This is f*cking insane. Anthropic dropped a free 58-minute Claude Code course Going from autocomple

Transcript

0:01

All right, we'll go ahead and get started

0:03

Thank you everyone for joining

0:05

Welcome to today's webinar

0:06

This is an introduction to Cloud Code

0:09

We are live from New York City and

0:10

San Francisco

0:12

This session has been highly

0:14

highly requested

0:15

so we're super excited to walk you all

0:16

through it today

0:17

On theme with awards season

0:19

I will be your host

0:20

So my name is Emily Gray

0:21

I'm on our sales team here at Anthropic

0:27

Cool. So first

0:28

we're going to walk you through what Cloud Code is

0:30

the foundations of how to use the platform

0:33

features you should be aware of

0:34

like certain slash commands

0:35

and we're also going to cover

0:36

a live demo

0:41

Awesome. So with that

0:42

you guys know who I am

0:43

I'm going to quickly kick it over to

0:45

Harry

0:45

and Ali from our beloved Applied AI team

0:48

to quickly introduce themselves

0:50

Thank you, Emily

0:51

my name is Harry

0:52

I'm here in the Applied AI team at

0:53

Anthropic. I can't see all of you

0:55

but

0:56

I'm excited that all of you are here

0:57

I think there's almost

0:58

or more than 2,000 of you on

1:00

on the line right now

1:02

I sit at the intersection between go-to-market

1:04

engineering, product, and some other orgs

1:07

and really help

1:08

our customers and folks that are also interested

1:10

in Anthropic and Claude products get the most out of

1:14

what they're doing with Claude today

1:16

And I'll

1:16

go ahead and pass it over to Ali

1:18

for intro on his side as well

1:21

Thank you, Harry

1:22

hi all, I'm Ali

1:23

and I'm an Applied AI engineer here at

1:25

Anthropic

1:26

super excited to be here and to get things going

1:30

Awesome, thanks guys

1:33

Cool. So just a couple quick housekeeping

1:35

housekeeping items

1:36

So one, yes

1:37

this will be recorded and we'll make sure

1:39

to send everyone a copy after

1:41

And then next

1:42

please make sure to ask questions

1:43

We want this session to be

1:45

as live as possible. So

1:46

Ali and I will be monitoring the chat

1:48

and answering as much as we can

1:50

And then of course

1:51

please share your feedback with us

1:53

Our goal is to make these sessions as

1:54

valuable as possible. So we always love

1:56

to hear feedback from our customers

1:59

And one more housekeeping item

2:01

our team can

2:02

answer only so much as humanly possible

2:06

but the good news is that we have

2:07

our dear friend Claude. So on our website

2:10

you can see, at code.claude.com

2:13

which is a mouthful. If you go there

2:15

that's the best place to

2:17

get questions

2:18

That version of Claude scrapes all of our

2:20

documentation and will provide the fastest answers for you

2:23

But again

2:23

we'll be making sure that we monitor the

2:26

chat and answer as much as we can

2:28

Cool. So with that

2:29

I'm going to kick it over to Harry

2:30

to walk through the presentation

2:32

Thank you, Emily

2:33

so

2:34

before we get into the Claude Code

2:36

overview and

2:37

a little bit of an introduction on what

2:39

Claude Code is and how you can get

2:40

the most of it

2:41

I want to set the stage and give

2:42

a little bit of context of how we

2:43

got here. So

2:45

everyone knows

2:46

or has at least heard of the word

2:48

agents by now. So you've seen

2:50

billboards if you're here in the Bay Area

2:52

or you've seen them mentioned online

2:54

you've seen demos of them on LinkedIn

2:56

or other platforms

2:58

But what are actual agents?

2:59

Are they workflows? Are they

3:01

you know, autonomous robots that live in our systems and

3:04

and do things for us?

3:06

so what they actually are

3:08

are, A system that is powered by AI

3:11

where a human can give the agent

3:14

some kind of goal to accomplish

3:16

and the agent can use all of its tools

3:18

all the systems and harnesses and layers underneath to think

3:24

to do

3:25

and then to iterate on what it's thinking

3:27

to ultimately accomplish that goal

3:28

So I want to

3:29

really set the stage for what agents are

3:30

because Claude Code

3:32

truly is an agentic platform versus an AI-powered one

3:35

and I want to make that difference

3:37

fairly clear

3:38

as we walk into history of

3:39

what AI coding has looked like for the last few years

3:43

right? So let's rewind five years from today

3:47

So back from 2026 all the way to 2021

3:50

we started with autocomplete

3:51

So if you're a developer

3:52

and you've been in this space for a while

3:54

you've probably seen

3:55

in some platforms that you've been using a few years ago

3:59

the ability to code and then the

4:03

IDE

4:03

or the development environment or the coding platform

4:06

will actually autocomplete the rest of the

4:08

Coding line for you

4:10

So

4:11

that's where AI is going to work

4:13

for seconds for you. So it's going to

4:15

think about the line that

4:16

that you're working on

4:17

It's going to think about the file that you're in

4:19

and then it's going to auto-complete the rest

4:21

of that line. So really that's where we were years ago

4:24

and then we

4:25

evolved a couple years down the line

4:27

to maybe being able to modify whole files

4:30

right? Where AI

4:31

is able to take in the context

4:33

of what you're working on

4:34

and then maybe some of the other modules

4:36

and methods and libraries that you have in your project

4:40

and then modify a file at large

4:43

Where are we today

4:44

and where are we going in the future?

4:45

So

4:46

Cloud Code came out in 2025

4:48

as a research preview

4:49

created within our labs here at Anthropic

4:53

And you're seeing Cloud Code become

4:56

from an AI assistant

4:57

more to an agentic coding tool

4:59

And Cloud Code's actually able to work for hours on end

5:03

being able to complete tasks that you give

5:04

it at a very general level

5:06

whether it's a complicated deep dive into your code base

5:09

and then being able to modernize it or refactor it

5:13

or whether it's

5:14

a task that has a lot of edge

5:15

cases that you really want to code develop with Cloud on

5:18

right? So this is now an agentic platform

5:21

And as we go into the future

5:22

we're seeing AI working for not just hours

5:25

but days on end

5:26

whether it's really sitting with you and prioritizing your user feedback

5:31

and then just working on its own until

5:33

it has a question and it comes back

5:34

to you. So as we go into the future

5:36

you're going to see this curve

5:38

go up and up and up

5:39

And this follows

5:41

What you're seeing on the screen

5:42

in terms of this diagram

5:44

In terms of the technology itself

5:46

we've evolved from being able to get

5:47

just simple outputs to maybe automated workflows

5:51

which I talked about earlier

5:53

I want to

5:53

stress the difference between workflows and agents

5:56

Workflows are deterministic or specific paths that you go through

6:00

but an agent truly is

6:02

something that is a loop

6:04

So a human interacts with an agent

6:06

by giving it a goal

6:07

like an open-ended problem

6:09

something that's a little difficult in terms of predicting

6:14

something that doesn't have a deterministic path

6:17

such as search, coding, support, et cetera

6:21

And then it works in a loop to

6:23

use all the tools

6:24

in the environment that it has access to

6:26

to ultimately complete it without

6:28

a human needing to intervene or steps needing

6:30

to be defined specifically

6:32

Okay. So

6:33

that was

6:34

a little bit of context before we get

6:35

into cloud code. So

6:37

Would love to introduce Cloud Code

6:38

So

6:39

a lot of you on this call have

6:40

probably seen Cloud Code in some fashion already

6:42

even if you haven't used it

6:44

Again

6:45

would encourage you to look in the

6:46

docs tab here on the platform

6:48

If you want to

6:49

start getting an overview on Cloud Code using

6:52

some of the documentation

6:52

you can install it even while I'm talking

6:55

But it truly is an agentic

6:57

what we call a CLI or a command

6:59

line

7:00

interface tool

7:01

to allow you to

7:03

start doing development or even non-technical use cases with

7:06

with Cloud Code

7:08

The reason we've chosen this form factor you're

7:10

seeing on the screen here on the right

7:11

is because we want

7:14

we wanted developers to be able

7:15

be able to actually interact with parts

7:18

of the

7:19

scripts and the

7:20

CI CD pipelines that are native to their

7:22

terminal today

7:24

So

7:25

you can certainly use it in other

7:28

forms and fashions and we'll

7:29

we'll take a look at what that looks

7:31

like later down in this presentation

7:34

I want to showcase all of the You know

7:37

different buckets that you can actually use Cloud

7:39

Code in

7:40

So let's take the software development lifecycle

7:42

right? So we have

7:43

maybe five different steps that folks are used to

7:46

especially if you're a developer that's tuning in today

7:49

all the way from discover

7:51

to supporting and scaling your environments

7:54

So you might go through a typical workflow like this

7:57

right? You might onboard to a new project

7:59

you might develop certain technical specifications

8:02

and then you might actually

8:03

write tests and then implement code according to

8:06

those tests

8:07

and then really automate

8:09

you know

8:09

the CICD pipeline and then

8:11

be able to debug and monitor usage from

8:13

that

8:14

So Cloud Code, I want to emphasize

8:16

doesn't just take care of the middle bucket

8:18

which a lot of people think it ends up doing

8:21

but it can start helping you

8:22

discover parts of your code base

8:25

develop those technical specifications by using plan mode

8:29

automating code review

8:30

and then ultimately scaling out

8:32

debugging and then

8:33

refactoring your code at the very end as

8:35

well. So as we're going This presentation

8:38

I just wanted to really emphasize

8:40

that folks can really start thinking about

8:42

how to use Cloud Code

8:44

in different parts of your development lifecycle

8:47

and then before we actually go

8:49

to the rest of the content

8:50

I wanted to

8:52

kind of leave this section

8:53

on how Anthropic actually uses Cloud Code internally

8:57

You'll

8:57

you'll see in the top row

8:59

are the usual development use cases

9:01

right? Whether you're developing

9:03

your data infrastructure pipeline with a support workflow

9:05

whether you're doing some kind of product development

9:07

whether you're doing security engineering or data science

9:10

for sure, these are great fits for Cloud Code today

9:13

However

9:14

a lot of our different internal teams are

9:16

also using Cloud Code. So Emily

9:18

I'm going to turn it back to you

9:19

I would love to hear some of your input

9:21

because I know you're talking to a lot

9:22

of our customers

9:23

How, how do folks that are non-technical use Cloud Code today

9:27

or how do, you know

9:28

people even within Anthropic that are non-technical use

9:31

Cloud Code today?

9:33

Yeah

9:33

it's a really fun tool to kind of

9:34

dive into and see what you can build

9:36

if you're a non-technical person like myself

9:39

here on the sales team at Anthropic

9:41

we actually have had a couple of folks

9:43

who

9:43

have built out things like a pricing calculator or account plans

9:47

and those are just some sales use cases

9:49

I actually work with a hedge fund here

9:51

in New York City where

9:53

their CEO mandates that every person all the

9:54

way down to HR

9:57

uses cloud code. So there's a lot you

9:58

can do there with any kind of analysis

10:01

working across the board with cost modeling

10:03

P&Q assumptions. It's kind of fun to

10:05

to

10:06

get your hands on it and see what

10:07

you can build

10:08

Amazing. Thank you for sharing

10:09

So hopefully that gives you some ideas

10:11

and you can see on the bottom row

10:12

as well how some internal anthropic employees

10:16

use Cloud Code for things like growth marketing

10:19

or product design or legal for their different workflows

10:22

internal applications, and prototyping

10:24

Okay, so we're going to have our first poll

10:26

so

10:27

you're going to be able to start getting interactive truly with

10:30

you know

10:31

some of the content that you're seeing here

10:33

So I would love to know

10:34

you know, where, if you're using Cloud Code today

10:36

where you're using it

10:38

Is it in the terminal?

10:39

Is it in, you know, an IDE?

10:41

Is it in the web?

10:42

Is it on mobile?

10:42

So feel free to participate in the poll

10:44

Would love, the anthropic team here would love to know

10:47

you know

10:48

kind of your

10:49

user experience and where you're using

10:51

Cloud Code today

10:52

But we'll keep moving on

10:53

And as you can see within the slide you're seeing here

10:57

you can take Cloud Code wherever you need to

10:59

right? So if you're

11:00

familiar with the terminal and you don't feel

11:02

you know, as

11:04

it's as daunting of an environment to be in

11:06

you can certainly stay there

11:07

However, if you're Used to a code editor

11:10

you can bring it there

11:11

We have plugins that work natively with

11:13

a lot of the primary IBEs out there

11:16

and then you can

11:16

access Cloud Code on the web and iOS

11:19

We even have a feature where you can

11:21

pretty much teleport or remote control your

11:24

terminal on your laptop from your phone

11:26

So if you want to code embed

11:28

you can code embed

11:29

All right

11:30

perfect

11:30

I don't know if we have any questions

11:32

that are coming up so far

11:33

but I would love to pause for

11:35

one question

11:36

if we can answer something from the crowd so far

11:41

Yeah, maybe we can start with this one

11:44

Is there

11:45

an objective function set for an agent based

11:47

on the user's prompt

11:48

or is it all direct

11:50

is it all directed

11:51

just very directed, predictive probability

11:53

or a combination of both?

11:56

interesting

11:57

so I think there's a lot of parts

11:58

to that question

11:59

Ali, I don't know if you have any

12:02

thoughts on your side on this

12:04

but in, in terms of prompting in general

12:07

predictability in ter, versus kind of prompt techniques and prompt engineering

12:12

a lot of what we internally Anthropic advise customers on is

12:18

you know, there's, with LLMs and with generative AI in

12:23

at scale or at

12:24

at a whole

12:26

it's not going to be able to be

12:27

predictable in terms of the next

12:28

next steps that it takes

12:29

right? It's truly what word comes next is how the technology

12:34

works in terms of LLMs

12:36

So when you're prompting an LLM

12:38

especially in cloud code

12:39

you don't have to be perfect

12:40

there can be typos

12:41

there can be ways of

12:42

you know, getting to the same goal overall

12:45

but I would say

12:47

some things that

12:48

will really help is we do have prompting guides

12:50

online at Anthropic

12:52

that you can definitely search up

12:53

If it's not in the docs

12:54

we can certainly link them

12:56

and then we also have something called skills

12:57

which

12:58

I don't know if I'll cover

12:59

in too much of a detail today

13:01

but

13:02

they are tools that you can use in cloud code to

13:05

that are already pretty much packaged up

13:07

to get you further than

13:09

pretty much a blank slate or a blank text box

13:11

Ali, anything else to add there

13:13

that, that I Yeah, a couple of things that help me

13:17

on top of what you said

13:19

adding examples in instructions

13:22

helps steer Claude, well, as well as constraining the output format

13:27

as XML tags or

13:29

or JSON schema, that also, helps bring, some predictability, some structure

13:34

to the response

13:36

Amazing. Thank you, Elaine

13:38

And, and if we can just answer one more question

13:40

cause we're getting this a lot in the

13:41

chat here. I think there's

13:42

there are a lot of questions around how

13:43

to get started. Like

13:45

is it directly in the terminal?

13:47

Can people do it on their phone?

13:48

If you can just quickly cover how to get started

13:50

I think that'd be helpful

13:51

Yeah. In the docs view

13:53

there should be a link

13:54

that, opens up, I think

13:56

learn more about Cloud Code

13:57

It should be the Cloud Code landing page

13:59

You can install this directly on your terminal

14:01

right away and get started

14:03

if you want to

14:03

you can download the Cloud Desktop application

14:06

and there is a code tab in there

14:09

You do have to link it to a GitHub repo

14:12

so that is a requirement to actually get

14:13

started there

14:14

and then you can

14:15

obviously take it on the go

14:16

in mobile as well. You do also have

14:18

to link that to a GitHub repo

14:19

But the easiest way

14:21

I know it's a little daunting

14:22

especially if you haven't worked out of a terminal before

14:24

but if you have access to your terminal

14:26

you can go to the overview page

14:28

Just go ahead and

14:29

there's going to be a little copy

14:30

button

14:31

on that page

14:32

Go ahead and copy and paste that into the terminal

14:34

and you can get started right away with that

14:37

Awesome, thanks

14:38

Amazing. Okay, let's keep moving forward

14:40

So, we're going to start our foundation section

14:43

It's, it's a fairly short section

14:45

but I wanted to go over the

14:46

the very kind of basics

14:47

but maybe most important knowledge of

14:49

starting to use Cloud Code

14:50

So

14:51

we've talked a little bit about where you

14:53

can access Cloud Code. I

14:55

this point I'd like to emphasize because a

14:58

a lot of developers or folks that are

15:00

you know

15:00

writing code on a day-to-day basis have

15:03

their environments that they like to work out

15:04

of

15:05

Whether it's an IDE like VS Code or JetBrains

15:08

whether it's another platform that they

15:11

they like to, to work out of

15:12

you can bring Cloud Code

15:13

again, with you wherever, wherever you need to

15:15

And

15:16

the great thing about Cloud Code is

15:17

it will interact with certain native parts of

15:20

those code editors as well

15:22

So I

15:22

I definitely don't want to skip that

15:25

ClaudeMD

15:27

So

15:27

you might have heard of this concept called

15:30

ClaudeMD. If you haven't

15:31

I'm totally okay. But if you look on the screen here

15:35

what ClaudeMD is

15:36

it's pretty much like a

15:37

memory file for our models

15:40

So when you're using Claude Code

15:42

some of you that have

15:43

already been using it might notice that

15:45

if you get to the end of a conversation

15:49

the model has a limited

15:50

what we call context window

15:52

I'm going to refer to it as the brain

15:53

right? So

15:54

if the brain runs out of space as

15:56

you're back and forth with this model

15:58

it's going to need to

15:59

clear the brain to keep moving forward and

16:01

take additional actions

16:02

So what ClaudeMD does is when the brain is cleared and

16:07

you know, it's a blank slate

16:09

it's wiped, it's going to keep moving forward

16:11

ClaudeMD keeps the most important things

16:14

in the brain saved so that every new session

16:17

every time you clear the memory or the context

16:21

quote unquote

16:22

you're going to be able to save those important best practices

16:25

whether it's Project structure

16:27

common commands, styling, tips, et cetera

16:30

all of this can go into a CloudMD

16:31

CloudMD is fairly lightweight

16:33

so you're going to

16:34

only add 30 to 50 lines into it

16:38

Everything that you're seeing here in terms of

16:40

you know, whether it's CloudMD or any other tools

16:42

you can absolutely

16:44

have Cloud make for you on your behalf

16:45

So if you're using Cloud Code

16:47

you're going to use the slash init command

16:49

slash I N I T

16:51

to actually start a CloudMD

16:53

in your local directory as well

16:56

and I

16:57

I want to kind of connect it to

16:58

what folks might be seeing

17:00

If you're seeing maybe

17:02

Claude behave in a way that

17:03

is unrelated to the topic you're asking about or

17:06

development flows that you are working on

17:09

it might be because there's too much in

17:11

its brain and you need to clear it

17:13

and maybe keep the most important things in

17:15

the Claude MD

17:17

so

17:18

I wanted to illustrate this in

17:20

more of a pictorial fashion or

17:22

or a flow diagram. And as you can see

17:24

there is a repository here

17:26

a monorepo

17:28

and you can see you can have Claude

17:29

MDs at every single level

17:31

So Claude MDs are useful for

17:33

you know

17:33

at the top level describing

17:35

what this system is doing and maybe what

17:37

teams are mapped to

17:38

to the different

17:40

subproject folders or modules within this system

17:43

And then as Claude actually goes down and

17:45

explores these different folders

17:47

it starts pulling the different Claude MDs

17:49

into the brain as well

17:51

for, you know, the relevant context

17:53

If I'm exploring the front end folder

17:55

I want to pull in In the usual UI patterns

17:58

in the test strategy

17:59

in the components. I'm exploring the test cases

18:01

I want to understand how deeply

18:04

you know

18:04

the test coverage runs and my automation strategy

18:06

So

18:07

if there's one thing that I can leave folks with today

18:09

one of the things that are important

18:11

I would say maintain your context

18:13

aka the brain

18:15

and then using CloudMD to

18:17

give Cloud the context that it needs

18:19

and the best practices to

18:21

save over and over again as you have

18:22

new sessions with Cloud as well

18:25

Okay

18:25

perfect. So we're going to have poll number

18:27

two come up. So if you are using Claude Code today

18:30

or even Claude in chat

18:32

in cowork

18:33

you might see that you're able to

18:35

select different models. So we'd love to know

18:37

you know, which models you're using today

18:40

And you can even

18:41

put

18:41

in the Q&A if you want to

18:43

say why you're using certain models

18:44

you can do that as well

18:46

So when you're selecting a model in Claude Code specifically

18:49

you might see a few different models come

18:50

up

18:51

We have what are called our model families

18:53

So model families include Opus

18:56

Sonnet, and Haiku

18:58

And Opus, I'm going to start from the top

19:00

is our most performant model family

19:02

We have Opus 4.6 or 4.6 out now

19:05

And it's very capable in terms of very complex tasks

19:09

like exploring an archaic code base deeply

19:13

understanding architectural changes that you're requesting

19:16

how to refactor that or modernize that

19:19

and then

19:19

provide a plan for you to implement

19:22

As we go down

19:23

Sonnet is our, we, we

19:24

Like to call it our daily driver here

19:26

at Anthropic

19:27

So if you as developer

19:29

you know

19:29

have a plan already put together that's highly

19:32

specific and highly complex

19:33

and you need Sonnet to do

19:35

the heavy lifting of just implementing that specific plan

19:38

you can go ahead and use Sonnet

19:40

And then Haiku is used for things like writing tests

19:43

or maybe more specific point

19:46

you know

19:47

solutions or point features

19:49

that you might just want to do a one-off of

19:51

Again, as you're using Cloud Code

19:53

I would actually encourage

19:55

you to flip between these models to see

19:56

what makes sense for parts of your workflow

19:59

A lot of people ask me

20:00

which models should I use every single day?

20:03

I always say it depends on what you're

20:05

doing in Cloud Code

20:06

Some people leave it on at Opus all the time if

20:09

you know

20:10

they don't mind kind of

20:11

some of the thinking that it does and

20:13

the way that actually

20:14

you know, gets to the answer and outputs

20:16

And some people will flip between the models

20:18

based on

20:19

the workflow that they have as well

20:23

Okay

20:24

so I think we can pause here for another question

20:27

Emily, Ali

20:28

any questions in the chat that we can

20:30

help answer again?

20:32

Yeah, this is a good one

20:34

so the, the most upvoted question is

20:36

do the terminal and desktop applications provide the same experience

20:39

or are they designed for different purposes?

20:42

It's a really good question

20:43

So the terminal is where

20:46

a lot of our native Cloud Code features live

20:49

and it's where you probably get the

20:51

highest amount of kind of experience that

20:55

developers often need for their day-to-day workflows

20:58

What I mean by that is there's a

20:59

lot of slash commands

21:01

and it has the ability to work

21:03

obviously, with your local desktop and your files and folders

21:06

and use all of these

21:08

what we call agentic tools

21:09

to actually interact with your code base

21:12

The desktop application, as well as Cloud Code on web

21:17

has to be connected to a GitHub repo

21:19

to actually operate. So you do have to

21:22

you know, authenticate to GitHub

21:24

and then sync it with a specific repo

21:26

and then it will

21:27

launch pretty much a web version of Cloud Code

21:30

and then do operations on your behalf there

21:32

Certainly, I would say this

21:35

If you are maybe onboarding quickly onto a code base

21:39

and maybe you're a new developer on the team

21:41

you're getting used to Cloud Code

21:42

Code

21:43

start

21:43

using Cloud Code on desktop or web is

21:45

a great experience to start trying it out

21:47

and testing it

21:47

However

21:48

if it's

21:49

part of your daily workflow and you're really

21:51

deeply developing in Cloud Code

21:53

and you have the repository cloned to your computer already

21:56

I would

21:57

suggest definitely using terminal for Cloud Code

22:02

Perfect. And Ali, Emily

22:03

anything you would add there as well?

22:07

I would say what I typically hear in

22:08

the field is that it's preferential to developers

22:11

depending on where they like to work

22:15

Amazing

22:16

Perfect. Okay

22:17

Let's keep moving on. And

22:19

just for, to remind folks in this room

22:22

we are going to have a demo at

22:24

after this portion. So if you're on the

22:26

edge of your seat

22:27

you're not going to have to wait any

22:28

longer. So we will get

22:29

get through this

22:30

and then we're going to get to a

22:31

live demo

22:32

Fingers crossed that it works

22:34

cool. Okay. Let's

22:35

let's get back to this concept of context

22:37

So we talked a little bit about this

22:38

word context

22:39

What does it actually mean?

22:40

So when you think about context

22:42

just think about pretty much it equating to

22:44

words that the models actually understand

22:47

right? So

22:48

the more context or words you put in a model's brain

22:52

the more it's going to have to think

22:53

through and process

22:54

to understand what your end goal as the

22:56

user is

22:57

so if you curate that context

22:59

like you feed the specific files or the

23:02

you know

23:02

you're a little bit more specific with your

23:04

ask in terms of what you want Claude

23:06

or Claude code to do

23:07

it's going to lead to more efficient and

23:09

better performance. And likewise

23:11

as you bloat the brain or the context window

23:14

the model needs to clear it

23:16

and it

23:16

might lead to performance degradation as well

23:19

so how do you curate your context then?

23:22

Let's, let's go through some commands

23:25

so if you're in Cloud Code

23:26

you can use

23:27

what's called the slash context command

23:29

The slash context command

23:30

is going to give you an

23:32

inside view on the model's brain

23:34

pretty much. So what is taking its context?

23:36

Is it the prompt

23:38

the system prompt underneath the surface

23:40

that Claude Code is operating on to understand

23:42

how to actually interact with the user?

23:44

Is it the messages that you've

23:45

you know, talked back and forth with Claude?

23:48

Is it any of the tools that it's calling

23:50

like

23:51

connectors or, you know, other sub-agents that

23:54

that it might be summoning?

23:56

is it, anything else that

23:58

that might be loading the context?

23:59

Or is it even the Claude MD files

24:00

that you're loading in? So

24:02

the more understanding you have of what's in context

24:04

you can actually curate

24:06

obviously what's called

24:08

and then you can use the slash compact command

24:10

So compact, if I were to explain it very simply

24:12

is when

24:14

the brain is almost getting to the point where it's overloaded

24:17

you can actually, as a user

24:18

trigger a command to summarize everything that the

24:22

the, the model has seen so far

24:24

and then clear the model's brain

24:26

and then give the summary

24:28

only keep the summary

24:29

In the new session and then keep going

24:31

So it, it

24:32

it gives Claude Code

24:34

an understanding of what's been done so far

24:36

in a summarized fashion

24:37

and then you can keep moving forward to

24:39

accomplish your task

24:40

so I would use these two commands

24:42

very

24:44

frequently

24:44

to understand what's going on in your context

24:46

and then to clear and summarize your context

24:49

as needed

24:51

Okay

24:52

So, at file mentions

24:54

this is another way to curate your context

24:56

So

24:56

if

24:57

you as the user know

24:58

exactly what file you want to refer to

25:01

while you're talking to Claude Code

25:03

instead of having Claude Code search

25:05

and it's going to use some local tools

25:07

to actually search your computer for the relevant files

25:10

you can actually do what's called at file

25:12

mentions. You can at mention

25:13

hey, app.py. You can at mention

25:16

you know, a JSON file or a folder

25:18

et cetera. So

25:18

if you want to pull things specifically into context

25:20

this is a great way to specifically do

25:22

that as a user if you know what you need to

25:26

Claude to start working on

25:29

Okay, so next poll, we're going to

25:33

get a little information on plan mode

25:34

So it's going to be a

25:36

do you use plan mode today

25:37

yes or no?

25:38

And then if you do or if you don't

25:41

feel free to participate in the Q&A

25:43

you know

25:43

how you find it useful or if you

25:44

have any questions about plan mode

25:46

feel free to ask them there

25:49

the, the reason that some

25:51

some folks might use plan mode

25:53

and I'll talk a little bit about what it is

25:55

is

25:56

it is a

25:57

way to actually allow you as the developer

25:59

or even someone that's non-technical to

26:01

use Claude Code

26:02

to have Claude

26:04

create a full end-to-end implementation plan for you as you're going

26:09

and then have it outline stepwise what it's planning on doing

26:12

maybe some edge cases that it's seeing

26:14

as well as maybe some assumptions that

26:17

aren't clear. It's going to ask you some

26:19

questions to clear those up

26:21

so that it doesn't guess while it's planning

26:24

folks that have used a lot of other tools

26:27

you know, will find out maybe When I'm asking something of

26:31

an AI coding tool

26:33

it's gonna get to an end goal

26:34

but it might not have all the details that I want

26:36

right? So

26:37

plan mode is a great way to start

26:38

if you're implementing something complex

26:40

if you need to explore your code base

26:42

if it's a huge refactor

26:43

it's a really tricky item that you're tackling

26:46

I would highly suggest

26:47

going in plan mode first

26:49

Ali, I would love to know maybe some

26:52

you know, information from your side

26:53

How have you seen plan mode used

26:57

either, you know, in your day-to-day work at

27:00

Anthropic or even in the field as you're

27:02

talking to customers?

27:04

Both in my work at Anthropic

27:05

as well as in the field and in

27:07

my personal projects also

27:09

essentially

27:10

I see plan mode being used before

27:13

any task

27:14

It could be a small task

27:16

big task, a big project, a refactor

27:18

Before any of them

27:20

it's a great idea and we recommend it

27:22

to use plan mode. Instead of directly jumping into the code

27:26

results in much better

27:27

results

27:29

and also

27:30

it's not just meant to be for really large changes

27:32

small changes are, are good too

27:34

Think of

27:35

treat the plan mode like a

27:36

PR description that you're writing before the code

27:39

If the plan doesn't read so good

27:41

the code most likely won't

27:43

So that's a good way for you to correct Claude

27:46

as well as give more inputs

27:48

Or sometimes Claude also starts to interview you

27:50

to clarify on things that

27:52

aren't clear to it yet

27:55

Amazing. Thank you, Ali

27:57

so

27:57

I highly encourage everyone to try out plan

27:58

mode if you haven't already

28:00

certainly for very well-scoped prompts

28:03

maybe you can start with

28:04

just normal mode or accept edits on

28:06

but plan mode is especially useful for

28:10

the topics that Ali and I mentioned

28:12

Okay

28:13

So for the next four or five slides

28:15

I'll go through fairly quickly

28:18

I would love to

28:19

get to the demo and leave some time

28:20

at the end for some questions

28:22

But I do want to emphasize the ability

28:24

to not just send text

28:26

to the screen here within Cloud Code

28:29

but you can also drag and drop screenshots

28:32

as well. So

28:33

if you are especially a front-end engineer or a developer

28:36

or you're working with maybe visual elements and design

28:39

you can actually screenshot

28:41

what's happening on your browser or an application

28:43

et cetera

28:44

and then drop that directly into Cloud Code

28:48

What that allows Cloud to do

28:50

is it processes the image

28:52

and then

28:52

understands

28:53

the prompt that you're pairing with it to help either debug

28:56

improve, or modify what you're working on within your project

29:00

right? So it's as simple as dragging and dropping

29:03

or if you want to copy

29:05

a screenshot or a photo directly from your desktop

29:08

you can do that as well

29:09

So an example flow might be

29:12

you might want to Squash a bug from a screenshot

29:14

maybe there's elements within a front-end UI that are overlapping

29:18

you can screenshot that

29:20

paste that into Cloud Code

29:22

You can use plan mode like Ali and

29:23

I talked about to plan the fix to

29:25

make sure

29:26

that Cloud Code is going to operate it on it correctly

29:28

and then you can certainly review and verify

29:30

that within your code editor or your terminal

29:34

Okay

29:35

So I'm not going to

29:36

spend too much time on bash mode

29:37

However, bash mode is a way that

29:40

folks that are using the terminal can interact with the terminal

29:44

in

29:45

pretty much a raw way where you can run terminal commands

29:48

scripts directly within Cloud Code

29:50

So no need to open a new tab

29:52

You can run your terminal commands

29:54

also directly in Cloud Code as well

29:57

And then

29:58

the last two commands are very useful in

30:00

the day-to-day operations and

30:02

and workflows. This is why this section is labeled workflows

30:05

but

30:05

hopefully this helps you in your workflows as

30:07

you're using Cloud Code. But if you press the escape button

30:09

you can actually interrupt

30:11

what Cloud Code is doing immediately and redirect it

30:14

So myself, as well as the rest of

30:17

you know

30:17

our internal and anthropic employees and our customers

30:19

tend to use this a lot

30:20

especially if you need to

30:23

you know

30:24

you see Cloud going in a direction that you know it

30:27

should not be going

30:28

and you can give it additional context

30:30

I want to stress that as you press escape

30:32

Cloud does not

30:33

lose any of The previous conversation or context

30:36

that it's thinking through

30:37

you can just redirect it

30:38

It's going to understand

30:39

your new prompt as well

30:41

as the direction it was going on previously

30:43

If you hit escape twice quickly

30:46

it will rewind

30:47

any changes that it's made

30:48

in terms of the last

30:50

kind of code modification

30:51

So if there's anything that you didn't want it to modify

30:54

you can actually

30:54

do what's called rewind the change as well

30:59

And then resume

31:00

So

31:01

if you

31:01

somehow exit out of your terminal window or

31:04

your session in your IDE with Cloud Code

31:06

you can actually just

31:07

hit slash resume after you enter into Cloud

31:10

Code to pick up where you left off

31:12

This is great for

31:14

you know

31:15

going back into a session where you were

31:17

implementing something deeply and you got interrupted

31:19

or internally in Anthropic

31:21

we actually use this

31:22

to test maybe different approaches or implementations

31:25

on a feature that we're working on

31:26

and then maybe choose the best session to

31:28

resume as well

31:30

Okay

31:30

So before we hop into the live demo

31:33

let's maybe do one question

31:34

and then pivot over the demo so everyone

31:36

can see

31:37

a little bit of Claude code on the

31:38

screen as well

31:40

Awesome. Thanks, Harry

31:41

The top question that we have

31:42

and we get this one all the time

31:44

on customer calls as well

31:45

does Claude store information about what it's done?

31:48

Should I be concerned about having it analyze

31:50

data with PII in it?

31:53

that's a good question. So again

31:55

Cloud Code, is, operating locally on your computer

31:59

and we do not store anything in the

32:01

cloud. So we will send

32:02

some of

32:03

what

32:04

the prompts are and some of what

32:05

Cloud is analyzing to the cloud just to

32:08

to actually, you know

32:10

run what we call inference on it

32:12

and actually have it go through our models

32:14

but we don't actually store anything on our

32:15

side. And there's actually a lot of security

32:18

specific

32:19

features baked into Cloud Code that I won't go over in

32:22

in this introduction session

32:24

that you can actually configure to have it

32:25

stay away

32:26

from certain security elements or files or folders

32:29

that you actually don't want it touching

32:30

So that's a really good question

32:32

but within the Cloud Code product

32:34

we definitely take security and compliance seriously

32:37

which is why we bake some of those things

32:38

within the product as well

32:41

Awesome. Thanks

32:41

Harry. I also just linked our trust center

32:44

in the chat here in response to that

32:45

question

32:46

That is a wonderful resource

32:48

It has an FAQ in there

32:50

It has a lot of documentation

32:52

It's probably a good place to get started

32:53

if you have specific security questions

32:55

Amazing. Okay

32:57

And with that, we will go to the demo

32:59

So let's, I'm going to stop sharing on slides

33:04

and then let's start sharing

33:07

Screen, okay

33:09

one of the, my co-presenters

33:10

can you verbally tell me if

33:12

my screen is showing perfect?

33:14

Okay

33:15

So we are starting in a demo repository

33:17

If you're a developer

33:18

this might look familiar to you

33:19

We're starting in GitHub. So I'm a developer

33:22

I'm a front-end developer

33:24

and I'm working on

33:26

a landing page that

33:27

might look familiar to some of you

33:29

So this is the

33:31

web registration page

33:33

that you yourself actually

33:34

use to register for the webinar that you're

33:36

on right now. So

33:38

very meta, I know, but I think it illustrates nicely for

33:42

for the purpose of the demo today

33:44

So

33:44

right away you might see some things that

33:46

are maybe a little off

33:48

I'm going to point out

33:49

this hour should not be here

33:50

Obviously this, this webinar is in the future

33:54

It's July 15th of this year

33:56

However, it should just be 133 days

33:59

and the hours looks awfully

34:01

you know, different than what it should be

34:03

So there might be a bug in this code

34:04

right?

34:05

and as I scroll down

34:06

I'm previewing this page. It looks okay to me

34:08

but

34:09

maybe there's some other modifications we'd want to

34:11

make

34:12

and if we take a look at the issue in GitHub

34:14

in the repository that I'm working out of

34:16

it looks like there's already been a bug

34:17

report that's been filed

34:18

right? So when viewing the landing page

34:21

the countdown timer displays an incorrect number of

34:24

hours

34:25

so we've already

34:27

been made aware of the bug

34:28

How do we now fix that

34:30

with Claude Code

34:31

by working directly in the terminal and not

34:33

needing to copy and paste things and

34:35

and work back and forth?

34:36

So let's, let's go ahead and

34:37

and take a look at what Claude Code looks like

34:39

Okay

34:40

so now we're in our

34:42

IDE or our development environment

34:44

I'm using VS Code. Again

34:45

you can use a lot of different platforms for this

34:48

and I'm within my project folder

34:50

I've cloned it. It's on my local laptop

34:52

I'm ready to go

34:54

So you can see Claude Code

34:55

I'm in Opus 4.6. You can choose between different models

34:58

certainly

34:59

What I want to introduce first is the

35:01

Claude MD file

35:02

So

35:03

we talked a little bit about this in

35:04

the foundation section within this webinar

35:07

but what the Claude MD file is

35:10

again, is it's the readme for the Claude model

35:13

right? It understands what the project is doing

35:15

some of the architecture

35:16

how it's structured, some of the components of the project

35:20

how you run it

35:20

how you test it

35:21

maybe some of the best practices in terms

35:23

of conventions. And you'll notice

35:24

again

35:25

it's very short. It just has that architecture

35:27

It just has how you run it

35:28

the conventions

35:30

and some of the other things that might be important

35:33

we will cover in a more intermediate and

35:35

advanced section of how you save some of those components

35:38

but these are the main best Practices

35:40

So let's say I'm onboarding onto this issue

35:42

I just want Claude to go ahead and

35:44

explain this project and what it does

35:47

how it's structured, how do I run it

35:48

etc.

35:50

so I'm going to go ahead and hit enter

35:52

and I haven't done anything else but ask

35:54

this one question

35:55

And Claude, go ahead

35:56

go ahead and start thinking

35:58

So it looks like

35:59

this is a local recreation of that registration

36:01

page. It's going to give me that structure

36:03

It's going to give me certain constraints

36:05

how to run it

36:07

so very similar to that ClaudeMD

36:09

And you'll notice

36:10

this is because the ClaudeMD is already loaded

36:12

into context. So if I hit context again

36:16

you'll be able to see exactly what's

36:18

in the context right now

36:19

So we have some MCP tools

36:21

some memory files, some skills, etc.

36:23

Right?

36:24

Okay, perfect

36:25

So

36:26

why don't we go ahead and do this?

36:27

I'm going to go ahead and clear just so

36:30

The, the screen isn't loaded

36:31

and I'm going to go on to

36:33

to the next prompt

36:34

So you'll notice I am not calling any kind of connector

36:38

I'm not calling, you know

36:40

any kind of API

36:41

I'm literally just saying

36:43

there's an open issue on this repo

36:45

about a countdown bug

36:46

Can you pull up all the details?

36:48

Can you investigate the issue itself?

36:50

And then can you run the tests?

36:52

So right away you can see it's using a

36:55

a command to actually reach out to the issue

36:58

which is, again, remind everyone it's here

37:01

And then it's running the tests within this environment

37:04

and then if you see above

37:05

and I'll, I'll pull the terminal window down

37:07

it's seeing

37:08

it's already telling me what it's going to

37:09

change. So Claude was a little too quick

37:12

The demo is

37:13

my talk track is going slower than Claude

37:15

is working

37:16

But you can see it ran the tests

37:18

and if I hit control O

37:19

I can see

37:20

everything that's outputting in terms of the tests

37:22

So there are two failures

37:24

And then it found the issue on GitHub

37:26

you know, this is what the issue looks like

37:28

this is what the bug looks like

37:30

and can I make these changes now?

37:31

So you'll notice that

37:33

Claude Code does not

37:34

make the changes until I ask it to make the changes

37:37

because I'm in normal mode

37:39

I will talk about another mode later

37:41

as we get to the next prompt

37:42

But I'm going to go ahead and say

37:44

yes

37:44

I want it to make the edit because

37:45

the change looks correct. Okay?

37:48

So it's going to go ahead and make the edit

37:50

and it's going to even verify the fix

37:52

So it's going to rerun the test

37:53

and you'll notice

37:54

I did not ask it to rerun the test

37:56

but it's going to

37:56

go ahead and do that

37:58

And then now you'll see that all

38:00

17 tests pass. If we go back to the page

38:03

if everything looks correct or went correctly

38:06

you'll now see that the hours looks

38:09

on par with what we should be getting

38:11

in terms of the countdown

38:12

So big thumbs up there

38:13

We went through our

38:14

our, our first workflow

38:16

Okay

38:17

So let's now go on

38:18

I'm going to hit shift tab now

38:21

to accept edits on. What this mode does

38:23

is

38:24

I want Claude Code to generally accept

38:26

most of the edits that it's doing without

38:28

me as the user needing to approve it

38:29

On some

38:30

maybe permissions or operations that it

38:33

still needs to prompt me on

38:34

it will prompt me

38:35

but for a lot of them

38:36

now I just want Claude Code to pretty much cook

38:39

right? So

38:40

let's go ahead and do what's called

38:43

Call a skill now. So I want to emphasize

38:45

I didn't cover this yet

38:47

but in the next section

38:48

we will actually cover what a skill is

38:51

What a skill means

38:52

and I'll talk about this very shortly right now

38:55

is it's a tool

38:57

that you can use in Cloud Code

38:58

to

38:59

maybe complete the same task over and over

39:01

and over again. And

39:02

it's just described in natural language to Cloud

39:04

so you'll see this one

39:06

is adding an animation. So creating an

39:08

animation is going to ask for a photo

39:09

analyze the image, create the pixel art animation

39:12

structure the scene in a certain way

39:14

right? So

39:15

what I'm going to do

39:15

is I'm going to go through a couple

39:17

of photos we have on our side

39:18

So you see Ali

39:19

you see myself and Emily

39:20

I'm not going to pick on my co-presenters

39:22

so I'm going to

39:23

go ahead and use my own photo

39:25

And again, I just drag and drop it into Cloud Code

39:28

and I hit create animation

39:29

Again

39:30

I think we had the question earlier about

39:32

How do I prompt Claude well?

39:34

Is it probabilistic? You know

39:35

et cetera

39:36

So

39:37

you'll notice that

39:38

this skill takes all the prompting knowledge out of it

39:41

right? You don't need to know how to

39:42

prompt. You can just accomplish a workflow

39:45

codify that into a skill

39:46

and we'll talk about skills in a bit

39:48

And then Claude will go ahead and

39:51

be able to surface that skill so you

39:52

can call it over and over again

39:54

without needing to think of the right prompt

39:56

to operate on. So what it's doing now

39:59

and if I hit control O

40:00

is it's thinking, it's reading the relevant files

40:03

You can see it even processed my photo

40:05

right? So it said my hair is black

40:08

short, styled up, that which is true

40:11

Skin tone, glasses, clothing, clean shaven, friendly smile

40:14

I would hope so

40:15

hopefully

40:16

But

40:17

you can see now it's already making changes

40:18

It's adding specific code to

40:22

I believe this is the CSS

40:24

file within the actual project folder

40:27

And it's not asking me now for permission

40:29

because I clicked accept. Edits on

40:32

So it's not going to

40:34

ask me for the edits

40:35

It's going to go ahead and add those

40:36

to the code base

40:37

And now it looks like the animation's already

40:39

live. So the pièce de résistance

40:43

if we refresh this

40:45

we have a side scrolling animation

40:48

and it's a little mini representation

40:50

Doesn't look entirely like me

40:52

but as close as it can get via one photo

40:54

right? So, and I'm just walking with my laptop and saying

40:58

let Claude cook

41:00

Okay, so we've done a lot here already in this

41:02

you know, five-minute demo

41:04

Let's now push our changes

41:05

Let's, you know, kind of close this issue

41:08

So why don't we now tell Claude

41:10

can you commit all these changes?

41:12

Can you make sure that you save all of them

41:14

push that to the repository on GitHub

41:18

and then open a pull request

41:20

that closes the

41:21

previous issue as well as

41:23

pretty much takes note of the animation that we added

41:25

right? So we're going to go ahead and press enter here

41:28

and then it's going to think

41:29

it's going to be able to

41:30

look at the status of the repository

41:33

and then ultimately sync it to

41:36

our GitHub repository

41:37

So it's thinking through that process

41:39

it's taking a look at the log

41:41

it's going to add all the relevant files

41:43

to the staging area

41:45

and then it's going to go ahead and

41:46

push and create a PR

41:48

So you can see all of that happening

41:50

sequentially in the thought process here

41:53

and all of the tools that it's calling

41:54

So whether it's a bash tool

41:57

so I guess I want it to proceed

41:58

because it's doing something you know

42:00

that is a little more sensitive

42:01

It's going to ask me permission for that

42:03

And then now the PR is open

42:05

So if I'm going to go ahead and copy this

42:07

and then let's go to our browser

42:10

Let's see if this opens

42:12

And now we have our poll request

42:14

We have, you know, it

42:15

we fixed the countdown timer

42:16

We've added the animation now

42:19

and visually I have checked this off

42:23

now

42:23

the last thing that I want to do

42:25

in this demo is

42:27

now I want to have Claude review the

42:29

code

42:30

So I can actually just go ahead and say

42:32

Claude, make sure I tag Claude correctly

42:35

I think this is Claude

42:36

tagged correctly. So I can say

42:38

Claude, review this

42:40

pull request for correctness

42:41

and anything you'd flag in a real code

42:43

review

42:43

So it's going to have little eye emojis that pop up

42:47

perfectly queued, and Claude bot has seen this pull request

42:51

and it's going to go ahead and go

42:52

through all the code and review it

42:53

and if it comes back with thumbs up

42:55

then I'm going to go ahead and merge that

42:58

Obviously, after reviewing the code myself as a human

43:01

merge that afterwards

43:02

So

43:03

that was

43:04

an end-to-end demo of Claude Code of pulling from GitHub

43:07

working on some issues

43:08

adding some content, opening a pull request

43:11

and then using Claude to actually review the

43:12

code

43:14

Let's hop back into our slides

43:18

Okay

43:18

so now we are

43:19

on the power features section

43:22

And we'll have time for questions at the

43:23

end

43:24

I want to be sensitive of kind of

43:25

where we are. I want to leave probably

43:27

seven to ten minutes

43:28

So

43:28

what we're going to do is go through

43:29

this last section

43:30

and then get to some of the questions

43:31

at the end and then close from there

43:34

So in terms of power features

43:36

so this is

43:37

I think another poll is going to

43:38

probably go off around this section

43:42

What

43:43

MCPs do you

43:45

currently find useful today and what category does

43:47

that usually fall into?

43:49

So you'll see a poll come up shortly

43:50

If it doesn't come up now

43:51

it will come up later

43:53

But the MCP section

43:55

I want to explain this very shortly

43:57

You can think of

43:57

if you haven't heard of the word MCP before

43:59

they are known as connectors

44:02

So they connect to external data sources

44:04

or they

44:05

connect to maybe even internal data sources within

44:08

your company

44:08

to pull information from or to send

44:11

information to from cloud code

44:13

So we here at Anthropic use

44:16

MCPs or connectors to

44:17

Maybe test some of our code in terms of

44:20

you know, Python automation

44:21

or to pull directly from somewhere like a GitHub

44:24

or from a task management software

44:27

or from Canva, or from Jira, etc.

44:30

So think about what

44:32

where you'd like information to start

44:34

coming into Cloud Code or sending it from Cloud Code

44:36

and that's where you'll

44:38

you'll find a lot of value

44:41

Okay, perfect. Let's keep moving forward

44:43

So skills, you saw what a skill did within the demo

44:47

so I was able to

44:48

call the create animation skill

44:51

towards the very end there to

44:52

add something new to the page

44:53

But what skills are

44:55

are really natural language files

44:58

bundled with a couple of different other files like scripts

45:01

like examples, to allow Claude, not just Claude Code

45:06

but our Claude ecosystem to operate on the same task

45:10

a specific task over and over again

45:11

So think about documenting your API when you're creating new endpoints

45:15

think about creating PDFs

45:17

think about styling your UI in a specific

45:19

way

45:20

You can actually create a skill for that

45:22

The best way to create a skill

45:23

I always tell folks

45:25

that are curious about how to create skills

45:27

are to actually have a conversation with Claude

45:30

maybe complete a workflow

45:32

and then just tell Claude at the end

45:33

hey, I want to create

45:35

use everything that I've done so far

45:37

and create that into a skill

45:38

And you can certainly do that

45:40

and then reuse that from then on

45:43

And then plugins

45:45

So, plugins are

45:48

a tool that you can use to

45:50

ultimately package up a lot of what we

45:52

talked about. So we've talked about connectors

45:53

we've talked about skills

45:55

and then there is a couple of

45:56

other concepts we didn't talk about

45:57

but

45:58

a lot of the other custom tools and

46:00

custom components that you create within Cloud Code

46:03

you can actually package up into what we call a plugin

46:06

and you can label it something like a front-end engineer plugin

46:09

or a back-end plugin

46:10

or, you know, a security plugin

46:13

and then start to share those within your

46:15

own organization or your company

46:17

to allow other people to get to the

46:19

same value that you are

46:21

with the skills and connectors and other

46:24

custom components that you're using today within your workflow

46:26

So, a lot of people ask

46:28

hey, what's the difference between

46:30

again, a CloudMD, a skill, an MCP

46:33

and a plugin?

46:34

You can see plugins are just a

46:36

an easy way to distribute skills and connectors

46:39

and then CloudMD

46:40

are these little memory files that always live

46:43

in a project folder

46:45

Okay

46:46

so

46:47

we're hitting the tail end of our presentation

46:49

I think we're, very good on time

46:51

so I'm very happy about that

46:52

but as we close

46:53

I'd like to remind folks

46:56

one thing that we talked about a lot

46:57

during the overview today is managing your context

47:00

So again, another review is, you know

47:02

how do you actively manage your context?

47:04

How do you view what that looks like?

47:05

How do you compact and make sure that

47:07

Claude is clear of irrelevant context before you

47:10

move on to the next

47:12

topic or the next task?

47:13

So again, use the context and commands

47:15

compact commands

47:16

and then even use clear

47:18

when you want

47:18

Claude's brain to completely clear and work on

47:21

something new

47:22

and then you can

47:23

use your Claude MD to save things over

47:26

across multiple sessions

47:28

and then of course at mentioning certain files

47:30

So, Ali, I'd love to, you know

47:32

hear from your side

47:34

as you're working on internal projects or advising our customers

47:37

what are some tips you have

47:39

for others on how to manage context well in

47:42

in terms of interacting with Claude Code?

47:45

Yeah, I'd say two things

47:47

One, to manage context

47:48

I often use the slash clear command

47:51

between tasks so that when I'm done with one task

47:55

I

47:55

I clear the context and then start a

47:57

new context to work on another task

48:00

and then the other one that I'll say is sub-agents

48:02

I, I started using sub-agents a lot

48:05

they're

48:05

I think they're a good cheat code for

48:07

big codebases

48:08

So if you need to read a lot of different files

48:10

a lot of different documents

48:12

or there needs to be a lot of

48:13

different smaller tasks that need to be done

48:15

I often instruct Claude

48:17

to spin up sub-agents to do all those tasks

48:20

and that helps keep the main agent's context window

48:23

cleaner

48:24

and then those sub-agents

48:26

report back

48:27

all of their results to the main agent

48:28

and it continues with the

48:29

with the main task

48:31

Amazing. I love that. So

48:33

sub-agents I know we didn't talk about today

48:34

but, if you're curious, they're, they are in our docs

48:39

and, if we host

48:40

host another webinar that's a little bit more advanced

48:42

we will absolutely cover that topic

48:44

So, again, another plug for AskClaude

48:48

if there's anything that we don't answer today

48:49

as the Anthropic team

48:50

you can certainly go to the docs

48:52

and get them answered there

48:53

They are oftentimes more updated than the knowledge that we have

48:56

even internally as employees here at Anthropic

48:59

but I will turn it now over to

49:00

some questions. I think we have

49:02

nine minutes for questions

49:03

which is perfect

49:04

so Ali and Emily

49:06

let's sort through some questions and hope

49:07

hopefully we can get some questions answered

49:09

on the session

49:11

Yeah. Hey, this one's, upvoted

49:14

what is the difference between workflow and skills?

49:20

so

49:20

workflow

49:21

and skills can actually mean the same thing

49:24

workflow is not necessarily a terminology within Claude

49:27

that is an official tool

49:28

let's say

49:29

But when I was referring to workflows in

49:31

the context of this presentation

49:33

I was meaning kind of some of the

49:35

you know, tasks and

49:37

workflows that you would do day-to-day as a developer

49:40

right? So if you are

49:42

exploring your code base in a very specific

49:44

way and taking notes on it

49:46

on that, that might be a workflow

49:47

And would you want to

49:50

or do you find yourself

49:52

doing that same workflow on a regular basis

49:54

whether it's multiple times a day

49:55

whether it's multiple times a week or a month?

49:58

If so, that's a great example of how you

50:01

what you can actually

50:02

port over into a skill

50:04

So

50:04

hopefully that gives a little differentiation on workflows and skills

50:07

Ali, Emily, any input that you'd like to

50:10

to share there as well?

50:15

That was good

50:16

Awesome, cool

50:18

next question

50:20

Yeah, let's jump to the next one

50:21

And let's spend a couple minutes talking about

50:23

this cuz the question is around best practices

50:25

and I think, you know, we internally, obviously, dog food, we call it antfooding

50:30

our own product

50:31

And, this question is from Naveen

50:33

At some point

50:34

we would love to learn

50:35

how successful teams coordinate when multiple people are

50:37

using cloud code on the same code base

50:40

Any best practices or patterns you guys would

50:42

recommend? That's a really good question

50:45

Ali, do you want to take this one first

50:46

and then I will come

50:48

come with an answer right after

50:50

Sorry, I could still the question again

50:51

just the last part

50:52

Yeah, all good

50:53

basically the question is around best practices for

50:55

folks who are

50:56

using cloud code on the same code base

51:00

Oh, yeah. we use branches and git worktrees a lot

51:05

within our teams

51:06

as well as even

51:07

if we want to launch multiple cloud code sessions

51:10

if, if one user wants to launch

51:11

let's say, five cloud code sessions

51:13

it helps to

51:15

Have each cloud or

51:17

have each team member

51:18

work on their own projects or work tree

51:20

And then in case there are any

51:22

merge conflicts that come up

51:23

also have cloud resolve some of them

51:25

This has been the main

51:26

main thing that's helped us

51:28

the most

51:30

Yeah

51:30

and

51:31

what I'd also add on to what Ali

51:32

is saying is

51:34

don't be afraid to use multiple cloud code

51:36

sessions in parallel. We didn't talk about it today

51:38

but that's how a lot of developers as well as

51:42

our own employees are using cloud code

51:44

So

51:44

if you're seeing

51:45

a part of the code base that you

51:46

can start implementation on

51:48

and then maybe another part that you can

51:50

actually start a plan on

51:51

and then maybe another part that you can

51:52

actually learn more about and explore

51:54

you can actually do all three of those at once

51:56

Obviously, it means, you know, multitasking, juggling a few different sessions

52:00

but especially if you have those guardrails in place

52:02

like, you know, you're on normal mode

52:05

so it can't do any edits without asking you

52:07

it's a great way to actually

52:09

make your workflow a little bit more efficient

52:11

as you're working on the same code base

52:13

as others that are

52:14

that are, doing the same tasks

52:18

Yeah

52:18

Thanks, Ali and Harry

52:20

And for folks on the line

52:21

please make sure you're continuing to upvote the

52:23

questions you want us to answer

52:24

because we're going to make sure we cover

52:25

those in the last five minutes here

52:28

Alright, let's talk about remote control

52:30

How can I use it on my phone

52:32

to keep it connected to my computer and keep building?

52:37

so, I actually did this this morning

52:39

I was working on a side project before

52:41

I got to the office

52:43

but

52:44

if you have Claude Code on your terminal in your computer

52:46

you can start a session

52:49

I believe, using Claude dash dash remote

52:53

Dash control

52:54

Don't cite me on that

52:55

Please look up the proper syntax around what

52:57

that looks like

52:58

But it's going to start a Claude code

53:01

Claude code remote session

53:03

in whatever working directory you activate Claude in

53:06

and then you're going to be able to

53:08

see it on your phone pop up

53:09

So when, when it pops up on your phone

53:11

if you have the Claude app installed

53:13

which if you don't

53:14

please

53:14

would love for folks to install that as

53:16

well on your mobile

53:17

you can go to the code section of the Claude app

53:20

and then be able to

53:22

see that pop up and then

53:23

operate

53:24

as if you were at your home computer

53:27

Some slash commands are not going to be

53:28

the same on mobile as they are in

53:31

you know, the terminal when you're at your computer

53:33

but that's how you take

53:34

some of what you're doing on the go

53:36

as you're on the subway

53:37

as you're, having lunch

53:39

you know, as you're, you know, kind of, I don't know, anywhere

53:42

anywhere you want to code on the go

53:45

Don't let your agents wait

53:47

Exactly

53:49

There's one more, this one got uploaded

53:51

how are skills different from cloud.md?

53:54

A lot of interest around skills

53:56

Yeah

53:57

So, skills are, again, when we talk about skills versus workflows

54:03

think about skills as a

54:05

custom tool that you can make as a user to

54:08

again

54:08

let you

54:09

accomplish a specific task over and over again

54:11

Whereas

54:12

a cloud.md is just what you want cloud

54:14

to always remember

54:16

about your projects

54:18

as you go from one cloud session to another

54:20

Again, if you remember, cloud sessions don't have

54:23

you know, shared memory across each other

54:25

So what you want to do is

54:28

Bake some of those best practices

54:29

like, hey, how should my coding

54:31

you know, methodology look or best practices look?

54:34

How should I

54:35

style this UI in this project?

54:37

Some of those things can be saved to

54:38

CloudMD

54:39

So think about CloudMD as best practices

54:42

for cloud code while working in projects

54:45

and think about skills as the actual task

54:47

that you might be doing over and over again

54:50

while you're actually implementing or exploring a code base

54:53

So hopefully

54:54

that gives a little bit more differentiation on

54:56

skills versus CloudMD

55:01

Cool. Next one

55:02

is there a repo with plugins or skill guidelines?

55:07

Yeah, I think we have

55:09

landing pages with plugins on them

55:12

if they're not linked

55:14

we can certainly search for them and link them

55:16

Although if, I think Emily, you

55:19

you linked a blog with skills

55:20

I don't know if that takes

55:22

the user to Cloud Code Skills specifically

55:24

but there are repositories that will host online with

55:27

you know, certain skills or

55:28

or plugins that might be useful to you

55:29

as a developer

55:30

Yeah. I also just pasted the

55:32

GitHub link to

55:34

some of the skills and plugins that we have

55:36

publicly available on Anthropic's GitHub

55:38

Nice

55:41

There's, yeah, and you go for it

55:44

I was just going to say

55:45

this is kind of a fun one

55:46

is there anything surprising a team has done

55:48

with Cloud Code that you didn't expect?

55:51

do you two have any

55:52

any interesting tidbits to share?

55:58

I would just say

55:58

well, first of all

55:59

thanks for not using my image in that

56:00

demo. I was surprised you did that there

56:04

I was pretty impressed. There's someone on our

56:05

team who's built out a very functional pricing application that we

56:10

like the entire sales team uses

56:12

and he, you know, is not an engineer

56:14

he is on our sales team

56:15

And so I think that

56:18

the philosophy around using Cloud Code is really

56:20

just to kind of dive into it and

56:21

see what you can build

56:23

That's been surprising on my side

56:24

What about you two?

56:29

it's a good question. I think

56:32

Some of what I see also

56:33

is definitely sales and marketing flows

56:35

One of the friends that I have here at Anthropic

56:37

he built a

56:40

marketing calendar that now all internal Anthropic employees use

56:44

and he doesn't have a development background at all

56:46

but he not only built the application himself

56:49

but he actually pushed it

56:51

and hosted it internally using Cloud Code as

56:53

well. So not just the development side

56:55

but the hosting and infrastructure side and the testing side

56:57

he did that all with Cloud Code

56:58

So there's a lot of use cases

57:00

especially as someone that's maybe even non-technical can

57:03

actually use with Cloud Code too

57:08

Ali, anything from your side?

57:11

I think the one that surprised me

57:12

almost every day is how like

57:14

the non-technical folks that I work with after

57:16

every call have

57:18

Really good prototypes and demos and ideas

57:21

to inform me as an engineer on how

57:23

to go about a task

57:24

that we want to work with for

57:26

for our customers

57:28

and so every day there's ideas that I

57:30

I'm not even thinking about

57:31

but they're, they're sending me either like single HTML files or

57:35

or literary zip files

57:37

with, with an agent that they made or

57:39

or a mini app that they made

57:40

and then I go off of that for

57:42

for building the actual bigger thing

57:45

Yeah

57:45

it's amazing. It's really amazing what you can

57:47

do with Cloud Code

57:48

and we've absolutely loved having you all join

57:50

us today

57:51

Thank you so much. We'll make sure to

57:52

send the recording and a list of all

57:54

the questions that we weren't able to answer

57:56

but, go have fun with Cloud Code and

57:59

ask your questions on cloud.com

58:02

Amazing. Thank you all

58:03

Thanks, everyone