KEVIN ROSE & KIERAN KLAASSEN (True Ventures / Every) - Compound Engineering

=== KEVIN ROSE & KIERAN KLAASSEN (True Ventures / Every) - Compound Engineering ===

(05:29:51): sweet kieran how are we doing doing good

(05:29:56): How are you?

(05:29:57): This is amazing.

(05:29:58): I have to say,

(05:30:00): I had no idea what we would talk about the whole day,

(05:30:04): but this inspires me to hear all of these perspectives.

(05:30:07): And it's so fun how similar or different,

(05:30:12): there is just something going on and that energy is just so cool.

(05:30:17): Everyone is trying to give words to it and figure out what it is and integrate it

(05:30:22): in their flows.

(05:30:23): And it's amazing.

(05:30:25): Yeah, love it.

(05:30:26): I'm learning stuff.

(05:30:27): Me too.

(05:30:28): I really am.

(05:30:29): It's so fun.

(05:30:31): And I'm really excited to have you here to host this conversation with our next guest.

(05:30:35): I'm adding to the stage, Kevin Rose.

(05:30:37): Kevin, welcome.

(05:30:38): Hello, hello.

(05:30:39): How's it going?

(05:30:40): Dan, you need some rest, man.

(05:30:42): This is insane.

(05:30:44): I don't know how you pull this off.

(05:30:46): I'm staying hydrated and I've got some I've got cashew butter.

(05:30:51): This is also a really good way to keep fueled.

(05:30:54): But thank you.

(05:30:55): I'm going to let you guys chat.

(05:30:57): Kevin just want to say like huge fan was a was an early dig user back in the day to

(05:31:01): see you back at it.

(05:31:03): And I'm super excited.

(05:31:04): You guys are going to chat about compound engineering here and take it away.

(05:31:09): Thank you.

(05:31:11): Dan, hopefully Dan takes a nap now.

(05:31:13): He has a routine of doing naps.

(05:31:15): He needs it.

(05:31:16): Thank you.

(05:31:18): Welcome.

(05:31:19): So cool.

(05:31:20): Welcome, Kevin.

(05:31:22): As Dan, obviously, I'm also a gigantic fan.

(05:31:27): Really cool thing, looking backwards.

(05:31:29): You're a big inspiration why I am in tech and really pushed me...

(05:31:35): with just everything you did from your newsletter to Dig to any podcast to build things.

(05:31:42): And that energy, I feel that today with many people.

(05:31:47): And it was just so cool to see you this week on X where someone said,

(05:31:53): oh,

(05:31:53): Kevin Rose is using your plugin.

(05:31:55): I was like, what?

(05:31:55): What?

(05:31:57): Why is he using my plugin?

(05:31:58): And then you said something like,

(05:32:00): I'm just sharing this for my friends because I'm excited and this is cool.

(05:32:05): And that just felt like this energy of creating, of building things.

(05:32:12): And it just felt like a full circle moment for me where I was like,

(05:32:15): yeah,

(05:32:15): you inspired me,

(05:32:16): I inspired you back.

(05:32:19): It's really cool.

(05:32:20): So...

(05:32:22): And thank you for all that you've done.

(05:32:24): And you've kind of blown my mind over the last couple of weeks.

(05:32:28): So a huge fan of your work as well.

(05:32:31): The feeling is mutual.

(05:32:33): Amazing.

(05:32:34): Yeah.

(05:32:34): So really what we're going to talk about is this concept of compound engineering.

(05:32:39): And it's a philosophy that many of us feel and are applying.

(05:32:44): And like everyone is like kind of understanding this is how it should be done.

(05:32:49): But like, how do we do it is always the question.

(05:32:52): And I created a plugin around it that I just...

(05:32:57): made by building Cora, which is my product for a year and a half.

(05:33:01): And now there are like tens of thousands of people using this plugin, which is insane.

(05:33:09): And it's not only using,

(05:33:10): but it's also like people like Ryan Carson this morning was on and he got inspired

(05:33:15): by creating like a product compounding flow for himself.

(05:33:19): And like he's also building, he just started full time on his own.

(05:33:24): uh app and just back at building again so uh and it's like this this is called vibe

(05:33:31): coding but like we're we're kind of at a moment where it's not vibe coding again

(05:33:36): it's not slop anymore we're actually building real companies and and that's kind of

(05:33:42): what you're doing as well like you're building you're you're doing stuff and

(05:33:48): And I would love to know,

(05:33:49): since you have all of this experience from building so many things,

(05:33:54): working with so many founders,

(05:33:56): how do you apply all that knowledge into this new world and relearning how we

(05:34:02): build?

(05:34:02): Because we can take inspiration from all the knowledge we have.

(05:34:07): Like Ryan said as well, I just ran companies with many engineers and did things.

(05:34:11): And now suddenly AI can do those things.

(05:34:14): So I would love to hear just your

(05:34:17): take like why did compound engineering resonate with you like why did you share it

(05:34:21): with your friends and like let's screen share and show stuff as well like it's just

(05:34:26): very interesting what's happening all this for sure yeah there's so much to talk

(05:34:31): about how much time do we have actually do you know we have 25 minutes left but we

(05:34:36): can do we

(05:34:38): Yeah, I mean, let's pick some things and we'll talk more at other points.

(05:34:43): But today we have 25 minutes because there is a lot of stuff going on.

(05:34:47): Oh, there's so much to talk about.

(05:34:48): So I will say that,

(05:34:51): yeah,

(05:34:51): I'm in agreeance with you in that this world of kind of vibe coding is

(05:34:56): transitioning from vibe to actually just coding,

(05:34:58): right?

(05:34:59): And when you see Linus Torvalds commit something via Claude or whatever he was

(05:35:05): using,

(05:35:05): I think it's safe to say this is kind of coding now.

(05:35:09): The difficult thing is that it is still a little bit of a kind of dark arts,

(05:35:14): kind of very confusing because there's two ways that people look at this largely

(05:35:19): that I see on the entrepreneurial front,

(05:35:20): which is they think of it as a fun little tool,

(05:35:23): which is a one-shot or type situation where they're like,

(05:35:26): oh,

(05:35:26): I just,

(05:35:27): you know,

(05:35:27): my workout app should be slightly different.

(05:35:30): I want a different functionality and I just want to say five sentences and then the

(05:35:34): app should just work and be awesome,

(05:35:35): right?

(05:35:37): And yeah, that's fun.

(05:35:38): It's a fun thing to try on some of these self-hosted do-it-all platforms.

(05:35:43): But really where things are getting serious is when you actually get into the

(05:35:47): terminal and you actually fire up VS Code or your favorite editor and

(05:35:52): And you're kind of in the weeds a little bit more.

(05:35:54): So for those of us like myself that,

(05:35:57): you know,

(05:35:57): I studied computer science in school,

(05:35:59): but I dropped out my second year and I didn't realize actually until a few weeks

(05:36:04): ago that I have this thing called aphantasia,

(05:36:06): which is a condition where people cannot visualize something in their mind's eye.

(05:36:11): And I was always wondering,

(05:36:13): and apparently hundreds of thousands of people have this,

(05:36:16): and I was always wondering,

(05:36:16): why can I just not remember syntax and certain things the way that my friends can?

(05:36:20): They were just so much faster than I was.

(05:36:23): So I understood technically what was going on, but my speed and efficiency was just shit.

(05:36:29): And so I realized I'm never going to be a developer, right?

(05:36:31): And so I kind of went more of the product design route,

(05:36:33): which I had a lot of fun doing and built a ton of different products.

(05:36:37): And now we're at this time where I can kind of do it all, right?

(05:36:41): And so,

(05:36:42): but to get high quality code out,

(05:36:44): I mean,

(05:36:45): yes,

(05:36:45): this changes every three months,

(05:36:47): but I take a look at my GitHub commits over time.

(05:36:51): I call it eight months ago.

(05:36:52): I was like, okay, I'm going to give this a shot.

(05:36:54): And I was not really impressed with the results.

(05:36:57): Like there was like endless bugs and I just couldn't quite get into it.

(05:37:00): And then I would jump into a function and try and figure it out on my own and try

(05:37:03): and like really navigate with the AI to get it to work through an issue.

(05:37:08): and then as the models matured and got better we were getting closer and closer to

(05:37:12): you know shippable products buggy sure uh production ready for scale probably not

(05:37:19): but we were getting closer right and i started developing kind of my own little

(05:37:23): internal system to do this which was i would go and turn on you know insert your

(05:37:28): favorite um you know top tier model here on their pro like deep thinking mode and i

(05:37:33): would just rant

(05:37:34): I would sit there and I'd hold down my function key and I would just go and I'd

(05:37:39): just rant with voice flow or whisper flow and just go for about 20 minutes or maybe

(05:37:46): 10 to 15 minutes and say,

(05:37:48): this is what I'm dreaming up.

(05:37:50): And then now go do all the deep research on how you would structure this, the database schema.

(05:37:55): You know,

(05:37:55): I would give it basics like I want this to be in Next.js,

(05:37:57): I want to deploy it on Vercel,

(05:37:59): I want it to be super based on the backend,

(05:38:00): and I would have kind of a,

(05:38:01): you know,

(05:38:02): overarching framework for what I wanted to do.

(05:38:04): And I'd get back this really massive document.

(05:38:08): And I found that when I could kind of distill that down and then send that,

(05:38:12): and yes,

(05:38:12): I'm front loading the model with a shit ton of tokens,

(05:38:16): I would have more success in terms of those initial first builds.

(05:38:20): And then there's all that weird stuff that happens around the compacting and trying

(05:38:24): to summarizing.

(05:38:24): And it gets to that 80% where it starts writing shitty code and all the stuff that

(05:38:29): you've gone and largely fixed.

(05:38:32): And then what I was doing is I was taking the work.

(05:38:34): It was funny because inside of Cursor,

(05:38:37): I would have two side panes.

(05:38:39): One would be Opus 4.5 and the other would be GPT-5.

(05:38:42): And what I would do is with GPT-5,

(05:38:44): I would act like it was just a friend that was looking at my code.

(05:38:48): And I'd let Opus do all the heavy lifting and I'd be like, hey, friend, how do I do?

(05:38:51): Check it out.

(05:38:52): What do you think?

(05:38:53): And it would go in and be like, oh, I wouldn't have done this or this or this.

(05:38:56): And then I would go and pass it back and be like,

(05:38:57): hey,

(05:38:58): my other friend looked at the code and said they would have changed these things.

(05:39:00): Don't agree with me.

(05:39:02): Don't just wholesale go in and make changes.

(05:39:05): Tell me why you're about to make those changes,

(05:39:07): why this is a better design decision than what you've done.

(05:39:10): And I was manually doing these multi-step,

(05:39:12): multi-agent process back and forth and back and forth.

(05:39:15): And I was taking my code quality from about,

(05:39:17): you know,

(05:39:17): call it like first compile,

(05:39:18): first build working to,

(05:39:19): you know,

(05:39:21): from call it 80% to like maybe close to 90% or something like that,

(05:39:23): right?

(05:39:24): Yeah.

(05:39:26): But it was very laborious.

(05:39:28): It was a lot of work.

(05:39:30): And then somebody told me about what you had created.

(05:39:32): And I'm like, oh, this changes everything.

(05:39:36): Because I would explain my process to friends that were tech savvy.

(05:39:40): And they're like, Kevin, this is too much work, man.

(05:39:42): I don't want to do all this shit.

(05:39:43): Like, come on.

(05:39:44): I have to say, yeah.

(05:39:45): Like, that's too much work, yeah.

(05:39:47): Yeah.

(05:39:47): And so then you went and created this.

(05:39:50): and i heard about i saw it on twitter if you were x few places and i was like okay

(05:39:55): well how do i i want to get a crash course on this i want to learn about this

(05:39:58): really quickly like what is this actually doing so actually are you familiar with

(05:40:03): deep wiki

(05:40:05): No.

(05:40:06): All right.

(05:40:07): Yeah.

(05:40:07): Is it,

(05:40:08): is it that from,

(05:40:09): is it the X one or is it the,

(05:40:11): it's where you can paste in any GitHub link and it gives you like the full.

(05:40:16): Okay.

(05:40:17): So it's awesome.

(05:40:17): So basically I use it and you can give it any GitHub link and it will build out a

(05:40:23): full,

(05:40:23): like deep document and you can ask questions of that GitHub and,

(05:40:28): and all of it.

(05:40:29): Right.

(05:40:29): So I actually, you know, I can, I can do a,

(05:40:33): Oh, it's the one from Devin.

(05:40:34): Yeah, yeah.

(05:40:34): Okay.

(05:40:35): I know the deep wiki then.

(05:40:36): Yeah.

(05:40:36): Yeah.

(05:40:37): So I went in there.

(05:40:38): Exactly.

(05:40:38): It's from Devin.

(05:40:38): Yeah.

(05:40:39): So I went in there and I did that with your plugin.

(05:40:42): And then I took back like this big technical brief and then I dump it in a notebook

(05:40:48): LM and I said,

(05:40:49): make me a 15 minute podcast on this.

(05:40:51): Love it.

(05:40:52): This is great.

(05:40:52): And then I listened back at 2X and in seven minutes,

(05:40:57): I'm like,

(05:40:57): oh yeah,

(05:40:58): I got to give this a try.

(05:40:58): Right?

(05:40:59): Yeah.

(05:41:00): And that's my flow because there's so much coming at us right now,

(05:41:04): like so many new products and features.

(05:41:06): How can we then take what multiple people are saying we should go spend hours on

(05:41:11): and compress it down into something that we can quickly grok and understand,

(05:41:14): does this make sense to take some of my precious time and go investigate whether

(05:41:19): there's a there there or not,

(05:41:21): right?

(05:41:21): And so that's what I did with – That's really funny.

(05:41:25): I love it, yeah.

(05:41:27): Can you share that with me?

(05:41:29): I would love to hear the podcast that got you into it, AI Generated.

(05:41:35): Oh, for sure.

(05:41:35): Yeah, I'll send it over.

(05:41:36): That would be fun.

(05:41:37): Nice.

(05:41:37): Yeah.

(05:41:38): So what I realized is that what you had done was the planning feature was phenomenal.

(05:41:43): I love this idea of deepening the plan as well,

(05:41:48): because you can look at a plan and pop it open and be like,

(05:41:51): ah,

(05:41:51): I know if I just added a little bit more structure here,

(05:41:53): my agent's probably going to have a better hit rate.

(05:41:57): And so having the optionality to deepen the plan is great.

(05:42:00): And then you have this whole loop.

(05:42:03): that is just beautiful that,

(05:42:04): I mean,

(05:42:05): obviously you can explain it better than I can,

(05:42:06): but everything that I've touched since then has just made my engineering efforts

(05:42:10): probably speed up.

(05:42:12): I'd say my code quality went from like 80 to like 97%.

(05:42:15): And then the time I've spent between jumping between different models,

(05:42:20): like I just don't have to.

(05:42:22): I just live in cloud code now with your plugin and I'm like good to go.

(05:42:27): Yeah.

(05:42:28): So it's kind of...

(05:42:31): Yeah,

(05:42:31): it's funny,

(05:42:32): like,

(05:42:32): if you say,

(05:42:33): is there any magic to this or something like that,

(05:42:35): like,

(05:42:36): I'm always thinking,

(05:42:37): like,

(05:42:37): this is just the obvious,

(05:42:39): because this is how we build software forever.

(05:42:42): And with the same with planning,

(05:42:45): like,

(05:42:45): for some reason,

(05:42:46): let me share my screen as well to show the screen,

(05:42:53): to show the plugin here.

(05:42:56): Yeah, so I think planning is very important.

(05:43:00): So this is the plugin,

(05:43:01): and this is the main workflow,

(05:43:03): planning,

(05:43:03): working,

(05:43:04): reviewing,

(05:43:04): compounding,

(05:43:05): and repeating.

(05:43:06): And if you think about normal engineering work, the first thing you do is you go plan.

(05:43:13): And for some reason, when we started AI coding, we just kind of forgot to plan.

(05:43:18): Yeah, we just one-to-one shot everything, yeah.

(05:43:21): Yeah, we're like, oh, I want this feature, do it.

(05:43:23): And we complain.

(05:43:26): And I'm like,

(05:43:26): yeah,

(05:43:26): but if you just think about it,

(05:43:28): how would you do this if you had a team of capable people?

(05:43:31): It's just, you go plan.

(05:43:33): So it is so obvious to me,

(05:43:35): and obviously everyone is doing planning now,

(05:43:38): but a year ago,

(05:43:39): not everyone was doing planning.

(05:43:42): So,

(05:43:43): yeah,

(05:43:43): it's just a realization of like,

(05:43:44): hey,

(05:43:45): actually,

(05:43:45): we have to look back at what we've always done,

(05:43:49): or maybe somewhere different where people already figured this out.

(05:43:54): Because it could be that we look at manufacturing, for example.

(05:43:58): Maybe we can learn there.

(05:44:00): Like,

(05:44:00): this is something Nate Bergetoff,

(05:44:03): he said,

(05:44:03): like,

(05:44:03): I'm being inspired because software engineering has become more of like a

(05:44:06): manufacturing industry.

(05:44:08): process where things are automated.

(05:44:10): So can we learn things from that industry?

(05:44:12): But also from like, if you are an IC, maybe you've never planned.

(05:44:18): So maybe that is also foreign to you.

(05:44:20): And you need to relearn what it is to be an IC.

(05:44:22): Because suddenly being an IC means planning.

(05:44:25): But also being a tech lead maybe now is also making

(05:44:30): like doing planning and managing things you've not really done as a tech lead ever.

(05:44:36): And I think the people that know all of this are CEOs and leaders.

(05:44:43): And you can see with Toby from Shopify,

(05:44:45): for example,

(05:44:46): he's just jamming,

(05:44:47): he's going and he gets it.

(05:44:49): And he recently said the ideal team now is a team of one.

(05:44:54): Yeah.

(05:44:54): Which is extreme.

(05:44:55): But if you just extend that.

(05:44:58): So I'm like, okay, planning is clearly what we need to do.

(05:45:03): But from your perspective as managing gigantic companies and things like that,

(05:45:08): what else can we learn from things that we already know?

(05:45:12): Where do we go?

(05:45:13): Because we have planning working with you and compounding.

(05:45:15): But yeah, there's more.

(05:45:17): Let me ask a couple questions.

(05:45:19): yeah i'm curious to get your take on this because in and this is just kind of my

(05:45:23): overarching thesis here and it's a very much a straw man and it's probably wrong

(05:45:27): but i'd love to get your take on it since we're on here in real time i feel that as

(05:45:32): i've coded with these models over time like they're being trained on high quality

(05:45:37): code and they're being trained on shitty code right yeah and so in some sense

(05:45:41): you're kind of getting some type of

(05:45:45): average of it all you know when you when you go to invoke something and say hey go

(05:45:49): do this thing and if you just say hey go build me this prototype you're not really

(05:45:54): somewhere hidden in there is the advanced engineer and somewhere hidden in there is

(05:45:58): the junior engineer in that big large language model right yeah and if you plan

(05:46:02): like a senior engineer in some sense it almost has to like call those engineers to

(05:46:08): the table

(05:46:09): to go and actually execute against that work.

(05:46:12): Is that wild to think or to say?

(05:46:14): Or would you disagree with that?

(05:46:16): I agree.

(05:46:17): Yeah, I think it's very true.

(05:46:19): Like if your plan is good.

(05:46:22): So I think it even goes to like if you as a person...

(05:46:26): talking to the computer because no one types, duh.

(05:46:30): But if you do that and plan and your plan gets really good because you know what's good.

(05:46:37): Like it kind of matches you.

(05:46:39): I think there's even research from Anthropic that came out this week that alludes

(05:46:44): to that where like if you are

(05:46:47): very knowledgeable, like it will respond differently to you than if you're not.

(05:46:52): So yes,

(05:46:54): and then on top of that,

(05:46:55): like the grounding you do in these flows is like you're pulling the things in that

(05:47:01): you want.

(05:47:01): So you're pulling the context in that you want, you're pulling the research in like your style.

(05:47:07): And you're just surrounded with all the stuff you wanted to do good with.

(05:47:11): And the planning phase, that's important.

(05:47:14): Because if the plan is not good, the work is also not going to be good.

(05:47:19): You're going to steer like a gigantic ship.

(05:47:23): It's way easier to catch mistakes in the planning phase than after that.

(05:47:30): Yeah,

(05:47:30): so I think it's one of the things that is for people just to catch them up real

(05:47:35): quick with 10 seconds of you install this plugin.

(05:47:38): And then when you're in cloud code, you do like slash forward slash workflows plan.

(05:47:43): And so it'll look like what you're seeing right now.

(05:47:46): And then you start

(05:47:48): you know, kind of pontificating into the AI, what you want to create.

(05:47:52): And then you hit enter and it goes,

(05:47:53): but it's important to know that you have to use that slash command to actually kick

(05:47:58): off this whole process.

(05:47:59): Now,

(05:48:00): when it's done,

(05:48:01): sometimes it asks follow-up questions and then I know I'm still working with it.

(05:48:06): Other times it kicks me back to a plain prompt and I don't know,

(05:48:09): do I need to re-invoke with the slash command or does it continue forward?

(05:48:14): Yeah.

(05:48:15): So the beauty is,

(05:48:16): if this plugin is loaded and you're in the same flow,

(05:48:19): in closed code,

(05:48:21): you don't need to invoke it.

(05:48:23): So actually you can say, can you use the compound?

(05:48:27): Okay, I'm typing here, that's against the rules.

(05:48:30): But you can even do this, create a plan.

(05:48:35): And it will do it as well.

(05:48:36): And this is the beauty of Claude.

(05:48:40): You can call slash command.

(05:48:41): So you don't need to do that.

(05:48:42): I'm always doing slash every time because I'm scared I'm going to lose it.

(05:48:47): You can do it.

(05:48:48): Yes, you can do it.

(05:48:49): You don't need to do it.

(05:48:52): And the beauty is also, it is flexible.

(05:48:55): So if you, for example, are...

(05:48:58): making a plan,

(05:48:59): if your job is to create plans,

(05:49:00): you can just use the plan mode and then hand it off to people to work on it instead

(05:49:05): of AI,

(05:49:05): which is...

(05:49:06): But it is composable.

(05:49:08): Like, you can just push it up to GitHub and say, this is the plan.

(05:49:11): I did my job here.

(05:49:12): So you can use it in different pieces as well.

(05:49:15): So many questions for you.

(05:49:17): To try to save some time,

(05:49:18): I want to get to a couple that I'm personally just so curious about,

(05:49:21): if you don't mind.

(05:49:23): Yes, let's do it.

(05:49:23): Throw stuff back at me as well.

(05:49:24): Yeah.

(05:49:26): the compound portion of it yeah can you tell people exactly what that is doing on

(05:49:32): the back end because I think this is the secret sauce here because planning we all

(05:49:36): get launching multiple agents like really cool they're all they all have really

(05:49:40): deep context windows they can do their best work that's fantastic uh

(05:49:45): The reviewing makes a ton of sense.

(05:49:46): Like you're just checking for security vulnerabilities.

(05:49:48): You're launching very specific agents that go out and look for performance.

(05:49:52): Like it's all kind of all in one, which is great.

(05:49:55): It's all the stuff you would normally want to think to do kind of already.

(05:49:59): But the compounding part is like you kind of don't know what's going on.

(05:50:03): You see it working, but technically what is happening there?

(05:50:07): Yeah, that's a good question.

(05:50:08): So like the planning and the working and the reviewing is in most systems or people do that.

(05:50:15): But the compounding is mostly the part people are not doing and the compounding part is really

(05:50:21): like either there is something that went wrong and you want to say,

(05:50:26): hey,

(05:50:26): next time,

(05:50:27): don't make this mistake again.

(05:50:28): So this is like a specific context you are in at a time,

(05:50:32): for example,

(05:50:33): it made a design decision that you think is very ugly.

(05:50:35): And what you then can do is run slash workflows compounds.

(05:50:41): And what it will do is we'll try to extract

(05:50:44): what it needs to learn from this.

(05:50:46): And what it does is it creates documentation.

(05:50:49): So there's a folder called docs,

(05:50:52): and it will create a markdown file with frontmatter in there with some keywords.

(05:50:57): And it will say, hey, like in this specific pattern, like we should do this instead of that.

(05:51:03): It's more or less what we used to do to say,

(05:51:06): hey,

(05:51:06): can you add this to ClaudeMD so we never make this mistake again?

(05:51:10): This is more of the way,

(05:51:11): if you do it like this,

(05:51:13): you can do thousands of files without bloating your context because we're not

(05:51:17): always loading all these things in.

(05:51:19): How do you know where the load is in?

(05:51:22): You don't need to load this in.

(05:51:23): How do you know, as the software creator, when to load them in?

(05:51:29): So how I know to load them in is I just have them baked into the planning and the review.

(05:51:37): So there is a specific agent that is the best practices,

(05:51:41): and that agent will start to do agentic search

(05:51:47): finding all the relevant files and maybe saying hey these are related and this uses

(05:51:52): haiku so it's faster it doesn't like it doesn't spend a lot of tokens and just says

(05:51:59): hey i found a few look at these because they're super applicable to the plan you

(05:52:03): have or the codes you have uh in the review and then it will just learn from that

(05:52:09): so that's auto shoves those into the context window as you're doing the planning

(05:52:14): yeah yeah okay so it's it's kind of similar to it's it's it's a way better because

(05:52:19): traditionally what would happen is they do like a a grep or a search and try and

(05:52:23): find bits and pieces of the code that might be relevant and inject them and you're

(05:52:27): like no no no these are some of the best practices and learnings that we've had

(05:52:32): let's push that into the window instead as a way to kind of prime the system and in

(05:52:36): essence yeah

(05:52:38): Yeah, it's priming the system.

(05:52:39): And it's also, like, I have many people now pushing PRs.

(05:52:43): And,

(05:52:43): like,

(05:52:43): someone pushed one that said,

(05:52:45): hey,

(05:52:46): actually,

(05:52:46): I ran the compound as,

(05:52:49): like,

(05:52:49): an initializer,

(05:52:50): basically.

(05:52:51): I said, okay, you just, like, extract everything from this repo and, like,

(05:52:55): create it now because you can learn stuff.

(05:52:57): And he said, that works very well.

(05:53:00): So I'm like, yeah, we need to have an init and a setup and all of that.

(05:53:03): So we're adding there.

(05:53:04): But the point is just having really good documentation about your code.

(05:53:10): And it's inspired by Context 7, for example.

(05:53:13): Like they just have very good.

(05:53:14): Yeah, Context 7 is great.

(05:53:16): Yeah, so like a context 7, but then for your own code.

(05:53:19): It's kind of something like that.

(05:53:21): And it's just files.

(05:53:23): So it's very clear what's happening.

(05:53:25): And you can make changes very easily as well.

(05:53:29): When you're coding and you're looking down and you're seeing that little wheel get

(05:53:32): to 80% on your context window,

(05:53:34): what do you do?

(05:53:35): Do you compound?

(05:53:35): Do you start a new window?

(05:53:37): Do you start freaking out?

(05:53:40): I'm pretty lazy.

(05:53:42): I know there are purists here that are very good at managing.

(05:53:46): I just think that's not my job.

(05:53:48): I think Anthropic should just figure that shit out,

(05:53:51): which we'll talk after you,

(05:53:54): Tariq from Anthropic,

(05:53:55): will be,

(05:53:55): so I'll just ask him to make it work.

(05:53:59): But I do think you need to manage your contacts, but they're doing lots of work.

(05:54:03): I think my job is not...

(05:54:06): like my job should be more on the like, what are the flows?

(05:54:09): What are like the next things there?

(05:54:12): And I think the next thing with compound engineering that like suddenly popped up

(05:54:17): is like with Opus 4.5,

(05:54:20): we can now run the entire

(05:54:22): a compound engineering flow.

(05:54:23): Like you were saying, I had to copy and paste.

(05:54:26): I felt that as well with the planning and the working because I had to do a plan

(05:54:30): and then it was like,

(05:54:30): yeah,

(05:54:31): it looks good.

(05:54:31): And then hit work and then deepen.

(05:54:34): I was like, why do I do all of these things manually still?

(05:54:37): So I introduced a new command called slash LFG.

(05:54:41): And basically that's just capturing all of those things in a relv loop.

(05:54:46): And what it does, it does everything.

(05:54:49): So it's just like, you have an idea.

(05:54:51): until PR is up with screenshots and videos and test it in a browser.

(05:54:56): It's just me trying to see where does this go.

(05:55:01): And it actually kind of works already.

(05:55:03): With Opus 4.5, we're at a point where you can do end-to-end flows like that.

(05:55:09): And then we get into having your Mac mini run 24-7.

(05:55:13): to just build features because suddenly that is an option because you don't need to

(05:55:19): do these separate commands anymore.

(05:55:21): It's funny you say that.

(05:55:21): I literally just bought a Mac Mini.

(05:55:24): I probably can vary it.

(05:55:25): Just around 24.7.

(05:55:26): Everyone.

(05:55:30): If anyone is not running a Mac Mini or something... You can fire up with BCP Ensense, right?

(05:55:36): And do the same thing, essentially.

(05:55:37): I think a Mac Mini is kind of cool because you have access to the Mac ecosystem and

(05:55:45): you can just use Apple Scripts.

(05:55:47): So I have it linked to my iMessage because it's a Mac.

(05:55:52): So I don't need to do anything.

(05:55:53): I just have Apple Scripts send me iMessages and stuff like that.

(05:55:56): So it's cool, but...

(05:55:59): So what's next?

(05:56:01): What does the next couple releases look like for you?

(05:56:05): So I think something you were sharing with me that you were vibe coding and working

(05:56:12): on prototypes before putting it into...

(05:56:18): I spent a lot of my time in V0 just to get...

(05:56:22): for me, a product has to feel right before I want to build it.

(05:56:26): So even though I kind of have to like,

(05:56:28): I'm a very visual person and I,

(05:56:31): and I want to see it and touch it.

(05:56:33): And,

(05:56:34): and I just like,

(05:56:34): I just crank on V zero,

(05:56:36): like,

(05:56:36): you know,

(05:56:37): I just let that go in multiple windows with multiple projects.

(05:56:39): And then when I get that gut feeling like this needs to exist,

(05:56:42): then I jump over and I'll just download that,

(05:56:45): that throw it into its own folder and then let compound engineering go to,

(05:56:49): go to town on it,

(05:56:50): you know?

(05:56:51): Yeah, so I'm doing exactly the same, which is cool.

(05:56:55): I just, yeah, I call mine Baby Cora, which is just my mock-up version.

(05:57:00): It's like Ryu inspired from Cursor.

(05:57:05): And I just Vibe code whatever I want.

(05:57:10): And this is a new version for Cora on iOS that I'm working on as well.

(05:57:15): So yeah, I think that's kind of next in compound engineering.

(05:57:19): I think we'll push a little bit like towards the design and the iterative sides

(05:57:24): because I think before the planning,

(05:57:27): you can do so much more like discovery.

(05:57:30): We have also a new command since yesterday that is brainstorm,

(05:57:35): which if you don't have a plan yet or need some more thinking,

(05:57:39): you can brainstorm.

(05:57:40): So that's really cool.

(05:57:41): And also, we are now on OpenCode.

(05:57:44): So if you use OpenCode, you can install it as well, which is very cool.

(05:57:48): I will be running experiments to see different models, how they work.

(05:57:51): And we're on Codex.

(05:57:52): So if you use Codex... By the way, you make DHH one of the reviewers when they do the review.

(05:57:59): Yes.

(05:58:00): You wrote that, right?

(05:58:01): Yeah, yeah, yeah.

(05:58:03): So I sent him his review of my code, and he had a good chuckle at it.

(05:58:09): He said, that's actually something I would have said.

(05:58:10): Oh, yeah.

(05:58:13): DHH reviewer is hilarious.

(05:58:15): I have people send me messages that say, just delete the whole plan.

(05:58:19): This is crap.

(05:58:20): Like that was his review.

(05:58:23): Just delete the plan or something like that.

(05:58:24): Yeah.

(05:58:25): There's something about like, I know, like I love how opinionated he is.

(05:58:28): Like that's the whole thing about,

(05:58:30): you know,

(05:58:32): 37signals and Jason Fried and DHH is like,

(05:58:34): you got to respect,

(05:58:36): like they do it their way.

(05:58:37): You know, there's something about that that is just like, no, no, no, this is the way we do it.

(05:58:41): You don't have to sign up for it, but this is the way we do it.

(05:58:44): Like I do love that.

(05:58:46): Yeah,

(05:58:46): and that's also kind of the spirit of the compound flow is like just having strong

(05:58:51): opinions about how to do software because otherwise

(05:58:55): it's just getting mediocre.

(05:58:57): And it doesn't mean you have to agree with DHHs all the time,

(05:59:00): but like sometimes the DHH reviewer has a point and it makes you think like,

(05:59:04): hey,

(05:59:05): maybe this is the complete wrong thing I'm building.

(05:59:07): And that's good.

(05:59:08): Like you want like opposites.

(05:59:11): Yeah,

(05:59:11): it's like a table side of mentors,

(05:59:14): you know,

(05:59:14): like it's just the table and they're just like,

(05:59:17): hey,

(05:59:18): this is the way I kind of think about it.

(05:59:19): And

(05:59:20): who doesn't want that you want people to be able to bounce around and have ideas

(05:59:24): with you know so but next time we have more time I'm going to show you some wild

(05:59:29): prototypes of weird stuff that I've built yeah I want to see it I want to see it

(05:59:33): Kevin I'm so curious about the prototypes this was an incredible conversation I

(05:59:39): love that little DHH anecdote that makes me so happy to hear um

(05:59:43): Well, and I sent him that you ported it, Kieran, this morning.

(05:59:45): I sent him a text that you ported it over because,

(05:59:47): you know,

(05:59:47): he's on Amarki,

(05:59:49): which is his version of Linux.

(05:59:50): And so the fact that it'll work on there will be great.

(05:59:54): So that's awesome.

(05:59:56): Great.

(05:59:57): Would love to have you back to demo some of the stuff you're working on.

(06:00:00): And thank you so much for joining.

(06:00:02): Yeah.

(06:00:02): Well, thanks for everything you guys do.

(06:00:03): Like I'm,

(06:00:04): I'm really,

(06:00:05): um,

(06:00:06): you know,

(06:00:06): it was so cool to discover your site and I signed up and I put my email in there

(06:00:10): and,

(06:00:11): you know,

(06:00:11): it's,

(06:00:12): uh,

(06:00:13): it's important that someone is like putting a lot of structure around this and

(06:00:16): educating people because,

(06:00:17): um,

(06:00:17): it is certainly the future.

(06:00:19): And I see a lot of teams that are like,

(06:00:21): they're still believe it or not on the fence of whether they should actually go

(06:00:24): into AI or not,

(06:00:25): which is just for us.

(06:00:26): It's like, once you, once you see it, you can't unsee it, you know?

(06:00:29): So it's very important that people have a sources they can trust of high quality information.

(06:00:34): So thanks for everything you do.

(06:00:35): Thank you.

(06:00:36): That makes me so happy to hear.

(06:00:37): I really appreciate it.

(06:00:39): Awesome.

(06:00:39): Have a great day.

(06:00:40): Have a good one.

(06:00:42): You heard it here, folks.

(06:00:44): Kevin Rose signed up for Every, and you should too.

(06:00:48): Every is the only subscription you need to stay at the edge of AI.

(06:00:51): We host stuff like this all the time.

(06:00:53): If you're an Every subscriber, you get access to ideas, apps, and training.

(06:00:58): On the idea side, we give you the best of what we know about how to build an AI.

(06:01:02): We do day of vibe checks of new models and new products.

(06:01:06): written videos live uh we also have a suite of apps that we build kieran is the gm

(06:01:12): of quora not monologue uh which is an ai email agent that he builds with his

(06:01:17): compound engineering plugin it's an entire email product that he builds by himself

(06:01:21): he has he has one other uh engineer who helps him but it's it's mainly his brain

(06:01:25): child um and he's also building the compound engineering plugin on the side

(06:01:30): So we're taking everything that Kieran learns building Cora and pushing it into

(06:01:33): all the ideas that all the things that we write and all the things that get

(06:01:36): published on every.

(06:01:37): We also have training, so we do these kinds of live streams a lot.

(06:01:40): We call them camps.

(06:01:41): This one is a free camp for everyone.

(06:01:43): We have about 15,000 people on here right now, which is so cool.

(06:01:48): I'm so happy.

(06:01:52): But we also do them for paid subscribers.

(06:01:53): So it's a more intimate setting.

(06:01:55): You get to talk to us, ask questions.

(06:01:57): We talk to you about what the latest of our workflows are.

(06:02:00): We bring in people from places like OpenAI and Cursor and other big companies to

(06:02:04): tell you about how they use these kinds of tools so you can stay at the edge with

(06:02:08): us.

(06:02:09): And that's all for one subscription.

(06:02:11): Go to every.to to subscribe.

(06:02:13): All right,

(06:02:14): now we have a special guest that I'm super,

(06:02:17): super,

(06:02:17): super excited to have,

(06:02:18): Dorek from Anthropic.