What made it work really well is that Claude Code has access to everything that an engineer does at the terminal. Everything you can do, Claude Code can do. There's nothing in between.

There's actually an increasing number of people internally at Anthropic that are using like a lot of credits, spending like over a thousand bucks every month. We see this like power user behavior. This is something that they teach in YC: if you can solve your own problem, it's much more likely you're solving the problem for others.

There's this really old idea in product called latent demand. You build a product in a way that is hackable, that is kind of open-ended enough that people can abuse it for other use cases it wasn't really designed for, and you build for that 'cause you kind of know there's demand for it.

Do you think the CLI is the final form factor? Are we going to be using Claude Code in the CLI primarily in a year or in three years, or is there something else that's better?

This podcast is sponsored by Google. Hey folks, I'm Omar, Product and Design Lead at Google DeepMind. We just launched a revamped vibe coding experience in AI Studio that lets you mix and match AI capabilities to turn your ideas into reality faster than ever. Just describe your app and Gemini will automatically wire up the right models and APIs for you. And if you need a spark, hit "I'm feeling lucky" and we'll help you get started. Head to aistudio.google.com/build to create your first app.

Cat, Boris, thank you so much for being here.

Thanks for having us.

Yeah. Um, so for people who don't know you, you are the creators of Claude Code. Thank you very much from the bottom of my heart. It's uh, I love Claude Code.

That's amazing to hear. That's what we love to hear.

Okay, I think the place I want to start is when I first used it. Um, there was like this moment, like I think it was around when Sonnet 3.7 came out, where I was like, I used it and I was like, "Holy [shit], this is like a completely new paradigm. It's a completely new way of thinking about code."

And the big difference was, um, you went all the way and just eliminated the text editor. And you're just like, all you do is talk to the terminal and that's it. Um, and you know, previous paradigms of AI programming, previous harnesses have been like, you have a text editor and you have the AI on the side and it's kind of like, or it's a tab complete. So, take me through like that decision process, that process of architecting this new paradigm. How did you think about that?

Yeah, I think the most important thing is it was not intentional at all. [Laughter] We sort of ended up with it. So at the time when I joined Anthropic, um, we were still on different teams at the time. There was this previous predecessor to Claude Code. It was called Clyde. And it was this like research project, you know, it took like a minute to start up. It was this kind of really heavy Python thing. It had to like run a bunch of indexing and stuff.

And when I joined, I wanted to ship my first PR and I hand-wrote it like a, you know, like a noob... I didn't know about any of these tools.

Thank you for admitting that.

I didn't know any better. And then I put up this PR and Adam Wolf, who was the manager for our team for a while, he was my ramp-up buddy and he just like rejected the PR. And he was like, "You wrote this by hand? What are you doing? Use Clyde." Cuz he was also hacking a lot on Clyde at the time.

And so I tried Clyde. I gave it the description of the task and it just like one-shot this thing. And this was like, you know, Sonnet 3.5. So I still had to fix a thing even for this kind of basic task and the harness was super old. So it took like 5 minutes to turn this thing out and just took forever. But it worked. And I was just mind-blown that this was even possible. And that just kind of got the gears turning. Maybe you don't actually need an IDE.

And then later on I was prototyping using the Anthropic API and the easiest way to do that was just building a little app in the terminal, 'cuz that way I didn't have to build a UI or anything. And I started just making a little chat app and then I just started thinking maybe we could do something a little bit like Clyde. So let me build like a little Clyde. And it actually ended up being a lot more useful than that without a lot of work.

And I think the biggest revelation for me was when we started to give the model tools. It just started using tools and it was just this insane moment. Like, the model just wants to use tools. Like we gave it Bash and it just started using Bash, writing AppleScript to like automate stuff in response to questions. And I was like, this is just the craziest thing. I've never seen anything like this. 'Cuz at the time I had only used IDEs with like, you know, like text editing, a little like one-line autocomplete, multi-line autocomplete, whatever.

So that's where this came from. It was this kind of convergence of like prototyping but also kind of seeing what's possible in kind of a very rough way. Um, and this thing ended up being surprisingly useful. And I think it was the same for us. I think for me it was like kind of Sonnet 4, Opus 4—that's where that magic moment was. I was like, "Oh my god, this thing works."

That's interesting. So like tell me about that tool moment because I think that is one of the special things about Claude Code is it just writes Bash and it's really good at it. And I think a lot of previous agent architectures, or even anyone building an agent today, your first instinct might be: "Okay, we're going to give it a 'find file' tool and then we're going to give it an 'open file' tool," and you build all these like custom wrappers for all the different actions you might want the agent to take. But Claude Code just uses Bash and it's like really good at it. So how do you think about what you learned from that?

Yeah, I think we're at this point right now where Claude Code actually has a bunch of tools. I think it's like a dozen or something like this. We actually like add and remove tools most weeks. So, this changes pretty often. Um, but today there actually is a tool for searching. Um, and we do this for two reasons. One is the UX, so we can show the result a little bit nicer to the user because there's still a human in the loop right now for most tasks. Uh, and the second one is for permissions. So, if you say in your Claude Code settings.json "this file you cannot read," we have to kind of enforce this. Uh, we enforce it for Bash but we can do it a little bit more efficiently if we have a specific search tool.

Um, but definitely we want to like unship tools and kind of keep it simple for the model. Um, like last week or two weeks ago we unshipped the LS tool because in the past we needed it but then we actually built a way to enforce this kind of permission system for Bash. Um, so in Bash, if we know that you're not allowed to read a particular directory, Claude's not allowed to LS that directory. And because we can enforce that consistently, we don't need this tool anymore. Um, and this is nice because it's a little less choice for Claude, a little less stuff in context.

Got it. And how do you guys split responsibility on the team?

Um, I would say Boris sets the technical direction and has been the product visionary for a lot of the features that we've come out with. I see myself as more of like a supporting role to make sure that, um, one, that like our pricing and packaging resonates with our users. Um, two, making sure that we're shepherding all our features across the launch process. So from like deciding, "All right, like these are the prototypes that we should definitely ant-food," to like setting the quality threshold for ant-fooding through to communicating that to our end users.

And um, there's definitely some new initiatives that we're working on that, uh, I would say historically a lot of Claude Code has been built bottoms-up. Like Boris and a lot of the core team members have just had these great ideas for to-do lists, sub-agents, hooks... like all these are bottoms-up. As we think about expanding to more services and bringing Claude Code to other places, I think a lot of those are more like, "All right, let's talk to customers. Let's bring engineers into those conversations and prioritize those services and knock them out."

Got it. What is ant-fooding?

Oh, ant-fooding is...

Oh, "ant fooding."

Oh, um, it means dog-fooding. [Laughter] So, Anthropic Ant.

I got it.

Yeah. Our nickname for internal employees is Ant. And so "ant-fooding" is our version of dog-fooding. Uh, internally over I think 70 or 80% of Ants—technical Anthropic employees—use Claude Code every day. And so every time we are thinking about a new feature, we push it out to people internally and we get so much feedback. We have a feedback channel—I think we get a post every five minutes. And so you get really quick signal on whether people like it, whether it's buggy, um, or whether uh it's not good and we should unship it.

You can tell um, you can tell that someone that is building stuff is using it all the time to build it. Uh because the like its ergonomics just make sense if you're trying to build stuff, and that only happens if you're like ant-fooding. [Laughter] Um, yeah. And I think that that's a really interesting paradigm for building new stuff, like that sort of bottoms-up "I make something for myself." Um, tell me about that.

Yeah. And Cat is also so humble. Um, I think Cat has a really big role in the product direction also. Like it comes from everyone on the team. And like these specific examples, this actually came from everyone on the team. Like to-do lists and sub-agents, that was Sid. Hooks, Dixon shipped that. Plugins, Daisy shipped that.

So like everyone on the team... like these ideas come from everyone. Um, and so I think for us like we build this core agent loop and this kind of core experience and then everyone on the team uses the product all the time. Uh, and so everyone outside the team uses the product all the time. And so there's just all these chances to build things that serve these needs.

Like for example, Bash mode. You know, like the exclamation mark and you can type in Bash commands. This was just like many months ago, I was using Claude Code and I was going back and forth between two terminals and just thought it was kind of annoying. Uh, and just on a whim, I asked Claude to kind of think of ideas. It thought of this like exclamation mark Bash mode. And then I was like, "Great, make it pink and then ship it." [Laughter] It just did it. And like that's the thing that still kind of persisted. And you know, now you see kind of others also kind of catching on to that.

That's funny. I actually didn't know that. And that's extremely useful because I always have to open up a new tab to like run any Bash commands. So you just do an exclamation point and then it just like runs it directly instead of filtering it through all the Claude stuff.

Yeah. And Claude Code sees the full output too.

Interesting. That's perfect. [Laughter]

So anything you see in the Claude Code view, Claude Code also sees. And this is kind of a UX thing that we're thinking about. Like in the past, tools were built for engineers, but now it's equal parts engineers and model. And so like as an engineer, you can see the output, but it's actually quite useful for the model also.

And this is part of the philosophy also, like everything is dual use. Um, so for example, the model can also call slash commands. So like you know, I have a slash command for /commit where I run through kind of a few different steps like diffing and generating a reasonable commit message and this kind of stuff. I run it manually, but also Claude can run this for me. Uh, and this is pretty useful because we get to share this logic. We get to kind of define this tool and then we both get to use it.

Yeah. What are the differences in designing tools that are dual use from designing tools that are, you know, used by one or the other?

Surprisingly, it's the same. So far.

Okay.

Yeah. I sort of feel like this kind of elegant design for humans translates really well to the models.

So, you're just thinking about what would make sense to you and the model generally... it makes sense to the model, too, if it makes sense to you.

Yeah. I think one of the really cool things about Claude Code being a terminal UI, and what made it work really well, is that Claude Code has access to everything that an engineer does at the terminal. And I think when it comes to whether the tool should be dual use or not, I think making them dual use actually makes the tools a lot easier to understand. It just means that: okay, everything you can do, Claude Code can do. There's nothing in between.

Yeah, that's interesting. Yeah, there are a couple of those decisions. So, um, no code editor, it's in the terminal, so it has access to your files. Um, and it's on your computer versus like in the cloud in a virtual machine. So you get like repeated... you get to use it in a repeated way where you can like, you know, build up your CLAUDE.md file or you know, like build slash commands and all that kind of stuff where it becomes very composable um, and extensible from a very simple starting point.

And I'm curious about how you think about, you know, for people who are thinking about, "Okay, I want to build an agent, I want to build probably not Claude Code, but like something else," how you get that simple package that then can extend and be really powerful over time.

For me, I start by just thinking about it like developing any kind of product where you have to solve the problem for yourself before you can solve it for others. And like this is something that they teach in YC is you have to start with yourself. So like if you can solve your own problem, it's much more likely you're solving the problem for others.

And I think for coding, starting locally is the reasonable thing. And you know, now we have Claude Code on the web. So you can also use it with a virtual machine and um, you know, you can use it in a remote setting. And this is super useful when you're on the go, you want to take that from your phone. And this is sort of... we started proving this out kind of a step at a time where you can do @claude in GitHub.

And uh, I use this every day like on the way to work I'm like at a red light—I probably shouldn't be doing this—but I'm like you know on GitHub at a red light and then I'm like @claude you know fix this issue or whatever. And so it's just real useful to be able to control it from your phone. Um, and this kind of proves out this experience. I don't know if this necessarily makes sense for every kind of use case. For coding, I think starting local is right. Um, I don't know if this is true for everything, though.

Got it. What are the slash commands you guys use?

/pr. [Laughter]

Yeah.

Um yeah, I think the pr command makes it a lot faster for Claude to know exactly what Bash commands to run in order to make a commit.

And what does the pr slash command do for people who are unfamiliar?

Oh, it just tells it like exactly how to make a commit.

Okay.

Um, and you can like dynam... you can say like, okay, these are the three Bash commands that need to be run. And what's pretty cool is also we have this kind of templating system built into slash commands. So we actually run the Bash commands ahead of time. They're like embedded into the slash command. Um, and you can also pre-allow certain tool invocations. So for that slash command we say "allow you know git commit, git push, gh" and so you don't get asked for permission after you run the slash command because we have like a permission-based security system.

Um, and then also it uses Haiku which is pretty cool. Um, so it's kind of a cheaper model and faster. Um, yeah and for me I use like commit, uh commit PR... feature-dev we use a lot. So like Sid created this one. It's kind of cool. So it kind of like walks you through step by step building something. So we prompt Claude to like first ask me what exactly I want, like build the specification, and then um, you know, kind of like build like a detailed plan and then make a to-do list, walk through step by step. So it's kind of like more structured feature development.

And then I think the last one that we probably use a lot... so we use like security-review for all of our PRs and then also code-review. Um, so like Claude does all of our code review internally at Anthropic. Um, you know, there's still a human approving it, but Claude does kind of the first step in code review. That's just a /code-review command.

Got it. Yeah. One of the things I would love to go deeper into [is] like how do you make a good plan? So, the sort of the feature dev thing, because I think there's a lot of like little tricks that um I'm starting to find, or people at Every are starting to find that work, and I'm curious like what are things that that we're missing.

So for example, one um step in the one unintuitive step of the plan development process is: even if I don't exactly know what the thing that needs to be built is, I just have like a little sentence in my mind like "I want feature X," I have Claude just like implement it just without giving it anything else. And I see what it does. And that helps me understand like, okay here's actually what I mean because it made all these different mistakes or like it did something that I didn't expect that might be better.

And then I use that like the learning from the sort of throwaway development—I just clear it out—and then that helps me write a better plan spec for the actual feature development, which is something that you would never do before because it'd be too expensive to just like yolo send an engineer on a feature that you hadn't actually spec'd out. But because you have Claude going through your codebase and doing stuff, you can like learn stuff from it. Um, that helps inform the actual plan that you make.

Yeah. I feel maybe I can start and I'm curious how you use it too. I think there's like a few different modes maybe for me. Like one is prototyping mode. So like traditional engineering prototyping, you want to kind of build the simplest possible thing that touches all the systems just so you can kind of get a vague sense of like what are the systems, where are the unknowns, and just to kind of trace through everything.

Um and so I do the exact same thing as you, Dan. Like Claude just does the thing and then I see where it messes up and then I'll ask it to just throw it away and do it again. So just hit escape twice, go back to the old checkpoint and then try again. I think there's also maybe two other kinds of tasks. So one is just things that Claude can one-shot and I feel pretty confident it can do it. So I'll just tell it and then I'll just go to a different tab and I'll shift-tab to auto-accept and then just go do something else or go to another one of my Claudes and tend to that while it does this.

Um, but also there's this kind of like harder feature development. So these are, you know, things that maybe in the past it would have taken like a few hours of engineering time. And for this, usually I'll shift-tab into plan mode and then align on the plan first before it even writes any code. Um, and and I think what's really hard about this is the boundary changes with every model, and in kind of a surprising way, where the newer models, they're more intelligent so the boundary of what you need plan mode for got pushed out like a little bit. Like before you used to need to plan, now you don't. And I think it's this general trend of like stuff that used to be scaffolding, with a more advanced model it gets pushed into the model itself and the model kind of tends to subsume everything over time.

Yeah. How do you think about like building an agent harness that isn't just going to like... you're not spending a bunch of time um building stuff that is just going to be subsumed into the model in 3 months when the new Claude comes out? Like, yeah, how do you know what to build versus what to just say "it doesn't work quite yet, but next time it's going to work, so we're not going to spend time on it"?

I think we build most things that we think would improve Claude Code's capabilities, even if that means we'll have to get rid of it in 3 months. If anything, we hope that we will get rid of it in three months.

Interesting.

Yeah. And an example of this is something like even like plan mode itself. I think we'll probably unship it at some point when Claude can just figure out from your intent that you probably want to plan first. Um, or you know, for example, I just deleted like 2,000 tokens or something from the system prompt yesterday just cuz like Sonnet 3.5 doesn't need it anymore. Um, but Opus 3 did need it.

What about um, you know, in the case where uh the latest frontier model doesn't need it but you know you're trying to figure out how to make it more efficient because you have so many users that you know... you're maybe you're not going to use Opus or Sonnet 3.5 for everything. Maybe you're going to use Haiku. So there's a trade-off between having a more um elaborate harness for Haiku versus just like not spending time on it, using Sonnet, eating the cost and working on more Frontier type stuff.

In general, we've positioned Claude Code to be a very premium offering. So, our north star is making sure that it works incredibly well with the absolutely most powerful model we have, which is Sonnet 3.5 right now. Um, we are investigating how to make it work really well for like future generations of smaller models, but it's um, it's not the top priority for us.

Okay. What do you think about, um, you know, one thing that I notice is we get models um often—and thank you very much for this—we get models a lot before they come out and it's our job to kind of figure out is it any good. And over the last six months, when I'm testing Claude for example in the Claude app with a new frontier model, it's actually very hard to tell immediately whether it's better.

Um, but it's really easy to tell in Claude Code because the harness matters a lot for the performance that you get out of the model. And you guys have the benefit of building Claude Code inside of Anthropic. So there's like a much tighter integration between um the fundamental like model training and the harness that you're building and and they seem to kind of like really impact each other. So how does that work internally and what are the benefits you get from having that like tight integration?

Yeah, I think the biggest thing is like researchers just use this and so you know, as they see what's working, what's not, they can improve stuff. Um we do like a lot of eval to kind of communicate back and forth and understand where exactly the model's at. Um, but yeah, there's this frontier where you need to give the model a hard enough task to really push the limit of the model. And if you don't do this, then all models are kind of equal. But if you give it a pretty hard task, you can tell the difference.

What sub-agents do you use?

Um, I have a few. I have like a planner sub-agent that I use. I have a code review sub-agent. Code review is actually something where sometimes I use a sub-agent, sometimes I use a slash command. So usually in CI [it's] a slash command, but in synchronous use I use a sub-agent for the same thing.

Um, it's a good question. Yeah, maybe it's like a matter of taste. Yeah, I don't know.

I don't know. Um I think it's maybe when you're running synchronously, it's kind of nice to fork off the context window a little bit because all the stuff that's going on in the code review, it's not relevant to what I'm doing next. But in CI, it just doesn't matter.

Are you ever spawning like 10 sub-agents at once? And for what?

For me, I do it mostly for like big migrations. This [is] like the big thing. Um, actually we have this like coder slash command that we use, there's a bunch of sub-agents there. And so one of the steps is like find all the issues. And so there's one sub-agent that's like checking for CLAUDE.md compliance. There's another sub-agent that's looking through git history to see what's going on. Another sub-agent that's looking for kind of obvious bugs.

And then we do this like kind of deduping quality step after. So they find a bunch of stuff. A lot of these are false positives and so then we spawn like five more sub-agents and these are all just like checking for false positives. And in the end, the result is awesome. It finds like all the real issues without the false issues.

That's great. I actually do that. Um, so one of my non-technical Claude Code use cases is um expense filing. So like when I'm in SF right now, so like I have all these expenses. And so I built this little Claude project that uh in Claude Code that um uses one of these, you know, finance APIs to just download all my credit card transactions. And then it decides like these are probably the expenses that I'm going to have to like file.

And then I have two sub-agents, one that represents me and one that represents the company. And they like do battle to like figure out like what's the proper um like actual set of expenses. [Laughter] Uh, it's like an auditor sub-agent and like you know "Pro-Dan" sub-agent. So um yeah that kind of thing, the sort of like opponent processor pattern seems to be like an interesting one.

Yeah. Yeah. It's cool. I feel like when sub-agents were first becoming a thing, actually what inspired us... there's like a Reddit thread a while back where someone made sub-agents for like a front end dev and a backend dev and like a think it was like a designer...

Testing dev.

Testing dev, like there was like a PM sub-agent. And this is like, you know, it's cute, like it feels like a little maybe too anthropomorphic. Um maybe maybe there's something to this but I think like the value is actually like the uncorrelated context windows where you have like these two context windows that don't know about each other and this is kind of interesting. Um and you tend to get better results this way. What about you? Do you have any interesting sub-agents you use?

So, I've been tinkering with one um that is really good at front-end testing. So, it uses Playwright to like see, all right, what are like all the errors that are client side and pull them in and try to test more steps of the app. Um, it's not totally there yet, but I'm seeing signs of life and I think it's the kind of thing that we could potentially um, bundle in one of our plugins marketplaces.

Yeah. Um, definitely. I've used something like that just with Puppeteer and just like watching it build something and then open up the browser and then be like, "Oh, I need to change this." It's like, "Oh my god."

Yeah. It's really cool.

It's really cool. I think we're starting to see the beginnings of this like massive... like massive sub-agents. I don't know what they call this, like swarms or something like that. There's a bunch of people—there's actually an increasing number of people internally at Anthropic that are using like a lot of credits every month, like you know like spending like over a thousand bucks every month. Um and this percent of people is growing actually pretty fast. And I think the common use case is like code migration. And so what they're doing is like framework A to framework B. There's like the main agent, it makes a big to-do list for everything and then just kind of map-reduce over a bunch of sub-agents. So you instruct Claude like "Yeah, like start 10 agents and then just go like you know 10 at a time and just migrate all the stuff over."

That's interesting. What would be like a concrete example of the kind of migration that you're talking about?

I think the most classic is like lint rules. So there's like you know there's some kind of lint rule you're rolling out. There's no autofixer because it's like you know like AST analysis can't really... it's kind of too simplistic for it. Um I think other stuff is like framework migrations, like um we just migrated from like one testing framework to a different one. That's a pretty common one where it's super easy to verify the output.

One of the things I found is—and this is both for projects inside of Every and then just open source projects—it's like if you're someone building a product and you want to build a feature that's um been done before. So maybe like an example that people might need to implement a bunch is like memory. How do you do memory? Um, because we have a bunch of different products internally, you can just like spawn Claude sub-agents to be like "How do these three other products do it?"

And there's like possibility for just like tacit code sharing where you don't need to like have an API or you don't need to like ask anyone. You can just be like "How do we do this already?" And then use the best practices to um uh to build your own. And you can also do that with open source because there's like tons of open source projects where people are like, you know, they've been working on memory for like a year and it's like really really good. You can be like "What are the patterns that um people have figured out and which ones do I want to implement?"

Totally. You can also connect your version control system. If you've built a similar feature in the past, Claude Code can use those APIs like query GitHub directly and find how people implemented a similar feature in the past and read that code and um copy the relevant parts.

Yeah. Is there um, have you found any use for like log files of "okay, here's the full history of like how I implemented it" and like is that important to give to Claude and how are you making it useful for it?

Some people swear by it. Uh there are some people at Anthropic where for every task they do, they tell Claude Code to write a diary entry in a specific format that just documents like what did it do, what did it try, why didn't it work. And then they even have these agents that like look over the past memory and synthesize it into observations.

I think this is like the starting budding... like there's like something interesting here that we could productize. Um but it's a new emerging pattern that we're seeing that works well. I think the hard thing about like one-shotting memory from just one transcript is that it's hard to know how relevant a specific instruction is to all future tasks. Like our canonical example is: if I say "make the button pink," I don't want you to remember to make all buttons pink in the future. And so I think um synthesizing memory from a lot of logs is a way to um find these patterns more um consistently.

It seems like you probably need like... there's some things where you're going to know um you'll be able to summarize or synthesize in this sort of like top-down way like "this will be useful later" and you'll know the right level of abstraction at which it might be useful. But then there's also a lot of stuff where it's like you actually you know any given like commit log like "make the button pink" it could be useful for kind of an infinite number of different reasons um that you're not going to know beforehand. So you also need the model to be able to look up all similar past, you know, commits and surface that at the right time. Is that something that you're also thinking about?

Yeah, I think I think there could there could be something like that. And maybe I think one way to see it is this kind of like traditional memory storage work, like Memex-like kind of stuff where you just want to like put all the information into the system and then it's kind of a retrieval problem after that.

Um, yeah. I think as the model also gets smarter, it naturally... I've seen it start to naturally do this also with Sonnet 3.5 where if it's stuck on something, it'll just naturally start looking—like we talked about before like using Bash spontaneously—so just like look through git history and be like, "Oh, okay. Yeah, this is kind of an interesting way to do it."

Yeah. One of the things that like we were talking before we started recording, one of the um things that we're doing inside of Every... like I feel like it has really um changed the way that we do engineering because everyone is Claude Code-pilled, like CLI-pilled. And um we have this engineering paradigm that we call "compounding engineering" where in normal engineering every feature you add it makes it harder to add the next feature, and in compounding engineering your goal is to make the next feature easier to build um from the feature that you just added.

And the way that we do that is we try to um codify all the learnings from um from everything that we've done to build the feature. So like you know, how did we make the plan and and what parts of the plan needed to be changed or like when we started testing it like what issues did we find? What are the things that we missed? Um and then we codify them back into all the prompts and all the sub-agents and all the slash commands so that the next time when someone does something like this uh it catches it and that makes it easier. And that's why for me, for example, I can like hop into one of our code bases and start like being productive even though I don't know anything about how the code works because we have this like built-up memory system of um of all the stuff that we've learned as we've implemented stuff. But we've had to build that ourselves. I'm curious, are you working on that kind of loop so Claude Code does that automatically?

Yeah, we're starting to think about it. Uh it's funny. We were just uh... we heard the same thing from Fiona. She just joined the team. And you know, she's our manager. She hasn't coded in like 10 years, something like that. And she was landing PRs on her first day. And she was like, "Yeah, like not only did I kind of... I forgot how to code and Claude Code kind of made it super easy to just get back into it, but also I didn't need to ramp up on any context because I kind of knew all this."

And I think a lot of it is about like when people put up pull requests for Claude Code itself—and I think our customers tell us that they do like similar stuff pretty often—um if you see a mistake I'll just be like "At Claude add this to CLAUDE.md so that the next time it just knows this automatically." And you can kind of like instill this memory in kind of a variety of ways. So you can say like "At Claude add it to CLAUDE.md." You can also say "At Claude write a test." You know, that's like an easy way to make sure this doesn't regress. And I don't feel bad asking anyone to write tests anymore, right? It's just like super easy. And like I think probably close to 100% of our tests are just written by Claude. And if they're bad, we just won't commit it. And then the good ones stay committed.

Um, and then also I think lint rules are a big one. So for stuff that's enforced pretty often, we actually have a bunch of internal lint rules. Claude writes 100% of these. Um, and this is mostly just like "At Claude in a PR write this lint rule." And yeah, there's sort of this problem right now about like how do you do this automatically?

And I think generally how Cat and I think about it is we see this like power user behavior and the first step is how do you enable that by making the product hackable so the best users can figure out how to do this cool new thing. But then really the hard work starts of like how do you take this and bring it to everyone else. Um, and for me, I keep myself in the "everyone else" bucket. Like, you know, I don't really know how to use Vim. Like, I don't have this like crazy tmux setup. So, I have like a pretty vanilla setup. So, if you can make a feature that I'll use, it's a pretty good indicator that like other kind of average engineers will use it.

That is interesting. Like, tell me about that because like that's something I think about all the time is um making something that is extensible and flexible enough that power users can find like novel ways to use it that you would not have even dreamed of. But it's also simple enough that anyone can use it and it's and they can be productive with it and you can you can kind of pull what the power users find back into like the basic experience. Like how do you think about making those design and product decisions so that you enable that?

In general, we think that like every engine environment is a little bit different from the others and so it's really important that every part of our system is extensible. Um, so everything from your status line to adding your own slash commands through to hooks, which let you um insert a bit of determinism at pretty much any step in Claude Code. So we think these are the basic building blocks that we give to every engineer that they can play with.

Um, for plugins... plugins is actually our, um—so it was built by Daisy on our team—and this is our attempt to make it a lot easier for the average user like us um to bring these slash commands and hooks into our workflows. And so what plugins does is it lets you browse existing MCP servers, existing hooks, existing plugins and just like or sorry existing slash commands and just let you write one command in Claude Code to pull that in for yourself.

There's this like really old idea in product called latent demand which I think is probably the main way that I personally think about product and like thinking about what to build next is. It's a super simple idea. It's: you build a product in a way that is hackable, that is kind of open-ended enough that people can abuse it for other use cases it wasn't really designed for. Then you see how people abuse it and then you build for that cuz like you kind of know there was demand for it, right?

Um and like you know when I was at Meta, this is how we built kind of all the big products. I think almost every single big product had this nugget of latent demand in it. Um you know like for example something like Facebook Dating, it came from this idea that when uh we looked at who looks at people's profiles, I think 60% of views were between people of opposite gender, so kind of like traditional setup, that were not friends with each other. And so we're like, "Oh man, okay maybe there's like... maybe if we like launch a dating product we can kind of harness this demand that exists."

That's interesting.

And for, you know, Marketplace it was pretty similar. I think it was like 40% of posts in Facebook groups at the time were buy/sell posts. And so like: okay, people are trying to use this product to buy and sell stuff. We just build a product around it, that's probably going to work. And so we think about it kind of similarly, but also we have the luxury of building for developers and developers love hacking stuff and they love customizing stuff. And it's like as a user of our own product, it makes it so fun to build and use this thing.

Um, and so yeah, like I said, we just build the right extension points. We see how people use it and that kind of tells us what to build next. Like for example, we got all these user requests where people were like, "Dude, Claude Code is asking me for all these permissions and I'm out here getting coffee. I don't know that it's asking me for permissions. How could I just get it to like ping me on Slack?" And so we built hooks. Uh, Dixon built hooks um so that people could get pinged on Slack, and you could get pinged on Slack for anything that you want to get pinged on Slack for. Um, and so it was very much like people really wanted the ability to do something. We didn't want to build the integration ourselves. And so we exposed hooks for people to do that.

The thing that makes me think of is um you recently um released... you kind of moved or rebranded how you talk about Claude Code to be this like more general purpose agent SDK. Is that driven by some latent demand where you you sort of saw there's like a more general purpose use case for what you built?

We realized that, similar to how you were talking about using Claude Code for things outside of coding, we saw this happen a lot. Like um we get a ton of stories of people who are using Claude Code to like help them write a blog and like manage all the like data inputs and take a first pass in their own tone. Um, we find people building like email assistants on this. Um I use it for a lot of just like market research.

Um because at the core it's like an agent that can just go on for an infinite amount of time as long as you give it a concrete task and it's able to fetch the right underlying data. So one of the things I was working on was I wanted to look at all the companies in the world and how many engineers they had and to create a ranking. And this is something that Claude Code can do even though it's not a traditional coding use case. So we realized that like the underlying primitives were really general as long as you have like an agent loop that can continue running for a long period of time and you're able to like access the internet and write code and run code... pretty much you can, if you squint, you can kind of build anything on it.

Mm-hm. And and I think like by at the point where we like rebranded it from the Claude Code SDK to the Claude Agent SDK, there were already like many thousands of companies using this thing and a lot of those use cases were not about coding. So it's like both internally and externally, we kind of saw that... like health assistants, like financial analysts, legal assistants. Um, it was pretty broad.

Yeah. What are the coolest ones?

I feel like actually you had Noah Brier on the podcast recently. I thought like the Obsidian, like kind of mind mapping/note keeping use case is really cool. It's funny, it's insane how many people use it for this particular combination.

Uh I think some other like some coding or kind of coding adjacent use cases that are kind of cool is um we have this like issue tracker for Claude Code. The team's just like constantly underwater like trying to keep up with all the issues coming in. There's just so many. And so Claude dedups the issues and it automatically finds duplicates and it's extremely good at it. It also does first pass resolution. So usually when there's an issue it'll um proactively put up a PR internally. And this is a new uh thing that Inigo on the team built. Um so this is pretty cool.

Uh there's also like on call and kind of collecting signals from other places like getting like Sentry logs and getting like logs from BigQuery and kind of collating all this. Um Claude's really good at doing this because it's all just Bash in the end, right? And so these are all kind of these internal use cases that I saw.

Is it... so when it's you know collating logs or um you doing issues, is that like you have Claudes like continually running in the background and is that something that you're building for?

Um it gets triggered for that particular one. It gets triggered whenever a new issue is filed. So it runs once but it can choose to run for as long as it needs.

Got it. What about the idea of Claudes always running?

Oo, proactive Claudes. I think it's definitely where we want to get to. I would say right now we're very focused on making Claude Code incredibly reliable for like individual tasks. And you know, if you think about like if you think about like multi-line autocomplete and then like single turn agents and then now we're working on like Claude Code that can complete tasks.

I feel like if you trace this curve eventually you go to even higher levels of abstraction, like even more complicated tasks, and then hopefully the next step after that is a lot more productivity. So just understanding what your team's goals are, what your goals are, being able to say "Hey, I think you probably want to try this feature and here's a first pass at the code and here are the assumptions I made and are these correct?"

I can't wait. Um and I think probably right after that is um Claude is now your manager. [Laughter]

Um, that's not in the plan. [Laughter]

So, everyone on the team was like super excited that uh we were we were talking today and they gave me a bunch of questions and I want to make sure I hit all the questions. Um, uh, oh, here's a good one. Why did you choose agentic RAG over vector search in your architecture? And are like vector embeddings uh still relevant?

Um so actually initially we did use vector embeddings. Um they're just really tricky to maintain because you have to continuously reindex the code and they might get out of date and you have local changes. So those need to make it in. And then as we thought about what does it feel like for an external enterprise to adopt it, we realized that this exposes a lot more surface area and like security risk. Um we also found that actually Claude Code is really good and Claude models are really good at agentic search. So um you can get to the same accuracy level with agentic search and it's just a much cleaner deployment story.

Hm, that's really interesting.

Um, if you do want to bring semantic search to Claude Code, you can do so via an MCP tool. So if you want to manage your own index and expose an MCP tool that lets Claude Code call that, that would work.

What do you think are the top MCPs to use with Claude Code?

Puppeteer and Playwright are pretty high up there.

Definitely. Yeah.

Sentry has a really good one. Asana has a really good one.

Hm. Do you think that there are um any power user tips that you see people inside of Anthropic or you know other people who are you know big Claude Code power users that people don't know about but they should?

Um, one thing that Claude Code doesn't naturally like to do, but that I personally find very useful is, um, Claude Code doesn't naturally like to ask questions. But you know, if you're brainstorming with a thought partner, a collaborator, usually you do ask questions back and forth to each other. And so, this is one of the things that, um, I like to do, especially in plan mode. I'll just tell Claude Code like, "Hey, we're just brainstorming this thing. Um, please ask me questions if there's anything you're unsure about." Um I want you to ask questions and it'll do it. And I think that actually helps you arrive at a better answer there.

There's also like so many tips that we can share. I think like there's a few really common mistakes I see people make. One is like you said, like not using plan mode enough. This is just super important. And I think this is people that are kind of new to AI coding. They kind of assume this thing can do anything and it can't. It's like not that good today and it's going to get better but today it can one-shot some tests, can't one-shot most things. Um, and so you kind of have to understand the limits and you have to understand like where you get in the loop. And so like something like plan mode, it can like 2x, 3x success rates pretty easily if you like land on the plan first.

Um, other stuff that I've seen power users do really well is companies that have really big deployments of Claude Code—and now um, you know, luckily there's a lot of these companies so we can kind of learn from them—uh having settings.json that you check into the codebase is really important because you can use this to pre-allow certain commands so you don't get permission prompted every time and also to block certain commands. Let's say you don't want web fetch or whatever and this way as an engineer I don't get prompted and um I can check this in and share it with the whole team so everyone gets to use it.

I get around that by just using dangerous mode, they skip permissions. [Laughter]

Yeah, we kind of have this here but we don't you know we don't recommend it. It's like it's a model, you know, it can do weird stuff.

Um, I think another kind of cool use case that we've seen is people using "stop hooks" for interesting stuff. So stop hook runs whenever the turn is complete. So like the assistant did some tool calls back and forth with you know whatever and uh it's done and it returns control back to the user, then we run the stop hook. And so you can define a stop hook that's like "if the tests don't pass return the text 'keep going'" and essentially it's like you can just like make the model like keep going until the thing is done. And this is just like insane when you combine it with the SDK and this kind of programmatic usage. You know, this is a stochastic thing, it's a nondeterministic thing, but with scaffolding you can get these deterministic outcomes.

So you guys started this sort of CLI, this CLI paradigm shift. Um, do you think the CLI is the final form factor? Are we are we going to be using Claude Code in the CLI primarily in a year or in three years, or is there something else that's better?

I mean, it's not the final form factor, but we are very focused on making sure the CLI is like the most intelligent that we can make it and that's as customizable as possible. You can talk about the next form factors.

Yeah, I mean... [Laughter] Cat sees asking me to talk about because no one knows like this stuff's like it's just moving like so fast, right? Like no one knows what these form factors are. Like right now I think our team is in experimentation mode. So we have uh CLI, then we came out with the IDE extension. Now we have a new IDE extension that's like a GUI. It's a little more accessible. Um we have @claude in GitHub so you can just add Claude anywhere. Um, now there's @claude, there's Claude on web and on mobile, so you can use it on any of these places. Um, and we're just in experimentation mode, so we're trying to figure out what's next.

I think like if we kind of zoom out and see where this stuff is headed. I think one of the big trends is longer periods of autonomy. And so with every model, we kind of time how long can the model just keep going and do tasks autonomously and just, you know, in dangerous mode in a container, keep autocompleting until the task is done. And now we're on the order of like double digit hours. I think it's like the last model is like 30 hours, something like this. And I, you know, the next model is going to be days. And as you think about kind of parallelizing models, um there's kind of a bunch of problems that come out of this. So one is what is the container this thing runs in because you don't want to have to like close your laptop. I have that right now because I'm doing a lot of uh DSPY or disb prompt optimization and like it's on my laptop and it's like I don't want to close... I'm like in the middle like with my laptop open because I'm like I don't want to close it.

Yeah.

Yeah. That's right. Yeah. We've like visited companies before like like customers that everyone's just like walking around with their like Claude Codes. [Laughter] Is this running?

So, I think like one is kind of getting away from this mode and then I also think pretty soon we're going to be in this mode of like Claudes monitoring Claudes. And kind of I don't know what the right form factor for this is because as as a human you need to be able to inspect this and kind of see what's going on. Um but also it needs to be Claude optimized where um you're optimizing for kind of bandwidth between like the Claude-to-Claude communication. Um so my prediction is terminal is not the final form factor. My prediction is there's going to be a few more form factors in the coming months, you know, maybe like year or something like that. And it's going to keep changing very quickly.

What do you think about, you know, I teach a lot of Claude Code to a lot of Every subscribers and...

Thank you.

You're welcome. Doing your work for you. [Laughter] Um uh and I think the like one of the big things is just the terminal is intimidating and uh just like being on a call with subscribers being like "here's how you open the terminal and you're allowed to do this even if you're non-technical" is like a big deal. [Laughter] How do you think about that?

Yeah, I um... one of the people on our marketing team uh started using Claude Code because she was writing some content that touched on Claude Code and I was like "You should really experience it" and she got like 30 popups on her screen where she had to accept various permissions because she'd never used a terminal before. So I completely see eye to eye with you on that. It's definitely um hard for non-engineers and there's even some engineers we've found who aren't fully comfortable with working day-to-day in the terminal.

Um, our VS Code GUI extension is our first step in that direction because you don't have to think about the terminal at all. It's like a traditional interface with a bunch of buttons. Um, I think we are working on more um graphical interfaces. Uh, Claude Code on the web is a GUI. I think that actually might be a good starting point for people who are less technical.

Yeah. Yeah. There was this like magic moment maybe like a few months ago where like I walked into the office and some of the data scientists at Anthropic like sit right next to the Claude Code team and the data scientists just had like Claude Code running on their computers and I was like "What is this? Like how did you figure this out?" I think it was like Brandon uh was like the first one to do it and he was like, "Oh yeah, I just like installed it. Like I work on this product so like I should use it." And I was like, "Oh my god." So he like he figured out how to like use a terminal and JS like you know he hasn't really done this kind of workflow before. Obviously like very technical.

Um so I think now we're starting to see all these kind of like code-adjacent uh like functions. People use Claude Code and um yeah it's kind of interesting like from a latent demand point of view these are people hacking the product so there's like demand to use it for this and so we want to make it a little bit easier with more accessible interfaces. But at the same time, for us, for Claude Code, we're laser focused on building the best product for the best engineers and so um we're focused on software engineering and we want to make this like really good but we want to make it a thing that other people can can hack.

Sometimes Claude Code will write code that's a bit verbose. Um, but you can just tell it to simplify it and it does a really good job.

Interesting. And so, and how are how and when are you doing that? So, you're using a slash command or you're...

I just say it.

Sometimes you're like, "Hey, this should be a one-line change" and I'll write five lines and you're like, "Simplify it" and it understands immediately what you mean and it'll fix it.

Yeah. I think a lot of people on our team do that, too.

Um, that's that's interesting. Why do you like why not then if you're saying that all the time why not then you know push that into like a slash command or the harness or something like that to yeah make it just happen automatically.

We do have instructions for this in the CLAUDE.md. I think it impacts such a low percentage of conversations that we don't want it to like over-rotate in the other direction. Um and then the reason why not a slash command is because you actually don't need that much context. I think slash commands are really good for situations where you would otherwise need to write two, three lines, but for simplify... even for plan mode you actually can use a few words but sometimes but it actually takes two or three lines to capture the entirety of what you want in plan mode. Um for simplify it you can just write "simplify it" and it gets it.

Yeah. Yeah, that makes sense. Cool.

Yeah.

Um okay, now we're we can... [Laughter] Um that's interesting. Yeah, but this stuff like you know it still feels just so early. You know, like we we were talking before before the recording about like kind of where are we on the adoption curve and it still...

The hype curve or whatever. [Laughter] Whatever that term was.

Exactly. And it just feels it just feels like we're you know like first 10% still like the stuff is going to change so fast it's going to keep changing.

Even when I talk to researchers outside of Anthropic who abuse Claude Code, um they also get stuck on things like this, like not realizing that they can just tell the LLM to simplify it. And I think that just goes to show that even for people who are like working in this industry they don't always realize that you can just talk to the model.

That's the thing is like I think that there's this underlying expectation that using AI shouldn't have to be a skill like because it just does whatever you say and you're like well I mean whatever you say is going to matter for what it does. So if you can say things better it's going to do better. [Laughter]

Yeah. I mean it changes with every model though. That's the that's the hard part. Like you know prompt engineer was a job and now famously it's not a job anymore and there's going to be more jobs that are then like not not jobs anymore of these kind of like little micro skills that you have to learn to use this thing and as the model gets better it can just like interpret it better.

But I think that's also like for us this is part of this kind of humility that we have to have building a product like this that we just really don't know what's next and we're just trying to figure it out kind of along with everyone else we're just here for the ride.

And that's why it's cool that you're building it for yourself cuz I think that's the that's the best way to know that is just like you're and this is what we do too is like you're sort of living in the future. You're using it all the time. And uh it's pretty clear what's missing. You're like I just want this thing and you can just do the next thing rather than being like "Hm let me ask like some enterprise product manager at like some gigantic company like what kind of AI feature do you want?" And they're like "I don't know like you know put a little chatbot on the side of my you know IDE" and you're like okay. [Laughter]

Yeah. This is like the luxurious thing about building dev tools, right? You're your own customer.

I think it's also really um a unique thing about AI because um it sort of reset the game board for all software. So um you know we have Kora this like email assistant and we have like Sparkle which organizes your files and it's like anything that you do for something that you want to use on your computer, if you're building it with AI there's a good chance that hasn't been done before because like the whole landscape has been reset. And so it's a uniquely exciting time to build stuff for yourself.

Totally. I think it totally opens the playing field, too. It's like any individual can now build an app to fill their need and then distribute it to everyone else.

Yeah, it's really cool. I've been prototyping all these like random pet projects. Um... I just moved into a new apartment and it's empty. And so I've been um I've been building this like shopping advisor assistant on like the Claude Agent SDK cuz who has time to like read all the reviews and like look at all the options and find their pricing and everything's like really hard to discover. And so it just like asks me a bunch of questions and I tell it what I want and it shows me a bunch of...

Yeah, exactly.

...and it shows me a bunch of photos of like different sofas and options and what people say online and then I tell it what I don't like and it's literally feels like working with a shopping assistant. And it's been really cool. Um I also have my little email response agent that like drafts responses for me but I don't use email that much so...

Oh, and I knew it wasn't you responding. [Laughter] That's why it's seven days delayed.

The agent's just taking doing a very thorough job. [Laughter] Yeah, Agent SDK is cool though.

Yeah, Agent SDK is cool.

Yeah, it always just feels amazing like how much we're able to build with such a small team.

Yeah.

So, I feel like the other thing that's really cool is that I think people are just shifting their mindset from docs to demos. Like internally, our currency is actually demos. It's like you want people to be excited about your thing? Yeah, show us like 15 seconds of what it can do.

And we find that everyone on the team now has this kind of indoctrinated demo culture for sure. And I think that's better because there's a lot of things that you might have in your head that if you're a great writer, maybe you could figure out how to explain it, but it's just even then it's just really hard to explain. But if someone can see it, they like get it immediately.

And I think that's happening for product building, but it's also happening for like all sorts of other types of creative endeavors like making a movie for example. Like you had to pitch it, but now you can just be like "I made this Sora video and like you know check..." like you can kind of see like the glimmer of the thing you're trying to make for very cheap. And so that means you don't have to spend time convincing people as much. You can just be like here I made it.

Yeah. And and also as a builder like you can just make it and then like make it again and then make it again [Laughter] until you're happy. Like I feel like that... like the flip side is like you used to make a doc or you know like whiteboard something or you know like I would draw stuff in like Sketch or Figma or whatever and now we'll just like build it until until I like how it feels. And it's just like so easy to get that feeling out of it now. And I think it's like you could see it visually before or you could describe it in words but it's like you could never get the vibe. And now like the vibe is really easy.

Yeah. And you built plan mode like three times.

Yeah. Yeah.

Because of this.

Like you you built it and then you threw it out and rebuilt it and then threw it out and rebuilt it. Yeah. Or like Todos... like Sid built the original version like also like three or four he built like three or four prototypes and then I prototyped maybe like 20 versions after that like in like a day. Yeah. I think this is like a lot of pretty much everything we released there was at least a few prototypes behind it.

How do you like um keep track of and carry forward the things you learn from prototype to prototype? And especially if it's like, you know, some one person is prototyping it and then you're like, I'm going to take it over. I'm going to do 20 more. Like how do you how do you maximize what you get out of that?

You know, it's like there's maybe a few elements of it. One is the style guide. So there's like some elements of style that we discover. And I think a lot of this is like building for the terminal or like we're kind of discovering a new design language for for the terminal and kind of building it as we go. Um, and I think some of this you can codify in a style guide. So this is our CLAUDE.md, but then there's this other part of it that's like kind of product sense where I don't think the model totally gets it yet. And I think maybe we should be trying to find ways to like teach the model this kind of product sense about like this works and this doesn't, right? Because in in product, you want to solve the person's problem in the simplest way possible and then delete everything else that's not that and just get everything out of the way. So you kind of you align the product to the intent as cleanly as possible. And maybe the model doesn't totally get that yet.

Yeah. It's never it doesn't really feel what it's like to use Claude Code. Like the model doesn't use Claude Code.

Yeah. Yeah. And so I think like when you know like Claude Code can like test itself and it can kind of use itself. Um and like we do this when developing and it can see like UI bugs and things like that. I don't know maybe we should just try prompting it though. It could like honestly a lot of the stuff is as simple as that. Like when there's some new idea usually you just prompt it and often it just works. Maybe we should just try that.

A lot of the prototypes are actually the UX interactions. Um, and so I think once we discover a new UX interaction like shift-tab for auto-accept—I think Boris figured out... um...

That was Igor actually.

Oh, Igor. Yeah, we went back and forth... can like fit into that.

We did like dueling prototypes for like a week. [Laughter]

Yeah, shift-tab felt really nice. And then one of the the now current plan mode iteration um uses shift-tab because it's actually just like another way to tell the model how agentic it should be. And so I think as more features use the same uh interaction, you form like a stronger mental model for what should go where.

Yeah. Or like thinking I think is another really good one. Like first we were like before we released Claude Code or maybe it was like the first thinking model was it like 3.7? I forget what the first one was. Yeah. But yeah and it was like it was able to think and we're like brainstorming like how do we like toggle thinking? And then someone was just like what if you just like ask the model to think in natural language and it knows how to think and we're like okay sweet let's do that. [Laughter]

And so like we did that for a while and then um we realized that people were accidentally toggling it. So they were like "don't think" and then the model was like "oh I should think." It just started thinking. And so we had to kind of like tune it out so you know "don't think" didn't trigger it but then it still wasn't obvious. But then we made a UX improvement to like highlight the thinking. That and that was like that was so fun and it felt really magical.

When you do ultra think it's like rainbow or whatever?

Exactly. [Laughter] And then with uh with Sonnet 3.5 we actually find like a really really big performance improvement when you turn on extended thinking um and so uh we made it really easy to toggle it because sometimes you want it... because you you kind of for a really simple task, you don't want the model to think for like five minutes. You want it to just do the thing. And so we used tab as the interaction to toggle it. And then we unshipped a bunch of the thinking words. Although I think we kept ultra-think just for like sentimental reasons. [Laughter] It was such a cool UX.

Interesting. Do you think there's some... there's some new metric that's about what you deleted? And I think programmers have always felt like, you know, deleting a bunch of code feels really good, but there's something about because you can build stuff so fast, it becomes more important to like also delete stuff.

I think my favorite kind of diff to see is a red diff. [Laughter] This is the best whenever I'm like, "Yeah, bring it on. Another one. Another one." Um, but it, you know, but it's hard because like anything you ship, people are using it. And so you got to keep people happy. And so I think generally our principle is if we unship something, we need to ship something even better um that can kind of um that people can take advantage of that that kind of matches that intent uh even better.

Um and yeah, I think this is kind of back to like how do you measure like Claude Code and the impact of it and this is something like every company every customer asks us about and I think like in so internally at Anthropic I think we like doubled in size since January or something like that but then productivity per engineer has increased like almost 70% in that time.

Um, measured by?

Uh I think we actually measured it yeah in a few ways but kind of PRs are the the simplest one and the main one. But like you said like this doesn't capture the full extent of it because a lot of this is like making it easier to prototype, making it easier to try new things, making it easier to [do] these things that you never would have tried because they're way below the cut line. Um you're launching a feature and there's this kind of like wish list of stuff now you just do all it because it's so easy and you just wouldn't have done it.

So yeah, it's really hard to talk about it. And then there's this flip side of it where more code is written. So you have to delete more code. You have to code review more carefully and you know automate code review as much as you can.

There's also like an interesting like new product management challenge because you can ship so much that you end up... it ends up not feeling as cohesive because you could just like add button here and like a tab there and like a little thing here. Like it's just it's much easier to build a product that has all the features you want but doesn't have any sort of organizing principle because you're just shipping lots of stuff all the time.

I think we try to be pretty disciplined about this and making sure that all the abstractions are really easy to understand for someone even if they just hear the name of the feature. We have this principle that I believe Boris brought to the team that I really like where we don't want a new user experience. Everything should be so intuitive that you just drop in and it just works. And I think that's really set the bar really high for making sure every feature is really intuitive. How do you do that with um a conversational UI? Because um you know when there's not a bunch of buttons and knobs and it's just a blank text box to start, how do you think about making it intuitive?

Um there's a lot of like little things that we do like um we teach people that they can use the question mark to see tips. Um we show tips as Claude Code is working. We have like the change log on the side. Um we tell you about like oh there's a new model that's out or like we show you at the bottom we have a notification section for thinking. I think there's just like subtle ways in which we tell users about features. I think the other thing that's really important is to just make sure that all the primitives are very clearly defined, like hooks have a common meaning um in the developer ecosystem. Plugins have a very common meaning in the developer ecosystem and just making sure that what we build matches what like the you know the average developer would immediately think of when they hear that.

There's also this like progressive disclosure thing like you know to to any anytime in Claude Code when you run it you can hit Ctrl+O to see like the full raw transcript, the same thing the model sees. And we don't like show you this until it's actually relevant. So when there's a tool result that's collapsed then we'll say "use Ctrl+O to see it." So we kind of we don't want to put too much complexity on you at the start because this thing can do you know anything.

Um I think there's this other kind of new principle which we've just started exploring which is like the model teaches you how to use the thing. And so you can ask Claude Code about itself and it kind of knows to look up its own documentation to tell you about it but we can also go even deeper. Like for example slash commands are a thing that people can use but also the model can call slash commands and maybe you see the model calling it and then you'll be like "oh yeah I guess I can do that too."

Yeah. Yeah. Yeah. Interesting. How has it changed like you know when you first started doing this Claude Code was this sort of like singular thing this singular way of thinking about you know using AI through a CLI. Other people had stuff like this but it felt like this shift and now there's a whole landscape of everyone is like going CLI, CLI, CLI. Like how has that changed how you think about building? How it feels to build? And how are you dealing with the sort of pressure of the race that you're in?

I think for me like imitation is the greatest flattery. Mm-hm. Um, so it's like it, you know, it's awesome and it's just like it's cool to see all this other stuff that everyone else is building like inspired by this. And I think this is ultimately the goal is to kind of inspire people to build this next thing for this just incredible technology that's coming. And that represents just really exciting. Personally, I don't really use a lot of other tools. So, usually when something new comes out, I'll maybe just try it to get a vibe. Um, but otherwise I think we're pretty focused on just solving problems that we have and our customers have and kind of building the next thing.

Cool. Sweet. Um, I loved this part of the interview, too. [Laughter]

Did we answer all of your team's questions?

Questions? Oh, did we get through all my team's questions? Let's see. Uh, I think we did. Um, uh, I'm curious also how you would answer like the unshipping question cuz also like if you're doing this kind of like AI-driven development, you ship a lot. You have a small team, so it's a lot of operational load.

The reason I asked that is because I don't think we do a good job of that. Um, and I have this feeling that some of the products are like a little bit messy because of that. And I think particularly for Kora, um there's just a big product surface area and it can do a lot of different things. Like we have an email assistant so you can ask it like you know "tell me about the trip I'm taking" and it'll go through all your emails and you know summarize the trip. Um or we have this feature that it automatically archives any email that you don't need to respond to immediately. Um, and then twice a day you get a brief that summarizes all the stuff that you probably need to see but you don't need to like actually do anything with and you just scroll through it and you're done.

Um, and there's just like all this... there's all this complexity around you know for example how are emails categorized? So now we have a whole view of we have all these categorization rules and you can order them and whatever, but like it's just complicated and hard to communicate and and uh and I want to retain a lot of the like all the power and flexibility, but also you can't look at a screen and be like "I have no idea what's going on. This is like way too complicated." So that's I'm just like I'm processing all that stuff. So the the kind of like deletion, you know, unshipping idea feels like an interesting um cultural principle that we haven't really explored.

Yeah, it's really hard. I think there's like a social cost to it, too, where like you kind of want to be the person who tells your coworker to unship their thing. [Laughter] It's definitely tricky. It's more than just the code. Yeah, I definitely learned this at Instagram honestly cuz I think Facebook does a terrible job at unshipping. And we had this problem where...

Every time we I think even like unshipping Pokes was like really spicy cuz there's a bunch of these like old-timers. They're like, "No pokes, you're never going to take it away." But like if you look at the data, no one really uses it anymore. But for sentimental reasons, they were kind of tied to it.

And so like for Facebook, it always... maybe nothing ever got unshipped. It always got moved to like a secondary place like a, you know, like an overflow menu somewhere that no one looks at, like a graveyard. And I think Instagram was just very principled. There was like, you know, very strong product and design point of view that was like, if this thing isn't used by like half of people, you know, 50% of WAUs or whatever, we're just going to delete it and deal with it and then we'll figure out some next thing that's used by more people.

I love it. Um, well, thank you. This was amazing. I'm really uh glad I got to talk to you and uh keep building.

Thank you for having us.

Yeah. Thanks.

Oh my gosh, folks. You absolutely positively have to smash that like button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard, but instead of gold, it's filled with pure unadulterated knowledge bombs about ChatGPT. Every episode is a roller coaster of emotions, insights, and laughter that will leave you on the edge of your seat, craving for more. It's not just a show, it's a journey into the future with Dan Shipper as the captain of the spaceship. So, do yourself a favor, hit like, smash subscribe, and strap in for the ride of your life. And now, without any further ado, let me just say, Dan, I'm absolutely hopelessly in love with you.