YT YouTube2Text

YouTube2Text → Transcripts → Reflecting on a year of Claude Code

Reflecting on a year of Claude Code — Transcript

by Claude · 3,969 words · 560 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00When we first released Claude Code,
0:02it was like a little video
0:03and I remember posting it to Slack,
0:04and there was like two people that gave like the reactions
0:07and like people were like excited.
0:09I thought it was really cool,
0:10especially for my very easy engineering tasks.
0:13It was quite good at it.
0:14That's like a really nice way to say that
0:16it wasn't really good.
0:31I can't believe it's only been a year
0:32since we first launched Claude Code.
0:34It's hard to remember what what that was like.
0:36Like, it’s so different than what we're doing today.
0:40Like, now I just have, like, armies of agents
0:42that are doing stuff like I'm
0:44prompting one agent or I have like an agent
0:46that's like prompting agents, that's prompting agents.
0:48And it's like a tree of like thousands of agents.
0:50But I think it's just like the most important idea
0:53when working on this stuff
0:54is like, every single time Claude makes a mistake.
0:57I don't tell Claude to do it differently,
0:58I tell it to write it to the CLAUDE.md,
1:00or to like make a skill or
1:02or something to do it differently.
1:04And if you can do this, then Claude
1:06can just like run forever.
1:07And I think the other thing that we kind of realize is
1:09the verification is really important.
1:11Like we didn't realize that.
1:12I hear this come up a lot with developers
1:15and enterprises that we meet with.
1:17What are your tips for making a really good
1:19making Claude Code really good at verification?
1:21I sort of feel like this
1:22is this thing that
1:23just like everyone misunderstands
1:25because whenever we talk about verification,
1:27people are thinking like unit tests
1:28or they're thinking like lint or like type check.
1:31These are the things that are
1:32obviously really easy to automate.
1:33And these are the things that were already automated.
1:36But actually when we talk about verification
1:38for agents, it's something slightly different.
1:40It's like can the agent run the thing?
1:42It takes a little bit of mental work
1:44to figure out how exactly do you do this,
1:45because it's often not straightforward.
1:47And I think that's like,
1:48that’s one of the challenges. I remember
1:50I remember with Opus 4
1:52Claude tested itself.
1:54And we just like hooked it up to Opus 4
1:58And I was like, Claude build the feature
1:59and then test yourself in like bash.
2:02And it opened a little Claude CLI and tested
2:05its own feature.
2:07And I was just like, whoa.
2:10It's crazy!
2:10Like now, now we're so used to it.
2:12Like now, you know, now we have these loops going for,
2:14you know, like the iOS simulator
2:15and the Android simulator and like computers
2:18for desktop, like it's not surprising.
2:20But back then that was crazy.
2:22How are, like, how are you doing it?
2:24So I've been mainly
2:25hacking on the desktop app these days.
2:27And one of the engineers on the team
2:29actually added this desktop
2:31development skill
2:32that teaches Claude how to run the local desktop app.
2:35And I've been having it use it,
2:36and it still runs into issues or like bugs
2:40with the staging environment sometimes.
2:42And so what I have it do is in those cases,
2:44I have it read Slack and understand, hey,
2:46is staging down right now?
2:48Or has someone else already hit this?
2:51And then when it
2:52debugs the whole issue,
2:54I tell it to update the desktop development skill.
2:56What the skill does is
2:57Claude actually spins up a local desktop app,
3:00and it uses computer use to click around on it.
3:03And so when I add a new UX,
3:05it clicks around to invoke the new UX.
3:07It also tests edge cases,
3:09and when there's an issue it fixes it and re-checks.
3:13This is like honestly, one of my favorite things
3:14about this team is everyone codes.
3:17I've never been on a team where, like,
3:22my PM would code
3:23and it's like crazy and like your code is like really good.
3:26You’re too nice.
3:28But I also just feel like it's
3:30it's also just becoming easier
3:31because it's like essentially Claude writes the code.
3:34And so what matters a little more is like,
3:36what's the idea that you have?
3:38And I feel like
3:39if you're a person
3:40that has like the product context
3:41and the business context and you're thinking
3:42about the design and the user,
3:44you're just going to come up with better ideas.
3:46It's kind of like all the roles are merging.
3:48I remember seeing Megan our designer’s PRs
3:50and I was just horrified at the beginning
3:52I was like, oh my God, why is Megan putting up PRs?
3:54And then she was like, yeah, yeah.
3:56I'm just like, I'm fixing the button.
3:57And I was like,
3:58okay, all right, well, the code looks good,
4:00so maybe it's maybe it's fine.
4:02And I feel like now it's just like it's totally normal.
4:04Yeah, and we see this across
4:05all the enterprises we talk with.
4:07Like, it's the engineers adopt Claude Code first
4:10and then the, the eng adjecent roles look over their shoulder
4:14and they're like, whoa, this thing is very powerful.
4:16Let me try it out. And we found it's crazy.
4:19We found that, like,
4:20our designers are more productive making prototypes
4:23and making changes directly in the app
4:25instead of pinging an engineer,
4:26PMs are making changes in the app.
4:29Our finance team runs and in Claude Code,
4:32they do their projections there.
4:34Data science.
4:36Like if you talk with our data scientists, it's so cool.
4:38It's just like everyone just has
4:40Claude Codes up on their screens.
4:42I feel like it's remarkably versatile
4:46for different roles.
4:47What do you feel like nowadays,
4:48are the use cases that are pushing the limit?
4:51One that I'm super excited about is routines.
4:54There is one engineer on our team
4:56who launched voice mode across all of our products.
4:59And, he has his routine set up that just listens for
5:03every ticket that comes, every GitHub issue, every
5:07bug report about voice mode.
5:08And his Claude just picks it up,
5:10proactively puts up a fix, and then pings the PR to him.
5:14And when he got that working
5:15for voice, he thought, okay,
5:17we're getting a lot
5:18of other feedback
5:19that isn't being responded to.
5:21So, he also set up a routine to listen for that.
5:24So I ship this, small feature.
5:26And there was like an edge case in it that I didn't see.
5:29And so someone filed a bug for it,
5:31and I was going to get to the bug that night.
5:34And my Claude was working, it said, wait a second,
5:37another Claude has already fixed this.
5:39And I was like, how is this possible?
5:40Like, I've never talked to him
5:41about this feature before
5:43and so I pinged him,
5:43and I was like, how did you fix this
5:45so quickly?
5:46And he said, he has another routine
5:47that just looks for bug reports
5:49that haven't been responded to in five hours
5:51and puts up a fix, and he merges the ones that
5:53are easy to verify.
5:55Claude tells me this like all the time now.
5:57That someone else has already fixed it?
5:59There's always like another person’s
6:00Claude that's working on it.
6:01It's like, yeah, that's been one of the changes.
6:04I feel like we're,
6:06a while ago we were trying to figure out,
6:07like, how to use routines,
6:08and I feel like just like the agent SDK
6:10was this first idea that we could use Claude Code
6:14programmatically. But
6:15I feel like at the beginning, it just wasn't obvious.
6:18How do we use it? What do we use it for?
6:20And I think routines are
6:21the first really obvious application.
6:24And I don't know, like
6:26it just does like all the code review,
6:28it babysits like every PR,
6:30you remember back in the day you used to actually
6:32have to like respond to code review comments.
6:34You used to have to like fix CI. You used to have to rebase.
6:38Yeah. Like I haven't done that in a long time.
6:40Yeah.
6:41When you're in the CLI
6:42and you're synchronously working with Claude,
6:45what are your go to features?
6:47Okay.
6:47What they used to be is plan mode.
6:49I don't use that anymore.
6:51What do you use instead?
6:52Auto mode.
6:53Auto mode?
6:53It’s the best.
6:54Instead of plan mode?
6:55Instead of plan mode.
6:56Yeah because the newer models
6:58they don't actually need like a planning step anymore.
7:01I think this was really
7:02important for like Opus 4 through maybe 4.5.
7:05Then I think starting with four six and definitely with
7:07four seven, it just doesn't need that planning step.
7:09I think some people still use it.
7:10They like to have that artifact.
7:11I don't use it
7:13And I just do auto mode for everything
7:14because then I start my Claude,
7:16it starts to work
7:17and then I just like move on to the next Claude
7:19and I don't have to sit there and watch it.
7:21But from the very early stage we had this
7:22like permission prompts model for Claude Code, right?
7:25Like it runs a tool and then it asks you like,
7:27hey, are you okay running this tool?
7:29And you had to say yes or no.
7:31And at the time,
7:32that was kind of the best we had a year and a half ago
7:34because we didn't have, you know, classifiers.
7:36The model was not as well aligned as it is today.
7:38So auto mode was just such a
7:40it was such a big step up because actually
7:42you don't want to read most of these requests.
7:44Just routing it to a different model
7:46and having it check for security works so much better.
7:48Yeah.
7:49And if a thing like is a little suss or,
7:51you know, this isn't the command that
7:53you think you want to run or it's not safe,
7:56the model will just deny it.
7:57And then you can go back and you can allow it later.
8:00I think this has been one of those, like, step changes.
8:02We just, there's no way
8:03we could have done this a year and a half ago.
8:04It's just human nature,
8:05when you accept 99% of requests, that your eyes
8:09just glaze over when you read it.
8:11And so actually, we feel that auto mode is more safe
8:14than reading
8:15every single permission prompt,
8:16because it means
8:17that your only paying attention to the most important thing
8:20and not like being spammed
8:22a bunch of things that are just 99% yes.
8:24I think security is one of these things.
8:26Like you can talk about it
8:27and then
8:28it's a totally different thing to actually do it correctly,
8:30because it just doesn't always look
8:32the way that you think it's going to look.
8:33And it's just all about
8:34like always red teaming, always pentesting
8:37always looking,
8:38you know, always having a threat model
8:39and then using that to figure out,
8:41you know, how is this thing going to get attacked?
8:43How are people going to get prompt injected?
8:45And I just feel like like the team
8:47is just like obsessed with this.
8:48And it's so important because as a result,
8:52I just trust the agent to run
8:53and I can move on
8:55and I can just have like a second agent.
8:57And if I didn't trust it,
8:58then I just wouldn't have been able to do that.
9:00And internally,
9:02to actually get auto mode out to our users,
9:05we needed to really trust it first.
9:07And so what we did was we collected thousands
9:10of transcripts of like an entire agent
9:14trajectory and a permission prompt
9:15and had auto mode classify whether or not it was safe.
9:18And it was extremely good at this.
9:20So then we got red teamers,
9:21and we asked them
9:22to try to prompt inject, and try to hack
9:25the code base.
9:27And we use this to create evals
9:28and make sure that all of these were denied.
9:30And then we had our own internal teams
9:33try to prompt inject and hack Claude Code’s auto mode.
9:37And then we improved auto mode
9:39to make sure that we caught all of these.
9:40So it's not only just protecting you
9:42against the vulnerabilities
9:43that are out there in the wild today, but,
9:45the most intelligent attacks that we can construct.
9:50Yeah. I mean, it's like,
9:51it’s honestly like a weird approach.
9:52I feel like there's, like, all these features
9:54the last year
9:55where the first time someone pitched it,
9:57I was like, no way, that's not going to work.
9:59And I feel like over time I just learned,
10:00like I'm actually wrong, like so often now.
10:03Because, like, building on the model is so weird.
10:06It's just like all this,
10:07like, engineering stuff
10:08that I've learned over the years.
10:09So much of it I just have to, like, throw out.
10:11And this is just like part of what the job is now.
10:13We're building on a new thing
10:14and we just have to relearn it.
10:16And auto mode was definitely one of these.
10:18I was like, the first time I heard it,
10:19I was like, route the prompt for a model?
10:21No way. That's not going to work.
10:23And then it actually turns out empirically,
10:24it works really, really well.
10:26But I heard you also love loop.
10:28Yeah, I love loop.
10:30How do you use it?
10:31I think for loop, there's
10:32this transition that we went through
10:34like a year and a half ago
10:36where we were like, all right there’s source code.
10:39But actually the thing an engineer should interact
10:42with, maybe it's not the source code,
10:44maybe it's the agent.
10:45And so we made this leap of
10:47I don't write the source code,
10:48I talked to an agent, and the agent writes
10:50the source code for me.
10:51And I think right now what's happening
10:53is we're making the next leap.
10:55I don't talk to an agent anymore.
10:56I talk to loop or I talk to a routine
10:59and it prompts Claude for me.
11:02And it's just it's crazy.
11:04I mean, it's been like, it's a year and a half
11:05and this was like two big leaps.
11:07If you take like, a step back, how are you
11:09seeing entire engineering orgs change?
11:12I'm going to put on my business cat hat.
11:14I have this, like, favorite case study.
11:16This is like a Harvard Business Review from the 90s.
11:18And they were talking about, like, computers are here.
11:20Why are we not seeing the productivity benefits?
11:23And it's just this, like amazing snapshot into like,
11:25what it actually felt like at the time
11:27because, like, you know, people used to use mainframes.
11:29At some point companies switch to personal computers.
11:32It was sort of a new thing, and the companies were trying
11:34to figure out
11:34how to use it.
11:35The same way they're trying to figure out
11:37how to use AI right now.
11:38And it turned out that to get the productivity
11:41benefits from computers, what you had to do isn't like
11:44you have your paper
11:45filing cabinet and your, like, paper and pen
11:47business process.
11:48And then there's like
11:49a computer on the side that does something.
11:51Actually, what you have to do
11:52is you throw out the filing cabinet,
11:54you have to throw out all your paper and all your pens,
11:56and then you put a computer in the center
11:58and everything has to run through the computer.
11:59It has to be at the center of every business process.
12:01And I feel like at Anthropic
12:03we do this thing where when you on board,
12:05you don't ask people questions like no one asks me
12:08questions when they on board.
12:09You probably have the same thing, they ask Claude.
12:12And this is kind of weird, like,
12:14this is the first company I've been at like that.
12:17And I feel like for us, Claude
12:18is just at the center of everything.
12:19Whenever I have a question, I ask Claude.
12:21Whenever I write code, I use Claude.
12:22Whenever I need a code review, Claude does it,
12:25whenever I need a security review, Claude does it,
12:27whenever I need to
12:28you know, fill out a form or something,
12:30Co-work does it.
12:31So it's just like Claude is at the center of everything.
12:33And I feel like the companies
12:35that are really figuring it out,
12:36and there's a bunch of them now,
12:38they're just putting Claude at the center of it.
12:40And I think for computers,
12:41the transition took 10 to 15 years.
12:43But actually for AI, because so much of our work
12:46is already digitized and Claude can use a computer
12:49and it can write code and run code.
12:52This transition is happening a lot faster.
12:54I think it's just like, really, it's
12:55just really exciting.
12:56Like,
12:57I feel like
12:58now I don't have to bug people anymore
13:00and when I
13:01interact with people, it's
13:02because it's like fun
13:02and I get to collaborate with them on stuff
13:04and we get to create something together.
13:06It's not that like, I need them.
13:08I need something,
13:09you know, from them
13:10because, like, Claude can actually do
13:11a lot of that stuff now.
13:13And I also feel like as an engineer,
13:14I've just never had this much
13:15fun doing engineering because the
13:17like the tedious part I don't have to do.
13:19Like I'm just coming up with ideas.
13:20I'm talking to customers and every idea, like,
13:24I don't have a to do list anymore.
13:25Like Claude just builds everything.
13:27And so my job is to come up
13:28with these ideas and it's just so fun.
13:30Okay, so here's a question.
13:31Is the future product or engineering?
13:33Like, is everyone going to be a PM
13:34or is everyone going to be an engineer?
13:36Everyone's going to be both.
13:38I feel pretty strongly that these roles are merging.
13:41Like when we look at our team, our product team
13:44all writes code.
13:45Our Devrel team all writes code. Our design team all writes code.
13:50And then we look at our
13:50engineers and a lot of them ship products end to end.
13:54They have an idea for what to build.
13:56They build it.
13:57They work with legal and marketing to figure out
14:00how we communicate this to the world
14:01and make sure it's safe and with security, too.
14:04And a lot of times they just see through
14:06this whole process end to end.
14:08So I think right now AI really benefits people
14:11people who have a lot of curiosity, have a lot of product tastes
14:14who love to have this like end to end ownership.
14:18And now a lot of people
14:19are running like hundreds of agents.
14:22What are the products that you think
14:23people should be adopting
14:24as they transition from single to multiple to hundreds?
14:29Until recently,
14:30the way that I wrote code was I had like six terminal tabs
14:34with six git checkout on the same repo,
14:36and then I would just like tab between them.
14:39Now it's pretty different.
14:40I have like one tab,
14:41I use the new agent view that we just shipped.
14:43It's like so good.
14:44And I'm so glad that we took a while
14:46to iterate on it to make that really good.
14:48And I also use the desktop app
14:50because I don't have to fiddle with checkouts that way.
14:53It just like, you know,
14:54it does the work tree cloning or the like, it
14:56creates the work trees for me.
14:58And the thing that I would not have expected
15:00six months ago is probably half my engineering now
15:03I do on my phone.
15:05So I just have like I have so many agents
15:07running that I just start for my phone.
15:09I use a remote control, which is like amazing now
15:12and like I will start something on my computer.
15:14And then I’ll just remote control in from my phone
15:16and I’ll just like, walk around
15:17I’ll get coffee,
15:18and then I'll check in on my agents
15:19and maybe I'll start another agent.
15:21And sometimes I'm like, talking to someone
15:23and we come up with a new idea.
15:25I’ll just start an agent on the spot.
15:26I like talk to it with voice mode
15:28and just have it build something,
15:30and I don't even have to go back to my computer anymore.
15:32I remember when you started doing this
15:33because you would actually
15:35leave work,
15:36have your computer on your desk
15:37open, plugged in, screen locked,
15:40and I just thought you would, like, come back to the office
15:42at some point to get your computer,
15:44but then it would be like pretty late and I was like,
15:46maybe he just left it here by accident.
15:48And then it happened again the next day.
15:50And then it happened again the next day.
15:52And I was like,
15:52wait, it's so weird because you're landing PRs
15:54but your computer is right next to me,
15:57and I remember you responding and you're like, yeah,
15:59I'm coding from my couch.
16:01Yeah,
16:02that was the week the remote control got really good.
16:04Yeah.
16:04So another thing that users
16:06are asking about all the time is
16:08how do you do context engineering,
16:10especially in a large enterprise?
16:12This is a thing.
16:12You know, people used to talk about prompt engineering.
16:14They used to like work context engineering.
16:16This is sort of matching where the model was at the time.
16:20Back in the days of Sonnet 3.5,
16:22you had to prompt engineer back in the days of Opus 4,
16:25you had to context engineer.
16:26But with the models of today, you don't do any of this.
16:29You give it the minimal possible system prompt,
16:32the minimal possible tools,
16:34and then you let the model figure it out.
16:36Like you just have to give
16:37the model some way to pull in the context.
16:38I think that's the most important thing.
16:40How do you think about it?
16:41I see things very similarly.
16:42I'm a context minimalist, so my general philosophy is
16:46tell the model only what it needs to know
16:50and let it figure out the rest of it.
16:53I think when you give the model
16:54too much context, it's kind of like
16:56you're micromanaging it.
16:58And sometimes the model knows a better way
17:00to get to the same outcome.
17:02And I personally prefer to give the model that freedom
17:05to do that.
17:07And then in general, we're
17:08also making our harness more lean
17:10so that you have more room for your own prompts.
17:13And so that follows your prompts better.
17:14There's all these different ways to Claude now,
17:17but I feel like in a year it's
17:18going to be a totally new set of things,
17:20and it's going to be so surprising
17:22if it's still these same things,
17:24because I think, like
17:25we're seeing these giant trends happening
17:27right now, agents are running for longer.
17:29They're more autonomous.
17:31Very rarely am I running one agent at a time.
17:33It's usually like a few agents
17:35or dozens or hundreds or thousands.
17:37And so like the form factor for that, it's
17:39going to be really different than what came before.
17:41And I don't know what it's going to be.
17:42And I think in a large part, it's
17:44going to be up to the team to figure it out.
17:46And this is,
17:47this is why I'm like,
17:48so happy we run the team that the way that we do,
17:51where everyone just comes up with ideas
17:52and everyone is able to think about the product.
17:55Everyone talks to users all the time
17:57because I don't think these ideas
17:58are going to come from us.
17:59It's going to come from the team.
18:00Totally, and from everyone
18:02in our community building with us.

About this transcript

This page contains the full transcript of Reflecting on a year of Claude Code by Claude, generated from the public captions YouTube serves with the video. The transcript has 3,969 words across 560 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.