Skip to content
ClaudeClaude

How the Claude Code team uses Claude Code

A year ago, using Claude Code meant prompting, giving feedback, and accepting permission prompts. What does it look like now? Thariq Shihipar, Sid Bidasaria, and Robert Boyce on the Claude Code team share how they use Claude Code in their day-to-day work, including why they do most of their coding through Claude Tag, why they give Claude goals rather than tasks, and why they delete harness features as the models outgrow them. They trace how the product expanded from the terminal into a multi-surface tool that incorporates primitives like auto mode, workflows, and routines. They also share what they miss most about their work as developers pre-Claude Code. 0:00 Intro 0:35 - Working through Claude Tag: from tool calls to goals 2:17 - Building on technology that evolves every two months 4:48 - How the Claude Code team uses Claude Code: AskUserQuestion, artifacts, and Claude Tag 6:41 - Running loops and routines remotely 8:52 - How code review inspired dynamic workflows 14:04 - Building Claude Tag with Claude Tag: verification and feedback loops in Slack 18:37 - What they miss about the old way of software engineering Get started with Claude Code: https://code.claude.com/docs/en/overview Follow ClaudeDevs on X for product updates and best practices from the Claude Code team: https://x.com/ClaudeDevs

Robert BoyceguestSid BidasariaguestThariq Shihiparhost
Sep 2, 202622mWatch on YouTube ↗

EVERY SPOKEN WORD

  1. 0:00 – 0:35

    Intro

    1. RB

      Do you guys have any thoughts about y- your next year?

    2. SB

      I feel like it's hard for me to think about the next two months [laughing]

    3. RB

      [laughing] Yeah.

    4. SB

      Let alone the next year.

    5. TS

      [upbeat music] I don't know if you guys know, it's been almost a year since I joined Claude Code. It's crazy to think about how far we've come from like a year ago when we were like prompting Claude Code, giving it feedback, accepting permission prompts, to now. So, like, tell me about how you work with Claude Code

  2. 0:35 – 2:17

    Working through Claude Tag: from tool calls to goals

    1. TS

      right now.

    2. RB

      Yeah, now I'm doing so much of my work through Claude Tag, actually.

    3. TS

      Yeah.

    4. RB

      Um, and the kinds of problems that I'm giving to Claude Tag are just worlds more complicated and challenging than, you know, implement this class or write this function, uh, which is kind of where things started at with, with Claude Code. I think that's really complementary to sort of the shape of, like, Claude Tag living in Slack as, like, a Slack-native agent because it means they can go and look up all this additional context around like, hey, what's the product context? Like, what are the decisions that we as a team have made about what this product should be? And that just makes its decision-making a lot better, um, having access to all that information and understanding how to incorporate that into the, the choices that the agent is making.

    5. SB

      Yeah. M-my work was pretty similar, too. I think seventy to eighty percent of my work happens on Claude Tag now, and for twenty percent of the work, I will maybe open up the TUI or the desktop app to, like, refine something, or if I w-want to be micromanaging my Claudes, I will, I will do that. I, I think we've gone from just, you know, caring about the transcript so much, and, like, caring about each individual tool call and each individual decision that the model makes to just zoomed-out view of like, "I have a goal, and I'd like to achieve this goal," um, and we-- and you just kind of give that goal to the model to, to achieve for you.

    6. TS

      You know, what I've been reflecting on is just how much the software development, like, life cycle has changed and, like, how we've sort of, like, learned all these lessons along the way, you know, like building Claude Code and building all these, like, primitives that, like, tie together eventually into Claude Tag, which is like, yeah, how we're doing everything from, like, verification to code review, like brainstorming, uh, monitoring.

  3. 2:17 – 4:48

    Building on technology that evolves every two months

    1. TS

      Tell me about how working on Claude Code compares to working on, uh, previous software products.

    2. SB

      Uh, I mean, there's always a shelf life-

    3. TS

      Yeah

    4. SB

      ... on any product or the technology underneath the products, rather.

    5. TS

      Yeah.

    6. SB

      And I think the shelf life for the technology used to be in the order of years.

    7. TS

      Yeah.

    8. SB

      Right? It used to be like, "I'm gonna make this product, and I have high confidence that this is going to be a thing for, you know, a few years or more." With models and the AI models, it's like the technology fundamentally shifts underneath you every two months.

    9. TS

      Yeah.

    10. SB

      And it's like maybe it's g- it's getting compressed as time moves forward, too. And so it's, it's kind of hard because you have to stay at the frontier, and you have to kind of go past the frontier really to, like, feel that edge.

    11. TS

      Yeah.

    12. SB

      Uh, but at the same time, you have to provide value to people using the models today-

    13. TS

      Yes

    14. SB

      ... uh, and in the now. And so it's this, like, balancing act, which is like part art, part science of like how do I, um-

    15. TS

      That's true

    16. SB

      ... how do I build for this technology that's moving so fast underneath me?

    17. TS

      So was there like a, a moment where you felt like you just caught the wave in the right part?

    18. SB

      I think the first time I felt the, the ride the wave feeling was maybe to-do lists.

    19. TS

      Oh, okay. Yeah.

    20. SB

      Like, that was... This was, like, maybe Sonnet 3.5, uh, and the models just were not able to do any kind of long horizon work.

    21. TS

      Yeah.

    22. SB

      And it w- You give it, give it, like, five things to do, and it would do three things and just give up. And so that's when we were like, "Oh, maybe you give it a to-do list, and it'll do really well." And it did really well. It was just like, whoa, like this, this is exactly the thing we need for this moment in time. But, you know, a year from then, like, it's disappeared. Like, you don't really need the to-do list anymore. Like, there's like, you know, more complex memory states and things that you can rely on. Uh, and I think that's like the perfect example of, like, how... Like, you gotta be very unattached to the things you're building-

    23. TS

      Yes

    24. SB

      ... uh, because they're gonna go away fairly quickly.

    25. RB

      I think that, like, the fact that we build these things into the Claude Code harness, and a lot of that is just to cover for some of the failure modes that the current state of the model has.

    26. TS

      Yeah.

    27. RB

      Um, and as the models get better, we have the freedom now to say, like, "Oh, actually all these, like, features that we built, like, it's not relevant anymore. We don't need those."

    28. TS

      Yeah.

    29. RB

      "We can get rid of them." And, like, the thing that, like, balances that out is, okay, well, now we're doing even bigger, more challenging tasks in Claude Code, and what are the tools that the model needs to be able to stay coherent and do a good job on those even bigger tasks? And it's like the, the tools that we need to give it, um, start to look different as the scope of the task changes.

  4. 4:48 – 6:41

    How the Claude Code team uses Claude Code: AskUserQuestion, artifacts, and Claude Tag

    1. TS

      A lot of people ask me, like, "How does the Claude Code team work," right? And I think one thing that's unique about, like, the Claude Code team is, like, we have to keep up with the models really closely, so we need to, like, you know, sometimes make big changes very quickly, where we're like, "Oh, the model is so good. Now we need to make a whole new proactive agent," and it needs to be-

    2. SB

      [laughs]

    3. TS

      ... like, a small team that does it, while it's also-- while we're also, like, uh, rewriting the permission system and creating artifacts and things. So I think in some ways we're this, like, hyperbolic time chamber of, like, where software development is going. Like, I don't think most products have this as short a shelf life-

    4. RB

      Mm-hmm

    5. TS

      ... you know, and have to keep up on the edge as much. But definitely as an engineer, this-- you kind of have to do some of this, right? You have to sort of like, you know, understand where the capabilities are going. And, um, obviously we try and help with that as much as possible through the product. But yeah, it's also good to, like, keep, like, your mental model of Claude as well.

    6. SB

      I mean, I think when you added, like, the AskUserQuestion tool-

    7. TS

      Right

    8. SB

      ... like, that was another kind of example of, like, riding through the wave. Like, what was-- what were you thinking when you, like, created that tool, and what, like, what was going on through your head?

    9. TS

      I wanted Claude to be interactive, and initially I had it after planning basically. Um, and, and then everyone's like, "Oh, wait, what if it's just a tool that it can call?" And it took me so long to design the tool. Like, it was, like, really hard to get Claude to- Call the tool well.

    10. SB

      Mm-hmm.

    11. TS

      You know, and then eventually I, I designed it, and then it just started getting really good at calling the tool. And then recently, I'm just like, I don't... I even use the tool as much. I just create an artifact, and the artifact asks me questions, you know?

    12. SB

      Yeah.

    13. TS

      And it's like in HTML and has diagrams and mock-ups and things like that, and like it happens so quickly, right-

    14. SB

      Yeah

    15. TS

      ... where it was like, oh, like the tool could barely do it, and now it can. But yeah, I think this happens a lot in Claude Code. I, I think that, like, the way I think we go about it is, like, we create these primitives, right?

    16. SB

      Mm-hmm.

    17. TS

      Where it's like, okay, um, you know, how do permissions work? Yeah, how do, like, uh, visualizations work, like all of these things, and then they sort of layer

  5. 6:41 – 8:52

    Running loops and routines remotely

    1. TS

      together. Tell me about loops, like how do you use them?

    2. SB

      Yeah, I mean, I think like it's, it's interesting to like track the journey of like how we arrived at loops in some ways, right?

    3. TS

      Yeah.

    4. SB

      It's like we... Like, I used to run everything locally on my laptop, and at some point I, I had to go back home, you know, from the office.

    5. TS

      [laughs]

    6. SB

      Uh, and so I would have to like shut down my laptop, and I'd be like, "That's, that sucks," like my agents aren't running anymore.

    7. TS

      Yeah, you don't like leave it open with like-

    8. SB

      Yeah.

    9. TS

      [laughs]

    10. SB

      I mean, you could, you could do that, but then I'd have to go to the parking lot, and like people would look at me funny, like it's just like-

    11. TS

      Yeah, yeah, yeah.

    12. SB

      Um, but you know, what I started doing then was I started using like our, like remote hosted devel-developer boxes.

    13. TS

      Yeah.

    14. SB

      And so like, those were my way of like, okay, I don't need my laptop on at all times. Like it's just, it's running, and it's running in a hosted environment, so I can just do my thing.

    15. TS

      Yep.

    16. SB

      Uh, but that's also unsatisfying in some ways because then like you gotta like, you know, SSH into your environments and kind of get back to your hosted surface. Uh, and so that's when I guess like Claude Code on the web was created, right? So it's like hosted containers that you have, uh, that can just keep running in the background. Um, but even then, like it was, it was a little tricky to get started because if you wanna use a container to host your, uh, host your dev environment, you need to give it access to your dev environment.

    17. TS

      Oh.

    18. SB

      And setting that up is a bit of a pain.

    19. TS

      Yeah.

    20. SB

      Uh, it's totally worth it in my opinion. I think it's like, you know, 10X to my productivity ever since we did that, but that was kind of the challenge there. And so I think that's the origin story of like how we kind of came across this, like Claude running locally to Claude needs to run on the cloud.

    21. TS

      Mm.

    22. SB

      And now if it's running on the cloud, you can do really interesting things.

    23. TS

      Yeah.

    24. SB

      Like you can have routines, and you can be like, you know, every day go and fi- uh, look at all the feedback that we're getting-

    25. TS

      Yeah

    26. SB

      ... and look at, you know, bucket them into buckets of importance and fix the ones that, um, that it actually has high confidence in fixing. This is the whole like loop journey-

    27. TS

      Yeah

    28. SB

      ... that Boris talks about a lot too. It's like you, you go from, you know, prompting a model in a session, uh, and you... to like breaking that boundary of like a session-

    29. TS

      Yeah

    30. SB

      ... and instead prompting something that's one level above that.

  6. 8:52 – 14:04

    How code review inspired dynamic workflows

    1. SB

      for you.

    2. TS

      Yeah. I actually wanted to talk more about code review, you know-

    3. SB

      Yeah

    4. TS

      ... and like security. Like, I, I think that's one of those things where like, okay, you're writing a lot more code, more code is h- like, you know, being generated, and you don't want your job to be like reviewing code, right? So, uh, yeah, obviously we have like, um, our code review bot right now. But generally, I wanted to ask you guys about like how, you know, code review has changed inside of the, the team and like how we do it right now.

    5. RB

      Uh, I think one big thing actually has been Claude really helps human reviewers-

    6. TS

      Mm

    7. RB

      ... uh, choose where to focus their time and energy. I think like a common thing that I saw happen in traditional, like human code review-

    8. TS

      Yeah

    9. RB

      ... is you would have someone go and review the, the code, and they would like pick out like, you know, three little nitpicks to post, and it was kind of just like I'm signaling to you that-

    10. TS

      Sure

    11. RB

      ... I read this code.

    12. TS

      Yeah. [laughs]

    13. RB

      And it's like, here's my like three little things that I wanna point out to prove that I actually was like reading through this carefully.

    14. TS

      Yeah.

    15. RB

      Now, A, you can have Claude figure out those little things and address them autonomously, so it's like that's not really the important part of human code review anymore.

    16. TS

      Yeah.

    17. RB

      Um, what is important is understanding maybe the bigger picture. It's like maybe Claude doesn't have a full understanding of why is this API structured the way it is or, you know, the service boundaries are drawn in a particular spot for, for a reason that maybe Claude, uh, hasn't fully internalized when it was writing the PR. So being able to pull information like that into code review, um, and then hand that information off to a human, I think is, uh, a really valuable thing that the Claude can do, and it keeps you from getting pulled into that little like, "Oh, I'm just gonna like nitpick something because I'm like reading through the code line by line," and, uh, just level up the like abstractions that you're thinking about as a human code reviewer.

    18. TS

      I think code review was the first time where we did like this big like fan-out, right? Where we're like, okay, like how do you find, you know, every bug, right? Or how do you like search across potential bugs, right? And we do this big fan-out and then, you know, coalesce it, which is like, you know, what we call test time compute, right?

    19. RB

      Mm-hmm.

    20. TS

      Like how we use like a lot of inference time to, uh, to put a lot of w- thought, uh, and compute into a problem. So, and this led to workflows, right? So could you, could you tell me about that, Sid, and how, how you built workflows?

    21. SB

      Code review was a big part of, uh, how workflows came about in the first place.

    22. TS

      Yeah.

    23. SB

      Like we, we have this, as you were saying, like massive, uh, fan-out thing where, uh, it was kind of like a, almost like a custom harness where-

    24. TS

      Yeah

    25. SB

      ... you tell Claude to go out and fan out and find bugs, and then for each bug you might do an adversarial review where you ask it to, uh, look at the bug from three different, you know, opinions or perspectives and, uh, see if the bug is actually real. What that does is like it filters out all these bugs for you and gives you just the, the most important ones that require or need your attention.

    26. TS

      Totally.

    27. SB

      And it's not just code review. Like, you could do the same thing for, for many, many different types or classes of problems.

    28. TS

      Yeah.

    29. SB

      Uh, like you could do the same thing for performance bugs or performance issues.

    30. TS

      Yeah.

  7. 14:04 – 18:37

    Building Claude Tag with Claude Tag: verification and feedback loops in Slack

    1. RB

      do here.

    2. TS

      Uh, wanna talk about Claude Tag. Like, I know, Robert, you're like building Claude Tag right now. Um, but yeah, how are you using it? Like, tell us about it.

    3. RB

      Yeah. Yeah, I think, uh, so I mentioned earlier, but we're using Claude Tag to build Claude Tag really aggressively.

    4. TS

      Yeah.

    5. RB

      Um, and that's actually been a lot of my focus is making sure that like Claude Tag is unblocked on being able to develop Claude Tag itself, making sure that the development environment and dev loop are really, um, really easy for Claude to use. It can just like do everything that I, as a human, need to do to build the software and test that it's working end to end. As we build more and more, uh, complex, heavily integrated pieces of software, that job gets harder.

    6. TS

      Right.

    7. RB

      Um, so that making sure that Claude is able to, to do that sort of thing has been a big focus for, for me. This is the first time we've had the user interface-

    8. TS

      Mm

    9. RB

      ... m- more fully decoupled from the transcript.

    10. TS

      Okay, yeah.

    11. RB

      So it's like we're actually like-- we are working now at a level, one level of abstraction away from like the actual tokens that are being output from Claude.

    12. TS

      Yes.

    13. RB

      So when you see a message from Claude Tag in Slack, it's calling a tool to send that message to you.

    14. TS

      Yeah.

    15. RB

      Uh, but its internal monologue essentially is, is not visible to you in Slack. Like, there is a link to view the full transcript, uh, from Claude.

    16. TS

      Yeah.

    17. RB

      But the primary user interface is through a messaging tool, and that abstraction was like a little scary to get through at first. We're like, oh, like, I can't see everything that Claude is like thinking as it's like thinking it.

    18. TS

      Yeah.

    19. RB

      Um, but it's also been, I think, really freeing in terms of just like, oh, Claude is choosing what to say to me and when.

    20. TS

      Mm.

    21. RB

      And I'm not thinking about the details of like, which tools is it calling? What parameters is it sending to the tools as it's using them?

    22. TS

      Yes.

    23. RB

      Um, and like being forced even-

    24. TS

      Yeah

    25. RB

      ... to let Claude cook.

    26. TS

      Mm-hmm. Yes.

    27. RB

      Um, and seeing that the models now are good enough-

    28. TS

      Yeah

    29. RB

      ... that I'm getting back good results-

    30. TS

      Yes

  8. 18:37 – 22:22

    What they miss about the old way of software engineering

    1. TS

      I, I think looking back at the evolution over the past year, to me, it's sort of like we started with Claude Code and then we sort of, like, started adding these primitives that would work, like auto mode and memory and, you know, workflows. There was a lot to keep track of-

    2. RB

      Mm-hmm

    3. TS

      ... as a developer, right? To just be like, oh, like, what, what's... Like, all this stuff happening, right? And but you kind of have to do that. I do think that's, like, the hard part about being a software engineer these days is, like, it's like freeing your mind, kind of. Like, like, there's so much... I think you always had to do this. You always had to be like, "Okay, what is possible now? How do we, like, automate more?" Um, but I think, like, how do we, you know, uh, step back and think bigger, right, I think is a, like, really important part. But I am curious if there's, like, anything you guys miss about, like, the old part of software engineering, you know, and, like, how things used to work.

    4. RB

      For me, I used to get a lot of joy from, uh, performance engineering.

    5. TS

      Yeah.

    6. RB

      Uh, and so that means, like, really digging into a system and, like, increasing its performance. Um, I think Claude is much better at that than I am now.

    7. TS

      Yeah.

    8. RB

      Uh, but I still get to reap the benefits of the actual performance improvements, which is nice. Uh, the other thing that I think I've been focusing a lot on is just creating new ideas quicker. So if I have an idea, going from idea to prototype to production is much faster.

    9. TS

      Yeah.

    10. RB

      And so I think more of my attention is now going on, like, new ideas and how quickly I can bring them to life.

    11. TS

      Yeah.

    12. RB

      And so it's like I've had to shift my kind of, like, paradigm from, uh, you know, getting really deep into the weeds to, like, just more zoomed out over time.

    13. TS

      Yeah, yeah.

    14. RB

      Yeah.

    15. TS

      Like, less mastering that, like, specific thing of, like, performance that you're spending a ton of time on.

    16. RB

      Exactly.

    17. TS

      Right. Yeah, yeah.

    18. RB

      I, I think kind of a similar thing for me. I'm a very, like, detail-oriented, love to get into, like, the weeds of, like, UI design. Um, for my personal website, like, I made this, like, super detailed CSS recreation of the macOS 10.4 Aqua, like, buttons, and I probably spent, like, a day, like, getting the exact layer of, like, radial gradients, like, stacked up to make-

    19. TS

      Right

    20. RB

      ... this perfect recreation of it. And, like, that's the type of thing where I'm like, oh, Claude can just do that for me now, and, like, I wouldn't ever go and do that manually now. It, it kind of brought me back to thinking about when I was probably seven or eight, I really wanted to make a video game.

    21. TS

      Yeah.

    22. RB

      And, like, I didn't know how to code yet.

    23. TS

      Totally.

    24. RB

      But I was just like, you know what? I have, like, PowerPoint-

    25. TS

      [laughs]

    26. RB

      ... and I can use, like, PowerPoint to, like, make, like, clickable things and, like, draw these shapes and stuff.

    27. TS

      Yeah.

    28. RB

      Because, like, th- that was a tool that I was able to use at that age. Thinking now about how Claude kind of makes the entirety of software engineering accessible to me, it's like-

    29. TS

      Yeah

    30. RB

      ... whatever idea I have, I can... I don't have to be like, "Oh, I don't have the technical skills," like, "I'm not able to, to solve this problem." Like, I can say, "Here's what I wanna accomplish. Let's, like, break it down and figure out how to, like, work with Claude to, to make it happen." Um, yeah, just everything kind of inside of that, uh, realm of software engineering feels like it's, it's within my grasp now, which is really exciting.

Episode duration: 22:22

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Transcript of episode S-sYlFiGFv8

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

iOSAndroidClaudeChrome