NAT ELIASON (Author) - ClawdBot 24/7 agent

=== NAT ELIASON (Author) - ClawdBot 24/7 agent ===

(02:30:23): Good friend.

(02:30:23): And Nat, welcome to the show.

(02:30:27): I've not seen you in a while,

(02:30:28): so I feel like the hair has gotten bigger and more and back as you've entered your

(02:30:34): daddy era and gotten...

(02:30:36): You have three kids now,

(02:30:37): so...

(02:30:39): Yeah, excited, excited to have you.

(02:30:41): The one thing Claude can't do is give me a haircut.

(02:30:45): So I've pretty much given up on that at this point.

(02:30:47): And we're just going to let it let it keep going until I can put a little robot in

(02:30:52): my office with me that can give me a haircut while I yell at the machine god.

(02:30:55): I love it.

(02:30:56): If you're if you're looking for a side project startup idea,

(02:30:59): a Claude bot to give Nat a haircut would be a really good one.

(02:31:03): It's coming.

(02:31:04): It's coming.

(02:31:06): Nat, we've known each other for a long time.

(02:31:08): We've been really good friends for a long time.

(02:31:09): You've been in the every universe for a while.

(02:31:12): You actually were one of the original authors of a really great every column on

(02:31:17): crypto called Almanac.

(02:31:19): You're also a fantastic author, the author of many books at this point.

(02:31:24): And you're a...

(02:31:26): inveterate vibe coder and note taker and systems person.

(02:31:31): For people who don't know of you, can you just give a little bit of your background?

(02:31:37): Yes.

(02:31:37): So obviously, thank you.

(02:31:39): That was a very nice introduction.

(02:31:41): And yeah, I mean, like Dan said, we've been in each other's

(02:31:46): life for, gosh, at least six or seven years now.

(02:31:49): And I've somehow managed to make a career out of just getting super interested and

(02:31:53): obsessed in things and going super deep on them and sharing what I'm learning along

(02:31:58): the way.

(02:31:59): And so I think beginning of 2024,

(02:32:05): uh i started using cursor to uh we didn't have this term yet we didn't know you

(02:32:11): know we didn't know it was going to be called vibe coding but using cursor to write

(02:32:14): code uh using like claude sonnet 3 5 i think was the model back then and it wasn't

(02:32:20): very good it was it was good uh it was it was way better than than having to write

(02:32:25): everything myself but the it was clear that the opportunity was there and i kind of

(02:32:30): Got pretty into it at the beginning of 24,

(02:32:33): stepped away from it for a while because it wasn't,

(02:32:35): you know,

(02:32:36): it couldn't quite do what I wanted it to do yet.

(02:32:39): Came back at the end of 2024 when, gosh, what was that, like 3.7 and maybe 4 were launching.

(02:32:46): And now there were agents,

(02:32:48): there was replit,

(02:32:49): and cursor had more of an agentic mode,

(02:32:53): and that was kind of becoming the thing.

(02:32:54): And then in January,

(02:32:55): basically a year ago,

(02:32:56): I launched a course called Build Your Own Apps,

(02:32:59): teaching other people how to vibe code.

(02:33:01): Unfortunately, missed the term vibe code by like a month.

(02:33:04): And so I didn't brand it around that, although it's such a good term.

(02:33:08): And I think for most of 25,

(02:33:09): vibe coding had this negative connotation to it,

(02:33:13): where it was like,

(02:33:14): oh,

(02:33:14): you're going to make slop.

(02:33:15): It's not going to make it to production.

(02:33:17): It breaks down when the apps get really large.

(02:33:20): And to a certain extent, that was true.

(02:33:23): unless you were quite skilled as an engineer or decently skilled as an engineer going into it.

(02:33:28): But really in the last two months,

(02:33:30): three months since Opus 4.5 and Codex 5.2 launched,

(02:33:34): we've crossed that threshold.

(02:33:40): where it really does feel like you can just go all the way with this stuff now.

(02:33:44): And I think we're in this really interesting point too,

(02:33:46): where you've got some of the most senior best engineers in the world coming out

(02:33:51): publicly and saying like,

(02:33:52): yeah,

(02:33:53): this is good enough now.

(02:33:56): It works as well as a real engineer.

(02:33:58): And when people like that are saying it,

(02:34:01): I think it's worth taking more seriously than somebody like me who's just been a

(02:34:05): hobby engineer all of my life.

(02:34:07): And so I know enough to be dangerous with these tools,

(02:34:10): but I'm certainly not like a 15-year software engineer,

(02:34:14): DHH,

(02:34:15): Creative Ruby on Rails,

(02:34:16): or whatever,

(02:34:16): saying that like,

(02:34:17): no,

(02:34:17): this is really good now.

(02:34:18): And now it is really good.

(02:34:19): And now you can do these incredible things.

(02:34:21): And now all this infrastructure has been built around it where you can basically be

(02:34:25): building 24-7 with the right orchestration setup.

(02:34:30): And that's basically my life now.

(02:34:34): So I've been deep on ClawdBot, which I think we're going to talk about today.

(02:34:39): And I literally went out on Monday and bought a separate Mac Mini to set up in my

(02:34:44): office so I could have ClawdBot running 24-7.

(02:34:46): And I'm literally sending it requests in bed at 10 PM,

(02:34:51): like,

(02:34:51): hey,

(02:34:51): Claude,

(02:34:51): I'm going to sleep,

(02:34:52): but why don't you start on this and just get it as far as you possibly can.

(02:34:56): And then I'll show you guys how I'm doing this.

(02:35:00): And then I just like wake up and I've got like a whole new feature built and it's been tested.

(02:35:04): And it's, I mean, it's magical.

(02:35:06): It's just really cool that we get to live in this world now.

(02:35:09): And I don't think most people realize how much you can do with it now,

(02:35:14): because unless,

(02:35:16): you know,

(02:35:16): you're like me and Dan and you haven't seen the sun in three months and you're just

(02:35:19): like locked into X getting the latest AI updates,

(02:35:22): you might still be in the mental mind.

(02:35:24): I mean, if you're here, you're probably not in this mental model, but a lot

(02:35:26): of people are still in the mental model of like oh it's it's going to be slop like

(02:35:30): it's not good they you know they they haven't used codex 5 2 they've only used

(02:35:34): codex 5 they haven't used opus 4 5 they've only used sonnet 4 5 right they're like

(02:35:39): two months out of date those two months make a big difference and i mean even like

(02:35:43): claude bot this didn't really exist as a thing two weeks ago i mean it was out but

(02:35:47): it wasn't like as known and really in the last two weeks we've seen this new it

(02:35:51): feels like um

(02:35:53): seismic shift in what you can do where,

(02:35:56): you know,

(02:35:56): I felt a little bit crazy before paying $200 a month for Claude,

(02:35:59): but now I have two $200 a month Claude subscriptions because like it's worth it

(02:36:03): just because you can do so much now.

(02:36:05): I love that you have two,

(02:36:06): and I also love that one of the unintended consequences of being near AGI is that

(02:36:13): Mac minis are the thing to buy right now.

(02:36:15): Who would have thought?

(02:36:19): Yeah.

(02:36:19): It's like,

(02:36:20): oh,

(02:36:20): this $500 device with a $200 a month subscription is literally like hiring another

(02:36:26): engineer.

(02:36:28): And maybe in some ways better,

(02:36:31): in some ways worse,

(02:36:31): because it's not quite as independent,

(02:36:34): although it sort of is.

(02:36:35): I'll show you something I've built in a minute here.

(02:36:37): But if I hired somebody,

(02:36:40): I couldn't text them at 1030 and ask them to have something built for me by the

(02:36:44): time I wake up.

(02:36:45): I mean, I could, but they probably wouldn't keep working for me for very long.

(02:36:48): But Claude doesn't care.

(02:36:49): He really doesn't sleep.

(02:36:52): And so it's pretty neat that we're in this phase shift.

(02:36:56): A hundred percent.

(02:36:56): So now we're all freaking psyched to see your setup and what you're doing and what

(02:37:01): you're building.

(02:37:01): Take it away.

(02:37:02): Yeah, cool.

(02:37:03): So,

(02:37:04): um,

(02:37:05): I,

(02:37:05): I'm hesitant to actually like show this because it's definitely not ready for prime

(02:37:08): time,

(02:37:09): but you know,

(02:37:09): the,

(02:37:10): the app that I've been building the last couple of months.

(02:37:12): Oh, wait, I have to show my screen.

(02:37:14): Mm-hmm.

(02:37:18): Okay, and we'll just do the entire screen and hope I don't accidentally leak anything.

(02:37:22): Whoa, we got the, okay.

(02:37:24): So a lot of,

(02:37:26): some of you may know,

(02:37:27): I used to run a content marketing agency called Growth Machine,

(02:37:30): and we were working with a lot of like,

(02:37:32): like Combinator startups and e-commerce businesses to like do their content

(02:37:36): marketing and SEO for them.

(02:37:37): And that was a great business.

(02:37:38): I ran that for four years.

(02:37:40): I hired my COO into the CEO role and stepped out in 2020.

(02:37:45): And I kind of had this thought

(02:37:47): six to eight months ago that like,

(02:37:48): maybe I could rebuild that whole content agency as an AI first solution.

(02:37:54): And so I started hacking on this two and a half months ago and have been building

(02:37:58): it out quietly and not really talking about it publicly.

(02:38:01): Um, and kind of just like bringing people in one by one.

(02:38:04): And it's the world premiere net, like the international premiere.

(02:38:08): It's kind of like the secret premiere.

(02:38:10): I'm going to put a big asterisk on this.

(02:38:13): If anybody wants to go try it out, go for it.

(02:38:16): And I did make a coupon code that's just every that gets you a month for free if

(02:38:21): you want to try it out.

(02:38:22): Because it's not completely ready.

(02:38:24): There's still a lot of stuff that I'm fixing.

(02:38:26): And you're going to hit bugs.

(02:38:27): You're going to break things.

(02:38:29): But it's getting pretty good.

(02:38:31): And it's been really cool to work on it and to start to get more people in it.

(02:38:37): They're like, there was nothing else that I wanted to demo.

(02:38:39): So we're going to talk about Tegan.

(02:38:40): So Tegan is an agent specifically focused on content marketing and content creation.

(02:38:46): And so this is how I work on it.

(02:38:48): Did you design this?

(02:38:50): Did Claude design this?

(02:38:51): It looks good.

(02:38:52): Claude designed this.

(02:38:53): It looks good, right?

(02:38:54): It looks good, yeah.

(02:38:55): Yeah, so there's a Claude skill called Frontend Design.

(02:39:00): It's one of the most popular Claude skills.

(02:39:02): I highly recommend installing it if you haven't.

(02:39:05): And then I picked a primary color, which is the fallow green.

(02:39:11): If you look up Phthalo Green, you'll find it.

(02:39:12): It's the color that's often the perfect green.

(02:39:15): It's just beautiful.

(02:39:17): I literally just said,

(02:39:18): hey,

(02:39:18): Claude,

(02:39:18): use your front end UI skill and create a landing page and use this green as the

(02:39:23): primary color and go.

(02:39:24): It came up with this.

(02:39:25): I rather like this.

(02:39:29): Anyway, the way we work together is I've got basically two things going.

(02:39:34): I have my computer where I'm using conductor.

(02:39:38): This is conductor build, which is basically a wrapper around Claude code and codex.

(02:39:46): And what's really neat about Conductor is that,

(02:39:49): one,

(02:39:49): this is just a way better UI than having to use cloud code in the terminal.

(02:39:53): It just looks nicer.

(02:39:55): It's more pleasant.

(02:39:56): It's formatted better.

(02:39:57): You can see all of the changed files over here.

(02:40:00): You can one-click open PRs.

(02:40:02): You can one-click do PR reviews.

(02:40:04): And if you want to do multiple things on a repository in parallel,

(02:40:09): Every time you open a new workspace,

(02:40:12): and Content Buddy was the working title for Tegan,

(02:40:15): every time you open a new workspace,

(02:40:17): it creates a Git work tree to do all the work for you automatically so that your

(02:40:22): work isn't conflicting until it gets to GitHub and you start doing PRs,

(02:40:26): and then you might have merge conflicts.

(02:40:28): But this way you can start working on six things at once,

(02:40:32): and they're not going to be on top of each other messing each other's work up,

(02:40:35): which is really cool.

(02:40:37): And when they,

(02:40:38): when they finish something,

(02:40:40): you can,

(02:40:40): uh,

(02:40:41): run,

(02:40:41): you can,

(02:40:42): you can have it open a PR and then that will,

(02:40:44): you know,

(02:40:45): go and open a PR on GitHub.

(02:40:47): So if I go to pause you there,

(02:40:49): Nat,

(02:40:49): because I,

(02:40:50): some people are,

(02:40:51): are in the chat,

(02:40:52): uh,

(02:40:52): talking about this and I actually have this feeling too,

(02:40:54): like it's fighting words to be like conductor and a GUI versus a,

(02:40:58): versus a terminal.

(02:41:00): Tell us, tell us more about that.

(02:41:01): Cause like,

(02:41:02): I feel like a lot of us right now are very terminal pillar and like,

(02:41:04): I would never go back to a,

(02:41:07): regular user interface.

(02:41:08): What makes you like this?

(02:41:10): I mean, I just... I have a hard time understanding why you would prefer like...

(02:41:18): know this and then it's like okay if i want to do like another instance i like have

(02:41:24): to open another one and do this and then i have to like run the git work tree

(02:41:30): command i mean it's like this this feels like or in the stone age this is like two

(02:41:34): months ago like why would you want right like i have to i have to type like open a

(02:41:40): pr right or i have to like use the i have to like

(02:41:43): do the slash commands for skills.

(02:41:45): It's just like, I don't know.

(02:41:46): I mean, it's fine.

(02:41:47): It works, but like Conductor is so nice for managing all of this.

(02:41:52): And especially if you're doing multiple projects, right?

(02:41:54): It's like, here's the Tegan website.

(02:41:56): So if I want to like update something on the Tegan website,

(02:41:59): I can just pop open work tree here and,

(02:42:01): you know,

(02:42:02): type in my commands and whatever,

(02:42:03): and it does it and I can open a PR,

(02:42:05): right?

(02:42:05): Like you can manage all of your projects in one interface and bop around between them.

(02:42:10): really easily i do think like it's so funny that we're almost like speed running

(02:42:14): the dos era and instead of it yeah yeah yeah 15 years to get to guis like it's just

(02:42:19): we we had dos for two months and now we're into guis you know

(02:42:24): I will say to another really big thing with Conductor that I do a lot that is again

(02:42:29): very annoying to do in the terminal is I actually have Codex 5.2 do a lot of my

(02:42:36): initial implementations now.

(02:42:39): I'll give the initial prompt to Codex and it'll code up what I need it to do.

(02:42:43): Then I can just open a new tab.

(02:42:46): that's using Opus 4.5 and either have that one write up the PR and open it or I'll

(02:42:52): have it review Codex's work.

(02:42:55): This is within the same Git work tree.

(02:42:57): It's just looking at what happened in this isolated work tree.

(02:43:00): It's just looking at Codex's work and it's getting a second eye on it before I send

(02:43:04): it up to GitHub and do my GitHub review process.

(02:43:07): Again, this is just like a pain in the ass to do in the terminal.

(02:43:10): Like it's so much simpler to do here.

(02:43:13): And you're still using your Claude Pro Max, your Codex Max subscriptions.

(02:43:17): You know, it's got that full command line access.

(02:43:20): You're just getting it in a much nicer interface.

(02:43:23): Also, if you're doing something, like if I'm working on the Tegan site,

(02:43:27): and I want to run it and see how these changes work,

(02:43:30): it's just one click for it to spin up the dev server,

(02:43:33): and then I can pop it open and see how its changes look on localhost,

(02:43:36): right?

(02:43:37): And it does all the NPM install.

(02:43:39): It does everything that I need it to do.

(02:43:41): Again, I'm sorry, but this is just better than working in terminal.

(02:43:45): I can't think of any reason we'd want to be in terminal besides to look cool in a coffee shop.

(02:43:49): All right, you heard it here first, folks.

(02:43:53): The CLI is the Stone Age from two months ago, and GUIs are back.

(02:43:58): The rest of the world will probably figure that out in a month or two,

(02:44:00): but you heard it here first.

(02:44:01): Nat, continue with your workflow.

(02:44:04): Okay,

(02:44:04): so once I have some work that's good here,

(02:44:09): I'll push it up to GitHub,

(02:44:10): open a PR,

(02:44:11): and then you can see I've got these three PRs open for Tegan right now.

(02:44:16): one thing that i've built and this is kind of like uh it's like a lighter version

(02:44:22): of the ralph loop or a lighter version of what you guys have with compound

(02:44:26): engineering where basically i have git workflows that i've configured that kind of

(02:44:32): expand how clauds uh built in github review works because the problem that i saw i

(02:44:39): was running into

(02:44:40): is that if you open Claude code and you type like slash GitHub install or

(02:44:46): something,

(02:44:46): it'll install the GitHub workflows for Claude.

(02:44:50): And it adds this PR review where it uses Claude to review the PR and make

(02:44:55): suggestions,

(02:44:57): improvements,

(02:44:58): review it,

(02:44:58): and things like that.

(02:45:00): And then I found myself like,

(02:45:01): OK,

(02:45:01): well,

(02:45:01): I'm just going to copy and pasting this and going back to Claude code and telling

(02:45:04): it to implement it.

(02:45:05): And that's annoying.

(02:45:06): So I expanded it to add another step where Claude reviews its own review.

(02:45:11): And it says, we should address this stuff in this PR.

(02:45:15): And these things we should create as separate issues.

(02:45:18): It then creates those follow-up issues on GitHub for the things that could be

(02:45:22): addressed in a future PR,

(02:45:24): and then it implements the stuff that needs to be implemented.

(02:45:27): All of this just runs automatically as soon as I open a PR.

(02:45:31): Did you set that up so that it runs automatically?

(02:45:34): I just asked Claude, honestly.

(02:45:39): I said, I opened the repo where I already had the Claude GitHub basic app installed.

(02:45:46): And I said, hey, I want to expand this so that it does these things.

(02:45:49): I want it to review its review.

(02:45:51): I want it to create follow-up issues.

(02:45:52): I want it to implement the stuff that needs to be implemented.

(02:45:55): And then when I close a PR, I want it to automatically go to those issues

(02:46:01): evaluate if they actually need to be implemented.

(02:46:04): And then if they do need to be implemented, implement them and open a new PR.

(02:46:08): So it's got this recursive system going where literally all I'm doing now is

(02:46:14): opening PRs,

(02:46:16): going and making sure the like,

(02:46:18): stuff ran properly.

(02:46:19): And then after 10,

(02:46:21): 15 minutes,

(02:46:21): all of the follow-up work has been done,

(02:46:23): and I can just merge it and move on to the next thing.

(02:46:25): It cuts down the back and forth significantly and reduces how often I have to go

(02:46:30): back to Claude and say,

(02:46:31): hey,

(02:46:31): fix this.

(02:46:31): Hey, adjust this.

(02:46:32): Hey, update this.

(02:46:34): It makes it a much more autonomous system without running the risk of the scope expanding.

(02:46:40): Because what this is doing is it's scope narrowing.

(02:46:43): It's saying, just fix the things that have to be fixed and then create these follow-up issues.

(02:46:47): Then it can look at those follow-up issues and it can say, we don't actually need to do this.

(02:46:51): This is scope creep.

(02:46:52): We're not going to address that.

(02:46:53): But this was actually important.

(02:46:55): We should do that.

(02:46:56): Then it runs that same process on the new PRs that it opens.

(02:46:59): Got it.

(02:47:00): Cool.

(02:47:02): The other things that I have are Cloud Code workflows that automatically fix lint

(02:47:08): and test failures,

(02:47:09): but you can see something went wrong here.

(02:47:12): It might be that there's merge conflicts that need to be resolved.

(02:47:16): I've got these three PRs open,

(02:47:18): and I don't really feel like going back to Cloud Code and resolving these lingering

(02:47:25): issues.

(02:47:26): Instead, I'm going to go to Cloud Bot.

(02:47:28): Now, ClawdBot, for those of you who don't know, is the hot new thing.

(02:47:33): This is basically a wrapper around Claude or really any agentic coding framework.

(02:47:40): It can use Codex,

(02:47:41): it can use Minimax,

(02:47:42): it can use whatever you prefer,

(02:47:43): but most of us are using it with Claude code.

(02:47:46): And this is what's living on my Mac Mini.

(02:47:47): And I primarily talk to it through Telegram.

(02:47:51): So before we started here,

(02:47:53): I asked it like,

(02:47:54): hey,

(02:47:55): we're going to do a live demo of how we work together.

(02:47:57): What should we show off to get people excited?

(02:48:00): And it came up with some ideas.

(02:48:02): And I said, oh, well, they're excited about building apps with you.

(02:48:05): And so it came up with some of these ideas.

(02:48:07): So what I'm going to tell it is I'm just going to say,

(02:48:11): We have three PRs open on Tegan right now,

(02:48:15): and there's some testing and linting issues and some merge conflicts.

(02:48:20): Could you pull all of them down and then resolve all of those issues,

(02:48:24): and once they're resolved,

(02:48:25): get them merged into staging?

(02:48:28): Obviously, I'm using monologue for my voice to text because it's a wonderful tool.

(02:48:33): Typing is also not something we're bringing into 2026.

(02:48:36): Nobody should be typing anymore.

(02:48:37): You should just be using voice to text because it's way better.

(02:48:40): Typing is over.

(02:48:41): Typing is over.

(02:48:42): Typing is done.

(02:48:44): The cool thing with ClawdBot too with this Telegram setup is you can also send it voice notes.

(02:48:50): I'll walk to the coffee shop and I'll just send voice notes to ClawdBot for things

(02:48:53): that I wanted to do or wanted to work on,

(02:48:55): and it'll start responding to me.

(02:48:57): You can see it already responded.

(02:48:59): It's like, all right, I'm going to see what we're working with.

(02:49:01): Here's the situation.

(02:49:04): It's got those two PRs are good,

(02:49:07): but this last one's failing,

(02:49:09): so it's going to pull them down and figure out what's going on and how to merge it

(02:49:13): all together and fix it.

(02:49:16): This is just really nice because a couple of things are happening in the background here.

(02:49:21): One,

(02:49:22): there's this direct conversation that I'm having with Claude,

(02:49:26): but I've also designed it so that it has access to spin-off codecs sub-agents.

(02:49:32): When it needs to do a little bit more of a coding lift,

(02:49:36): it actually creates sub-agents using Codex instead of just quad normal sub-agents

(02:49:41): for a couple of reasons.

(02:49:42): One, I do find that Codex does a little bit better with coding than Opus 4.5.

(02:49:47): It goes a little bit slower, digs a little deeper, and builds out slightly better solutions.

(02:49:52): But Opus is a much better general intelligence model.

(02:49:56): I trust Opus to be my engineering manager,

(02:49:59): and to control all of my codex instances to write the code.

(02:50:03): The other benefit of this is I'm not using up all my Claude credits on everything.

(02:50:09): Claude is the orchestrator, and then it's using codex credits for a lot of the programming.

(02:50:16): Once those Codex agents figure out these issues and implement them and fix them,

(02:50:20): it'll push them back up to this flawed instance,

(02:50:22): which will do a final review,

(02:50:24): and then it will go merge all the PRs into staging.

(02:50:27): And how are you using,

(02:50:29): because I find the same thing,

(02:50:30): like I want Opus to talk to Codex,

(02:50:33): but I often want Codex to do a lot of the more in-depth coding.

(02:50:36): How are you doing that orchestration?

(02:50:38): What tools are you using?

(02:50:40): It's just through ClawdBot.

(02:50:41): So when you set up ClawdBot, you can add other AI models for it to have access to.

(02:50:48): And it's just, it's a pretty simple, like, I mean, it's, okay, it's not simple.

(02:50:52): It's terminal based and there's like some finicky stuff with AI or with API keys.

(02:50:57): And so you've got to like wrestle with it a bit.

(02:50:59): ClawdBot is like frontier software.

(02:51:00): You've got to be comfortable.

(02:51:02): Can you just explain like ClawdBot just takes your own cloud code and runs it

(02:51:06): in a loop on a server and connects it to all the other services or like what is ClawdBot?

(02:51:12): Yeah, that's a good way to think about it.

(02:51:14): Think about ClawdBot as like an always running Claude Code instance on your computer

(02:51:20): that can run any other service on your computer that you could run through the

(02:51:25): command line.

(02:51:27): So you're giving it full computer control.

(02:51:32): And so it can do everything from kick off these codex agents and manage GitHub

(02:51:37): workflows to doing stuff in your email.

(02:51:41): So I gave it access to my email.

(02:51:42): And so if I said, hey, I need you to go

(02:51:46): find this email and draft a reply or, hey, I have this subscription that I need to cancel.

(02:51:50): Can you just go find that and reply to them telling them that I want to cancel the subscription?

(02:51:54): It can go do that too.

(02:51:56): And you can also build custom skills for it.

(02:51:58): Basically anything you could do from the command line or actually from the browser

(02:52:02): because it has browser automation.

(02:52:04): So it can open a Chrome browser and actually do things on the browser mostly in the

(02:52:10): same way that you would.

(02:52:15): One of the skills that I built for it is actually a full Tegan testing suite.

(02:52:20): And so I have it run this overnight, so it doesn't use up my daytime credits.

(02:52:25): But what it actually does is it opens the staging deployment, and it creates an account.

(02:52:31): So it'll do all of this stuff.

(02:52:38): And it will go through the onboarding flow.

(02:52:43): And it knows to look for any issues

(02:52:48): that need to be fixed or any weird behavior.

(02:52:51): And it'll write up a report on what it did going through the test deployment.

(02:52:58): And if it finds any bugs, it will just automatically start addressing those bugs.

(02:53:04): And actually,

(02:53:04): it doesn't even have to necessarily do that because I also have ClawdBot looped

(02:53:08): into Slack

(02:53:09): where anytime there's a bug on Tegan,

(02:53:12): Sentry sends that bug to this channel in Slack and it tags ClawdBot.

(02:53:18): And then you can see ClawdBot receives the error

(02:53:22): it looks into it and then it automatically spins up a codec sub-agent to code up a

(02:53:28): fix and then open a PR onto staging.

(02:53:32): Then it'll ping me on Telegram or just I can reply here.

(02:53:36): Sometimes if the bug is weird,

(02:53:38): it'll just ask me a question here in this channel and I can reply here and tell it

(02:53:42): like,

(02:53:42): oh,

(02:53:42): let's solve it this way before it starts coding up the solution.

(02:53:47): Very cool.

(02:53:48): I love this.

(02:53:49): Okay.

(02:53:51): so it says both pr is fixed and pushed let me check the ci status and off the bat i

(02:53:56): can see we have something weird going on here in the staging deployment right so

(02:54:00): i'm going to have to look into this probably something that we just added in these

(02:54:04): last tests but if it doesn't throw an error then i can just go into the logs find

(02:54:09): it and send it to clot afterwards but if we go look at the pull requests and

(02:54:15): refresh this okay we can see one of them's been closed

(02:54:18): This one's rerunning.

(02:54:19): This one's still failing,

(02:54:21): but it should be able to find those issues and start resolving them too.

(02:54:28): So I'm trying to make a list as you're doing this of what's over and what's next.

(02:54:31): So far I have over is CLIs, typing, and code diffs.

(02:54:36): And what's next is GUI's voice and vibing.

(02:54:39): Anything we're missing here?

(02:54:41): I think to me, the like what's next is moving from

(02:54:49): having a really competent engineer at your disposal to having a really competent

(02:54:53): engineering manager at your disposal,

(02:54:56): where the other skill that,

(02:54:59): you know,

(02:54:59): we obviously don't have time to demo right now,

(02:55:01): but that I've built for ClawdBot is whenever somebody,

(02:55:05): so like,

(02:55:06): you know,

(02:55:06): somebody signs up to try out Tegan and they go through the onboarding flow and

(02:55:10): they're like making articles and stuff.

(02:55:12): Obviously I have a log of those conversations on the servers.

(02:55:16): And so I built a custom API for ClawdBot to pull in the conversation history for the day.

(02:55:21): And so every night at 2 AM,

(02:55:23): it pulls in all of that day's conversations,

(02:55:26): and it just reads through them.

(02:55:28): And it looks for user frustrations,

(02:55:32): weird behavior,

(02:55:33): things that failed,

(02:55:35): and it writes up a report on everything that happened that day and ways we could

(02:55:39): improve the app to avoid that in the future.

(02:55:41): And then if it finds a bug, it automatically fixes it.

(02:55:44): But if it's an improvement, it just sends me the report in the morning.

(02:55:47): And it sends it to me in a Google Doc because I have it connected to my Google Docs.

(02:55:50): I just would rather read it in a Google Doc.

(02:55:52): So I'll go through the report and then I'll send it back a voice note and say,

(02:55:55): OK,

(02:55:56): these are all good findings.

(02:55:58): How do you think we should address A, B, and C?

(02:56:00): And then it'll come up with a plan.

(02:56:02): for how to improve the app to avoid that behavior in the future.

(02:56:05): And then once we're agreed on the plan,

(02:56:07): it'll kick off the codex agents to actually do the coding and implement it.

(02:56:13): And that's like a whole new level from me having to watch every single conversation

(02:56:18): and me having to figure out

(02:56:20): what should I tell Cloud Code to do?

(02:56:22): That's abstracted now because Opus is a better engineering manager than I am.

(02:56:27): I trust its judgment on a lot of these things, especially as we talk about it.

(02:56:32): Then it can kick off all of the Cloud Code and Codex instances to actually do the

(02:56:35): coding,

(02:56:36): report back to it.

(02:56:37): It can check its code against the plan that we created,

(02:56:41): and then it can start opening the PRs from there.

(02:56:43): That's this new layer that we now have access to that's pretty magical.

(02:56:47): I love that.

(02:56:48): So we only have a couple of minutes left.

(02:56:49): So I have two final questions.

(02:56:52): One,

(02:56:52): which I think is highly relevant to everyone here is,

(02:56:54): for everyone who's vibe coding,

(02:56:57): for Teagan,

(02:56:57): how did you get a celebrity sponsorship from Chrissy Teagan for your app?

(02:57:04): That's a good question.

(02:57:05): No, I just had fun with one of the AI photo generators.

(02:57:10): And I was like, oh, let's just make somebody who looks like a friendly content marketer.

(02:57:15): And I think there's going to be cool stuff you can do with that,

(02:57:18): too,

(02:57:18): where I think that the direction we're moving in with these AI agents is not a

(02:57:21): faceless one like Devin.

(02:57:23): but a faced one like tegan where it just looks like a person in your slack and you

(02:57:28): talk to it like a person and it just feels like another employee and that's going

(02:57:33): to be like an interesting new world i love it okay last thing is one of the things

(02:57:38): i think you're you're very you're a special person in a lot of ways but one of your

(02:57:41): most special characteristics

(02:57:43): is I feel like you're someone who has reinvented yourself a lot.

(02:57:47): You've lived many lives.

(02:57:48): You've run a marketing agency.

(02:57:49): You are a crypto author.

(02:57:51): You're deep in the crypto scene.

(02:57:53): You're vibe coding.

(02:57:54): You're doing all these different things and wearing all these different hats.

(02:57:58): And especially right now, there's this huge new wave happening.

(02:58:02): And I feel like a lot of people are maybe torn between

(02:58:05): their identities, like whether at that work or at home and sort of this whole new world.

(02:58:10): How do you navigate that?

(02:58:11): How have you managed to be so good at reinventing yourself,

(02:58:15): depending on like whatever is whatever you're interested in at any given time?

(02:58:18): It's a good question, and I don't have a good answer for you.

(02:58:22): And I will say,

(02:58:23): too,

(02:58:23): that I wouldn't actually,

(02:58:24): like,

(02:58:24): recommend it as a career path because,

(02:58:29): like,

(02:58:30): yes,

(02:58:30): it has worked out well for me,

(02:58:32): but most of the time,

(02:58:34): if you're this,

(02:58:35): like,

(02:58:36): non-committal to stuff that doesn't go well for you.

(02:58:39): And I'm actually trying to move away from that.

(02:58:41): I do switch too often.

(02:58:43): I do jump to new things too often.

(02:58:45): I don't stick with stuff for it to really hit the compounding gains that it could have.

(02:58:50): And with

(02:58:52): With this,

(02:58:53): with what I'm doing with Tegan,

(02:58:54): I basically just put everything else on hold for the most part.

(02:58:59): I'm just all in on mastering this new way of building and applying it to building

(02:59:04): something where I have a very strong competitive knowledge advantage,

(02:59:07): where it's a type of product that I care about and where I feel like I can build

(02:59:10): something really great.

(02:59:11): It's just been very fun to fully commit to this and go all in on it and see how far

(02:59:18): I can take it.

(02:59:20): Amazing.

(02:59:20): I love it.

(02:59:20): Thank you so much for the honest answer.

(02:59:23): Thank you, as always, for coming on and helping out and sharing everything that you know.

(02:59:30): If people want to find you on X or elsewhere, where should they find you?

(02:59:35): Yeah, just Nataliason on X. You can find me there.

(02:59:39): And then my occasionally updated at this point newsletter is blog.nataliason.com.

(02:59:45): And then Tegan, where's Tegan?

(02:59:49): Don't want to say, okay.

(02:59:50): It's hiretegan.com.

(02:59:52): Hiretegan.com.

(02:59:53): And then what's your, what's your latest book and where can, where can they get it?

(02:59:56): Oh, yeah.

(02:59:57): So this came out on Tuesday, and that is The Birth of Paradise.

(03:00:01): And this is a sci-fi novella.

(03:00:03): So if you go to Amazon and search my name, you'll find it or it's in my Twitter bio as well.

(03:00:10): See, this man is, you know, vibe coding, publishing sci-fi novellas.

(03:00:14): He's got three kids.

(03:00:15): He's doing it all.

(03:00:17): So thank you.

(03:00:18): Thanks for joining that.

(03:00:19): Thanks, Dan.

(03:00:20): See you soon.

(03:00:22): all right uh hello everybody if you are here we are four hours into the stream at

(03:00:29): this point or almost four hours or three hours and 15 minutes into the stream we

(03:00:32): are doing this all day this is vibe code camp uh from every if you want to know the

(03:00:37): agenda we have a stacked lineup we've had so many amazing people already we have a

(03:00:41): stack lineup coming up um check out uh the agenda at every.to slash agenda we're

(03:00:46): gonna be we're gonna be going for another probably five or six hours until like six

(03:00:50): or 6 30 eastern

(03:00:52): We're super excited.

(03:00:52): If you don't know what Every is,

(03:00:56): Every is the only subscription you need to stay at the edge of AI.

(03:01:01): We have three big components of the subscription we offer, ideas, apps, and training.

(03:01:05): On the ideas side, we keep you up to date with everything that's going on in AI.

(03:01:08): When new models come out,

(03:01:09): we do immediate day of vibe checks with hands-on testing from the builders inside

(03:01:13): of Every.

(03:01:14): And just keep you really up to date on what you need to pay attention to and what

(03:01:17): you don't need to pay attention to.

(03:01:18): You should go to every.to.subscribe to sign up for free.

(03:01:21): That's the idea side.

(03:01:22): On the app side,

(03:01:23): we have four apps that we build and build for ourselves internally that we make

(03:01:27): available to you to keep you working at the edge.

(03:01:29): We have Cora, which is an email assistant.

(03:01:32): It lets you do your email with AI, Cora.computer.com.

(03:01:35): We have Monologue.

(03:01:36): It's a speech-to-text app.

(03:01:39): It's a smart dictation app, sort of like Whisperflow or Super Whisper.

(03:01:42): That's at monologue.to.

(03:01:44): We've got Sparkle.

(03:01:45): It's an AI file system, a file organizing app.

(03:01:49): And we also have Spiral, which is a Gentic ghostwriter with taste.

(03:01:53): You get all of these.

(03:01:54): These are all tools that we built for ourselves that we use every day.

(03:01:56): You get all these as part of your subscription app.

(03:01:59): They're all run by one engineer.

(03:02:01): Each of them is run by one person.

(03:02:03): Really, really cool.

(03:02:04): Last part of the subscription is training.

(03:02:06): And as part of training, we do live streams like this one all the time.

(03:02:10): We have camps.

(03:02:12): twice a month where you get to sit down with every team and talk about our

(03:02:16): workflows,

(03:02:16): talk about how we're using Codex or Cloud Code or any of the new products that come

(03:02:20): out and ask direct questions.

(03:02:21): We also teach you our engineering philosophy from Kieran Klassen,

(03:02:24): compound engineering,

(03:02:25): and our philosophies on building in AI.

(03:02:29): So our agent native architecture philosophies and all that kind of stuff.

(03:02:32): If any of this sounds good, we have a 30-day totally free trial.

(03:02:36): Check it out at every.to slash subscribe.

(03:02:39): It's 30 bucks a month,

(03:02:39): but you get all of that for 30 bucks a month,

(03:02:41): and we're constantly adding new things.

(03:02:42): We have a bunch of actual new things coming.

(03:02:45): Welcome to the stream, Tina He.