Transcription
Open Core is the most powerful AI agent framework in the world right now, but 99% of people are still using it like it's just another chatbot. Used properly, you could be running a specialized team of AI agents working 24/7, remembering everything, and costing far less than a full-time employee.
Wait a minute. The problem is there are plenty of tutorials showing you how to install OpenClaw, but no one's actually showing you how to get it to do anything. Over the last month, I've spent over $1,000 and hundreds of hours stress testing it. And I'll show you how to configure OpenFlor so it actually becomes useful. I'll also show you how to run multiple models cheaply, how to deploy a coordinated team of agents, and how to build the infrastructure that runs tasks automatically while you sleep.
And if we haven't met yet, I'm Doobie. I'm a former marketer who now builds AI apps live on stream, currently documenting a challenge to reach 100K using AI. And in just one month, I've already generated $46,000 and grown a following of over 10,000 people. I also run a community called Booby Builders where 600 plus members are learning to build their own apps with AI.
Okay, so here we are inside of Open Claw. And what I really want to do is cover the most impactful things I've set up that actually help me run my business. things like remote control from my phone, auto routing to different LLM models to handle different tasks so I can save on costs, automated jobs that run while I sleep, and the multi- aent system that ties everything together. I'll also show you the practical ways I'm using OpenClaw right now, like this automated content pipeline that has generated hundreds of thousands of views or a marketing agent that runs a full funnel on its own with self-improvement loops baked in. By the end of this, you should have a clear blueprint for turning OpenClaw from a chatbot into something that actually works for you in the background.
OpenClaw starts as a baby. When you first install OpenClaw, it doesn't magically just start running your life. It's basically an empty shell. It doesn't have workflows, no specialization or memory of how you work. This is where most people get disappointed. They just treat it like chatbt, ask it a few questions, and when it acts like a chatbot, they just assume that's all it can do. But OpenClaw isn't designed to be a single assistant. It's designed to be a system, a team of agents that each handle specific tasks and coordinate with each other. Once you set that up, the whole thing feels completely different.
Now, one of the biggest mistakes I see people make is running everything on a single model. That might be fine for casual use, but at scale it gets expensive fast and wastes a ton of capability. Not every task needs to be a top tier model. Simple jobs can be handled by cheap, fast models, while complex reasoning or coding can be routed to more powerful ones. Once you assign the right model to the right task, you will cut your cost dramatically while actually improving performance because each agent is optimized for what it does best.
Here's the exact stack I'm running right now. Everyday coordination runs on Gemini 3 Flash because routing tasks doesn't need expensive intelligence and Flash can handle it for pennies. Building, debugging, and writing all run on Claude Sonet 4.6. It's significantly cheaper than Opus and currently performs better on many real world coding tasks. So, it's the best balance of speed, cost, and capability. research runs on Kimmy K 2.5 because it can process large amounts of data in parallel and costs less than onetenth of models like set or opus with basically 90% of the capability. And then I keep claudus 4.6 as a fallback. Basically the senior architect you call in when something truly breaks or requires reasoning.
Also a quick side note, you'll probably see people recommending this trick where you dynamically switch between the Claude Max web plan and the API to reduce your token costs. Honestly, I'm not really doing that. Anthropic has been cracking down on accounts using the Max plan through unofficial rappers, and I'm not risking my main account just to save a few dollars on tokens. If you're building something serious, just use the official API. And at the end of the day, if cost is the concern, just like I've shared in my setup, let cheap models handle most of the workload and escalate only when needed. That's the sustainable way to run this long term.
One more thing that made a huge difference is this autonomy directive thingy. I explicitly tell the system to figure things out instead of asking me every step of the way. If it gets stuck, it'll check local files and logs, search docs and APIs, run small tests, and then give me a bunch of options. So, if you want, just pause here and screenshot this, drop it into OpenClaw, and say, "Hey, this is how I want you to behave." This has actually saved me a bunch of time, so I highly, highly recommend it.
Now, if you try to run everything through one AI agent, it'll quickly fall apart. The context will get bloated. the response will drift and the system will become slow, expensive, and just stop being useful. So instead of treating AI like one assistant, I treat it like a team. And it makes sense because one agent doing everything is equivalent to hiring one guy for sales, marketing, engineering, and support. They're just going to be mid at all of them. So at the center of this team is something called the orchestration layer. basically turning your main open core agent, in my case Sheldon, into a coordinator that routes work to specialized agents. This keeps the overall context of your main agent relatively free as it sends tasks to specific agents. For example, if I say create a YouTube script about AI agents, the orchestrator doesn't do it itself. It spawns a research agent to gather ideas, then a content agent to write the script, then an image agent to generate a thumbnail, and finally hands me the finished result.
So, to make this system actually work, you need structure. If you thought OpenClaw ran off one giant prompt, you'd you'd be wrong. It runs on a set of markdown files that define what each agent is, what it knows, and what it's allowed to do. If you understand these files, you can build your own agent team instead of having to rely on pre-built templates. So these are the core files you need to know.
First you've got soul.md. This is the brain and arguably the most important file. It defines the agents identity, role, and operating rules. For an orchestrator, this is where you tell it not to do work by itself, but to route tasks to the right specialist agents. Without a clear soul file, agents either try to do everything or don't know what they're supposed to do. A good soul file is extra specific. It defines the role, what inputs to expect, what output format it produces, and what it should never do. The tighter this file is, the better the results will be.
Number two is agents.mmd, and this is the team list. This file tells your main agent which agents exist and what each one is good at. Think of it as a roster of specialists. It includes things like builder, researcher, writer, and when each should be used. Important to note, only the orchestrator needs this file. Specialized agents don't need to know about the whole team.
Third, you've got tools.mmd. And this is what the agent actually uses. This defines the external capabilities the agent has access to. Things like GitHub, Notion, APIs, web tools, or anything else it needs to complete its job. Make sure you only give an agent the tools it actually needs. More tools means more tokens, which equals more noise and more chances to go off track.
Okay. Four, you've got user.md. This is information your agent knows about you. And this file contains everything the system needs to know about how you work, your preferences, style goals, and constraints. When this is set up properly, the outputs start to feel consistent instead of generic.
Five, you've got memory.mmd, which is the long-term knowledge. This file grows over time as you use the system. It stores useful context like past decisions, working solutions, technical details, and ongoing projects. So, instead of starting from scratch every session, the agent can build on what already worked. One common mistake is sending too much context to every agent. Each agent should only get what it needs for that specific task. Otherwise, you waste tokens and increase the chance of bad output.
Now, if you're wondering how to fill these files quickly, I've actually made a data packet prompt that does it for you. Just paste it into the LLM that you talk to most, whether that's chat GPT or Claude, and it generates a structured profile which you can just drop into OpenClaw and say, "Hey, here's everything you need to know about me. What should we do with this?" If you want the prompt, it's completely free. I've got it saved in the Discord. And also, if you want to support a small creator like me, a subscribe and a like is more than enough.
So before we move into use cases, one of the first things you need to set up is remote communication. You don't want to be stuck at your computer to use this system. So connect it to a messaging platform like Telegram, WhatsApp, or Discord, which will let you send requests and receive updates from your phone. Last night, I was literally in bed. It was like 9:00 p.m. at night. I know I go to bed early. Shoot me. But I asked it to create the data packet prompt for all of you guys. And it did exactly that. Like honestly, this has greatly improved my productivity because for someone who isn't a night owl, I can never really work that late. And having open claw there on my phone, even when I'm on a break, watching a movie, maybe I'm out hanging out with friends, I'm still able to get work done through my phone.
Now, the main way I like to use this for my business, is there's a thing called chron jobs, which is just short for chronological jobs, meaning tasks that are set on a schedule. So, let me show you what this actually looks like in practice. So, one of the first things that I ever built was an automated content generation pipeline that runs while I sleep. And this past week, it has generated me over 300,000 views across all of my content, including this one video with 100,000 views. Now, coming from a guy who was scraping by with 50 to 100 views per post. This setup, it has honestly changed my life. And it's completely automated through OpenClaw.
It starts with X. I've got OpenClaw connected to the X API using a research assistant skill. I really like this one specifically because it lets you filter out tweets. So to cut through the noise, I make sure it only looks at tweets with 100 likes or more. Throughout the day, it scans my niche for anything interesting, trending, or heavily debated. When it finds something useful, it scrapes it and saves it into my Obsidian vault as a markdown file inside a clippings folder. I don't touch anything. It just builds a library of ideas in the background all day. Then at 4:00 a.m., a scheduled cron job kicks in. My content agent reads everything in that folder and turns those posts into script drafts. But it doesn't just summarize them. I've given it a memory.mmd file with reference material, including a full transcript of a YouTube tutorial on how to write viral scripts. So, it drafts using a proven structure. No guesswork.
Now, after drafting, everything goes through a humanizer. This guy basically pointed out that Wikipedia literally has a detailed list called signs of AI writing. He had Claude read that article and build a skill specifically designed to avoid those patterns. It automatically updates as Wikipedia adds to the article. And it currently has 24 different identifiers to remove from my open clause writing. By the time I wake up, I've got a stack of viral scripts ready to review. Now, it's not perfect. I'd say around 70 to 80% is actually usable. But editing a rough draft is infinitely easier than starting from nothing. And the more clippings the system collects over time, the better the outputs get because it has more high signal material to work from.
So this one blew up recently. A guy on X shared that his OpenClaw agent Larry generated over 8 million views in one week. And instead of gatekeeping it, he released the entire skill for free. I set it up myself to see if it was legit. And what the system does is it generates Tik Tok slideshow posts automatically, images, hooks, captions, everything, then saves them as drafts, so you just add trending audio and publish. The creator uses it as a funnel to his apps, and the system is now generating around $670 per month in recurring revenue with almost no manual content creation. The coolest thing about this system is that it doesn't just post and forget. It actually tracks the performance and has automated self-improvement baked in. It logs which hooks get views, which ones flop, and which posts actually convert into downloads or paying users. Over time, it updates its own rules, so the content keeps improving instead of repeating the same mistakes. Under the hood, it's actually just a bundle of markdown files and scripts that teach your agent how to generate consistent slideshow images, write viral style captions, upload drafts automatically, track analytics and conversions, adapt future posts based on what worked. I noticed it still requires a little bit of involvement. You still need to pick the audio and hit publish, but the agent will handle everything else. I've only had this running for a short time, but the concept is super powerful. Your agent becomes this full-time growth operator instead of just a chatbot.
Now, this one is still in progress, but it's probably the craziest thing I've built so far. For the past couple of weeks, I've been working on a system that can come up with app ideas, build them, test them, and get them ready for release completely autonomously. Like check out this app that I built recently. It's a warranties app where you can literally put in the warranty date of anything that you purchase. So let's say you bought a chair recently from IKEA, bought it today, but it expires in 3 months. You can add your receipt and you can also get a reminder before expiry. Then you just save it and boom, you are now tracking your warranties. And this entire app was built in less than 30 minutes through my app factory. And my goal is to create this system that is able to build multiple apps daily for me to submit to the Apple App Store. I call it Sheldon, the name of my openclaw. And I want you to think of it like a small factory for apps. You give it ideas and it pulls them from a queue one by one. Each idea gets turned into a real project built into a working app. Check for quality and prepared for the app store. If something goes wrong, it stops and tells me. If everything looks good, it just keeps going. Different parts of the process use different AI models depending on the task. uh Codex 5.3 and Opus 4.6 for hard problems, sonet 4.6 for simple stuff, so it's fast without being insanely expensive. There's also a simple control panel where I can see what it's building, pause it, approve submissions, or drop in new ideas. There's even a button that generates new app ideas automatically.
Now, for the home stretch, the last 10%, I still need to submit the app into the app store that still requires a human to, you know, press a button because the Apple API doesn't allow for app submissions.
Now, if you think OpenClaw is cool, I spent weeks researching the most profitable OpenCore agents out there. You can check that video out here. And make sure to subscribe and leave a like if you got something out of this.