A conversation on Claude Code

- Is this, you know, secret sauce? Are we sure we want to give it to people? 'Cause this is the same tool everyone at Anthropic uses every day. - Hey, I'm Alex, I lead Claude relations here at Anthropic. - And I'm Boris. I'm a member of technical staff and the creator of Claude Code. - And today, we're gonna be talking about Claude Code. Boris, to start, what is Claude Code and how did it come about?

- Yeah, Claude Code, it's a way to do agentic coding in the terminal. So, you don't have to adopt new tools, use new IDs, a particular website or anything. It's just agentic coding, and it works wherever you work. And I think that actually came out of the way Anthropic engineers and researchers use tools to get the job done, 'cause people have all sorts of different stacks. It's super weird. There isn't one normal stack everyone uses. There's people that use the Zed ID, and VS code, and then there's people that are like, "You're never gonna take my Vim from me. "Pry it from my cold, dead hands. " And we wanted to build something that works for everyone, and that's how we ended up in the terminal. - I see, so the terminal is almost like the most universal of all the interfaces in that it's flexible, and it's incorporated into everybody's workflow already. - Exactly, exactly. And it also just happens to be the simplest, and because it's so simple, we get to iterate really fast. And that's, in hindsight, it turned out to be a good thing, but definitely wasn't the intent from the start. - Interesting, so if I'm a new developer

and I wanna be using Claude Code, what does that look like to actually get this product working? - Yeah, it's pretty simple. So you just download it from NPM. It's npm install -g @anthropic-ai/claude-code, so it's a little bit of a incantation. So you download it, you just need no JS on your system, and a lot of people have that and that's it. You open it up, and it'll walk you through everything else. - Wow, so you just type Claude into the terminal, hit enter and then that's it, and it's just like Claude will guide you through the rest of the process, and then you can just start talking to it, and it'll start coding. - Exactly. Exactly, yeah. You install it and then you run Claude. One of the cool things is Claude works in any terminal, like you said, so if you use iTerm2, or Apple terminal, or whatever terminal you use, even in a SSH session or TMUX session, it'll work. One of actually the top ways that people use Claude Code is within ID terminals. So one thing you can do is run Claude within, for example, VS code terminal, and it'll become more powerful. So, instead of seeing file edits in the terminal, you're gonna see them nice and big and beautiful in the IDE itself. - Okay, interesting. - And we use more signals from that IDE also to make Claude more intelligent, but the experience is the same, you just run Claude in the terminal. - Okay, so there's a lot there that I wanna get to in a second. But before we get to that

so we released Claude Code in February. - Yeah. - So it's been about a little over three months. What's it been like? What's the reaction been from the community? - Yeah, just insane, so unexpected. But before we released it, we weren't sure if we wanna release it. It was this tool that internally is just, it makes our engineers and researchers so productive, and we were having this debate. We were like, "Is this secret sauce? "Are we sure we want to give it to people? " 'Cause this is the same tool everyone at Anthropic uses every day. And yeah, I think it turned out to be the right decision, 'cause it makes people more productive, and people like it. - What was the moment that you knew we had to ship it? - So it started in a small group, and it was just a few people in our core team using it. And then, at some point, we gave it to all Anthropic employees, and there was this DAU chart, so the daily actives, just looking at employees, and it was vertical for like three days straight. - Wow. - And we were like, "All right, this is crazy. This is a hit. " And then at some point, we give it to a few people externally just to see are we crazy, is this useful? And all the feedback was just super positive, and I think it was pretty obvious. - Okay, so it really took fire within Anthropic first and then all the engineers, all the researchers were using it themself, and that made it pretty obvious for us that we should put it out into the world as well. - Yeah, and that's a big way that we developed this thing. Claude Code is written using Claude Code. Almost all the coding Claude Code has been written and rewritten using Claude Code. And yeah, we're very big believers in dogfooding. It's so important, 'cause when you use a product that has obviously been dogfooded, you can feel it. And in the products I use every day, I can feel this is something that the team uses all the time and this is not, and yeah, we just wanted this to be one of those products that when you pick it up and try it, it's obvious, a lot of love went into it, and it's something that we use ourselves.

- Who do you think is the ideal Claude Code customer right now? Who is using Claude Code? What type of person? What type of developer? - Yeah, so I think the biggest thing is Claude Code is pretty expensive. So, if you're coding on the weekends, you can try it for a little bit. So, you get an API key and put in five bucks, and you can just try it. But if you want to use it for more serious work, it'll cost you $50, $100, 200 bucks a month, something like this. There a big range. It depends what you're using it for. But generally expect maybe like 50 bucks a month. There's a lot of enterprises that are using it, so if you're in a big company, it tends to be a really good fit. It's amazing at big code bases. There's no indexing step. There's no extra stuff to set up. You just run it, and it works outta the box for pretty much every big code base in any language. - And what's this integration with Claude Max? How does that work? - Yeah, so one thing that we found is when people were using API keys to pay for this, they were a little worried about their usage. And they didn't use it as much as they wanted. And so, we shipped Claude Code as part of Claude Max, so you pay for the Max subscription. It's like 100 bucks a month or 200 bucks a month, you get to pick the price point, and there's different usage limits. And you get pretty much as much Claude Code as you want. And practically, you're not gonna hit any rate limits. Very few people do. It's unlimited Claude Code. - Wow, so unified between your Claude. ai and then your Claude Code accounts is just this one subscription package. - Exactly. - Okay. So if I'm a developer and I'm using Claude Code

and I have my code base that I'm working on my computer, I get into my terminal, I type "Claude," I hit enter. What happens next? - Yeah, Claude gets to work. - Okay. - So, it's gonna use tools, it's gonna go out and do its thing. It's gonna take many steps. So if you've only used coding assistants in IDs before, and you're used to an experience where what the assistant does is it completes the line or completes a few lines or something, it's not this at all. It's super, duper agentic. So, Claude will understand your query, and it's gonna use all the tools at its disposals so that's Dash, file editing and so on to explore the code base, read files, get the context it needs, and then edit files, make any changes that you want. - Wow, so this is maybe a new form factor of coding compared to how we've been doing it for the past 20 or 30 years or so. - Yeah, for me, my journey for coding goes way back. So, I've been coding for a while, but my grandpa was actually one of the first computer programmers in the Soviet Union back in the 1940s or something. - Wow. - And he programmed using punch cards, 'cause programming with software wasn't a thing yet. And what he would do is he would take these big punch cards and in the US there was this IBM thing that was sort of the ID of the time, and he would use that to program these paper punch cards, and that's how he programmed, and he would bring these home every night. And my mom would tell me stories growing up about how she would draw over these with crayons and for her, that was part of her experience growing up. And since then, programming has evolved. So it was punch cards and then we had assembly, these first high-level languages, so COBOL and FORTRAN. And then we got, in the eighties, into Java and these typed languages in Haskell, and that was really exciting. And then in the nineties, we got into JavaScript and Python. So, these are interpreted languages that still give you a lot of safety. And I think about the programming language and the experience of using the programming language as evolving in lockstep. Because around the time that Java started appearing, you saw, for example, the Eclipse IDE, and it had the first type of head features. Like you could type a character and then you get a dropdown and the idea is like, do you mean this or this and this? And it was just incredible. 'Cause as a human, you didn't have to read the code anymore. And so I see this as the evolution, so this is the... Languages have sort of, I think, leveled out. All the modern languages belong to similar families. There's a few big families of languages, and they're pretty similar if you glance. But the experience is now really evolving where now you don't have to deal with punch cards or assembly or even code, you deal with prompts, and the model figures out the coding part, and this is just hugely exciting to me as a programmer. - Yeah, I love that. We've gone from punch cards to prompts basically. Well I have a couple questions for you on that front later, but before we get into that

I wanna talk a little bit on the model front. So, up until just recently, Claude Code was powered mainly by Claude 3. 7 Sonnet. So now with the Claude 4 models powering Claude Code under the hood, what has this unlocked and where do you think we're headed? - Yeah, maybe a couple months or so, before the models came out, we started trying them out internally, and I just remember my jaw dropping at how much more capable it feels. So, I think there's all these new use cases that are unlocked. when you're using Claude Code synchronously in a terminal, I think one of the big changes is Claude is just much better at holding your instructions. And so you tell it to do something either in a prompt or in Claude. md, and it'll tend to just do that and stick to it. And this is a big change, 'cause 3. 7 was kind of a beast. It's an amazing coding model, but man, it's hard to steer. Like you try to write tests, and it would just mock all your tests and you're like, "No, that's not what I mean. " And usually, you say that once or twice, and it figures it out, but it was just so powerful that it was worth it. And I feel like now with this new generation of 4 models, you don't have to do that anymore. They typically do what you want the first shot. And Opus is, it feels like this next level above Sonnet where not only does it understand my intent really well, but also, it's able to one-shot a lot of things that previous models just couldn't. So, for example, I haven't written a unit test in months, because Opus just writes my tests, and almost every time, it'll one shot it perfectly the first time. And this is pretty useful in a terminal. It makes it so it can be a little bit more hands off. But I think one of the coolest use cases is running it in GitHub Actions and other environments where you can give a task and then the model would just go off and do its own thing, and when it comes back with the right result the first time, that feels amazing. - So, with GitHub Actions now, we can, within GitHub, @Claude and then have it go off and work on the task in the background and then bring it back with a result and a new PR. - Exactly. - Okay. - Yeah, you open up Claude in the terminal, like normal, just run Claude, and then you run /install GitHub Action, and that'll walk you through this install step. There's a few steps. It's all automatic. You just have to click a button or two, and it'll install the Claude app in your GitHub repo. And yeah, the experience is pretty cool. So in any issue, you can @mention Claude, just @Claude. I use it in PRs every day. A coworker will put up a pull request, and instead of asking them, "Hey can you fix this thing? " I'll just say, "Hey @Claude, fix this thing," and it'll fix it. - Wow. - And instead of asking, "Can you write tests? " I always feel kind of guilty when I have to do that. I'll just say, "Hey @Claude, write tests. " It's just not a thing anymore. - I mean, that feels incredible to me. That is like an entirely new aspect of programming right there, where we can basically pull in your always on-demand programmer to go fix these issues for you, not even on your computer but operating in background. - Yeah, and I think it's the beginning of interacting with a model like you would with a fellow programmer. So instead of at mentioning a coworker, I would at mention Claude. - How does this change software engineering, when we move into that model? We're managing all these Claude codes in the background.

- I think there's a little of a mental, there's a bit of a mental shift that has to happen where some people really love controlling the code, and if you're used to handwriting code, I think now, the industry is shifting to a place where you're orchestrating agents that write your code, and it's more about reviewing code than handwriting code. And yeah, I think people have to deal with this transition, and I think as a programmer, it's incredibly exciting, 'cause you can do so much more so much faster. And there's still some stuff where I'll have to dip down and hand write code, but now I dread it, 'cause Claude is just so good at it. - Interesting. - And I think increasingly as models get more capable, these windows where you have to hand write code either 'cause it's a complex data model, or it's just something really sophisticated, like the interaction between a bunch of system components or something that's hard to type out in a prompt, I think this will keep receding, and more programming will be about orchestrating agents. - So I wanna dive into that a little bit more

on your type of workflow. So how are you currently using the combination of all these things from the IDE integration to just Claude Code as it is in the terminal to these background actions in GitHub? - Yeah, I think there's two kinds of work that I do. Some things are really easy. So, for example, writing some tests or making a little bug fix or something. And usually, I'll ask Claude to do it in GitHub issues. Or the other thing I'll do is I have a couple Claudes running in parallel usually. checkouts of our code base, and so in one of these terminal tabs, I'll ask Claude to do something. I'll hit shift-enter to enter into auto accept mode, and then I'll come back in a few minutes, and I'll get a terminal notification when Claude is done. - Oh wow. - There's a second kind of work where you have to be a little bit more involved, and I think this is still the majority of engineering. Most engineering, you can't one shot, it's still hard. And so what I'll do is I'll run Claude in my ID terminal, and I'll ask it to do something, and at some point, it'll get stuck, or the code won't be perfect or something. And so I'll go in and edit in my ID to get that last mile of edits, and- - I see, so there's this spectrum almost of difficulty of task compared to where you're interacting with Claude. - Yeah, there's a learning period when you first start using this kind of tool. I think something people do sometimes is try to use it for too much, and you give it something too hard, and it gets choked up and you're not happy with the result. And this is a learning that everyone has to go through to get internally calibrated on what can Claude do, what can it one-shot, what can it two-shot, and what's that interaction like? And unfortunately, it changes with every model, so you can't just run it once. Every time there's a new release, the capability grows, and Claude's able to do more stuff correctly the first time, and so you can ask it for a little bit more every time. - Right, I've noticed that just generally, even outside of code where these models are changing so fast and improving so fast that if you were to have tried a model six months ago and wrote it off for a task, it's not correct to still assume that frame in the present day. You almost have to reset your intuitions every single time. - Yeah, exactly. - Yeah. I'm curious for other maybe tips or tricks

that you're seeing from usage from devs or from people within Anthropic, what are some cool things that folks are doing with Claude Code? - Yeah, I'd say the biggest thing that I've seen power users start to do in an out of Anthropic is ask Claude to make a plan before it starts to code. And something people will do sometimes when they first start using Claude Code is they'll be like, "Hey, write this really big complicated feature," and then they get frustrated when it doesn't do it the way that they imagined in their mind. And a really good way to align what you want to do with what Claude wants to do is ask it to make a plan and run it by you first. And I'll explicitly say sometimes, "Here's the problem I wanna solve. "Before you code, brainstorm some ideas "and make me a list of ideas for how to approach it "and don't write any code yet. " And Claude will, you know, it'll give me option one, option two, option three. And I might be like, "Okay, option one and three sound good. "Let's combine it. Now you can start coding. " And it's generally pretty good at listening. Another thing to take this to the next level

is ask Claude to use extended thinking, and this works best if Claude has some context already. It doesn't work very well if Claude doesn't have any context yet, and it just thinks. And it's sort of like a human, right? You can think as long as you want, but until you go in and read the code, you won't actually know what it is you're doing. And it's the same thing with Claude, just ask it to read files first, then pause, and then I'll ask it to think and brainstorm some ideas, code. - Interesting, so that sort of interleaved approach where it's able to call out to a tool, think about the results, think about what it needs to do next, and then call another tool and continue that back and forth. - Yeah, exactly, exactly. And we actually see this on internal benchmarks too when we do internal benchmarks of different kind of evals. Generally, if you get context first and then think and then use tools to edit and use Bash and so on, the results are a lot better, and this is what it feels like too as a user. - Yeah, tell me about Claude. md files.

These seem really powerful. - Yeah, Claude. md, we use it for everything. It's Claude's memory. It's instructions to Claude that you want to share across our team. It's instructions that you wanna share across all your projects. So yeah, it's very powerful. There's a lot of different of Claude. md's. And so, the simplest kind is a file called Claude. md, and you put it in the root of your repository. - Just a markdown file. file, yeah. CLAUDE in all caps, md, lowercase. And Claude will automatically read it when you start Claude in that folder. So it'll be automatically read into context, any kind of instructions you want Claude to do every time, so Bash commands that you want it to run frequently or files that it should really know about when making changes or big architectural decisions, anything like that, MCP servers, just put it in the Claude. md. There's a second kind of Claude. md, and this one you check in, and so you wanna share it with your team. You wanna write it once and then share it with everyone across your team, so people don't have to write it themselves. - Yeah, interesting. - The second kind of Claude. md is just for you, and that's called Claude. local. md. And this one, it also goes in the same place, and it's just for you. You don't share it with your team. You can ignore it so you don't check it in. The third one is a global Claude. md, and that goes in your. Claude folder in your home directory. Most people actually don't use this one, but if you want, you can put any kind of instructions you wanna share across Claudes there. And then the final one is you can put Claude. md's in any nested file in any directory in your code base. - Oh wow. - And Claude will pull it in automatically when it thinks it's relevant just to get instructions about working with that part of the code base. - So these are, yeah, specific instructions or even your preferences for just coding style or anything like that, about how Claude should interact, what it should know about you, how you like to work, could be anything. - Exactly. And sometimes when I see Claude do something really good or really bad during a conversation, what I'll do is I'll hit the pound sign, and this goes into memory mode and I'll tell Claude, "Hey, you should memorize this. " And it could be an instruction, for example, "Whenever I make code changes, always around the linter," and I'll tell it that, and it'll incorporate it into the right memory file. - Interesting. I need to do more of that, I think.

What's next for Claude Code? - Yeah, I think there's two directions we're thinking about. So, one is how to make Claude work even better with all your tools. And it started with working with every terminal. Now it can work with many IDs, and it can also work with a lot of CI systems. And we've been thinking about what's next there, just to make sure it works with all the tools you use, Claude should know how to use them, and it should work with the tools natively. The second thing is how to get Claude better at these sort of easy tasks that you may not want to open a terminal for. So what if I could tag Claude in a chat app or something like this and have it fix an issue for me the same way that I can do on GitHub. And what does this mean, what feels good to use? So we've been trying a bunch of options here. We wanna make sure that it feels really good before we give it to users. - That's exciting. Well, I'd love to see Claude Code everywhere, so looking forward to it. And thank you, Boris, for the conversation. - Yeah, thank you.