Transcript for "Claude Code in an Hour: A Developer's Intro": Alright. We'll go ahead and get started. Thank you everyone for joining. Welcome to today's webinar. This is an introduction to Claude Code. We are live from New York City and San Francisco. This session has been highly, highly requested, so we're super excited to walk you all through it today. On theme with awards season, I will be your host. So my name is Emily Gray. I'm on our sales team here at Anthropic. Cool. So first, we're gonna walk you through what Claude Code is, the foundations of how to use the platform, features you should be aware of, like certain slash commands, and we're also gonna cover a live demo. Awesome. So with that, you guys know who I am. I'm gonna quickly kick it over to Harry and Ali Shazal from our beloved applied AI team to quickly introduce themselves. Thank you, Emily. My name is Harry. I'm here in the Applied AI team at Anthropic. I can't see all of you, but I'm excited that all of you are here. I think there's almost or more than 2,000 of you on on the line right now. I sit at the intersection between go to market engineering products, and some other orgs and really help our customers and folks that are also interested in Anthropic and Claude products get the most out of, what they're doing with Claude today. And I'll go ahead and pass it over to Ali Shazal for intro on his side as well. Thank you, Harry. Hi, all. I'm Ali, and I'm an applied AI engineer here at Anthropic. Super excited to be here and to get things going. Awesome. Thanks, guys. Cool. So just a couple quick housekeeping I housekeeping items. So one, yes, this will be recorded, and we'll make sure to send everyone a copy after. And then next, please make sure to ask questions. We want this session to be as live as possible, so Ali and I will be monitoring the chat and answering as much as we can. And then, of course, please share your feedback with us. Our goal is to make these sessions as valuable as possible, so we always love to hear feedback from our customers. And one more housekeeping item. Our team can answer only so much as humanly possible, But the good news is that we have our dear friend, Claude. So on our website, you can see, at code.claude.com, which is a mouthful. If you go there, that's the best place to, get questions. That version of Claude scrapes all of our documentation, and we'll provide the fastest answers for you. But, again, we'll be making sure that we monitor the chat and answer as much as we can. Cool. So with that, I'm gonna kick it over to Harry to walk through the presentation. Thank you, Emily. So before we get into the Claude Code overview and a little bit of an introduction on what Claude Code is and how you can get the most of it, I wanna set the stage and give a little bit of context of how we got here. So everyone knows or has at least heard of the word agents by now. So you've seen billboards if you're here in the Bay Area or you've seen them mentioned online. You've seen demos of them on LinkedIn or other platforms. But what are actual agents? Are they workflows? Are they, you know, autonomous robots that live in our systems and and do things for us? So what they actually are are a system that is powered by AI where a human can give the agent, some kind of goal to accomplish, and the agent can use all of its tools, all the systems and harnesses and, layers underneath to think to do, and then to iterate on what it's thinking to ultimately accomplish that goal. So I I wanna really set the stage for what agents are because Claude Code truly is an agentic platform versus an AI powered one. And I wanna make that difference fairly clear as we walk into the history of what AI coding has looked like for the last few years. Right? So let's let's rewind five years from today. So back from 2026 all the way to 2021, we started with autocomplete. So if you're a developer and you've been in this space for a while, you've probably seen in some platforms that you've been using a few years ago, the ability to code, and then the IDE or the development environment or the coding platform will actually autocomplete the rest of the coding line for you. So that's where AI is gonna work, for seconds for you. So it's gonna think about the line that that you're working on. It's gonna think about the file that you're in, then and it's gonna autocomplete the rest of that line. So, really, that's where we were years ago. And then we evolved a couple years down the line to maybe being able to modify whole files, right, where AI is able to take in the context of what you're working on and then maybe some of the other modules and methods and libraries that you have in your project and then modify a file at large. Where are we today, and where are we going in the future? So Claude Code came out in 2025 as a research preview, created within our labs here at Anthropic. And you're seeing Claude Code become, from an AI assistant more to an agentic coding tool. And Claude Code is actually able to work for hours on end, being able to complete tasks that you give it at a very general level, whether it's a complicated, deep dive into your code base and then being able to modernize it or refactor it or whether it's, a task that has a lot of edge cases that you really wanna, codevelop with Claude on. Right? So this is now an agentic platform. And as we go into the future, we're seeing AI working for not just hours, but days on end, whether it's, really sitting with you and prioritizing your your user feedback, and then just working on its own until it has a question and it comes back to you. So, as we go into the future, you're gonna see this curve, go up and up and up. And this follows what you're seeing on the screen in terms of this diagram. In terms of the technology itself, we've evolved from being able to get just simple outputs to maybe automated workflows, which I talked about earlier. I wanna stress the difference between workflows and agents. Workflows are deterministic or specific paths that you go through, but an agent truly is something that is a loop. So a human interacts with an agent by giving it a goal like an open ended problem, something that's a little difficult in terms of, predicting, something that doesn't have a deterministic path such as search, coding, support, etcetera. And then it works in a loop to use all the tools in the environment that it has access to to ultimately complete it without a human needing to intervene or steps needing to be defined specifically. Okay. So that was a little bit of context before we get into Claude Code. So would love to introduce Claude Code. So a lot of you on this call have probably seen Claude Code in some fashion already even if you haven't used it. Again, would encourage you to look in the docs tab here on the platform if you wanna start getting an overview on Claude Code using some of the documentation. You can install it even while I'm talking. But it truly is an agentic, what we call a CLI or command line, interface tool to allow you to start doing development or even nontechnical use cases with with Claude Code. The reason we've chosen this form factor you're seeing on the screen here on the right is because we want we wanted developers to be able to actually interact with parts of the scripts and the CICD pipelines that are native to their terminal today. So you can certainly use it in other forms and fashions, and and we'll we'll take a look at what that looks like later down in this presentation. I wanna showcase all of the, you know, different buckets that you can actually use Claude Code in. So let's take the software development life cycle. Right? So we have maybe five different steps that folks are used to, especially if you're a developer that's tuning in today, all the way from Discover to supporting and scaling your environments. So, you might go through a typical workflow like this. Right? You you might onboard to a new project. You might develop certain technical specifications, and then you might actually write tests and then implement code according to those tests and then really automate, you know, the CICD pipeline and then be able to debug and monitor usage from that. So Claude Code, I wanna emphasize, doesn't just take care of the middle bucket, which a lot of people think it ends up doing, but it can start helping you discover parts of your codebase. develop those technical specifications by using plan mode, automating code review, and then ultimately scaling out debugging and then refactoring your code at the very end as well. So as we're going through this presentation, I just wanted to really emphasize, that folks can really start thinking about how to use Code in different parts of your development life cycle. And then before we actually go to the rest of the content, I wanted to, kind of leave this section on how Anthropic actually uses Claude Code internally. You'll you'll see in the top row are the usual development use cases. Right? Whether you're developing your data infrastructure pipeline with a support workflow, whether you're doing some kind of product development, whether you're doing security engineering or data science, for sure, these are great fits for Claude Code today. However, a lot of our different internal teams are also using Claude Code. So, Emily, I'm gonna turn it back to you. I would love to hear some of your input because I know you're talking to a lot of our customers. How how do folks that are nontechnical use Claude Code today, or how do, you know, people even within Anthropic that are nontechnical use Claude Code today? Yeah. It's a really fun tool to kinda dive into and see what you can build if you're a nontechnical person like myself. Here on the sales team at Anthropic, we actually have had a couple folks who have built out things like a pricing calculator or account plans, and those are just some sales use cases. I actually work with a hedge fund here in New York City where their CEO mandates that every person all the way down to HR uses Claude Code. So there's a lot you can do there with any kind of analysis, working across the board with cost modeling, P and Q assumptions. It's kinda fun to to get your hands on it and see what you can build. Amazing. Thank you for sharing. So, hopefully, that gives you some ideas, and you can see on the bottom row as well how some in, internal Anthropic employees use Claude Code for things like growth marketing or product design or legal for their different workflows, internal applications, and prototyping. Okay. So we're gonna have our first poll. So, you're gonna be able to start getting interactive truly with, you know, some of the, content that you're seeing here. So I would love to know, you know, where if you're using Claude Code today, where you're using it. Is it in the terminal? Is it in, you know, an IDE? Is it in the web? Is it on mobile? Feel free to participate in the poll. Would love, the Anthropic team here would love to know, you know, kind of your user experience and and where you're using Claude Code today. But we'll keep moving on. And as you can see within the slide you're seeing here, you can take Claude Code wherever you need to. Right? So if you're familiar with the terminal and you don't feel, you know, as, it's as daunting of an environment to be in, you can certainly stay there. However, if you're used to a code editor, you can bring it there. We have plugins that work natively with a lot of the the primary IDEs out there, and then you can access Claude Code on the web and iOS. We even have a feature where you can pretty much teleport or remote control your terminal on your laptop from your phone. So if you wanna code embed, you can code embed. Alright. Perfect. I don't know if we have we have any questions that are coming up so far, but, would love to pause for one question, if we can answer some something from the crowd so far. Yeah. Maybe we can start with this one. Okay. Is there an objective function set for an agent based on the user's prompt, or is it all direct is it all directed, just very directed predictive probability, or a combination of both? Interesting. So I think there's a lot of parts to that question. Ali, I don't know if you have any, thoughts on your side on this. But in in terms of prompting in general, predictability in kind of prompt techniques and prompt engineering, a lot of what we internally Anthropic advise customers on is, you know, there's with LLMs and with generative AI in at scale or at at a whole, it's not gonna be able to be predictable in terms of the next steps that it takes. Right? It's truly what word comes next is how the technology, works in terms of LLMs. So when you're prompting an LLM, especially in Claude Code, you don't have to be perfect. There can be typos. There can be ways of, you know, getting to the same goal overall. But I would say, some things that will really help is we do have prompting guides online at Anthropic, that you can definitely search up. If it's not in the docs, we can certainly link them. And then we also have something called skills, which I don't know if I'll cover in too much of a detail today, but they are tools that you can use in Claude Code to that are already pretty much packaged up to get you further than, pretty much a blank slate or a blank text box. Ali, anything else to add there, that that I didn't cover? Yeah. A couple of things that helped me on top of what you said. Adding examples and instructions helps steer plot well as well as constraining the output format as XML tags or or JSON schema that also helps bring some predictability, some structure to the response. Amazing. Thank you, Yeah. Liu. And and if we can just answer one more question, because we're getting this a lot in the chat here. I think there's there are a lot of questions around how to get started. Like, is it directly in the terminal? Can people do it on their phone? If you can just quickly cover how to get started, I think that'd be helpful. Yeah. In the docs view, there should be a link, that, opens up I think, learn more about Claude Code. It should be the Claude Code landing page. You can install this directly on your terminal right away and get started. If you want to, you can download the Claude desktop application, and there is a code tab in there. You do have to link it to a GitHub repo, so that is a requirement to actually get started there. And then you can obviously take it on the go in mobile as well. You do also have to link that to a GitHub repo. But the easiest way, I know it's a little daunting, especially if you haven't worked out of a terminal before, But if you have access to your terminal, and you can go to the overview page, just go ahead and there there's gonna be a little copy button on that page. Go ahead and copy and paste that into the terminal, and you can get started right away with that. Awesome. Thanks. Amazing. Okay. Let's keep moving forward. So, we're gonna start our foundation section. It's it's a fairly short section, but I wanted to go over the the very kind of basics, but maybe most important knowledge of starting to use Code. So, we've talked a little bit about where you can access Claude Code. I this point, I'd like to emphasize because a lot of developers or folks that are, you know, writing code on a day to day basis have their environments that they like to work out of. Whether it's an IDE like VS Code or JetBrains, whether it's another platform that they they like to to work out of, you can bring Claude Code, again, with you wherever wherever you need to. And the great thing about Claude Code is it will interact with certain native parts of, those code editors as well. So I I definitely don't wanna skip that. CLAUDE.md. So you might have heard of this concept called CLAUDE.md. If you haven't, totally okay. But if you look on the screen here, what CLAUDE.md is, it's pretty much like a memory file for our models. So when you're using Claude Code, some of you that have already been using it might notice that if you get to the end of a conversation, the model has a limited, what we call, context window. I'm gonna refer to it as the brain. Right? So if if the brain runs out of space as you're talking back and forth with this model, it's going to need to clear the brain to keep moving forward and take additional actions. So what CLAUDE.md does is when the brain is cleared and it's a blank slate, it's wiped. It's going to keep moving forward. CLAUDE.md the most important things, in the brain saved so that every new session, every time you clear the memory or the context, quote unquote, you're gonna be able to save those important best practices, whether it's project structure, common commands, styling, tips, etcetera. All of this can go into a CLAUDE.md. CLAUDE.md is fairly lightweight, so you're gonna only add 30 to 50 lines into it. Everything that you're seeing here in terms of, you know, whether it's CLAUDE.md or any other tools, you can absolutely have Claude make for you on your behalf. So if you're using Claude Code, you're gonna use the /init command /init I n I t to actually start a CLAUDE.md in your local directory as well. And I I wanna kind of connect it to what folks might be seeing. If you're seeing maybe Claude behave in a way that is unrelated to the topic you're asking about or, development flows that you are working on, it might be because there's too much in its brain and you need to clear it, and maybe keep the most important things in the CLAUDE.md. So I wanted to illustrate this in more of a pictorial fashion or or a flow diagram. And as you can see, there is a repository here, a mono repo. And you can see you can have CLAUDE.md. at every single level. So CLAUDE.md. are useful for, you know, at the top level describing what this system is doing and maybe what teams are mapped to to the different subproject folders or modules within the system. And then as Claude actually goes down and explores these different folders, it starts pulling the different CLAUDE.md. into the brain as well for, you know, the relevant context. If I'm exploring the front end folder, I wanna pull in the usual UI patterns and the test strategy and the components. If I'm exploring the test cases, I wanna understand how deeply, you know, the test coverage runs and my automation strategy. So if there's one thing that I can leave folks with today, one of the things that are important, I would say maintain your context, aka the brain, and then using CLAUDE.mds to give Claude the context that it needs and the best practices to save over and over again as you have new sessions with Claude as well. Okay. Perfect. So we're gonna have poll number two come up. So if you are using Claude Code today or even Claude in chat, in Cowork, you might see that you're able to select different models. So we'd love to know, you know, which models you're using today. And you can even participate in the q and a if you wanna say why you're using certain models. You you can do that as well. So when you're selecting a model in Claude Code specifically, you might see a few different models come up. We have what are called our model families. So model families include Opus, Sonnet, and Haiku. And Opus, I'm gonna start from the top, is our most performant model family. We have Opus 4.6 or 4.6 out now, and it's very capable in terms of very complex tasks, like exploring an archaic code base deeply, understanding architectural changes that you're requesting, how to refactor that or modernize that, and then provide a plan for you to implement. As we go down, Sonnet is our we we we like to call it our daily driver here at Anthropic. So, if you, as developer, you know, have a plan already put together that's highly specific and highly complex and you need Sonnet to do the heavy lifting of just implementing that specific plan, you can go ahead and use Sonnet. And then Harry Liu is used for things like writing tests or maybe more specific, point, you know, solutions or point features that you might just wanna do a one off of. Again, as you're using Claude Code, I would actually encourage you to flip between these models to see what makes sense for parts of your workflow. A lot of people ask me, which models should I use every single day? I always say it depends on what you're doing in Claude Code. Some people leave it on at Opus all the time if, you know, they don't mind kind of some of the thinking that it does and the the the way that actually, you know, gets to the answer and outputs. And some people will flip between the models based on the the workflow that they have as well. Okay. So I think we can pause here for another question. Emily, Ali, any questions in the chat that we can help answer again? Yeah. This is a good one. So the the most upvoted question is, do the terminal and desktop applications provide the same experience, or are they designed for different purposes? It's a really good question. So the terminal is where, a lot of our native Claude Code features live, and it's where you probably get the highest amount of kind of, experience that, developers often need for their day to day workflows. What I mean by that is there's a lot of slash commands, and it has the ability to work, obviously, with your local desktop and your files and folders and use all of these, what we call agentic tools to actually interact with your codebase. The desktop application as well as, Claude Code on web has to be connected to a GitHub repo to actually operate. So you you do have to, you know, authenticate to GitHub and then sync it with a specific repo, and then it will launch pretty much a web version of Claude Code and then do operations on your behalf there. Certainly, I I would I would say this. If you are maybe onboarding quickly onto a codebase, and maybe you're a new developer on the team, you're getting used to Claude Code, start using Claude Code on desktop or web is a great experience to start trying it out and testing it. However, if it's part of your daily workflow and you're really deeply developing in Claude Code and you have the repository cloned to your computer already, I would suggest definitely using terminal for Claude Code. Perfect. And, Ali, Emily, anything you would add there as well? I would say what I typically hear in the field is that it's preferential to developers, depending on where they like to work. Amazing. Perfect. Okay. Let's keep moving on. And, just for to remind folks in this room, we are gonna have a demo at, after this portion. So if you're on the edge of your seat, you're not gonna have to wait any longer. So we will get get through this, and then we're gonna get to a live demo. Fingers crossed that it works. Cool. Okay. Let's let's get back to this concept of context. So we talked a little bit about this word context. What does it actually mean? So when you think about context, just think about pretty much it equating to words that the model's actually understanding. Right? So the more context or words you put in a model's brain, the more it's gonna have to think through and process to understand what your end goal as the user is. So if you curate that context, like you feed the specific files or the you know, you're a little bit more specific with your ask in terms of what you want Claude or Claude Code to do, it's gonna lead to more efficient and better performance. And, likewise, as you bloat the brain or the context window, the model needs to clear it, and it might lead to performance degradation as well. So how do you curate your context then? Let's let's go through some commands. So if you're in Claude code, you can use what's called the slash /context command. The slash /context command is gonna give you an inside view on the model's brain, pretty much. So what is taking its context? Is it the prompt, the system prompt underneath the surface that Claude Code is operating on to understand how to actually interact with the user? Is it the messages that you've, you know, talked back and forth with Claude? Is it any of the tools that it's calling, like connectors or, you know, other sub agents that that it might be summoning? Is it, anything else that that might be bloating the context, or is it even the CLAUDE.md files that you're loading in? So the more understanding you have of what's in context, you can actually curate, obviously, what's called. And then you can use the slash /compact command. So compact, if I were to explain it very simply, is when the brain is almost getting to the point where it's overloaded, you can actually, as a user, trigger a command to summarize everything that the the the model has seen so far and then clear the model's brain and then give the summary, only keep the summary in the new session and then keep going. So it it it gives Claude Code an understanding of what's been done so far in a summarized fashion, and then you can keep moving forward to accomplish your task. So I would use these two commands, very, frequently to understand what's going on in your context and then to clear and summarize your context as needed. K. So, @-file mentions. This is another way to curate your context. So if you as the user know exactly what file you want to refer to while you're talking to Claude Code, instead of having Claude Code search and it's gonna use some local tools to actually search your computer for the relevant files. You can actually do what's called @-file mentions. You can @ mention, hey. App.py. You can at mention, you know, a JSON file or a folder, etcetera. So if you wanna pull things specifically into context, this is a great way to specifically do that as a user if you know what you need to Claude to start working on. Okay. So next poll, we're gonna, get a little information on plan mode. So it's gonna be a, do you use plan mode today? Yes or no. And then if you do or if you don't, feel free to participate in the q and a, you know, how you find it useful. Or if you have any questions about plan mode, feel free to ask them there. The the reason that some some folks might use plan mode, and I'll talk a little bit about what it is, is it is a way to actually allow you as the developer or even someone that's nontechnical to use Claude Code to have Claude create a full end to end implementation plan for you as you're going and then have it outline stepwise what it's planning on doing, maybe some edge cases that it's seeing, as well as maybe some assumptions that, aren't clear. It's gonna ask you some questions to clear those up so that it doesn't guess while it's planning. Folks that have used a lot of other tools, you know, will find out maybe when I'm asking something of an AI coding tool, it's gonna get to an end goal, but it might not have all the details that I want. Right? So plan mode is a great way to start if you're implementing something complex, if you need to explore your codebase. if it's a huge refactor, it's a really tricky item that you're tackling. I would highly suggest, going in plan mode first. Ali, I would love to know maybe some, you know, information from your side. How have you seen plan mode used, either, you know, in your day to day work at, at Anthropic or even in the field as you're talking to customers. Both in my work in Anthropic as well as in the field and in my personal projects also. Essentially, I see plan mode being used before any task. It could be a small task, big task, a big project, a refactor. Before any of them, it's a great idea, and we recommend it to use plan mode instead of directly jumping into the code. Results in much better results. And, also, it's not just meant to be for really large changes. Small changes are are good too. Think of treat the plan more like a PR description that you're writing before the code. If the plan doesn't read so good, the code most likely won't. So that's a good way for you to correct Claude as well as give more inputs. Sometimes Claude also starts to interview you to clarify on things that aren't clear to it yet. Amazing. Thank you, Ali. So I highly encourage everyone to try out plan mode if you haven't already. Certainly, for very well scoped prompts, maybe you can start with just normal mode or accept that it's on. But plan mode is especially useful for, the topics that Ali and I mentioned. Okay. So for the next four or five slides, I I'll go through fairly quickly. I would love to get get to the demo and leave some time at the end for some questions. But I do wanna emphasize the ability to not just send text to the screen, here here within Claude Code, but you can also drag and drop screenshots as well. So if you are especially a front end engineer or developer or you're working with maybe visual elements and design, you can actually screenshot what's happening on your browser or an application, etcetera, and then drop that directly into Claude code. What that allows Claude to do is it processes the image and then understands the prompt that you're pairing with it to help either debug, improve, or modify what you're working on within your project. Right? So it's as simple as drag and dropping. Or if you wanna copy a screenshot or a a photo directly from your desktop, you can do that as well. So an example flow might be, you might wanna squash a bug from a screenshot. Maybe there's elements within a front end UI that are overlapping. You can screenshot that, paste that into Claude Code. You can use plan mode like Ali and I talked about to plan the fix to make sure that Claude Code is gonna operate it on it correctly, and then you can certainly review and verify that within your code editor or your terminal. Okay. So I'm not gonna spend too much time on bash mode. However, bash mode is a way that, folks that are using terminal can interact with the terminal, in, pretty much a raw way where you can run terminal current command scripts directly within Claude Code. So no need to open a new tab. You can run your terminal commands also directly in Claude Code as well. And then the last two commands are very useful in the day to day operations and and workflows. This is why this section is labeled workflows. But, hopefully, this helps you and your workflows as you're using Claude Code. But if you press the escape button, you can actually interrupt what Claude Code is doing immediately and redirect it. So myself as well as the rest of, you know, our internal Anthropic employees and our customers tend to use this a lot, if you need to you know, you see Claude going in a direction that you know it, should not be going, and you can give it additional context. I wanna stress that as you press escape, Claude does not lose any of the previous conversation or context that it's thinking through. You can just redirect it. It's gonna understand your new prompt as well as as the direction it was going on previously. If you hit escape twice quickly, it will rewind any changes that it's made in terms of the last kind of code modification. So if there's anything that you didn't want it to modify, you can actually do what's called rewind the change as well. And then resume. So if you somehow exit out of your terminal window or your session in your IDE with Claude Code, you can actually just hit slash /resume after you enter into Claude Code to pick up where you left off. This is great for, you know, under, going back into a session where you are implementing something deeply and you got interrupted. Or internally in Anthropic, we actually use this to test maybe different approaches or implementations on a feature that we're working on and then maybe choose the best session to resume as well. Okay. So before we hop into the live demo, let's maybe do one question and then pivot over the demo so everyone can see a little bit of Claude Code on the screen as well. Awesome. Thanks, Harry. The top question that we have, and we get this one all the time on customer calls as well, does Claude store information about what it's done? Should I be concerned about having an analyzed data with PII in it? That's a good question. So, again, Claude Code, is, operating locally on your computer, and we do not store anything in the Claude. So we will send some of what the prompts are and some of what Claude is analyzing to the Claude just to to actually, you know, run what we call inference on it and actually have it go through our models, but we don't actually store anything on our side. And there's actually a lot of security, specific features baked into Claude Code that I won't go over in in this introduction session, that you can actually configure to have it stay away from certain security elements or files or folders that you actually don't want it touching. So that's a really good question. But within the Claude Code product, we definitely take security and compliance seriously, which is why we've baked some of those things, within the product as well. Awesome. Thanks, Harry. I also just linked our trust center in the chat here in response to that question. That is a wonderful resource. It has an FAQ in there. It has a lot of documentation. It's probably a good place to get started if you have specific security questions. Amazing. Okay. And with that, we will go to the demo. So let's stop sharing on slides, and then let's start sharing the screen. Okay. One of the my copresenters, can you verbally tell me if my. screen is showing? Perfect. Looks good. Okay. So we are starting in a demo repository. If you're a developer, this might look familiar to you. We're starting in GitHub. So I'm a developer. I'm a front end developer, and I'm working on a landing page that might look familiar to some of you. So this is the, web registration page that you yourself actually use to register for the webinar that you're on right now. So, very meta, I know, but I think it illustrates nicely for for the purpose of the demo today. So right away, you might see some things that are maybe a little off. I'm gonna point out this hour should not be here. Obviously, this this webinar is in the future. It's July 15 of this year. However, it should just be one hundred thirty three days, and the hours looks awfully, you know, different than what it should be. So there might be a bug in this code. Right? And as I scroll down, maybe I'm previewing this page. It looks okay to me, but maybe there's some other modifications we wanna make. And if we take a look at the issue in GitHub, in the repository that I'm working out of, it looks like there's already been a bug report that's been filed. Right? So when viewing the landing page, the countdown timer displays an incorrect number of hours. So we've already been made aware of the bug. How do we now fix that with Claude Code by working directly in the terminal and not needing to copy and paste things and and work back and forth? So let's let's go ahead and and take a look at what Claude Code looks like. Okay. So now we're in our IDE or our development environment. I'm using VS Code. Again, you can use a lot of different platforms for this, and I'm within my project folder. I've cloned it. It's on my local, laptop. I'm ready to go. So you could see Claude Code. I'm in Opus 4.6. You can choose between different models, certainly. What I wanna introduce first is the CLAUDE.md file. So we talked a little bit about this in the foundation section within this webinar. But what the CLAUDE.md file is, again, is it's the read me for the Claude model. Right? It understands what the project is doing, some of the architecture, how it's structured, some of the components of the project, how you run it, how you test it, maybe some of the best practices in terms of conventions. And you'll notice, again, it's very short. It just has that architecture. It just has how you run it, the conventions. And some of the other things that might be important, we will cover in a more intermediate and advanced section of how you save some of those components, but these are the main best practices. So let's say I'm onboarding onto this issue. I just want Claude to go ahead and explain this project and what it does, how it's structured, how do I run it, etcetera. So I'm gonna go ahead and hit enter, and I haven't done anything else but ask this one question. And Claude, go ahead go ahead and start thinking. So it looks like this is a local recreation of that registration page. It's gonna give me that structure. It's gonna give me certain constraints, how to run it. So very similar to that CLAUDE.md. And you'll notice this is because the CLAUDE.md is already loaded into context. So if I hit context again, you'll be able to see exactly what's in the context right now. So we've have some MCP tools, some memory files, some skills, etcetera. Right? Okay. Perfect. So why don't we go ahead and do this? I'm gonna go ahead and clear just so, the the screen isn't bloated, and I'm gonna go on to to the next prompt. So you'll notice I am not calling any kind of connector. I'm not calling, you know, any kind of API. I'm literally just saying, there's an open issue on this repo about a countdown bug. Can you pull up all the details? Can you investigate the issue itself? And then can you run the test? So right away, you can see it's using a a command to actually reach out to the issue, which is, again, remind everyone, it's here. And then it's running the test within this environment. And then if you see above, and I'll I'll pull the terminal window down, it's seeing it's already telling me what it's gonna change. So, Claude was a little too quick. The demo is my talk track is going slower than Claude is working. But you could see it ran the tests. And if I hit Ctrl+O, I can see everything that's outputting in terms of the test. So there are two failures. And then it found the issue on GitHub. You know, this is what the issue looks like. This is what the bug looks like. And can I make these changes now? So you'll notice that Claude Code does not make the changes until I ask it to make the changes because I'm in normal mode. I will talk about another mode later, as we get to the next prompt. But I'm gonna go ahead and say, yes. I wanted to make the edit because the change looks correct. Okay? So it's gonna go ahead and make the edit, and it's gonna even verify the fix. So it's gonna rerun the test, and you'll notice I did not ask it to rerun the test, but it's gonna go go ahead and do that. And then now you'll see that all 17 tests passed. If we go back to the page, if everything looks correct or went correctly, you'll now see that the hours looks on par with what we should be getting in terms of the countdown. So big thumbs up there. We went through our our our first, workflow. Okay. So let's now go on. I'm gonna hit shift tab now, to accept edits on. What this mode does is I want Claude Code to generally accept most of the edits that it's doing without me as the user needing to approve it. On some, maybe permissions or operations that it still needs to prompt me on, it will prompt me. But for a lot of them, now I just want Claude Code to pretty much cook. Right? So let's go ahead and do what's called call a skill now. So I wanna emphasize, I didn't, cover this yet, but in the next section, we will actually cover what a skill is. What a skill means, and I'll I'll talk about this very shortly right now, is it's a tool that you can use in Claude Code to maybe complete the same task over and over and over again. And it's just described in natural language to Claude. So you you'll see this one is adding an animation. So creating an animation is gonna ask for a photo, analyze the image, create the pixel art animation, structure the scene in a certain way. Right? So what I'm gonna do is I'm gonna go through a couple photos we have on our side. So you see Ali. You see myself and Emily. I'm not gonna pick on my co presenters, so I'm gonna go ahead and use my own photo. And, again, I just drag and drop it into Claude Code, and I hit create animation. Again, I think we had the question earlier about how do I prompt Claude well? Is it probabilistic? You know, etcetera. So you'll notice that this skill takes all the prompting knowledge out of it. Right? You don't need to know how to prompt. You can just accomplish a workflow, codify that into a skill, and we'll talk about skills in a bit. And then Claude will go ahead and, be able to surface that skill so you can call it over and over again without needing to think of the right prompt to operate on. So what it's doing now, and if I hit Ctrl+O, is it's thinking. It's reading the relevant files. You can see it even processed my photo. Right? So it said my hair is black, short, styled up. That which is true. Skin tone, glasses, clothing, clean shaven, friendly smile. I would hope so, hopefully. But you can see now it's already making changes. It's adding specific code to, I believe this is the CSS file within, the actual project folder, and it's not asking me now for permission because I clicked accept edits on. So it's not gonna, ask me for the edits. It's gonna go ahead and add those to the codebase. And now it looks like the animation's already live. So the pièce de résistance, if we refresh this, we have a side scrolling animation, and it's a little mini representation. Doesn't look entirely like me, but as close as it can get via one photo. Right? So, and I'm just walking with my laptop and saying, let Claude cook. Okay. So we we've done a lot here already in this, you know, five minute demo. Let's now push our changes. Let's, you know, kind of close this issue. So why don't we now tell Claude, can you commit all these changes? Can you make sure that you save all of them, push that to the, repository on GitHub, and then open a pull request that closes the previous issue as well as pretty much takes note of the animation that we added. Right? So we're gonna go ahead and press enter here, and then it's gonna think. It's gonna be able to look at the status of the repository and then ultimately sync it to our GitHub repository. So it's thinking through that process. It's taking a look at the log. It's gonna add all the relevant files to the staging area, and then it's gonna go ahead and push and create a PR. So you could see all of that happening, sequentially, in the thought process here and all of the tools that it's calling. So whether it's a Bash tool, so I yes. I want it to proceed because it's doing something, you know, that is a little more sensitive. It's gonna ask me permission for that. And then now the PR is open. So if I'm gonna go ahead and copy this, and then let's go to our browser. Let's see if this opens. And now we have our pull request. We have you know, we fixed the countdown timer. We've added the animation now. And, visually, I have checked this off. Now the last thing that I wanna do in this demo is, now I wanna have Claude review the code. So I can actually just go ahead and say, Claude, make sure I tag Claude correctly. I think this is, tagged correctly. So I can say, Claude, review this pull request for correctness and anything you'd flag in a real code review. So it's gonna have little eye emojis that pop up, perfectly queued. And Claude has seen this pull request, and it's gonna go ahead and go through all the code and review it. And if it comes back with thumbs up, then I'm gonna go ahead and merge that. Obviously, after reviewing the code myself as a human, merge that afterwards. So that was an end to end demo of Claude Code of pulling from GitHub, working on some issues, adding some content, opening a pull request, and then using Claude to actually review the code. Let's hop back into our slides. K. So now we are on the power features section. And we'll have time for questions at the end. I wanna be sensitive of kind of where we are. I wanna leave probably seven to ten minutes. So we're what we're gonna do is go through this last section and then get to some of the questions at the end and and then close from there. So in terms of power features, so this is I think another poll is gonna probably go off around this section. What MCPs do you, currently find useful today, and what category does that usually fall into? So you'll see a poll come up shortly. If it doesn't come up now, it will come up later. But the MCP section, I wanna explain this very shortly. You can think of if you haven't heard of the word MCP before, they are known as connectors. So they connect to external data sources, or they, connect to maybe even internal data sources within your company to pull information from or to send in information to from Claude Code. So we here at Anthropic use, MCPs or connectors to maybe test some of our code in terms of, you know, Python automation or to pull directly from somewhere like a GitHub or from a task management software or from Canva or from Jira, etcetera. So think about what where you'd like information to start coming into Claude Code or sending it from Claude Code, and that's where you'll you'll find a lot of value. Okay. Perfect. Let's keep moving forward. So skills. You saw what a skill did within the demo. So I was able to call the create animation skill towards the very end there to add something new to the page. But what skills are are really natural language files bundled with a couple of different other files like scripts, like examples, to allow Claude, not just Claude Code, but our Claude ecosystem, to operate on the same task, a specific task over and over again. So think about documenting your API when you're creating new endpoints. Think about creating PDFs. Think about styling your UI in a specific way. You can actually create a skill for that. The best way to create a skill, I always tell folks that are curious about how to create skills, are to actually have a conversation with Claude, maybe complete a workflow, and then just tell Claude at the end, hey. I wanna create use everything that I've done so far and create that into a skill. And and you can certainly do that and then reuse that from then on. And then plug ins. So plug ins are, a tool that you can use to ultimately package up a lot of what we talked about. So we've talked about connectors. We've talked about skills. And then there is a couple of other concepts we didn't talk about, but a lot of the other custom tools and custom components that you create within Claude Code, you can actually package up into what we call a plug in, and you can label it something like a front end engineer plug in or a back end plug in or, you know, a security plug in. And then start to share those within your own organization or your company to allow other people to get to the same value that you are with the skills and connectors and other custom components that you're using today within your workflow. So a lot of people ask, hey. What's the difference between, again, a CLAUDE.md, a skill, an MCP, and a plug in? You can see plug ins are just an easy way to distribute skills and connectors. And then CLAUDE.md are these little memory files that always live in a project folder. Okay. So we're hitting the tail end of our presentation. I think we're very good on time, I'm very happy about that. But as we close, I'd like to remind folks, one thing that we talked about a lot during the overview today is managing your context. So, again, another review is, you know, how do you actively manage your context? How do you view what that looks like? How do you compact and make sure that Claude is clear of irrelevant context before you move on to the next topic or the next task. So, again, use the context and commands compact commands, and then even use clear when you want Claude's brain to completely clear and work on something new. And then you can use your CLAUDE.md to save things over across multiple sessions and then, of course, at mentioning certain files. So, Ali, I'd love to, you know, hear from your side. As you're working on internal projects or advising our customers, what are some tips you have, for others on how to manage context well in in terms of interacting with Claude Code? Yeah. I'd say two things. One, to manage context, I often use the slash /clear command between tasks so that when I'm done with one task, I I clear the context and then start a new context to work on another task. And then the other one that I'll say is sub agents. I I started using sub agents a lot. They're I think they're a good cheat code for big code basis. So if you need to read a lot of different files, a lot of different documents, or there needs to be a lot of different smaller tasks that need to be done, I often instruct Claude to spin up sub agents to do all those tasks, and that helps keep the main agent's context window cleaner. And then the sub agents report back all of their results to the main agent, and it continues with the with the main task. Amazing. I love that. So sub agents, know we didn't talk about today. But, if you're curious, they're they are in our docs. And, if we host host another webinar that's a little bit more advanced, we will absolutely cover that topic. So, again, another plug for Ask Claude. If there's anything that we don't answer today as the Anthropic team, you can certainly go to the docs and get them answered there. They are oftentimes more updated than the knowledge that we have even internally as employees here at Anthropic. But I will turn it now over to some questions. I think we have, nine minutes for questions, which is perfect. So, Ali and Emily, let's sort through some questions, and hope hopefully, we can get some questions answered, on the session. Yeah. Hey. This one's uploaded. What is the difference between workflow and skills? So workflow and skills can actually mean the same thing. Workflow is not necessarily a terminology within Claude that is an official tool, let's say. But when I was referring to workflows in the context of this presentation, I was meaning kind of some of the, you know, tasks and workflows that you would do day to day as a developer. Right? So if you are exploring your code base in a very specific way and taking notes on it on that, that might be a workflow. And would you want to, or do you find yourself doing that same workflow on a regular basis, whether it's multiple times a day, whether it's multiple times a week or a month? If so, that's a great example of how you what you can actually port over into a skill. So, hopefully, that gives a little differentiation on workflows and skills. Ali, Emily, any input that you'd like to to share there as well? That was good. Awesome. Cool. Next question. Yeah. Let's jump to the next one. And let's spend a couple minutes talking about this because the question is around best practices, and I think, you know, we internally, obviously, dog food, we call it ant fooding, our own product. And, this question is from Naveen. At some point, we would love to learn how successful teams coordinate when multiple people are using Claude Code on the same codebase. Any best practices or patterns you guys would recommend? That's a really good question. Ali, do you wanna take this one first, and then I will come come with an answer right after. Sorry. If you're still requesting again, that's the last part. Yeah. All good. Basically, the question is around best practices for folks who are using Claude Code on the same codebase. Oh, yeah. We use branches and get work trees a lot within our teams as well as even if you wanna launch multiple Claude Code sessions. If if one user wants to launch, let's say, five Claude Code sessions, it helps to have each Claude or have each team member work on their own branch for Workweek. And then in case there are any merge conflicts that come up, also have Claude resolve some of them. This has been the main main thing that's helped us the most. Yeah. And what I'd also add on to what Ali's saying is don't be afraid to use multiple Claude Code sessions in parallel. We didn't talk about it today, but that's how a lot of developers as well as, our own employees are using Claude Code. So if you're seeing a part of the code base that you can start implementation on and then maybe another part that you can actually start a plan on and then maybe another part that you can actually learn more about and explore, you can actually do all three of those at once. Obviously, it means, you know, multitasking, juggling a few different sessions. But especially if you have those guardrails in place, like, you know, you're on normal mode, so it can't do any edits without asking you. It's a great way to actually make your workflow a little bit more efficient as you're working on the same code base as others that are that are, doing the same tasks. Yeah. Thanks, Ali and Harry. And for folks on the line, please make sure you're continuing to upvote the questions you want us to answer because we're gonna make sure we cover those in the last five minutes here. Alright. Let's talk about remote control. How can I use it on my phone to keep it connected to my computer and keep building? So, I actually did this this morning. I was working on a side project before I got to the office. But if you have Claude Code on your terminal in your computer, you can start a session, I believe, using Claude dash dash remote dash control. Don't cite me on that. Please look up the proper syntax around what that looks like. But it's gonna start a Claude Code Claude Code remote session in whatever working directory you activate Claude in, and then you're gonna be able to see it on your phone pop up. So when when it pops up on your phone, if you have the Claude app installed, which if you don't, please would love for folks to install that as well on your mobile, you can go to the code section of the Claude app and then be able to see that pop up and then operate as if you were at your home computer. Some slash commands are not gonna be the same on mobile as they are in, you know, the terminal when you're at your computer, but that's how you take some of what you're doing on the go as you're on the subway, as you're having lunch, you know, as you're, you know, kind of, I don't know, anywhere anywhere you wanna code on the go. Don't let your agents wait. Exactly. Yeah. There's one more. This one got uploaded. How are skills different from CLAUDE.md? A lot of interest around skills. Yeah. So, skills are again, when we talked about skills versus workflows, think about skills as a custom tool that you can make as a user to, again, let you accomplish a specific task over and over again. Whereas a CLAUDE.md is just what you want Claude to always remember about your projects as you go from one Claude session to another. Again, if you remember, Claude sessions don't have, you know, shared memory across each other. So what you wanna do is bake some of those best practices like, hey. How should my coding, you know, methodology look or best practices look? How should I style this UI in this project? Some of those things can be saved to CLAUDE.md. So think about CLAUDE.md as best practices, for Claude Code while working in projects, and think about skills as the actual task that you might be doing over and over again while you're actually implementing or exploring a codebase. So, hopefully, that gives a little bit more differentiation on skills versus CLAUDE.md. Cool. Next one, is there a repo with plugins or skill guidelines? Yeah. I think we have, landing pages with plug ins on them. If they're not linked, we can certainly search for them and link them. Although if I think, Emily, you you linked a blog with skills. I don't know if that. takes the user to Claude Code skills specifically, but there are repositories that we'll host online with, you know, certain skills or or plugins that might be useful to you as a developer. I. also just pasted the GitHub link to some of the skills and plug ins that we have publicly. available on the topics get up. There's. yeah. Andy, go for it. I was just gonna say this is kind of a fun one. Is there anything surprising a team has done with Claude Code that you didn't expect? Oh, do you two have any any interesting tidbits to share? I would just say, well, first of all, thanks for not using my image in that demo. I was surprised you you did that there. Yeah. I was pretty impressed. There's someone on our team who's built out a very functional pricing application that we like, the entire sales team uses, and he, you know, is not an engineer. He is on our sales team. And so I think that the philosophy around using Claude Code is really just to kinda dive into it and see what you can build. That's been surprising on my side. What about you two? It's a good question. I think some of what I see also is definitely sales and marketing flows. One of the friends that I have here at Anthropic, he built a, marketing calendar that now all internal Anthropic employees use. And and he doesn't have a development background at all, but he, not only built the application himself, but he actually pushed it and hosted it internally using Claude Code as well. So not just the development side, but the hosting and infrastructure side and the testing side, he did that all with Claude Code. So there's a lot of use cases, especially as someone that's maybe even nontechnical can actually use with Claude Code too. Ali, anything from your side? I think the one that surprised me almost every day is how, like, the non technical folks that I work with after every call have really good prototypes and demos and ideas to inform me as an engineer on how to go about a task that we wanna work with for for our customers. And so every day, if there's ideas that I I'm not even thinking about, but they're they're sending me either, like, single HTML files or or literally zip files with with an agent that they made or or a mini app that they made, and then I go off of that for for building the actual building. Yeah. It's amazing. It's really amazing what you can do with Claude Code, and we've absolutely loved having you all join us today. Thank you so much. We'll make sure to send the recording and a list of all the questions that we weren't able to answer. But go have fun with Claude Code and ask your questions on Claude.com. Amazing. Thank you all. Thanks, Thank. you.