How This 5x Founder Runs His Startup Solo With AI Agents (OpenClaw, Codex, Devin) | Ryan Carson

The big thing everybody needs to remember about agents is that they are cron jobs and markdown files. [music] I'm probably shipping at least 10 PRs a day, sometimes a lot more. It's almost easier to onboard and train agents than to train humans.

It's a million times easier. In startups we used to say just do the bare minimum to get the MVP out. Do not spend time on systems [music] or processes or documentation.

That's literally reverse now. So now startup founder, you have to spend a lot of time to set up your documentation, your reference images, build all that into a cron job, the skill file, and then you suddenly are unlocked and you're doing the work of 10 people. So let's talk about Open Claw first.

So I think both you and I have set up Open Claw. It's still a little bit janky, but I think you got to work with do a bunch of stuff. And you wrote this awesome article.

Maybe we can use that as a framing of walking through your setup. So like Peter said, the truth is is that this stuff is still really hard to use, right? So that anyone who tells you that they're, you know, waving a wand and getting Open Claw, you know, to change their entire life and their entire business is not being truthful to you.

>> [laughter] >> So basically, you know, I I've been fortunate to run a number of companies. I've I've had several real executive assistants that are humans in the past, and my goal was to turn R2, my Open Claw, into an amazing executive assistant or chief of staff. So I started working on that.

And I built a system called called Claw Chief. What Claw Chief is basically a set of markdown files, and that's basically it. It's really simple.

It's some cron jobs and some mark markdown files. And here's what it does. So R2 schedules meetings for me, parses booking links, so Calendlys.

He can actually, you know, click a Calendly and go in and book time for me. Every 15 minutes he does a sweep of my inbox, on my calendar, on my priorities, and then pings me in Slack. That's our primary communication channel.

R2 follows up on emails proactively. So, if he notices that an email that he sent to schedule a meeting hasn't got a reply, he'll ping that person. Does business development outreach.

So, one of the cool things that R2 does every day is he goes out and does research for me, puts potential business development contacts in a spreadsheet, and then reaches out to them as me to try to book a meeting. So, this is open source. So, if you want your OpenClaw to behave like an executive assistant, basically just go grab this repo, tell your OpenClaw to look at it, and then to install it.

And you'll find that there's a a couple basic things in here. One of the most important things is this thing called priority map. And so, the priority map essentially says these are the projects and the people in my life that are a priority.

And that way it can make good decisions. Auto resolver is something that if R2 can resolve, you know, a question or a task, then he does that. And this is the way of telling him how to do that.

How it works. Okay. So, what is going to happen here is Let me Let me switch over to another screen, which is how I actually manage my OpenClaw.

So, I use VS Code. I also have Code X open. I also have Claude Code open.

I also have Devon open, right? So, VS Code is what I use to mess with R2. And the way I do that, you'll notice that I'm SSH'd into the MacBook Pro, which is down here in my closet.

And that's where R2 lives. So, SSH'ing through Tailscale is like a perfect way to configure your OpenClaw, right? So, that's what's happening here.

Now, you'll notice as well, we'll start a new thread, that I'm using Code X here, right? So, I've installed a Codex plugin. I'm logged in with my uh ChatGPT Pro subscription.

Um cuz everybody knows this like right now OpenAI is really subsidizing tokens. Um so, for 200 bucks a month, I mean, I'm probably I'm probably getting I mean, probably two to $3,000 worth of tokens for 200 bucks, right? >> Yeah.

Yeah, take the moment while it lasts. Yeah, I'm like, "Hey, I have investors, man. Like, I need to be as frugal as I can with this money.

So, [laughter] I'm going to take it while I can. " Um so, you know, the way that I configure R2 is mostly by talking to Codex. Um you know, I could go in and I could mess with the actual configuration files.

I could go into um the Open Claw dashboard, but honestly, I don't want to. Um so, you know, let's just uh ask Codex to do something. Uh check on our Open Claw setup and make sure that it is healthy and operating correctly right now.

All right. So, Codex is going to go off and and check on Open Claw. So, I have an Open Claw workspace, and this is actually a clone of the open-source Open Claw project.

And the reason why I do that, and this is kind of a pro tip, is that you want to allow Codex or your Open Claw to inspect its own code, right? So, often what I'll what I'll say to Codex is go pull the latest changes on the Open Claw repo and look at the source code and tell me, you know, how to update yourself to do XYZ. And then, it can go and actually look at its source code, right?

And then say, "Oh, this is how I work. I'm going to go ahead and do this and update that. " Okay.

So, it doesn't break when it updates. Is that the idea? That's part of it, or if it does break, it can go and figure out why, right?

So, to zoom out, what's hap what you know, what's happening here in VS Code is this is essentially an agent that is fixing Open Claw, right? Because often Open Claw will break, or it will break itself and it can't fix itself. Um and so by having this kind of agent on top of it, you can you can >> I see.

Otherwise, you have to go open your Mac and go like Yeah. >> clean up yourself. Right.

And you know, and like I said, my Mac lives down there. Like it's it's on a laptop and it's I can't really get to it. And and also, as you know, I was in Japan and it's like, man, if OpenClaw goes down, I can't go fix it unless I can SSH into the machine with with uh you know, Codex and have it fix it.

That's very smart. Yeah, that's very smart. This episode is brought to you by WhisperFlow.

WhisperFlow saves me at least 3 hours a week and is one of my favorite AI apps by far. It's just so much faster to dictate to AI using your voice than to type. You just talk naturally and it outputs clean, ready-to-send text.

WhisperFlow even removes filler words and formats your sentences for you. I use WhisperFlow for everything, including drafting newsletter posts, writing product specs, replying on Slack, and more. It works on Mac, Windows, iPhone, and Android across all of your favorite apps.

Try it free at whisperflow. com and use my code Peter WhisperFlow to get 6 months free. That's Peter WhisperFlow.

Now, back to our episode. So, just to recap for everybody watching. So, talked about a couple things.

We've got ClawChief, which is my specific flavor of OpenClaw that allows OpenClaw to be a really great chief of staff and executive assistant. And you can see here this is the ClawChief directory. Right?

So, when you clone that repo that I open-sourced, this is where these files end up, right? >> Mhm. And I'm not going to show you priority map cuz it has some, you know, very personal details about what my priorities are and who the important people are in my life, but that is in the priority map file.

You know, one of the the key things to understand here like the big thing everybody needs to to remember about agents is that they are cron jobs in markdown files, right? And so, if if your agent has a good cron and a good skill, you know, markdown file, it can do a lot. And the priority map is just the is it kind of long-term priority or is it every week the priority is like a task list as long-term, right?

It's long-term. So it I would say long-term and like think of it as like quarterly. And so right now my main priorities for the business are XYZ.

So R2 knows that and then and then super long-term is the people in my life that matter, right? So my family, my friends, you know, key business contacts, etc. And then you have like some sort of agents.

md that loads it into context? Uh yes. So in agents.

md it says, "Okay, your main priority is the the priority map. " But in this cron job, so I'm going to close this so you can see this. Let's actually look at what crons we've got, right?

So this executive assistant sweep, right? This is the main cron that causes R2 to operate. You can see it happens every 15 minutes.

So executive assistant sweep use the existing executive assistant skill. Right? So you can see that this message is really minimal.

Um it should actually be, you know, I actually probably want this to be even shorter. I would probably delete this for instance and just say, "Use this skill. " Is this basically like the heartbeat?

You kind of made a custom heartbeat? Yes. So the thing to know about Open Claw is you can't really rely on heartbeats.

Like heartbeats are not actually a time triggered thing. Um it's basically Open Claw saying, "Whenever I think about it, you know, I should do this thing. " Whereas a cron is a deterministic, repeatable run, right?

So every 15 minutes I want this to run and I want it to run this executive assistant skill, right? >> Oh, I didn't know that you can rely on heartbeats. I thought like, you know, it's a heartbeat it triggers like every 30 minutes or something.

It supposedly does, but I've found it not reliable. Um and so by I found crons to be more deterministic. Now maybe someone from the Open Claw team will correct me or maybe that's gotten better, but I find crons are the bulletproof way to do something deterministically.

And then, you know, then you then you you want everything to be dry, right? Do not repeat yourself. You don't want You don't want a big message in this cron that is then sort of repeated in the skill, right?

Um so, let's go into our executive assistant skill. Um Uh but the idea is this is how to do an executive assistant sweep, right? So, check these email inboxes, um get these things done.

Why don't you try asking your Do you have your OpenClaw set up on Slack or something? Can you just ask it to talk about what the executive assistant skill does without, you know, like don't don't don't share any confidential information? Like Yeah, I noticed you tried to prompt inject R2, by the way.

I thought that was funny. Yeah, exactly. How did it do?

Uh it did pretty well. It only gave me some generic information. It didn't give me your credit cards or anything, so.

>> [laughter] >> Which is good. Okay, so uh let me let me open up Slack here. I'd actually told my sisters uh I was like, "Hey guys, try to hack my uh my OpenClaw.

" Okay, so what you're seeing this is another really good uh I think OpenClaw hack. So, I have a channel called Ryan R2. Um and the reason why I do that versus direct messages is it really it makes it really easy to have threads.

And so, what I can do in here is create a thread. So, um we'll call this chat with Peter. This is a little extra work, but now I've got a thread and it's just kind of nice like I can contain all my thoughts for something.

And so, I'll ask R2 um tell us how the executive executive skill works. And to set this up on Slack is just like hooking up to like there's like some sort of connector or something, right? Uh yes.

So, what R2 is is an app, right? So, you can see right here. This is a pain in in ass.

I'm not going to lie. Like the a lot of this stuff is very fiddly, so I had to go through and create a custom Slack app. Um you know, there was like a manifest.

It was a whole thing. And I think the reason why I'm persevering here and persisting is because I really want this to be as custom to me as possible. You know, for instance, Every just launched I think they call it like a plus one, which is like Open Claw in the cloud on demand, and you can just plug it in.

I think that people are going to build billion-dollar businesses doing that. Like I think that's absolutely smart, and everybody is going to want digital um you know, digital employees, and they're going to want to plug them in quickly. But I want to be as close to the metal on this as I can right now.

Yeah. And so by by booting this all up myself and and really getting to know how Open Claw works inside of a business, I think it's valuable. So, R2 is saying the executive assistant skill is the ops layer for inbox plus calendar.

It works roughly like this. It starts by checking the live priority and auto resolution rules. We talked about that.

It reviews current To Doist tasks. So, this is one little upgrade I made. So, um To Doist is just a great to-do app.

It's It's not you know, it's nothing rocket science-y, but it's a it's a good app. It's got a good API. It's free.

And so instead of using markdown files for my tasks, I use To Doist. Um so R2 can write to To Doist. He can update it.

He can change it. Um for email, he uses the Gmail helper. So, I'm using um the the Google CLI right now instead of glog.

It's a little better. You know, so this is basically how it works. Um So, every 15 minutes it kind of runs through this stuff.

>> [snorts] >> Like let's let's let's take example. Like for for email, it just checks if there's any new messages and then sends you any important stuff. So, it has its own email address.

Um and so at my domain, right? So, R2 has his own email. Um he's got his own GitHub account.

I I think it's good to treat um your Open Claw the same way you would treat a real human employee. So, really an email address, you know, uh real permissions on that email address. Uh a real GitHub account, as much as you can.

Um I'm starting to use Agent Mail a little bit more, which is a cool YC startup um that allows your agent to have an email address. But, I think that's important. It has his own email, but it can uh you give me a read access to your email?

Yes. Yep. So, um and again, like an EA, like you would give usually an EA access, read access to your email, to your calendar.

Um and then you would give it right access to your calendar as itself, right? So, it's it's never creating things as me, right? So, all the time, like I'll get an email from someone and they'll say, "Yep, I want to meet.

" And then I'll copy R2 in and I'll say, "Hey, R2, find us a time. " Right? And then what R2 does is, you know, he he looks at all my calendars and then sends uh a a list of times that work.

I mean, we did this for this call, right? Where Yeah, yeah. you know, he's checking stuff and and and then he books a meeting, but it comes uh the invite is from R2 and then I'm added as a attendee, just like a real assistant would.

Yeah, I think just that set up alone is probably going to save people a lot of time. Yeah. So, you know, we're going to get into how do you run a whole company by yourself now, right?

Yeah. And you know, so uh I I I've raised a $2 million seed round, like I have money to hire people, but I'm not going to hire anybody for a while. And and the reason why goes back to like the purest sense of a startup founder.

Like, you want to understand every job and do it uh before you hire to fill, right? Um and you want to feel the pain and know that you need to hire before you do. Right?

So, I'm you know, I'm using all of these agents to do these skills and to learn directly how to run them myself and then I'll hire an agent augmented person, you know, to then start doing that. Yeah, I mean it's almost easier. I mean I hate to say it, but it's almost easier to onboard and train agents for some of this.

It's a million times easier. It's like I mean, I love people, right? I am a person, but the the reality is like an agent is hundreds of times easier and faster to onboard and they retain all the training and you can improve it, right?

And then they retain the you know, throughout. So the problem is if a human leaves you, they exit with all the training, right? Whereas you know, the agent always retains his training.

So that's what I'm doing. Got it. Okay, so we covered um the executive assistant sweep, the the cron job every 15 minutes.

Maybe you can just talk about at a high level what other cron jobs you have going on. And and I I love how you're using Slack by the way cuz I see you use Telegram and like it's way harder to organize the conversations and also like I get random like porn bots messages and I I can't block it. You're like, "What am I doing here?

" It's not a very serious work environment. >> [laughter] >> Yeah. Yeah.

It's like working in the alleyway or something. You're like, "What what am I doing here? " Yeah, Slack is great.

It's very professional. I And the threading is important. I mean, I don't understand how you would not have threads and obviously uh you can't.

All right, so this uh is a nightly backup. Um so clearly you you know, if your if your MacBook goes down or wherever you're running your Open Cloud goes down, like you don't want to lose all of your setup. So uh R2 actually backs himself up uh to a simple Git repo every night.

Um that's pretty straightforward. Oh, yeah, so this is my prospecting skill. This is kind of fun.

Um so this runs once a day and this uh goes out and finds people for me to meet with. Um and so you can see there's a prospecting skill. Um and this is awesome, y'all.

Um, I literally have gotten I mean, already in the probably the last couple weeks, like 10 to 20 meetings this way. And these are cold outreach meetings. Uh, so basically what I've said here is add verified Connecticut prospects to Untangles outreach sheet.

So, we have a Google sheet. It's kind of I'm using Google Sheets as my main CRM. You know, eventually we'll probably have a Postgres database or something.

Um, but this works out pretty well. Everybody already knows my email address, so that's fine. Uh, this is basically saying, "Okay, we want to find out uh, folks that are, you know, family law attorneys, um, we want you to find their LinkedIn.

Um, and then we also want you to find uh, mediators. And I've also plugged in I use the API for Firecrawl on this. So, by default with Open Claw, I I think it tries to push you towards the Brave API for search and I just >> And I just don't think it's very good.

Um, so, you know, it's like for whatever it is, 20 bucks a month for Firecrawl, I think it's really worth paying for a really good API for agent search, agent crawl, right? So, the the unlock there is, "Hey, these tools are important, like augment your agent with a good tool. Mhm.

And, you know, make sure that it's equipped with the right thing. So, this goes out every night. So, Fire [clears throat] it's using Firecrawl scrape, you know, doing all sorts of stuff in there.

So, So, it's basically just doing web searches for the divorce attorneys and stuff and is adding it to a spreadsheet and then it drafts the email and actually sends the email. Yeah, so this this is one instance where I've given it the right to send an email as me. So, so my obviously ryancarson.

com I've had it probably literally for I don't know, 25 years. So, it's a good domain for deliverability. Um, and so I I decided to give it, you know, send access to that.

Um, Um, and it basically has a service uh key to do that, um, which is a major pain in the ass, but it's done. How do you prevent it from like sending, let's say I'm a divorce attorney, how do you prevent it from sending me the same email twice or something? Is it a spreadsheet, too?

I think it's it's pretty good. Like, so so if honestly I that hasn't happened once. Um, and and the reason why is because you you have to take the time to set up Claude Chief or something like it.

Like, okay, here is the system, right? So, you've got contacts in the spreadsheet, right? The rule is always check, you know, if you've sent an email before to somebody.

And and it's a little painful cuz like you wouldn't have to say that to a person. Right? So, I think unfortunately you still have to be overly specific and pedantic, I would say, when you specify these things, but you do it once and then you don't have to specify it ever again.

Got it. So, there's some weird bugs though. So, I noticed that it would it seemed like R2 would reply to an email by sending a fresh email instead of replying in the thread.

And I was like, why are you doing this? This is this is crazy. Like, why are you sending a fresh email instead of replying in the thread?

And I couldn't figure it out. It was driving me crazy. Um, and I figured out I had specified in one of these markdown files you should copy in my other account.

So, I have, you know, uh, my untangled account, my personal account. And and what was happening is that R2 was copying my other email address. So, it looked like uh, a new email to me.

And so basically it was doing exactly what I said and it was replying in the thread, but it was also it was also creating a new uh, email. So, the point is you have to pay attention to to what you're saying, um, in these markdown files cuz the agent will do it. It'll follow your your instructions, yeah, yeah.

Makes sense. Okay, so let's kind of wrap up the open call stuff. So, I guess basically you set this thing up to be your kind of chief of staff, but also your sales person to a certain extent.

And also helping you manage your tasks and other and other stuff, right? Is that kind of what it does? Yeah.

That's the basic and then the the unlock is if it really knows all these things about your priority map and what's going on, then you can have a conversation every morning about what are we doing today? Right? Like you would with a chief of staff.

And I think if you're you know, a solo founder or you're operating by yourself, it's very helpful to have that very knowledgeable very informed you know, entity to to chat with for a little bit when you're trying to figure out like are we focusing on the right things today? Yeah. So, it's very very good for that.

That's what I use mine for just to get therapy and chat with it. Yeah. It's good.

I mean I mean you have to know obviously it's going to be a little sycophantic and it's you know, going to try to please you more than you wish it would, but but overall it's pretty good. Um, I will say it's funny though cuz I'm going to show this briefly. This is a written piece of paper.

And as of this week I've decided to switch my weekly priority task to paper. Yeah. Which sounds crazy.

But what I was finding is is that it was becoming more and more difficult for me to have visibility on the week. Like am I doing the right stuff this week? I know about today, you know, I know about this quarter, but what about this week and it and is my resolution level correct for my tasks?

And honestly like having a list and you know, and being able to see did I do everything this week has been very cathartic and very good. And I want to give my wife props cuz that's her system. So.

I just use a whiteboard, dude. I just whiteboard where I can only list three tasks. Helps you focus, too.

Smart. So, you already figured this out. Like I think I was bouncing too much between today and this quarter and I just needed something that was more at the weekly level and it's been a big unlock.

Do you have to do anything to fix open claw's memory? Cuz I I I feel like it's perhaps spotty there, too. You know, I I haven't.

I've stuck with out-of-the-box memory. I I've been tempted to go with QMD and some other things. Um but I've decided not to mess with it.

Um and this kind of goes back to like a question I think all of us are are battling with. Now that we can build so easily and quickly, there's this there's this um you know, temptation to say screw it. Like I'm not going to use open claw.

I'm going to build my own. Like I'm going to use you know, I'm just going to build my own harness, make it exactly what I want, keep it really simple. But then you end up maintaining it instead of working.

Mhm. And so I've decided, as much as I can, I'm going to just do out-of-the-box open claw with custom skills um and custom crons. And and I'm just going to call it a day.

Like I haven't even tried Hermes yet. Yeah, that's that's not real work, like just trying to switch to these different agents and like, you know, trying to optimize them. It's very No.

You can You can make good YouTube videos off of it, but it's not real work. It's not real output. It's not real work, people.

You got to do real work. And so at some level, you have to find this perfect line between customization and personalization and power tools and not building all the tools. Like you got to let go and buy a couple tools off the shelf and then get a you know, get a custom drill, but not build a custom you know, machine.

Okay, cool. All right. >> Well, let's let's move on to your AI building stack, right?

Cuz I don't think you use open claw to code, right? So like, what what do you This is Devin. The whole thing that matters here is that Devin uses cloud environments.

It's only cloud, right? So you do no local development at all. And I think a lot of people used Devin like 1 or 2 years ago, and it wasn't very good.

Um but the truth is now it's freaking amazing um because this always works. So, you know, I can say uh you know, tell me about our code base. And what this is doing is saying setting up.

It's actually spinning up a VM. Um and it's a perfectly configured real production VM. And so, I never ever ever now have to be on my Well, not very often on my local machine trying to figure out like what port I'm on and if my dependencies are up to date or if I've got some weird conflict with Homebrew, it None of that matters anymore.

It's all in the cloud. I see. And this is a big unlock.

So, I I think what's going to happen here is uh most serious people are going to move to just use cloud engineering. Like I don't think it makes sense to really clo- code on your local machine very often, if ever. Um and so, I use uh Devin for all my engineering.

But what about like, you know, all the CLIs and stuff like all the stuff they can install locally? I I guess you can install it on here, too, right? It doesn't matter.

Yeah, you spend about an hour setting up your your development environment, right? So, um you know, you carefully make sure all your dependencies are installed, you make sure it all works, you make sure npm run dev works. Like you do all that, and then it works, and then you can use it over and over again.

Um Got it. So, the other big unlock here is this. So, I'm going to show you a lot of people are doing this.

Um these are called schedules. I think everybody again who's doing this stuff seriously is using automations. And so, what you'll find here is that if you're running your own startup or if you're, you know, trying to to to augment yourself or your team, you're going to be uh running things on schedules.

And so, for instance, weekly full case coverage. So, this is an automation that uh that literally runs the entire user experience through my app from sign up to completing a divorce case and it does it automatically once a week. And the way this works is it uses a playbook.

And again, so this is kind of I just want people to think whether you use Devin or not, I think the idea of having automations that use playbooks and then increase and then uh making those playbooks better and better is a very powerful uh mechanism to use. What what is a playbook? Yeah.

So a playbook is similar to a skill. Um so what this is telling uh Devin how to do is how to do this entire uh run-through process. So it's like what what's needed from the user, okay, read the skill, set up everything.

It's just how to do it and so that is being used by this automation every week. So it's uh it's like using some playwright kind of thing to click through? Yeah.

And that's the other thing about Devin that's just amazing is like the browser testing just works. So the agent can use its own browser in the cloud absolutely reliably and deterministically. Got it.

So those are just like cron jobs, right? Like they're they're basically just These are cron jobs with skills attached. Like I said, the whole thing with agents is cron jobs and skills, right?

Um so if you can nail that down, um you know, a couple others like I do a lot of these smoke tests where it's like sign up and use this feature in the app and make sure it works. Record your screen and then tell me if anything breaks. We're doing all that.

All right, so that is my engineering. I use Devin, I love it. Um I think it's uh I think it's I think it's state of the art.

So I I do a lot of this. Now, as you know, GPT-55 just came out yesterday and we know that it's really good in Codex, right? So I'm going to share my Codex screen.

Uh let me ask you this while you share screen. Since you're a solo founder, do you even do PRs and stuff? Like how do you review your code?

Okay, so what you'll find is that if you have a good code factory, and what I mean by code factory is that agents are writing as close to 100% of the code as as possible and reviewing and shipping 100% of the code. If you actually want to do that, you need to actually have a software development life cycle that feels like it's designed for a very large team. Because what you find is that your agents are just like humans and they need reproducible uh processes that don't rely on you, right?

And so PRs are absolutely key to that. So like if we go um So everything is a PR. Um what I do is is when a PR is ready to ship, so it's actually working on some stuff.

Um So I have this playbook called land PR. And land PR basically explains to Devin exactly how to review, resolve, um and then merge a PR without me. Got it.

And I use this probably, I mean, 10 to 20 times a day. And I I really don't look at code anymore. I mean, maybe occasionally.

Um but I'm reading a lot of PRs. I'm reading a lot of um I wouldn't even say I read the PRs as much as I read the markdown and I read what the agent is saying to me about it. Okay, Ryan.

So um I'd love to hear just like end-to-end, uh now that you're solo founder, how you ship a new feature for Untangled. Yeah, like what do you use to plan? What do you use to ship?

Yep. So um I start in Devin. Um and you know, I've spent a lot of time customizing that environment.

So there's uh there's a lot of agents that MD files, there's a lot of knowledge, there's a lot of skills. So I think everybody has to start there, right? So when now that that's set up properly, then I I go in and I basically use WhisperFlow like everybody.

And I just talk for a while. And I'll say, "Okay, you know, here's everything I'm thinking about this feature. " And I just brain dump.

And then I'll say, "Help me make a PRD. " Um and then I have a skill um then it's not rocket science, but it's basically how to turn something into a PRD. And uh really all that is is is encouraging the agent to ask me questions and clarify.

And so we'll go through that process for a while, and then I'll ask it to write uh the PRD as a markdown file. And then I'll say, "Build it. " And the interesting thing is I I think you know with Codex or Devin or some of the tools, they're starting to do auto compaction pretty well now.

So I actually work in one thread. And I know there's auto compaction happening magically in the background, but I don't really have to think about, you know, stopping a thread and restarting anymore. I think the really good agent harnesses do that for you now, and it's pretty it's pretty good.

Um and then honestly, you know, good tools like a Devin or a Codex or uh I don't use Cloud Code. I haven't used Cloud Code in a long time. Um uh but I think between Codex and Devin, I honestly I'll walk away at that point, you know, or I'll completely switch uh context to something else.

And you know, I'll let that uh you know, work for 10 20 minutes, and then I'll come back. I think the other important thing is the the reason why I ended up going with Devin is because of the testing feature. So it has a really good feedback loop where the agent can use a browser and try the feature.

And this is what I tried to set up locally where you're trying to get the browser to work, you're trying to get the feedback loop to work. It was hard, whereas it's just all built in in Devin. And so it will it will do a screencast, record it, and then it will often notice problems with the with the its own PR, and then it will fix it, and then it will come back.

Uh and then I'll do some feedback, and then I'll use the the land it or land PR skill, and then and then ship it. Um and I'm probably shipping at least 10 PRs a day, I would say. You know, sometimes a lot more.

So that's the primary way that I'm building stuff right now. Um and in addition to I mean, you don't just build stuff. You have to do marketing and and sales, right?

So, that's where I'm using Open Claw to do a lot of the biz dev outreach, land the meetings, and then I take the meetings and then move things forward. So. Do you like work on multiple features at the same time or are they kind of like like maybe two at a time?

It depends. I mean, I I think it it's probably one to two at a time. Um I think we all know it's easy to build features now.

The The problem is how do you get users? How do you build revenue? And so, mostly what I think about is marketing and sales and go-to-market.

And and so, the big unlock there is that I had Devin build out a skill to use the API for Google Ads. And so, it basically built a CLI for itself to use the Google Ads API. And now, I'll go and chat to it every day and say, "How's How's our Google Ads campaign doing?

How can we optimize it? " Um and so, I don't use the Google Ads UI at all. I just, you know, talk to Devin about it.

>> Yeah, probably not that good in the Google Ads UI, so. I mean, it's just very complex. You know, you used to have to hire these agency to do your Google Ads, right?

Because it's so freaking complex. And the great thing about using the agent is it can deal with that complexity. You use it to set up the ads for you, too, or just read the stats?

Everything. So, it it has It's basically has upsert capability. So, it can it can insert, update, delete any campaign.

And that sounds amazing, but all I did is talk to Devin about it and, you know, built the CLI, gave it an API key, uh created an app, which you have to do, and then bada bing. And how about the I was going to ask you about the ad creative and also the designs for your product, right? Okay, so I think we all know this now.

Image 2, I think is the formal name of the new um Open AI image model. It's just so good. I mean, it's so so good.

And so, I have a a workflow where, you know, content creation is obviously part of the key here for driving organic traffic. So, what I do is I go out and find an expert to interview. Um, I use Descript to do that.

And then, what I'll do is I'll quickly chop that uh interview up into 60-second segments. And then, the key is I upload those to a uh a Google Drve. And they're just MP4s.

And then, what I did is I set up, again, a uh a playbook in Devin to say, "Every night, see if there is MP4s in that folder. I want you to make an API call. " I use Gemini for this.

Um, uh make a call to Gemini, watch the video, and write a description of it. And then, I want um the new openAI model, image model, to create me a really nice cover image for that video. Um, and so, I've got a nice cover image cuz, as you know, that's like the whole key.

Then, it writes the copy with Gemini. And then, I publish it all to social using Publer. And Publer is just a has a good API for publishing to social.

So, this is the machine, right? Like, you're trying to build the machine, right? And then, eventually, I'll hire someone to run marketing.

But, I want to know how to do it and operate the machine myself first and make sure that it works well. And you're able to get the brand consistency uh from like just some prompts? It's crazy.

Yeah, you know, it it there's there's a set up cost, right? So, essentially, what I've done, and and this is similar with Claude design, which we're all playing with right now. The idea is to create a design.

md um and some reference images, right? So, what I I've actually paid a designer, his name is Brett from Design Joy. He costs about six grand a month.

Um, I had him do all the initial branding work and all of the initial social images. So, I have a library of good of good image. And now what I do is I use those as reference images with a very good design at MD, and then I find that the new image model from OpenAI is completely able to reproduce it.

All right, you just saved $6,000 a month, I guess. That's good. >> [laughter] >> Basically, I mean, you know, I think it's important to pay a designer to do the first cut, right?

I I we're very much in a world where you need the taste and experience of a very good designer to build out your brand, your layout, and some basic social images, but after that, I think I think honestly we're in a world now where you can use Cloud Design plus OpenAI's new image model to get unlimited, you know, perfectly branded imagery. It's wild. Awesome, man.

So, so I guess the takeaway from this interview is like I feel like a lot of the work we do these days is just like setting up the system to do the work, right? Is that a good takeaway from this interview? Or It is.

Take the time to set up the system to do the work, because then, number one, you are understanding how the work happens, and you're refining it. And then, you either bring on more agents to augment that, or you bring on a human that's augmented with agents to do that. But it's funny, cuz I think we used to in startups, we used to say just do the bare minimum to get the MVP out.

Like do not spend time on systems or processes or documentation. But it's literally the reverse now. So, now, as a startup founder, you have to spend a lot of time to set up your software development life cycle, your documentation, your reference images, build all that into a cron job with a with skill file, and then you find that, you know, then you suddenly are unlocked, and you're doing the work of 10 people.

But the setup is pretty brutal. I feel like, especially for your startup, maybe most of your data and is freed up to talk to other humans and like, you know, do the do the more human stuff, right? Yep.

Yeah. I mean, you know, I had it's interesting because I built Untangle um before we raised money. You know, I built basically the MVP um in my free time and and so the product was basically built and now it's all about go to market.

Um so I spent a lot of time thinking about marketing, uh which is fun. I think working companies tend to bifurcate between product people and marketing people and you got to do both, man. You can't just do one of it.

Now, you need to do it all. Like you need to think about well, which features will uh unlock revenue and then how do I quickly Building them is almost trivial now and now it's issue of go to market and adoption and product-led growth and like that's the skill is is it's almost everybody is a founder now uh even even if you're not. Awesome, Ryan.

Thank you so much for sharing so much uh information and also like resources for free online. And um I guess people can find you on Twitter at Is it Ryan Carson? Yep.

Any I'm everywhere as Ryan Carson, so just Google Ryan Carson and you will find me. >> [laughter] >> Okay, cool. All right, man.

Well, you're you're personally an inspiration to me, so thanks so much for chatting again. Thanks for having me on. Take care.

Cool.