool is Claude dashboards. This is phenomenal
12:52
for pulling interactive data charts and graphs
12:54
from raw spreadsheets. And then you have the
12:57
fifth tool, which is arguably the most impressive,
13:00
Claude Motion. It creates animated videos and
13:03
dynamic motion graphics, but it doesn't use standard
13:05
video generation models. It uses raw code. Let's
13:09
unpack the mechanics of that. Why does it matter
13:11
that it uses raw code instead of standard AI
13:14
video generation? Motion using code. is a massive
13:18
financial insight for Anthropic. Generating pixel
13:21
-by -pixel video using diffusion models requires
13:23
insane amounts of GPU rendering power. It is
13:26
incredibly expensive. By using code like SVG
13:31
animations or React components, it runs instantly
13:34
in the browser. It completely bypasses the need
13:37
for massive rendering farms. You don't have to
13:39
wait five minutes for a clip to render. And if
13:41
you don't like a specific movement, you just
13:43
tell Claude to tweak the code. The animation
13:45
changes instantly without regenerating the whole
13:47
file. Exactly. It's highly adjustable and significantly
13:51
cheaper. But the real paradigm shift happens
13:54
when we look at AI -powered artifacts. Right.
13:57
This is where things get a bit inception -like.
13:59
You can actually embed AI capabilities directly
14:02
into the artifact itself. Think about the implications
14:05
of that. You aren't just creating a static tool.
14:09
You're creating a smart tool. Let's use a concrete
14:11
example. Sure. Say you build a complex sales
14:14
dashboard for your entire sales team. Normally,
14:17
they'd just look at the numbers. But now, you
14:19
can tell Cloud to embed a functioning AI chat
14:22
box directly into that dashboard layout. Yeah,
14:25
and it passes a secure session state into that
14:28
chat box. Now, your teammates can query the dashboard
14:31
directly. They can ask, Why did Q3 sales dip
14:35
in the Midwest? And they don't even need to know
14:37
how to use cloud themselves. They don't need
14:39
an account. It's like stacking Lego blocks of
14:41
data to build a dashboard and then dropping a
14:44
tiny AI brain inside the dashboard itself. It
14:47
completely democratizes complex data access for
14:50
your whole team. It removes a technical barrier
14:52
entirely. Are all these creative tools available
14:55
to everyone right now? No, feature availability
14:58
depends on your tier. For instance, cloud motion
15:01
isn't on Pro or Max yet. So always check your
15:04
specific subscription plan before planning your
15:07
workflow out. It's always good to verify your
15:09
tools first. You don't want to design a workflow
15:11
around a feature you can't access. Now, we really
15:15
need to talk about optimization. This is critical.
15:18
If we're wielding these complex artifacts and
15:20
running background nested sessions, we are going
15:22
to burn through token limits incredibly fast.
15:25
We need shortcuts. We have to move from creation
15:27
to the actual optimization of this workflow.
15:29
To manage the system efficiently, you need to
15:31
master three specific power tools, slash commands,
15:35
custom mods, and keyboard shortcuts. Let's start
15:38
with slash commands. These are typed directly
15:40
into the prompt box, right? You just type a forward
15:42
slash followed by a specific word to trigger
15:44
an action. Exactly. It's a direct override command
15:47
to the system. The slash doctor. command. It's
15:51
incredibly useful when things start breaking
15:52
down. It actively checks your custom instructions
15:55
and settings for any conflicting rules. That
15:57
is brilliant. It is. Sometimes you forget you
15:59
told it to always write in bullet points three
16:01
weeks ago and now you can't figure out why it
16:04
won't write paragraphs. Right. Then you have
16:07
the slash skill doctor command. This one scans
16:10
your connected apps and tools. It finds skills
16:12
that are consuming massive amounts of context
16:14
memory but are rarely ever actually used. So
16:18
it's basically a spring cleaning tool for your
16:20
working memory. Now, my absolute favorite command
16:23
is slash advisor. This is a brilliant budget
16:26
hack for your daily workflow. The mechanics behind
16:29
slash advisor are so smart. When you use it,
16:32
a cheaper, faster model like Claude Sonnet handles
16:35
the basic conversation. But when it hits a complex
16:37
problem, it actually consults the more expensive
16:40
Opus model for the important decisions. It routes
16:43
the heavy computational lifting only when it
16:45
is actually necessary. It's basically like a
16:48
junior developer calling over the senior engineer
16:50
only when they get stuck, saving the company
16:53
massive amounts of money. That is exactly how
16:55
it functions behind the scenes. It optimizes
16:57
your token spend without sacrificing the quality
17:00
of the final output. Moving on from commands,
17:03
we also have custom mods. These actually change
17:06
how the desktop app looks and functions for you
17:09
personally, and you build them entirely in plain
17:11
English. You don't write scripts. You just describe
17:14
what interface change you want Claude to build
17:16
for you, and it modifies its own app. Give me
17:19
a practical example. What is a mod people are
17:22
actually building? A great example is the brain
17:24
dump mod. Let's say you use... voice dictation
17:27
a lot and your prompts are rambling, messy stream
17:31
of consciousness, this mod automatically intercepts
17:34
that messy prompt, cleans up the logic, formats
17:37
it cleanly, and then submits it to the main AI.
17:40
That removes so much friction. I also saw people
17:43
using a progress bar mod. It adds a persistent
17:47
live visual panel to your sidebar specifically
17:49
for tracking the completion percentage of your
17:52
nested sessions. It's all about surfacing the
17:54
data you need. And speaking of vital data, you
17:57
really want to install the usage ribbon mod.
18:00
What exactly does the usage ribbon track for
18:02
me? It places a small banner at the top of your
18:05
screen. It shows your remaining weekly usage
18:08
limits in real time. But more importantly, it
18:11
constantly tells you if your prompt cache is
18:13
currently warm or cold. Let's define that jargon
18:17
quickly for clarity. Prompt cache is saved AI
18:20
memory that makes repeated tasks faster and much
18:23
cheaper. Knowing your cache status is absolutely
18:26
vital for controlling your token costs, especially
18:28
on large projects. Why does it matter if my prompt
18:31
cache is warm or cold? A cold cache means Clot
18:34
has to reread your entire project history from
18:37
scratch, which costs you significantly more tokens.
18:40
Warm cache saves money. Cold cache means paying
18:42
full price to process it again. Exactly. If the
18:44
ribbon says cold, you might want to wait a minute
18:47
before firing off a massive query. Finally, let's
18:51
touch on some essential keyboard shortcuts to
18:53
speed everything up. Navigation is key. If you
18:56
hit the Escape key twice, it actually rewinds
18:59
the conversation to an earlier point. It's an
19:01
instant undo for bad prompts. Hitting Control
19:04
-Alt -Space is another massive time saver. It
19:07
immediately opens a floating cloud window from
19:09
anywhere on your computer, no matter what app
19:12
you are currently using. You don't have to break
19:14
your focus to switch windows. And if you prefer
19:16
talking, control D instantly starts your voice
19:19
dictation. By the way, you can just press command
19:21
slash at any time to see the full exhaustive
19:24
shortcut list on your screen. The goal isn't
19:27
to memorize all of these immediately. Just pick
19:29
two or three that fit your style and build your
19:31
muscle memory gradually over a few weeks. It
19:34
is all about systematically removing friction
19:36
from your daily workflow. Let's synthesize this
19:39
entire journey we just took today. We have watched
19:42
Claude evolve dramatically over a very short
19:45
period. It really is a paradigm shift. It went
19:48
from being a simple text chatbot that answers
19:49
trivia into a full -blown semi -autonomous operating
19:53
system. The introduction of the code tab completely
19:56
changes the boundaries of what is possible on
19:58
a local machine. It moves us from ideation to
20:01
actual execution. By combining the code tab,
20:05
master sessions, artifacts, and custom mods,
20:08
you have incredible power at your fingertips.
20:10
You can orchestrate highly complex, multi -keered
20:13
projects using nothing but plain English instructions.
20:17
It is genuinely like having an entire engineering
20:20
research and design team living inside your laptop,
20:22
but it can feel overwhelming at first. We strongly
20:26
encourage you to start very small today. Yeah,
20:28
don't try to build a massive app on day one.
20:31
Switch to the code tab and just try organizing
20:33
your recent chats into three logical sections.
20:37
Exactly. Or maybe... Pin one chat as a master
20:39
session just to try a basic nested workflow.
20:43
Build one small custom mod to see how it feels
20:46
to alter the UI. Take it step by step. Get comfortable
20:50
with the new mechanics before scaling up. I want
20:53
to leave you with one final provocative thought
20:55
to mull over. If Cloud Code can write, build,
20:58
and launch applications entirely on its own based
21:01
on plain English instructions today, what happens
21:04
tomorrow? What happens when it starts silently
21:06
monitoring your daily friction points and spontaneously
21:10
builds custom desktop apps to solve problems
21:12
you didn't even realize you had?