123: Isolate the Claude  🔗 **Source:** [123: Isolate the Claude](https://share.snipd.com/episode/d8221ddf-b143-4262-b146-cb130946be20) ## Highlights - **Philly Poker Trip Won Then Lost Wrapped Cash** - Ian won roughly $8,000 playing private and casino standup poker during a Philly trip, getting paid in wrapped cash which felt more “real.” - He then lost about $6,000 next night in a tougher 10-20 game, leaving him net up but reminded him about game selection. Transcript: Ian Landsman Love, there's almost nothing I love better outside of, like, family. Whatever, whatever, whatever. Then wrapped money, like, from a bank, it was wrapped the little paper wrapping that's just it's so gangster i love it so much it's so ties in with all the things i love and movies and like Gangster movies also just like piles of wrapped money so yes i sent you a picture of wrapped money um it was great. We have a lot of chaos. I'll just give you the very thin version because it's tied to the poker, which is oldest kid at college coming home Friday, last Friday. This was last week. Middle kid, Spanish trip to Costa Rica. He's in Costa Rica. Wife and daughter are like, we're getting the heck out of here. We're going to Florida to catch some sun. Ian, home alone. I grew up in a big family, a lot of people. I don't want to be home alone. I don't like being home alone. Being home alone is sad to me. Aaron Francis You called up the other D-Gen. Ian Landsman I called up Daniel Colborne, my other Laravel poker D-Gen buddy, and also Chris Morell lives near him. I went to Philly. Bossman Chris. I didn't have him tagged as a D-Gen. Aaron Francis Is he a D-Gen? Ian Landsman No, he is not a D-Gen, although he is learning poker. All right. So he's actually playing poker tonight, I think, with Daniel at a home game. He's been getting into learning. Okay, so full D-Gen arc. This morning he messaged me that from bed last night, he built his own Tinder like poker trainer. So a pre flop poker trainer. So we're getting him. We're getting him. We're hooking him in. Aaron Francis RIP boss man, Chris. Ian Landsman So anyway, I was like, I'll go to Philly, hang out, I'll have some dinners. I'll play some poker. It's going to be great. So I went to Philly. I had some dinners. I played some poker. It was great. Mission accomplished. Mission accomplished. Cool. A couple of cool. So when I travel for poker, I like to play bigger than I can play in my normal setting. Daniel happened to know some people who know some people kind of situation, got me in a big game, like a private-ish type game at the casino, which is a thing that happens now, but they Don't want a bunch of pros in there. I actually have an interesting corollary to the story. So anyway, I got in this big game. Me and Daniel played in it. I won like eight grand in the big game. They paid me in wrap money because I had my buy-in, so it's more than $10,000. So I got $10,000 wrapped, which is glorious. Aaron Francis You got a stack of $100 bills with a piece of paper around it that said $10,000. Yeah, it's glorious. Ian Landsman Oh, man. It's so much realer than like getting a transfer for $10,000. When someone just hands you wrapped cash, you're like, this is real. This is physical. This is fresh. It's crisp. It smells good. It is good. ([Time 0:00:56](https://share.snipd.com/snip/ea693624-bc75-4716-8f10-d600950f4a77)) - **Transcode Videos With Modal Sandboxes** - Use Modal sandboxes to run heavy batch work like FFmpeg transcoding instead of keeping long-lived VPS instances. - Aaron runs separate Modal apps per bitrate, deploys a tiny Python FFmpeg wrapper from GitHub Actions, and scales sandboxes to e.g., 16–32 vCPUs on demand. Transcript: Aaron Francis So I have a lot of videos that I need to get onto websites, right? Okay. Yep. Whether that's database school or faster.dev, whatever. And I need to transcode them into multiple different, I think bit rates is the technically correct term. So I transcoded into 2K, 1080, 720, and maybe 480. So we can do adaptive streaming, yada, yada, yada. That obviously is a very resource intensive process, taking a big video and turning it into two big videos and two small videos and chunked up with HLS segments so that the streaming Can work nice and smooth. I've been around and around on this. Okay. So I've done the Mac Studio, which rips through it super fast. So I've got the Mac Studio back in the closet that handles a bunch of stuff. The problem with that is obviously it's back in the closet. And so the web server, you know, receives a video and then has to phone home to the Mac Studio and like it works, but I don't feel confident. It's a little janky. It's a little janky. It's a little duct tape and bubble gum. Right. I have tried Laravel VPS. So Forge has their first party. It's like a DigitalOcean wrapped thing where they keep a warm pool of servers. Laravel VPS. We've gone through this whole, we've probably talked about this, where I spin up Laravel VPS boxes that are huge to rip through a Laravel queue. That works okay. I have recently had some issues with Laravel VPS where I'll spin up this big expensive box and then Forge says, hey, we can't connect to it. And I'm like, well, you spun it up. Why can't I connect to it? Which leaves it in this liminal state where I'm getting charged for it, but I can't use it. And I got to go in and kill it. And my queues kind of slow down because it's like we spun up the box, but we couldn't deploy the site. And now you don't have your queue worker. Right. Ian Landsman Nothing I hate worse than a weird state. I don't like a weird state. I hate a liminal state. I hate that. A weird state is just nothing worse. It's hours. State is everything. You're trying to figure out where things are at. It's brutal. It's brutal. Aaron Francis Also, in Laravel, when you have queues that are supposed to take like 90 minutes to complete per job, stuff gets just the whole... This has been my singular gripe with Laravel since like 5.4, whenever I started using it. Right my singular gripe is the cue retry timeout and retry after okay trifecta it's like you got to set the job to have a timeout and the cue to have uh like the cue worker to have a timeout Right there's several time out retryry after. And if you don't do it just right, you end up with overlapping jobs because the job's actually dead, but the Q didn't kill it. Or the job is still running, but the Q worker killed it. And so that's always just been confusing to me. So long journey through the darkness. I've now arrived at a false glimmer of hope. I think it's real, but I'm very, I'm emotionally prepared for it to be false. All right. Modal. Have you heard of Modal? I don't think so. Modal.com. Modal.com. Give me the H1. Go to Modal.com and give me the H1. Modal.com. Ian Landsman AI infrastructure that developers love. Okay, well, that's a very generic H1. Yeah, that means nothing. Yeah, what else do you see? Is there anything else on there? Okay, run inference training and batch processing with sub-second cold starts, instant auto-scaling, and developer experience that feels local. I still don't know what it does. Am I building my own model? I guess. Yeah, we don't know what it does. What does this have to do with what you're doing? Aaron Francis Yeah, that's a great question. All right. So the subhead is run inference training and batch processing. No, that's nothing. We don't want any of that. Designed to help AI teams deploy faster. No, we don't want any of that. All right. So we're going to go to product and we're going to click on modal sandboxes. Secure sandboxes. We're deep though. Yeah, we're deep. Secure sandboxes at massive scale. I think that's the one I'm using is sandboxes. Okay, so you create sandboxes. We handle resource orchestration. All right. That's what I want. Strip out any AI marketing nonsense. I don't want any of that. All I use, so you can scroll down, modal sandbox pricing, compute costs, point zero zero zero zero three nine four two dollars per core per second right so you have like minute uh high Resolution pricing which is awesome. You can spin up multiple sandboxes like via HTTP request, right? Okay. So now, and they're very focused on Python because it is like AI. I don't know if it was always AI or they're like pivoted to AI, but it's like you can run all this Python stuff to do whatever, flippy flab, I don't care. But what you can also do is you can run ffmpeg on these things right and so i had the clanker write me up basically a very thin like python uh python wrapper that all it does is it receives A little bit of input and then it basically calls FFmpeg. And so now what the flow is, is I have these, they call them apps. I have these apps on modal that scale independently. So I can run the 2K, the 1080p, the 720p and the 480p. I can run all of those at the same time. So I'm already getting better throughput because it used to be one massive FFmpeg script. Now it's four. And I can spin up as many of these things as I want and not have to like spin up new Forge VPS servers. Right. And FFmpeg is really, really, really good at utilizing the number of uh cores that it finds itself on per machine okay and so for this i can say like hey give me give ([Time 0:10:36](https://share.snipd.com/snip/2c6098c5-3559-44d3-8dbc-9bfed7460089)) - **Self-Host Course Video HLS To Avoid Vimeo Overages** - Avoid expensive hosted video platforms for high-volume course libraries; self-host HLS on R2 to cut costs. - Aaron moved encodings to HLS chunks on R2, using VideoJS for playback and paying only storage/low egress costs. Transcript: Aaron Francis Yes i can and i have used vimeo um so when we were doing it was one of the early ones maybe it was mastering Postgres or something um we were still on Vimeo and they came to us and they're like Hey your bandwidth is kind of high um we could do like a 16 or 20 thousand dollar a year enterprise uh plan for you. And we were like, what? We're just two guys with two courses and you want to charge us $20,000 a year? Right. I think Jeffrey, I don't know if Jeffrey's moved on to MUX or something else, but for a long time up until maybe including now, he's used Vimeo. And Vimeo does all of this for you. It does all the transcoding, all the streaming, everything. Right. The setup that I use is one that mostly Steve invented, which is we just use a simple script that does HLS encoding. And so it turns huge videos into tiny chunks. And then we throw those chunks on R2. And all egress out of R2 is totally free. And so instead of $20,000 a year, I incur this little bit of pain of like figuring out how to transcode these things. And then we serve videos for like four to five dollars a year. Because egress is free. So I'm just paying for like, you know, the cold storage or whatever. So yes, I could use video. ([Time 0:19:40](https://share.snipd.com/snip/3ca48dbc-1884-49aa-9e52-831ae9447dae)) - **Claude Agents Can Orchestrate Large Feature Builds** - Superpowers orchestrates Claude agents to brainstorm, write tests, and implement features in parallel, accelerating V1 work. - Ian used Claude-driven agents to generate a 6,000-line markdown spec and an implementing run that produced working code he then refined. Transcript: Aaron Francis Well, hang on. Let's talk about some super companies. Alright, what is superpowers? Tell me. Did you know about this before? No, this is the first you're hearing it. Ian Landsman No, not before you DM'd it. Okay, so it's got like 95,000 GitHub stars or something. It's a series of skills, and I'm not usually one for this type of like. I know that. The skills all connect together into a massive blob for the Gary Tan GStack operating system. Somebody just sent me this. I didn't even know about that one either until like this morning somebody sent me that. And I was looking at it. This is not working for me this is not for me um yeah and just like yeah it's just not the way i think about it um but the superpowers it's very interesting it's got like this brainstorming And it does like a really good job with the brainstorming with you um and then it's got like agent implementers and then those hands those build work trees again this like this is the part Where everybody's trying to build the work trees outside of the LLM. I'm like, the LLM just knows how to do it. Just let it do it when it wants to do it. And it's worked great. It's like the agents go off and they build the work tree. And it's all seamless. I don't even know why you'd need the harness to do any of it. So it seems like you don't. But anyway, it's all working good. I built the whole V1 of this side project just through the brainstorming session. There was like five hours of like did the brainstorm. Then like me and counselors did some heavy work on like the document that was created. Did all that, went back and forth there. Then like it ended up being like 6,000 lines. So we went through and made a simplified version that was like cleaned up but then that missed some stuff. The markdown file was 6,000 lines? Yeah yeah and you know I don't have simple ideas there and it's like it's always complicated so gosh we got that back down to like eight or nine hundred but then it had missed things because Part of why the sick it had 6,000 lines to begin with is because I had a lot of code samples but like didn't even write up the sort of purpose of a feature wasn't written up in prose. It was like, here is a code chunk. That is the purpose. So the translation to the streamlined version had missed some of those. So me and the counselors went back in and we made sure we extracted that stuff and got it all in there. So got it done. Then the superpowers have the ability to like implement it, which I've seen a lot of people doing more of like the gas townie type, like build a million issues implementation. This doesn't work that way. This like has one implementing thread and then it's spinning off sub agents and it's like building the whole thing at once. So it went for like four hours and, like, built the whole thing at once. And it built it, and it worked. I mean, it's definitely not perfect. With what model? This was Claude. Aaron Francis Okay, so Superpowers is orchestrating Claude to just keep going. Ian Landsman Yeah, it could do Coda. It could do all of them, but, yes, it is. I used Claude, and it was, yeah, so it's like the main Claude thread is going and it's like stuff's, you know, I've got six agents going and they're going and they return and it's like, It sends off more. ([Time 0:28:30](https://share.snipd.com/snip/074421a4-0ee1-408b-a1ef-f9c4135e4ccf)) - **Force High-Quality Models For Critical AI Work** - When using Superpowers, force the system to use the best model for critical work to avoid cheap-model degradation. - Ian explicitly configures Superpowers to run everything on Claude’s best model instead of cheaper defaults for reliability. Transcript: Ian Landsman So like you're really buying in fully. So like for this side project, it was fine because I'm like, whatever, I'm just want to try this thing. I got this idea. I'm just doing it from the start. We'll see if it works. If it doesn't work, who cares? Putting it in an existing project, you're kind of like buying into that doing that and it does it's not totally forcing on you have to say it's pretty interesting it seems to interject Itself correctly so far but it does it since it's setting itself up with the hooks i do think there's like you just have to be a little cautious that like if you were using it on something Aaron Francis Existing that you're kind of ready for it to interject itself more forcefully that's one of the things i don't get about like uh god forbid g stack or superpowers will say it's like i think It works great for going from zero to one but once you're in like if you're building laravel cloud right if you're jason bags who hasn't been mentioned on the show for at least one show Which is brutal so if you're jason if you're jason bags are you pulling in superpowers to work on cloud and let her rip for four hours or is this just like let's go from no no app or no feature To app or feature. Do you see what I mean? Ian Landsman Yeah, I do see what you mean. I actually think in, I don't know about the other ones, but in Superpower's case, I actually think it's, I was actually the one abusing it to go from zero to one. I actually think it's not really ideal for zero. It is, it's better than me manually doing it all. I think it actually worked nicely. It brainstormed, it came up with some good ideas. We worked through some stuff. So it was good for zero to one too, but I actually think that ended up being kind of a bigger chunk than it's meant to take on necessarily. I don't even know if that's true, but it was a big chunk. It was a lot of stuff. That's why we went, I had to like manually go through the markdown, I think like, you know, it was a bigger thing. So it was like zero to one and it was a bunch of stuff um i actually think it is more set up for the like let's brainstorm this one new feature okay it's going to brainstorm that feature with You okay it's going to figure out what all it needs to do it's going to send the agents out to do it to write the tests and build the things and it's going to be that one feature is done so i actually Think it is kind of more for that um day in day out and i think it would be good at that it's like it is a totally different way of working, right? Like it's not the like, I'm just looking around as me and kind of like rolling through stuff. But I feel like I'm ready to like leverage it up a little bit more and save some more time. But I also feel like whenever I do that with just a pure AI, sometimes things get a little off the rails and they don't, you know, the AI gets a little lost in there. And it's like, so then that makes me come down to smaller bite-sized chunks, which it does really well. But now I'm not actually saving as much time because I am like in the details on like lots of little chunks. Right. So I'm seeing if this strategy lets me give it bigger, chunkier tasks and have them come out at least mostly right and then, you know, just refining. So, so far so good. I put it into outro, but I haven't used it in outro yet. So that'll be interesting to see. Aaron Francis What are we building here? Ian Landsman Oh, we're not going to talk about that, Aaron. Aaron Francis No! You know we don't talk about that. Ian Landsman You know we don't talk about that. It's going to be something that works with outro. It's not, it has nothing to do like with what outro does, but it will be a company outro. It might be something I just use for me. It might not even make it. It will be another big product. So I don't know. We'll see. It will reveal itself when the time is right and the time is not right. In the fullness of time. In the fullness of time, yes. I have not had enough time to let it percolate. It doesn't even work work yet. I've got to get it work working. There's not a lot of time for that. Aaron Francis Well, that's the hardest part, if I'm being honest, is getting it to work work. Yeah, I can get it to work, but to get it to work work. Exactly. There's a lot of steps in this thing. That takes the same amount of time it took to get it to work. Exactly. Ian Landsman So there's a high probability I've been in this thing. Aaron Francis You mentioned outro. Yes. It is March 17th. Oh, God. Ian Landsman This poker trip really set me back. Aaron Francis Oh, it's always something. It's a gallbladder. It's a poker trip. It's always something. Ian Landsman I got excuses. Aaron Francis What is the status? Ian Landsman The status is it's coming along. It's going to be ready for you and me. End of the month, as I believe I've committed to. That's my goal. Aaron Francis End of the month. Or you have to get a Laravel new tattoo. Ian Landsman Oh, only if do it together. Aaron Francis No, there's nothing in it for me. I'll do it again. I'll do it for $750,000. Oh, that's right. So if it reaches the end of the month and you have $750,000, we can both get Laravel new tattoos. Ian Landsman Dude, I wish I was rich enough to have you do this. I just think that would be so much fun. I wish you were rich enough to have me do this. That would be great. Aaron Francis Take however much you wish and double it because that's how much I wish. All right, end of the month. Ian Landsman 75 wrapped stacks of $10,000. That would be fun. Show up with a duffel bag of that for you. Drop it off. Yeah, that'd be good. Aaron Francis I'm a simple man. All I want is $750,000 to get a small tattoo. Ian Landsman End of the month. End of the month. The sort of what we use Trello for now part is very, mostly there's like some AI generated bits of like functionality that I don't like how they work. It's a part I let it do go bigger on. So I want to like double back, clean some of that up. Aaron Francis Don't care. Let her rip. Ian Landsman You want? No, I don't like it. I don't like it. And so I'm going to clean that up. And then I would like to get the mailbag going too, but that's probably not going to happen before the end of the month, which is fine. That's not like a huge deal for us in our use. So we will get that in. Cut the scope, baby. The superpowers are going to help me because I'm going to give it some bigger, chunkier stuff. We're going to see how it does. So that'll be interesting to report back on that and see how it does with something that's a little, that I've been more intimately involved with and that is farther along and to see if You can inject this little framework situation in here and have it do a good job. So I'm kind of curious about that. One other thing I wanted to mention in that, if somebody does try it, this might just be me. But one of the things that Superpowers does is it says, hey, plan with the best model and use the cheaper model to implement. And I don't want any part of that. I don't want any cheaper models near anything. Aaron Francis I got stacks. I got wrapped stacks. If you want me to use Sonnet, get real. Ian Landsman No, I don't want Sonnet. Get the hell out of here with the Sonnet. So you can just tell when you kick it off, you can just say, hey, make sure you use the best model for everything. And then it does honor that. Because when agents start, at least in cloud code, it tells you what model they're on. It implemented it, and it did it right. So that was good. ([Time 0:32:13](https://share.snipd.com/snip/ad8845a7-ce64-4fc7-be68-7ede8bc74864)) - **Isolate Local Claude In A Separate Mac User** - Isolate risky or experimental LLM instances on a non-admin Mac user account and use Fast User Switching for convenience. - Ian runs Claude in a separate macOS standard user, uses touch‑ID finger assignments to instantly switch accounts while sandboxes keep running. Transcript: Ian Landsman Now I got to pull request things around and I hate all that. So I've landed on an alternate standard Mac user on my computer. I can go over to it. Can't sudo. It can't do anything. But on there runs the Claude in this sandbox environment that Apple has created for us, which we all forgot about user accounts. But your computer has user accounts. Aaron Francis Isn't it adorable that you can imagine sharing a computer with? Like the family computer has logins for the whole crew? That's the sweetest thing I've ever heard of. Ian Landsman No chance. 1980s energy right there. It is. Yeah. So you made a non-admin user. Aaron Francis Yes. Ian Landsman So in the 80s, they needed user accounts on computers and they built them and they still have them. And you can actually use them even though none of us think about the computer having multiple accounts. Like this didn't even cross my mind. And it's kind of foundational that- It's foundational. Aaron Francis That mom and dad and the kids don't share documents. And so that's like fully built in, fully isolated, baked for a thousand years. And so now your agents have their own little Mac login. Ian Landsman Yes. And so that's the thing. It's like, whatever. I'm not the guy who's worried about the like, they find zero day and prompt inject you. If you hack through my Mac OS user, so be it. You got me. You win. Like you win. So I'm not worried about that. I don't have anything that's targetable. China's not targeting me to get the secret keys to help spot cloud. Like they don't care. So like, and even then you're going to have to get through the max security layer to do it. So I feel very confident in that, that that's well tested. They're on it. So yes, I have the standard user. Everything's running over there and it's great because it's just a full on user. So I have my solo running over there. I have GitHub desktop running over there so I can look at the changes. I have, you know, inside solo and I have Ghosty for regular terminal stuff sometimes. And I have Chrome is there and I have everything's there. It's all running in that account and it's separate from the main account, but the cloud can run free, can go naked. Aaron Francis Talk to me about log in, log out. Ian Landsman So this is the key. So this is the thing. I set this up and it was clunky, but logging in, log out is bad. And there's this fast user switching, which I might even talk about on here last week. I don't remember. But you can have fast user switching, which Mac has. So you can go up to your top bar and you can switch users. Great. And you can touch ID. But all the apps continue to run. Everything's still running. So this is a key thing. If you have the fast switching on, you don't log out of the account. So everything's running everywhere, which is great because you go back and everything's right where you left it. You can kick off Claude. You can go back to your main account. You can go back there. It's still running. But it's clunky. Still a little clunky. To do it like 100 times a day was going to be a little bit annoying. But then I discovered a not well documented feature at all. Very, very badly documented feature of Apple, which is that you can train the accounts to specific fingers on your hand. And so then you can just put your finger on touch ID and click down with the correct finger and it will switch you to that account. What? So it's some freaking magic. And I just have my middle finger is the coder guy and my right finger is the main guy. Aaron Francis And you can just lay your finger on it for a tenth of a second then push down like this that keyboard right there yep this keyboard right here touch id I can it is a button which is I never press I don't know why it's a button I don't know why it's a button this is why so you're telling me I can put I can I can set it up such I can put my middle finger on there and press. Yes. And it will switch me to that account. Instantly. That is pretty normal. That is super cool. So you just go back and forth between them instantly. Ian Landsman So now it's like it's almost faster than like three finger swiping to the other desktop, which is like how I usually run things. That is crazy so it's crazy it works great i mean look at this old dog learning new tricks this is wild minutes to like figure it out there's like a cadence to it you know so just like you got To figure it out and you got to make sure by default if you set up your second finger it will be both fingers will be attached to like the main account you don't want that You have to go into Settings and each finger, each account has only one finger associated with it. And you can actually label the fingers, which it doesn't seem, the UI doesn't make it clear you can do that, but you can actually click on the text. It says like finger one. If you click on that, you can rename it. Aaron Francis So you can rename it to like dev account and the other one can be main account. I've never heard the word finger so much in my life. Finger. Strange word. All right. So finger one. Finger one. Can be Main account. Finger two can be fast cloud account. Yes. And it's just like, how does it how does it manifest? Is it just like all your apps fade and then come in? Is it like a fast switcher? Ian Landsman Oh, man, that's cool. It's cool. So if you're in this spot I'm in out there, people, I would try this. It was very nice. It's nice to have all your Mac apps there. It's like all my development stuff now is just in this developer account. In the main account, I just run like, you know, my communication stuff. Got Slack and Telegram, whatever. Got the main browser for businessy stuff, help spot, whatever. ([Time 0:48:16](https://share.snipd.com/snip/6c544aa8-7404-4768-8461-dd1fd09ab3d7)) - **Worktrees Add Orchestration Burden Despite Branch Isolation** - Worktrees are useful when LLMs or agents need isolated branch-level environments, but they introduce orchestration complexity like separate databases and setup/teardown. - Ian and Aaron prefer multiple full checkouts or copy-on-write solutions for simpler isolation when juggling parallel AI tasks. Transcript: Aaron Francis Last week, we provoked the ire of the WorkTree Maxis. Oh, they came out. They came out with their takes. God bless them. A lot of WorkT stands. I was surprised that people live by the work trees. I happen to be on your page, which is like, if Claude wants to work tree, go ye with Claude. Like, have fun. But I just, I haven't, I still haven't succumbed to the work tree pressure i'll be honest yeah i i think it's great when claude uses it itself i think the thing that breaks down is like if Ian Landsman You want to do the full-on like i am literally working on like seven different features in the same product because then you need a bunch of other stuff besides work trees right the work Tree max these are never just about work trees. It's the work trees. But now you have to have a whole setup script and a teardown script. You have to have separate databases, right? Because like they're going to clobber each other at the database level if you don't do something about that, right? So like there's all these other layers that you have to orchestrate. If you want to be like, okay, we're going to work on the same screen at the same time with two different AIs or five different AIs or whatever. And it's just like I don't, I'm dubious on the productivity of such a thing occurring. Like, I just don't know that I can effectively manage three AIs on the same screen at the same time working on the same feature, basically. Like, it just seems unnecessary. Like, yeah, if you just have the LLM do it, it'll use WorkTrees when it wants to use them. If it's running multiple agents for something it thinks it needs to use them for, that's fine. Aaron Francis You know what I did do this week? What did you do? I did check out Solo, the desktop app. I checked it out twice. So I have... You checked it out. I have a desktop. I have a folder named Desktop. Now I have a folder named Desktop-2. I love a Desktop-2. I just checked it out twice. And it turns out it worked great. It's fine. And you know what? I have node modules in both. And you know what? I just don't care. It just doesn't matter. Ian Landsman It's fine. Aaron Francis And so now I can work on, you know, I can work on, I can like fix a bug and implement a feature in different ones. And just be sure you're not clobbering. Frankly, I push them both to main. Like, I don't even care. That's fine. But it is nice to like have two discrete checkouts. I did, like you said, have to set up a.env in each checkout to make sure that they weren't reading from the same SQLite, because it's a desktop app, same SQLite database, and they weren't Running on the same Vite port. So there's like, there's definitely work to do. But now with desktop and desktop two, it's just always like, I just have two copies. So if I ever want to do something else, I can just switch over to desktop two. And it's already got the env.local or the.env And it's already running on a different port, I need to set up so it has different icons, because sometimes I'll have production solo, dev One, dev two, and they all have the same icon. And I'm like, Oh which one do i choose um but that part like having it having being able to work on two different things is pretty nice and so i'm starting to come around to the like working On the same code base but i still i don't know if i'm just a laggard i just can't i just can't abide a work tree mean, again, because there's just more to it. Ian Landsman Like the two checkout thing is the way I think I would definitely handle it. And if I feel the need at some point to like, I'm really working on the same thing so much, so often that like I am slowing down because I have to wait, which I guess we'll see. It's kind of interesting. It kind of ties back to being in a conversation with the superpowers thing because in the way I'm currently working with the AI, I'm doing such like smaller chunks. Like it's never running that long. Like it's like, oh, it's five minutes. Like whatever. I'm not even like by the time I like kind of get it going, I check back on Slack and I come back, it's done. Like it's not even a big deal. But if I have like maybe giving it bigger chunks, which I guess is where people hit this. It's like, okay, well now this thing's maybe running for 20 minutes. And well, I don't want to do nothing for 20 minutes. So yeah, having a second checkout makes sense. I don't know. Maybe we'll end up coming around to work trees. ([Time 0:55:34](https://share.snipd.com/snip/ccd9c037-65b4-419e-8432-8d9d4a90d774)) - **Use Multiple Local Checkouts For Parallel Development** - Keep separate local checkouts with distinct env files to run multiple dev instances safely. - Aaron runs Desktop and Desktop-2 for Solo, ensures separate .env and ports, and pushes to main when appropriate. Transcript: Aaron Francis What did you do? I did check out Solo, the desktop app. I checked it out twice. So I have... You checked it out. I have a desktop. I have a folder named Desktop. Now I have a folder named Desktop-2. I love a Desktop-2. I just checked it out twice. And it turns out it worked great. It's fine. And you know what? I have node modules in both. And you know what? I just don't care. It just doesn't matter. Ian Landsman It's fine. Aaron Francis And so now I can work on, you know, I can work on, I can like fix a bug and implement a feature in different ones. And just be sure you're not clobbering. Frankly, I push them both to main. Like, I don't even care. That's fine. But it is nice to like have two discrete checkouts. I did, like you said, have to set up a.env in each checkout to make sure that they weren't reading from the same SQLite, because it's a desktop app, same SQLite database, and they weren't Running on the same Vite port. So there's like, there's definitely work to do. But now with desktop and desktop two, it's just always like, I just have two copies. So if I ever want to do something else, I can just switch over to desktop two. And it's already got the env.local or the.env And it's already running on a different port, I need to set up so it has different icons, because sometimes I'll have production solo, dev One, dev two, and they all have the same icon. ([Time 0:57:11](https://share.snipd.com/snip/50fe49f0-e507-4a8f-bfce-d576fdd05b89)) - **Widespread AI Hacking Is Often Hobby Not Product** - Many AI-driven side projects are entertainment-driven iterations rather than products with validated demand; repeat builds of similar apps signal real demand only when users pay. - Ian notes Stripe and mature companies show willingness to pay, distinguishing hobby builds from viable products. Transcript: Aaron Francis Hobby horse is that this all allows us to go so much faster in the wrong direction. Ian Landsman We don't know where we're going. To nowhere. We're going faster than nowhere. That's what everybody's doing. They don't know where they're going. How many people have you seen on Twitter that are like, oh, build this thing. And then like, nobody's using it or like, or they never ship it. All of them. I literally haven't seen anybody ship anything, like to be fair. The majority of people haven't shipped anything. Aaron Francis Yeah. Ian Landsman So again, there's just so much else to the whole thing that like people just want to ignore because this is the fun part. And like, and this is like, everybody i was just talking to somebody but like basically i think it might have been bossman chris um but you know like doing less games and less other things Because like ai is your game now right like ai is what you are doing uh so you're building a new product or you're building a thing because this is your fun entertainment time which i'm On board with which is fine but like just realize that you are in fun entertainment time a lot of the time with this you are not in productively building a product that's going to pay you Aaron Francis A lot of money time you can be in escape the permanent underclass but you have to build something that people are going to use and you have to actually ship it and we also have to balance We also have to balance this with presumably stripe is merging thousands of prs a day with no human oversight so there is some like there's some way that this is real but i think a lot of What we see on twitter is larping um but there is some reality like if you build up enough if you build up enough guardrails and enough infrastructure, you can get massive gains from like Duplicate work. But I don't know that the average user, myself included, is there. Ian Landsman But there's even such interesting stuff around that. Like, first of all, Stripe knows it is a viable product that people will pay money for, right? They do know that, yes. Which is different than what everybody else on Twitter is talking about, which is like, I'm going to build thing X, which you don't know. You haven't given any thought if this is the thing people even want, right? So right there you have a fundamentally different thing going on on some level. And then, you know, is Stripe gotten better? Aaron Francis That is the open question. Is software getting better? I think maybe it's a closed question with the answer of no. Ian Landsman Yeah, I mean, I don't know about the AI part of Stripe, but even just in general, Stripe has gotten way worse. Like Stripe is far worse than it was five years ago. It's so much harder to use. It's so much harder to use. So like, I don't know. Maybe at the end of the day, is the AI 2000 commits from AI day going to fix that? I can't ever find anything in the Stripe interface. Like I can't figure out what my MRR is. Yeah. Like, so like, this is the thing. It's like, these are things that a human needs to go in there and do. Presumably they've tried and they failed because like these are not like, I would imagine thousands of people work in Stripe and they're like, boy, nobody can ever find these settings. ([Time 1:03:10](https://share.snipd.com/snip/07b84721-9910-4aee-81ae-0959fe053cf6))