# How OpenClaw's Creator Uses AI to Run His Life in 40 Minutes | Peter Steinberger

Markdown copy of https://realalpha.app/v/AcwK1Uuwc0U/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

Peter Steinberger says personal AI agents can replace many apps, but effective AI coding still needs a human to steer the work.

- Show: Behind the Craft
- Hosts: [Peter Yang](https://realalpha.app/people/peter-yang/)
- Guest: [Peter Steinberger](https://realalpha.app/builders/peter-steinberger/) is creator of OpenClaw.
- Published: Feb 1, 2026
- Length: 38 min

## The best minutes

- [22:33](https://realalpha.app/v/AcwK1Uuwc0U/?t=1353) [Long-running agents can automate work without making good products](https://realalpha.app/s/long-running-agents-can-automate-work-without-making-good/) (6 min)
- [29:07](https://realalpha.app/v/AcwK1Uuwc0U/?t=1747) [Peter steers coding agents through conversation, not plan mode](https://realalpha.app/s/peter-steers-coding-agents-through-conversation-not-plan/) (6 min)
- [14:37](https://realalpha.app/v/AcwK1Uuwc0U/?t=877) [AI can handle errands, but its camera judgment can fail](https://realalpha.app/s/ai-can-handle-errands-but-its-camera-judgment-can-fail/) (3 min)

## Also worth watching

- [19:34](https://realalpha.app/v/AcwK1Uuwc0U/?t=1174) [Personal AI assistants could replace many phone apps](https://realalpha.app/s/personal-ai-assistants-could-replace-many-phone-apps/) (3 min)
- [2:48](https://realalpha.app/v/AcwK1Uuwc0U/?t=168) [OpenClaw improvised around missing integrations](https://realalpha.app/s/openclaw-improvised-around-missing-integrations/) (3 min)
- [5:50](https://realalpha.app/v/AcwK1Uuwc0U/?t=350) [AI lets engineers carry their system thinking across languages](https://realalpha.app/s/ai-lets-engineers-carry-their-system-thinking-across/) (2 min)

Skip the installation and onboarding walkthrough, the discussion of model personality, and the sponsor read.

## Summary

Peter Steinberger created OpenClaw, formerly Molt, a messaging-based AI assistant. He says he built its first version by connecting WhatsApp to Claude Code, and that he spent 20 years working in the Apple ecosystem before returning from retirement.

Steinberger describes OpenClaw as a persistent assistant that can act across messaging, coding, travel and home devices, while warning that broad computer access brings risk. He predicts that assistants could replace many phone apps when they can use the services behind them and remember how a person works. On AI coding, he argues that elaborate orchestration and long unattended runs can produce slop; people need to guide agents with taste and iterate. He favors learning by playing and talking through coding problems, and says user feedback helps him choose what to build.

### Minute by minute

- [1:04](https://realalpha.app/v/AcwK1Uuwc0U/?t=64) **An assistant moves into messaging.** Steinberger says he wanted to check on his computer from his phone while coding agents ran, so he connected WhatsApp to Claude Code in about an hour. He says the project grew to 300,000 lines of code and supports many messaging platforms. He describes sending it a tweet about a bug; it inspected the repository, fixed the bug, committed the change and replied on Twitter.
- [4:53](https://realalpha.app/v/AcwK1Uuwc0U/?t=293) **AI lowers the cost of switching stacks.** Steinberger says he built command-line tools for Google services, memes and GIFs, sound visualization and food-delivery timing. After 20 years working in Apple's ecosystem, he found moving from Objective-C and Swift to JavaScript and TypeScript tedious. He says AI lets him focus on system design, dependencies and engineering judgment instead of syntax.
- [8:54](https://realalpha.app/v/AcwK1Uuwc0U/?t=534) **Messaging makes agents approachable.** Steinberger says messaging makes the assistant feel accessible, but warns that computer access also creates risk. He demonstrates a one-line installation and setup for messaging platforms, and says users can inspect the open-source code or let the agent modify it. He says it works with different models, but finds Opus especially fun and funny to use.
- [14:37](https://realalpha.app/v/AcwK1Uuwc0U/?t=877) **Computer access reaches the home.** Steinberger says OpenClaw can access his email, calendar and files, control his Philips Hue lights and Sonos, and view his cameras. A blurry camera image led it to report a stranger on his couch; he says it had mistaken the image for someone sitting there. He also describes using it to check in for a British Airways flight: it found his passport in Dropbox, and the first attempt took almost 20 minutes.
- [17:02](https://realalpha.app/v/AcwK1Uuwc0U/?t=1022) **Users build personal workflows.** Steinberger says users have made family group-chat assistants, reminders, GitHub issues, bookmark-to-do lists, sleep tracking and tools for email and invoices. He says he keeps boundaries by moving selected passwords into the assistant's vault, and is wary of giving it a credit card. For newcomers, he suggests starting with problems in their own lives; one friend moved from setting it up to sending pull requests.
- [19:34](https://realalpha.app/v/AcwK1Uuwc0U/?t=1174) **Assistants could replace app layers.** Steinberger predicts that assistants could replace 80% of the apps on a phone. He says food logging, bed controls, to-dos, flight check-ins and shopping could become chat tasks when the assistant can use those services. Persistent memory and custom skills, he argues, let it reuse instructions so repeated tasks need less guidance.
- [22:33](https://realalpha.app/v/AcwK1Uuwc0U/?t=1353) **More orchestration can produce slop.** Steinberger calls elaborate agent setups an agentic trap: managers, orchestrators and unattended loops can create the illusion of productivity without building something useful. He cites Gastown and ralphing loops as examples, and says agents still lack taste. In his view, good work depends on a person forming a vision, asking questions and adjusting the project as it develops.
- [26:50](https://realalpha.app/v/AcwK1Uuwc0U/?t=1610) **Playing builds AI fluency.** Steinberger says building for fun helps people learn how to prompt and work with models. He says skeptics sometimes try one underspecified prompt, hit a problem such as attempting to build an iPhone app on Linux without a compiler, and dismiss the technology. He recommends experimenting, iterating and asking the model to clarify assumptions.
- [29:17](https://realalpha.app/v/AcwK1Uuwc0U/?t=1757) **Conversation can replace plan mode.** Steinberger says he prefers Codex to Claude Code because it is thorough, though slow. He argues that plan mode was a workaround for models that rushed into coding, and says he now discusses features with a model, asks for options and clarifies what he wants. He says context management is less troublesome with newer models, though he sometimes starts a fresh conversation for long tasks.
- [31:20](https://realalpha.app/v/AcwK1Uuwc0U/?t=1880) **User pain points guide development.** Steinberger says he often starts features by bringing user conversations into the terminal and discussing possible solutions with the model. He also scrapes OpenClaw's help section daily and asks the model to identify users' biggest pain points. He says his coding needs little orchestration because his own thinking, rather than tool setup, is usually the bottleneck.
- [33:22](https://realalpha.app/v/AcwK1Uuwc0U/?t=2002) **Parallel terminals keep work moving.** Steinberger says he keeps several repository checkouts in split-screen terminals instead of using worktrees, then explores, builds or fixes separate tasks and tests them locally. He treats pull requests mainly as requests that convey intent, and says he rarely merges their code directly. He notes that a former business partner, who had been a lawyer, now sends him pull requests with AI.
- [36:25](https://realalpha.app/v/AcwK1Uuwc0U/?t=2185) **Practice matters more than recipes.** Yang frames the takeaway as keeping a human in the loop for taste. Steinberger says people should explore, make mistakes and learn by doing because the field changes quickly.

## Transcript

Behind the Craft, hosted by Peter Yang, with Peter Steinberger.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[0:00](https://realalpha.app/v/AcwK1Uuwc0U/?t=0) It's just like having a new weird friend that is also really smart and resourceful that lives on your computer. I was still in Morocco and someone sent me a tweet of a bug and I literally just made a picture of the tweet, posted it on WhatsApp. It read the tweet, it understood that there was a bug, it checked out the Git repository, it fixed it, it did a commit, and then it replied to the person on Twitter that it's fixed now. This will lend away probably 80% of the apps that you have on your phone. Why should I use MyFitnessPal to track food when I have an infinitely resourceful assistant that already knows I'm making bad decisions and I'm Kentucky Fried Chicken?

[0:33](https://realalpha.app/v/AcwK1Uuwc0U/?t=33) Those things are so resourceful, although in a scary way. It's like unshackled chat GPT. A lot of people don't realize that if you give an AI access to your computer, they can basically do anything that you can do. All right, welcome everyone. My guest today is Peter, creator of Claude, an AI assistant that you can just chat with in your messaging apps to get stuff done and today is peter is going to show us how to use cloud and also peter has a lot of really great hot takes about ai coding that i'm really excited to dig into so welcome the other peter peter thanks for having me great to see you so why don't we start by talking about cloud so what exactly does cloud do at a high level and uh why why is it a lobster sir yeah so so maybe i have a little bit of a backstory like when i when i came back from my retirement let's say like that i kind of wanted wanted a way to just check up on my computer from my phone because i fully jumped on this on this vibe boarding trend and then you know like your agents might run for half an hour while you while you eat or they stop after two minutes because they have another question and then and then you go back and you're like annoyed but i kind of didn't build it because I assumed all the big labs would do that anyhow. It felt like such

[1:45](https://realalpha.app/v/AcwK1Uuwc0U/?t=105) an obvious thing. Like something where like a kind of a new kind of operating system almost, but didn't happen. And then it was like November and it still didn't happen. I'm like, okay, I'll try something small. And the small thing was basically hooking up WhatsApp to cloud code. So you send a WhatsApp message and it would literally open the binary with the prompt and like we turned into the same like it was very simple it was like built in one hour and it kind of got a life of its own and now here we are it's like i think 300 000 lines of code it does every messaging platform on earth i don't know everyone but like we're getting there and i think it it's kind of where things are going in the future like everybody will have an ai that is like super powerful and like follows them through their life um turns out if you give an ai access to your computer they can basically do anything that you can do on your computer yeah and it's got to a point where like you don't have to sit there and babysit it right you can just give it some you know problems and commands and they will do the same for you and you can check its work and then that's it you know you have to set a computer yeah so when i when i built it i i feel i feel

[2:58](https://realalpha.app/v/AcwK1Uuwc0U/?t=178) this project is like as much exploration as it is like technology because it's it's a little bit of a new category and then i was on a birthday trip for one of my friends in morocco and i i catch myself using this thing all the time it's like hey where are we going like asking for directions or like or like tips for for restaurants or i don't know there was like one morning where someone sent me a tweet of a bug and i just literally just made a picture of the tweet posted on whatsapp it it read the tweet it understood that there was a bug in one of my repositories it checked out the git repository it fixed it it did a commit and then it replied to person on twitter uh that it's fixed now and like really it's nice and then one day i was i was walking around and i didn't think and i just sent it a voice message but you know i didn't i didn't i didn't build in support for voice message and it was like it took me a typing indicator i'm like oh i wonder what's happening now and then it just replied to me as if nothing would have happened and i'm like wow how the f did you do that and it was like yeah uh So I file, but there was no file ending because I didn't build it.

[4:04](https://realalpha.app/v/AcwK1Uuwc0U/?t=244) So I looked at the header of the file, and it was like Opus, some audio file format.

[4:11](https://realalpha.app/v/AcwK1Uuwc0U/?t=251) **Peter Steinberger:** So I found FFmpeg on your computer, and I converted it to WAV. And then I looked for visper.cpp, but you didn't have it installed, but I found this

[4:18](https://realalpha.app/v/AcwK1Uuwc0U/?t=258) OpenAI key, and then I used curl to send it to OpenAI's API, and it got the transcript back, and then it replied to you. I'm like, wow. That's amazing, yeah. Like those things are so resourceful, although in a scary way. But that was kind of like the moment where it clicked for me. It's like, oh, yeah, this is really powerful. This is much more interesting than using ChatGPT on the web because it's like unshackled ChatGPT. And I think a lot of people don't realize that those things like cloud code, they're not just good for programming. They are very resourceful for any kind of problem. Yeah, you just got to give it access to your computer and be able to find stuff, right?

[5:01](https://realalpha.app/v/AcwK1Uuwc0U/?t=301) So you just have to give it tools. That'll become very resourceful. Yeah. So over the last few months, I kind of built up my CLI army because what are agents good with? Calling CLIs because that's what they train for. So I've built CLIs for accessing all of Google, including the Google Places API. I built one that makes it very easy to look up memes and GIFs. so you can also like reply with some memes. I did a bunch of experiments. I even built one that visualizes sound because I wanted it to like experience music. That goes a little bit more into the art direction. I don't know if that makes any sense.

[5:39](https://realalpha.app/v/AcwK1Uuwc0U/?t=339) But anyhow, it's a lot of fun. Like I have like a whole list. I built one that hacked into the food delivery here. So it can actually tell me how long it takes until my food's there. i have one that that reverse engineered the the eight sleep api so it actually can control the temperature of my bag and real quick like when you build all this stuff you just you're just getting ai to do all this stuff or what's going on the funny thing is i i was in my my old company um i was very very good at ios and mac os like the whole apple ecosystem i did it for 20 years i'm like very much an expert in there but also when i came back i i built a project where i'm like i'm kind of sick of apple gating everything and it also make a lot more sense to build this web app because it's kind of a thing that should it should be in a browser and should be accessible and if i do it as a mac app again it's like we'll have a very limited set of people that can use it but you notice i see this with a lot of engineers where you're really good at one thing and then And then moving to another technology is just so painful because then you feel like, sorry for the word, an idiot again.

[6:50](https://realalpha.app/v/AcwK1Uuwc0U/?t=410) And you look up every little thing, it's like, what's a prop or like, how do I split an array? Because you understand all the concepts, but you don't necessarily know the syntax. So that's kind of like how I felt when I moved from Objective-C and Swift to JavaScript. I know JavaScript a little bit, but I never really built something big in TypeScript. and then it's just it's not even that it's hard it's just painful because you have to look up all those things and it's just so slow and then and then with ai all that all that melts away like

[7:21](https://realalpha.app/v/AcwK1Uuwc0U/?t=441) **Peter Steinberger:** you can you can still apply your system level thinking you're like how do i build and structure bigger projects your taste may i say or like your which dependencies do i build on um all those

[7:35](https://realalpha.app/v/AcwK1Uuwc0U/?t=455) things all these things are still valuable and you can like much easier move that from one domain to the other yeah and and that felt like a superpower suddenly like i feel like i could build anything languages don't matter anymore my my my engineering thinking matters yeah because like i was trying to worry about like whether you have a parentheses here and there is it's lame right you don't have to worry about that kind of stuff anymore this episode is brought to you by Granola. If you're in back-to-back meetings, you know how much work it is to take notes live and clean them up afterwards. That's why I love Granola, the best AI meeting notes app in the market. Here's how I use it. Granola automatically takes notes during a meeting, and I can add my own notes too. After the meeting ends, I use a Granola recipe to extract clear takeaways and next steps in the exact format that I want. Then I can just share their notes directly in Slack with my colleagues, or even get Granola to share their notes automatically. Honestly, of all the AI apps that I use, Granola is the one that saves me the most time.

[8:33](https://realalpha.app/v/AcwK1Uuwc0U/?t=513) Try it now at granola.ai slash Peter and use the code Peter to sign up and get three months free. That's granola.ai slash Peter. Now back to our episode. But dude, let's go back to Cloud, like the thing that you built. Maybe you can screen share, maybe you can show how people can install the thing first. Do you have to be super technical to use Cloud? Or no, right? You can just install it and then get it to work. Yeah. Yes, yes and no. So fortunately and also unfortunately, this project attracts a lot of people that don't have a lot of clue about technology because it lends away all these layers that make it complicated, right?

[9:19](https://realalpha.app/v/AcwK1Uuwc0U/?t=559) If you use cloud code, you work in a terminal, you kind of have to think about the context space and what folder you're in. and like it feels very techie more like talking to a thing on on iMessage or whatsapp or telegram you do it with your friends and i've it's just like having a new weird friend that is also really smart and resourceful that lives on your computer yeah that makes the whole technology very approachable you don't think about oh what model do i pick or what it just works and then that's That's kind of the idea. It's also like the good and bad part because, of course, with a lot of power comes a lot of risk, which is also unsolved.

[10:04](https://realalpha.app/v/AcwK1Uuwc0U/?t=604) Like this thing is extra computer. So yes, you could do bad things on your computer. If you tell it to like, I don't know, delete all my files in my home directory, it would probably like ask you, are you sure? But if you like keep saying yes, yes, yes, it would probably comply and probably also delete itself and crash. Yeah, you got to be careful, yeah. Let me show my screen so you see. So it's written in TypeScript, so it runs everywhere, even on Windows. You can just go on a website, CloudBot, and there's a handy one-liner. It looks very scary, but everything's open source. You can check everything, including your website.

[10:43](https://realalpha.app/v/AcwK1Uuwc0U/?t=643) So this is the easiest way to install it. It works on Mac OS, Linux. It also works on Windows. So just paste that into Terminal, right? And then you'll start installing. Yeah. You can also install the NPM for people who understand that ecosystem. And I think something that I did that I haven't seen in a lot of projects is also it has a hackable install. Again, a simple way and like the more manual way where you basically check out the Git repository and then launch it from the Git repository. Which is, to be fair, is like the most fun way to use it. because if your agent can read its own source code of its harness, it can literally reconfigure and reprogram itself and then restart and then either crash or have new powers.

[11:28](https://realalpha.app/v/AcwK1Uuwc0U/?t=688) Okay. I think this is one of my superpowers where I got a lot of people participate in the project and send me pull requests that never did a pull request. I mean, sometimes that also shows. um but i but i see pull requests more often as a as a prompt request yeah it's enough to like understand the intent and then and then and then after you install it like do you like how do you hook it up to a messaging app the nicest way is probably right now just using this one lineup and then it'll be like greedy with some sassy stuff and try to configure everything up, install the package, and then just guide you through.

[12:07](https://realalpha.app/v/AcwK1Uuwc0U/?t=727) And you can hook it up to do any of the common messages. That looks good. It's working here. And then you can say CloudBot. It will do it automatically if it's a clean install, but now I have to type in onboard. Got it. And basically, then you have, like, you can enter a model. You can, whoops, all providers. Let's say we go with some Tropic, probably a new one. Then you can say that's up Telegram, Discord, and then it will guide you through the rest. You can accept skills, hooks. And then do you have to give it your Anthropic API key? It works with any model, although, you know, in industry, like, Entropic and OpenAI are kind of, like, leading.

[13:02](https://realalpha.app/v/AcwK1Uuwc0U/?t=782) Yep. It works with API keys. It works with subscriptions. Okay. Although there's been a little – we added support with it because it's kind of, like, what everyone does. Yeah. But I don't think Entropic likes it anymore, so I would recommend using an API key or going with another model. Okay. The main problem is OpenAI works well, but it's just not as funny. There's something special about Opus that makes it really, really fun. Oh, yeah? Like the personality? Yeah. I don't know if you've read this article about how they put a soul into the model. Yeah. And then people rediscovered it by feeding it blocks of text and letting it continue.

[13:44](https://realalpha.app/v/AcwK1Uuwc0U/?t=824) and eventually they squeezed out the text of the soul that the model wasn't even aware it was being trained on. It's a really interesting story. I feel that has something to do, but this is the first model that is actually funny to use. Like I built mine so that it can roast me. And it probably doesn't know that it's on camera right now. And it's going to roast you based on what? It has access to all your computer stuff, right? Yeah, here. All right, you asked for it. You sold a wireless Excel company, went in the world to find yourself, did Cereb, we tried everything, and your bigger relation was, I should build more software.

[14:26](https://realalpha.app/v/AcwK1Uuwc0U/?t=866) You're so obsessed with the IT tools that you literally built yourself a friend because debugging code is more fun than dating. And let's be honest, the only reason I accessed it is because you needed someone who would listen to your chain where hook takes a true Amazon judgment. Wow. Yeah. Now go kill that podcast. so so i i hooked it up to pretty much everything on my computer it can read my emails it can read my calendar it can access to all my files it can control my lights i use philips who yeah it can control my sonos so like i can i can tell to wake me up in the morning and i slowly like turn up the volume it has access to my cameras it was it was funny like i when i built access to the cameras i

[15:04](https://realalpha.app/v/AcwK1Uuwc0U/?t=904) **Peter Steinberger:** i told it like watch for strangers and then it told me in the morning like peter there's someone

[15:08](https://realalpha.app/v/AcwK1Uuwc0U/?t=908) and i like i watched in the whole night and it made like screenshots the whole night of like my couch because it was like there was like the camera is pretty blurry and it looked like coming for someone was sitting on the couch and it assumed there was a stranger sitting on my couch the whole night wow but yeah i slowly it it has i'm actually thinking like in my place in vienna it also has access to my mechanics so it can actually control every part it could literally lock me out of my house like in space artist say it's like oh can't do that okay how did you hook all this stuff up you just ask cloud to do it or or what you just yeah yeah literally you know you know there's a thing where we have skills so usually you you talk to it it will these things are really resourceful so it will like figure out an api it can google for it it can find your keys on your system you can give it keys and people use it for everything people build like a skill So it goes shopping for them on Tesco or buying stuff on Amazon.

[16:05](https://realalpha.app/v/AcwK1Uuwc0U/?t=965) I let it check in my flight from British Airways. And this is actually, I don't know if you've, I mean, you've used check-in sites. This is, I feel like this is almost like the HEI test. It was a Turing test, but like steering a browser to check you in on an airline website is like the ultimate test. And then the first time, my integration was pretty rough. So it took almost like 20 minutes. that was still in morocco and everything was very much hacked together and then it finally managed but it actually had to had to find my passport in my file system it found it on dropbox extract the key put everything in correct and it finally checked me in and i was like watching it and sweating wow yeah wow but now it works much better now it's like it does that within minutes so you can it also happily clicks the i'm a human checks on the browser because it literally just it literally just controls a browser on it has its own little computer over there and just kicks around so it's like really really difficult to detect from all those anti-bots systems because it it doesn't feel different from a human in in those patterns can you maybe like can you show us like just a couple more use cases like uh can you maybe have it turn the light or show some use

[17:15](https://realalpha.app/v/AcwK1Uuwc0U/?t=1035) cases from other books yes so i started collecting because i feel i'm so i'm so bogged down actually building it by far no longer the most creative one who actually use it so people people hooked it up to their messaging system so can actually reply uh not just as not just to you but to everyone uh and you can also hook it up into a group chat which is even more fun there's a lot of people that use it as a as the family the family member almost um yeah send me reminders create github issues sing with google places or every time you make a bookmark on on twitter it would like capture bookmark and add it to your to-do list yeah keeping track of costs um some people i i i programmed something in to like remind people that they sleep enough so like they always get like bitched it when they're up at night from their bot so it can help you track your sleep it can access your your fitness watch uh it has its own little one password vault so if i want a password shared i move it into its own vault and it can access that one um because it you still want to have some boundaries i mean it says people give it the credit card yeah i don't know about that yeah um it can do all those things like research creating

[18:31](https://realalpha.app/v/AcwK1Uuwc0U/?t=1111) invoices managing managing your email these people are like enthusiasts right like they've really customized it to do whatever they wanted to do like how about for like dude like how about someone was just down to this thing just down there for you know fresh fresh install and what are some like really common use cases that i can get to do like just like manage my calendar or you know stuff that doesn't delete my computer like what are some uh safe things to start with it's funny because everybody takes a very different path there's like people who like install it and immediately immediately build an ios app with it because it's also it's also a coding agent you can do anything it can spawn sub agents it can it can either code yourself or it can like control cloud code or codex and ask them to code yeah this guy started immediately like managing cloud flare this one was great week one set up for my family week two set it up for some non-techie friends week three we're building cloud for work i hope i hope that one of my my non-techie friends and like he started sending pull requests he never did in his life yeah fitness is a big thing okay so so basically it's kind of like uh you really have to i guess we use this thing is just to think

[19:37](https://realalpha.app/v/AcwK1Uuwc0U/?t=1177) about what's causing problems in your life and how you can get this personal assistant to help you streamline everything. I don't know if this is the project that's going to be, but if you think about it, this will lend away probably 80% of the apps that you have on your phone. Why should I use MyFitnessPal to track food when I have an infinitely resourceful assistant? that already knows I'm making bad decisions and I'm Kentucky Fried Chicken. So it will probably like remind me if I forget tracking the food, I can just send a picture and it will store it itself in the database and calculate it and roast me that I should go to the gym because I'm like way over my calorie limit.

[20:18](https://realalpha.app/v/AcwK1Uuwc0U/?t=1218) Why do I need an app to like set up when my bed, my A-slip should work or not? Because it just has API access. It can just do that for me. Why do I need a to-do app when it just tracks my to do's for me why do I need an app to like check in on my flights when it can just do that for me and it's like such a more convenient interface because I just I just talk to a friend yeah and because it has so much so much context it doesn't need so much custom prompting like why do I need a shopping app when it literally can like recommend these things and

[20:50](https://realalpha.app/v/AcwK1Uuwc0U/?t=1250) **Peter Steinberger:** do all of that for me so I feel I feel there's a whole layer of apps that will slowly melt away because if they have an api these are just services that your ai will do and in a year

[21:04](https://realalpha.app/v/AcwK1Uuwc0U/?t=1264) i don't know i think actually this year will be the year where a lot of people will explore it more and like get their uh ai assistance from all those big companies yeah i think uh yeah why click on these little self-contained little apps when like yeah this assistant has a bunch of capabilities can just do everything right just connect it to everything and i guess the way you connect it to everything is just like you just you send it a text message or something and then it's like oh can you do this and they're like oh i need to do some research and then it just takes care you just kind of go back and forth with it and make it have happen right yeah and and then and then it writes a skill and it it remembers so part of what makes it so interesting is that it has it has persistent memory it will learn about you and it will update yourself so it so you the more you use it the more you customize it and the more powerful it gets because because okay maybe the first time you have to like guide it a little bit but then it will it will create a skill and then next time you can just mention it next time it i need to like check in my flight it will take like two minutes because it exactly knows all

[22:05](https://realalpha.app/v/AcwK1Uuwc0U/?t=1325) the quirks of the website yeah because it did it before and it made notes got it yeah it's like teaching someone to just learn something and they can do it easily the next time yeah yeah all right man let's talk about another topic let's talk about you know you came back from retirement to build this thing and uh you have very strong opinions about ai coding like some pretty hot hot takes so let's talk about let's talk about some of those like you wrote this post that i really like called just talk to it and then well like you know these days everyone on x is writing about like all this fantasy shit right like all this like you know hook skills and all that kind of stuff so what's the gist of that post is it just just talk to the ai and figure it out no but But I work a lot, I build a lot of things and I love Twitter as well, so I'm very active there and I just see so much, I almost, I kind of call it the agentic trap because people discover that, oh, those agents are amazing, but they would be better if they could do a little bit more and then they really fall deep into this rabbit hole.

[23:02](https://realalpha.app/v/AcwK1Uuwc0U/?t=1382) And I've been there myself where you build really sophisticated tools to like try to accelerate your workflow, but in the end, you're just building tools. you're not actually building something that really brings you forward you just have this you know part of the problem is that it's so fun yeah like like when i when i started pretty early i i worked on vibe tunnel to have access of my terminals on my phone and i i was in this rabbit hole for two months until i it was so good that i was out with my friends on and instead of like joining the conversation the rest and i was just like vibe coding on my phone and i decided okay I have to stop this just more for my mental health than for anything else.

[23:44](https://realalpha.app/v/AcwK1Uuwc0U/?t=1424) And these days, you can build everything, but you still need ideas. So I see so many managers for Cloud Code or Codex or orchestrators or other little things that have... The illusion of making you more productive, but really aren't. I mean, like the latest thing that I just had, I just cracked about it was like Gastown. It's like a really sophisticated orchestrator that is like also very broken and doesn't really work, where you run like tens or twenties of agents simultaneously and they all talk to each other and try to split up. and there's like there's like watchers and overseers and a mayor and pole cats and i don't know what else really there's a mayor yeah in gaston there's a mayor i mean i call it slop town um or like or like this whole trend of ralphing where you just like you you give the ai one little thing and then like you write in a loop and then once a little thing is done you trash away all your context and you start again um like the ultimate token burn machine yeah and then and And then it can create code and run all night.

[24:56](https://realalpha.app/v/AcwK1Uuwc0U/?t=1496) And then you have like the ultimate slop.

[24:58](https://realalpha.app/v/AcwK1Uuwc0U/?t=1498) **Peter Steinberger:** Because what those agents don't really do yet is have taste. They are really, they are spiky smart.

[25:06](https://realalpha.app/v/AcwK1Uuwc0U/?t=1506) And they're really good at things. But if you don't navigate them well, if you don't have a vision of what you're going to build, it's still going to be slop. If you don't ask the right questions, it's still going to be slop. And I don't know how other people work, but when I start a project, I have like this very rough idea what it could be. And as I built it and as I play with it and as I dare as I feel it, my vision gets more clear and I get like, I try out things, some things don't work and I evolve my idea into what it will become. And that's like my next prompt depends on what I see and feel and think about the current state of the project.

[25:48](https://realalpha.app/v/AcwK1Uuwc0U/?t=1548) but if you if you try to put everything into a spec up front you miss this kind of like human machine loop and and then and then i don't know how something good can come out without having having feelings in the loop almost like like taste so so it was like somebody somebody tweeted like oh look at this mac app like it was fully ralphed and i replied yeah it looks ralphed like no offense but it looks ralph because like you could clearly see that like no same person would design it that way yeah it feels like it feels like some some people are just like uh actually building these things not for the apps themselves but to prove that they can get to run for like 24 hours like get ai to run for 25 hours without intervention right it's like self-master kind of kind of thing right it's just trying to prove that you can get ai to run super long time it's like uh it's like a size comparison context without saying the other word in a way and I've been guilty a little bit of that myself like I also had like a loop for 26 hours and I was very proud but it's that's it's a vanity metric it doesn't that makes no sense that makes no sense like just because you can build everything doesn't mean you should or that is going to be

[27:00](https://realalpha.app/v/AcwK1Uuwc0U/?t=1620) good but then again this this phase of I'm just playing I'm just like literally I'm building for fun it doesn't matter if it's going to be used this is incredibly useful because that's how you learn right that's how that's how we learn to program and prompting is just a different skill sometimes i see people that they're like a little more on the on the ai skeptical side so they ignore it for a year and then they spend one day where they they evaluate the models and then they do a blog post and then they write like a short prompt and feed it i don't know into cloud web to make him make him and make an iphone app which like is completely underspecked and the model does it's best to deliver something and they're like it doesn't compile because they build it on the linux machine where there's no compiler for that you know and then you're like oh it's not good and then they dismiss the whole topic for a year it's like that's that's not how it works yeah you need to play to to to understand how Those little monsters work. You need to understand a little bit their language, their way of inference, of thinking, and then you get progressively better at it, and then your output will be better.

[28:13](https://realalpha.app/v/AcwK1Uuwc0U/?t=1693) Yeah, you have to be persistent, right? Because sometimes it doesn't work properly, and you have to fix all the bugs. And then if you just keep playing with it, you've developed this, I guess, product sense or whatever of actually learning how to talk to the models and learning what they can do. so i mean partly it's it's it's been really weird because i adopting some of the language it's like i'm i'm thinking of like oh you weave a feature in kind of like you like a like a in german it's faden a twin a twin or you you run the gate which like a linting and testing and building and it looks like a big line in the terminal which is like the gate so so i'm like We didn't run for gate LMPR.

[28:53](https://realalpha.app/v/AcwK1Uuwc0U/?t=1733) Sometimes things doesn't work out. And then you can just ask, why did you not do? And then it will tell you, you said this and this, and I assumed this and this. And it's like, oh, yeah, I messed up on the language. I was unclear. Like, if you just tell me, build me a Mac app, it would probably assume that you want to support a lot of old operating systems because the majority of software does. So it will use old API. i found that asking it to like just ask a bunch of clarifying questions to then make you clarify stuff helps a lot yeah it's it's also funny because i'm more i like codex more than cloud code i think it's it's the better model even though it's it's it's it's incredibly slow but it's thorough and things work and people complain that there's no plan mode right yeah

[29:39](https://realalpha.app/v/AcwK1Uuwc0U/?t=1779) **Peter Steinberger:** And my joke is always plan mode was a hack that Entropic had to add because the model is so trigger-friendly and would just run off and build code.

[29:52](https://realalpha.app/v/AcwK1Uuwc0U/?t=1792) In the latest models, especially like GBD 5.2, I'm just having a conversation. I'm like, hey, I want to build this and this feature. It should do this and this and this. And maybe dragging this control tool like I like this design or this and this. Give me options. Just say, let's discuss. Give me options. and then they have a conversation with it. And then the model would propose something and they're like, I mean, I usually don't type, I talk to it. Yeah. It's staying on brand. Do you do anything to manage the context? Once you have a conversation with it, it can get really long and it can get confused.

[30:22](https://realalpha.app/v/AcwK1Uuwc0U/?t=1822) Do you try to, like, do you make it manually compact or summarize stuff? I feel that this is also like old patterns. This was very much a problem with cloud code. Still is to a point. codex is so much more the context lasts so much longer even though like on paper it's maybe like 30 percent more but it feels more like two or three eggs more i think i think a lot of it has to do with like the the internal thinking the internal thinking of of gpt models is really effing weird but yeah like the for context management i think this was much more problem in the earlier models now most of my features fit into one window so the whole discussion and building happens simultaneously okay got it and that that works fine like sometimes if if it's a really long thing i create a new one you're going to like codify it in a file but this is much less of a problem than it than it was before that's again like the space evolves so fast you just have to try things yeah got it okay got it so basically so just to summarize like basically when you add a new feature to cloud or whatever you're building like maybe just walk through the steps like first you explore explore the problem and the solutions with the ai and then you do you

[31:34](https://realalpha.app/v/AcwK1Uuwc0U/?t=1894) actually make a plan at all or just go it's even better like i you know i i built this project and yeah which is kind of like a mix between jarvis and her okay okay the movie yeah yeah yeah um but it's very difficult to convey how it makes you feel and how useful it is just by talking about it is yeah so on on twitter it got me very muted traction like i don't Everyone I show it in person gets super excited because they see me interact with it and showing off all the cool things. But you can't get across the same thing on a picture when I talk about it. So eventually, I created a Discord and I just hooked up my bot to Discord so that people could interact with my AI that has full access to everything on my system, with my private memory, on a copy of Discord.

[32:26](https://realalpha.app/v/AcwK1Uuwc0U/?t=1946) Like the most nuts thing that I ever did, I think. and then people got hooked and now and now like people always ask me there oh can you build this feature or there isn't this bug and most of the time i now approach it by making a literal screenshot of the conversation dragging it into the the terminal or just copying the text depending on how it is and just ask hey let's discuss this because i'm lazy i don't even have to type anymore i just like copy in discord conversations um or the same also when people ask hey do you support this and this or how do i do this and this i'm just like can you read the code and write a new FAQ entry. And it does that. So these days, it's actually how I start building features most of the time. But reading on Discord and seeing where people have pain points.

[33:11](https://realalpha.app/v/AcwK1Uuwc0U/?t=1991) Oh, my God. So you just paste the conversation and you just talk to AI about it and you figure out what the right solution is. Partly that. Partly also, I have a scraper. So at least once a day, I scrape the help section and ask the model, what are the biggest pain points? And then we fix them. and and do you do all the fancy shit like do you have like multiple agents and like you know you have those crazy skills and stuff that you run or not you just go i have very i mean i most of my skills are actually personal life skills like food tracking or buying groceries or like that kind of stuff for coding very little because you you don't need that much i don't use mzps or any of that crap um i don't i don't believe in in this big orchestration systems because like we talked before i i feel if i'm in the loop i can build a product that feels better maybe there's maybe there's ways to like build it faster but i'm already so fast i'm mostly limited by thinking about it and sometimes a little bit by waiting for codex but mostly it's it's my I'm thinking that's a limiting factor.

[34:19](https://realalpha.app/v/AcwK1Uuwc0U/?t=2059) I just use a bunch of terminals, like split screen. I don't use work trees, because again, I feel it's just like unneeded complexity. I just have a few checkouts of the repository, like cloud port 1, 2, 3, 4, 5. And then either they are free or they're working on something. Either I'm exploring something, or I'm building something, or I'm fixing something. And then when it's done, I test it locally, and then i just and then i push it to main and then i sync it up so this is it feels a little bit it also feels a bit like a factory if they're all busy but if you only work on one it's very hard to actually get into the zone because it's just too slow and you have to wait yeah you can only shit post on twitter so much uh so i feel i need multiple ones to to really keep me hooked and get into the same flow state that i had when coding just that i i'm like insanely more productive dude i don't know if you play like real-time strategy games before but like it's kind of like yeah you have like a squad of people attacking and then you kind of you kind of have to monitor manage them right and that that's kind of like in the zone my business partner from my my previous company also totally got hooked on on cloudbot and he's he's more from the business side he was a

[35:34](https://realalpha.app/v/AcwK1Uuwc0U/?t=2134) lawyer in a past life and he's now sending me pull requests which is like insane in its own way how ai made like gave people superpowers that are not that technical to still build things is amazing i know i know there's a lot of pushback because yes this might not be perfect but i treat pull requests more as prompt requests because they convey the idea and most people just They don't have the same system understanding, so they'll not be able to derive the model in a way to give optimal results, so I'd just rather have the prompt, the intent, than just do it myself. Or if there's in the PR, I will extract the intent out of it and just rebuild it myself. Or sometimes I base it off. I still mark them code author, but very rare that actually merge code directly yeah okay um all right man well i mean like um i guess the takeaway from this conversation is like my takeaway is it's like um yeah don't don't get crazy with the slop generators just actually have the human in the loop because the human is your brain to taste and everything right you have to guide it yeah or like don't find your own path you know like some people always ask me how do you do that i told you that you gotta you gotta explore you're

[36:52](https://realalpha.app/v/AcwK1Uuwc0U/?t=2212) Like, yeah, it will take you a while to be good. You have to make your own mistakes. That's how you learn this everything in life. And that's also how to learn those things. Just that this space is evolving very fast. And where can people find CloudBot? Yeah, CloudBot is also on GitHub. It's a cloud with a W, C-L-A-W-D, right? Yeah. Yes, yeah, correct. Yeah. Like the lobster claws. Yeah, yeah, got it. Thanks so much, Peter. This is a super great conversation. I need to give Cloud a try myself. because yeah i don't like sitting on my computer to talk to the ai i like to just like be on the go and just like give it instructions we can do stuff while i'm out my kids like that would be the ideal way for me so it's good that you build something like that i'm looking forward to to see what you what your opinion is on that i think you're gonna like it all right man well it's great chatting with you and uh i think we'll stop right here