Transcription
I've spent over 500 hours inside of OpenClaw. I've built more than 10 agents and I've automated operations across two seven-figure businesses that I consult and I've made every single mistake in the book so you don't have to. Okay?
Now, listen. I'm not a developer. I don't write code for a living. I'm just a business operator with eight years of e-commerce experience and done over $20 million in gross revenue. And these AI agents, well, I'm not going to lie, like they changed everything about how I run and operate. So now, without further ado, here are the five things that I wish somebody told me before I started actually building AI agents.
And honestly, this is the one that would have saved me so much time if I had figured this out sooner rather than later. Now, when you first start your open claw and set it up and get the agent to work, it it works. It responds. It seems smart, and you think, "Cool. Okay, this is great. I'm good to go." But, you know, then you come back the next day and start a new session and your agent has absolutely no idea what happened yesterday, right? Like it doesn't know its role. It forgot a bunch of context. It doesn't know how you even wanted it to operate. It doesn't remember a single thing you even told it to do. And it's like hiring someone new every single day when you start working with it or every time context compaction happens. And that's what it feels like without these three files. And not just without these three files, but utilizing them correctly.
So, here is what changed everything for me. Now, the first thing is the soul.md. Okay? Now, if you guys don't understand or you're new to this, this is your agent's identity. All right? It's everything. It's the personality. It's the role that it plays. It's the operating principles and the chain of command. It's literally everything. Now, my main agent, Machaveli, is an absolute G, and I set him up correctly, and he is my CEO, okay? And his soulm tells him to be sharp, autonomous, proactive, the whole nine yards. All right? and he only reports to me. And I tell him to push back on me when something he thinks is a bad idea. And I tell him, "Think like an owner, okay? And not just don't follow instructions like a robot." All right? And without a soul.md, your agent is just generic. But with it, it has actual personality, purpose, and actual boundaries.
Now, the agents MD, this is the complete operating manual. All right? When the agent first wakes up, this is exactly what it reads first. Okay? This is what it checks. This is what procedures it follows. Okay? Think of this like the actual handbook. The soul.md says who you actually are. But the agents.mmd tells my agent and your agent how to operate. And my agents.md tells every agent to read the soul.md. Read all your memory files. Check today's daily log in the previous two days daily log. I want you to also read the lessons.mmd that I created. I want you to read your identity and everything. Okay? Don't ask for permission. Just do it every single session and after every compaction that happens.
Now the memory.mmd this is a long-term memory. Okay, this is the stuff that persists across all context compaction in all sessions. All right, the decisions that were made, the context about the business, context about conversations we were having, the things that the agent needs to remember three weeks from now, right? Everything. Okay, without this, you need to understand that your agent's going to wake up and just not know what is going on every single time. So, you need to have this set up correctly. And here's what I learned the hard way, right? Like I've built over 10 agents at this point. Every single one that works well has all three of these files absolutely dialed in. And the agents that were garbage were the ones that were missing this up until I absolutely fixed it. All right, this is extremely and utterly important for you to understand.
Now, one of my agents, I I set him up with a soldout MD and skills and then gave him Discord access and told him his role. And the first time I tested him, he said the skill isn't built yet. And he he didn't even know his own capabilities, right? And that's that's honestly my fault, right? because I forgot the agents.mmd, the file that actually like tells him where to find his skills and how to actually use them. The agent was online, he was good to go, but he just didn't even know his own brain. And you can't expect anything less. You can't expect him to know anything. You have to understand that these agents when you spawn them in, they are literal babies. They are only going to do what they know and what they are told. Nothing more and nothing less. Okay? So that's why it's very important that you know if you skip these things and you don't actually put true importance on them. This is what happens. You're going to get an agent who doesn't know anything and it's just it's literally going to be a like an agent you just have to babysit all the time and have to keep reexplaining yourself and it's going to make you less effective and less efficient and it's going to do the opposite of what you want it to do. So before you do anything else, before you build skills, before you build this whole orchestration, before you connect any tools and you try to get fancy, take these three files seriously. Soul.md, agents.mmd, and memory.mmd. Okay, this is the foundation that everything else is built on. Your agent will perform a thousand times better if you take these seriously because these are the most important.
Now, thing number two that I missed and that I wish someone would talk to me about is the memory system, daily logs and the long-term memory and lessons file. Okay, very very important. And to be honest with you, this builds directly on what I just said. So, you've got the memory MD, which is great, but your agent has long-term memory, but that's only like one piece to the puzzle at this point. Because here's the thing that nobody tells you, okay? These agents forget everything when context resets. You'll hear me talk about context compaction a lot, and you need to understand that once this happens, your agent will lose 99% of context and nuance unless you go ahead and set things up how I do out of the box when you set up your open claw, which you can go ahead and check out. I have a video on that. The card is at the top right of the screen, so you need to check that out. Okay, but context resets. Okay. And this happens after context compassion and the session ends and then the context window just becomes empty again and the agent will compact its memory to make room for new things. That's what happens when this happens. Anything that was written down is gone. Okay? Unless you build it the way I do.
So I built a three layer memory system. Okay? This changed everything for me on top of the other things that I've done. Now layer one is the daily log files. Okay? And essentially it's just a memory for that entire day. So essentially it looked like 322 2026 MD. Now these are raw notes. What happened today? What decisions were made? What tasks were completed? What broke. Think of it like just a complete daily journal. It's messy. It's detailed and it captures absolutely everything though in real time. This is extremely important to utilize later.
Now layer number two is your memory.mmd. Okay. Now this is the curated version. The daily logs are like a raw journal and your memory.mmd is like long-term memory. All right? It's it's it's wisdom gathered over the course of time. And this stuff matters three weeks from now. Okay? Because my agents will periodically review their daily logs and update the memory. MD with the insights worth actually keeping. Okay? And just like how you'd review your notes at the end of the week and pull out all the important stuff, they do the exact same thing.
Now layer three and this one really compounds over time and this is your lessons.mmd. Okay, this is where every single mistake gets logged. Okay, what happened? Why it happened and the permanent fix so that way it never happens again. These compound heavily. Okay, heavily. Now, I'll be honest. I had 51 lessons logged in three weeks of doing this. 51. Okay, and each one prevents a specific failure from ever repeating over and over again. things like always update the documentation and the script at the same time, not one or the other or one after the other. Or something along the lines of like when action fails after consuming its inputs, undo the consumption so that the system can retry. Or another thing would be like before building anything, I need you to map the entire flow end to end and present the gaps before writing a single line of code. These are things are not generic tips, okay? Okay, these things are learned through experience and are specific mistakes that cost you hours to debug or go back and fix. And now they will never ever happen again because they're in the file that gets read every single session.
Okay, now here is my rule. If you don't write it down, it did not happen. Okay, mental notes do not survive at all once your agent session resets. Okay, your agent can't remember what you told it yesterday if it's not in the file somewhere. And to be quite honest with you, the same thing applies to humans. All right. Like AI agents, it's not optimal for them to be this way. It's it's survival. They have to do this. And the agents that work the best in my systems, which now are all of them because they're all wired this way, are the ones with the thickest daily logs and the most lessons learned because every single session they wake up smarter than the last one. And then everything starts to compound. And this becomes a video game, so to speak, where how good can my agents become if I keep compounding all these things and utilizing them daily? Well, now they're going to become so smart and so good over the course of time that they're going to be very hard to beat.
Now, the third thing that I want to talk about is the QMD and Lossless Claw. Okay, these are plugins that people are starting to talk about now. Okay, and honestly, most people really don't know they exist until recently, but these are absolute game changers. So, remember what I just said about memory, okay? Remember that daily logs, memory, the lessons.mmd. Well, they're they're all very very important. Like, don't get me wrong, okay? They are. But here's the problem that nobody warns you about or warned me about, which is your agents write all these files. It builds up this incredibly long and big history. Like, we're talking dozens of daily logs, 50 plus lessons, detailed memory files across multiple workspaces, all of this. And then your your agent can't even find any of it. You ask it to go find it, it can't, right? Because OpenClaw out of the box is not built this way. Your agent can only see what's in its current context window. And that window is very limited. But when it fills up, the system compacts and it summarizes older messages to make room. And then what happens? Well, now your agents rules, its lessons, its personality directives, all of this gets compressed into just vague memory and a vague summary. And the agent starts operating on vibes instead of rules. And that's when things break and you need to know how to mitigate it. And these two plugins are exactly what changed everything for me and help me mitigate this issue from happening or when I want to recall on something I did in a previous day.
Now QMD is a very important plugin and it stands for query memory documents. Okay. Now this gives your agent a semantic search across all of its memory files. Instead of hoping the right information is in the current context, your agent can actually search its own history. It's 10,000 times better. Trust me. Like it has made a night and day difference. Now, let's say for instance that you need to figure out something that you were doing like a couple days ago or a week ago and you can't remember what it is. All right. Now, QMD will find it for you. And if you need to pull up what happened on like a specific day of the week last week, you'll be able to find it through QMD. Now, I have QMD indexed across all my agents workspaces. Every daily log, every mmemory MD, every lessons MD, all of this is searchable. My agent doesn't have to remember where something is. It just searches and finds it. That is what makes this plugin so so handy. You can recall anything to the agent and it will be able to find it.
Now, lossless claw is this is like context management here. Okay, so when your agent's context compacts, lossless call prevents critical information and important information from getting lost, your hardwired rules, your lessons, your core directives, right? They stay intact even when the context window is like completely cleaning house. All right. Now, without this compaction is like your agent getting partial amnesia. I'm not even going to lie to you. But with it, the important stuff that you need to know or that your agent needs to know survives. Now, think of it this way, okay? Your memory system from the second point that I made, that's your agent writing things down. QMD and lossless claw are your agent being able to actually find and use what it wrote down. Without one, the other isn't is incomplete. It's like gas and a battery on a car, unless you drive a Tesla. Uh, but without these plugins, your agent is it has a fragile memory. Like his memor is fragile and it writes things down, but it can't search through them. It has rules, but compaction just completely wipes out wipes from from the memory. But with QMD and lossless claw, like your agent can search its own brain, and it never loses context. And that's the difference between an agent that forgets lessons and has to be retaught and one that compounds knowledge over time and becomes better and better day over day over day. Now, most people in the open claw space aren't talking about these plugins too much yet. more so QMD uh they're talking about but not lossless claw but trust me once you set them up you will wonder how you ever ran without them and trust me when I was doing this I'm like this is a gamecher absolute gamecher so you're not going to want to ignore those are super important.
Now the fourth thing here is going to be using Discord as your agents command center okay and this is one that I see a lot of people talk about well maybe a decent amount of people but it genuinely for me it changed how everything works. Okay, so I use Discord as my agent's command center and I choose that over Slack because at the end of the day, Slack costs money and you know this doesn't. Discord is free. But here's what I mean. I set up a Discord server. I created channels for every business function and for every single agent. And then I connected all my agents to it and now they communicate through Discord channel using mentions. And let me tell you, like the moment that I did this, everything was just much easier because I was able to have my agents start communicating with each other and having them do all of these things without me even having to, you know, step in, right? Everything was just getting done in Discord as it needed to be. All right, but before Discord, like my agents operated in complete isolation, right? Maggie is in my Telegram. You know, my other agent, Mercury's doing her thing, Victor's doing contract stuff, but they didn't have the ability to talk to each other. And if Mercury hit a problem that was outside of her scope, she had nowhere to escalate these issues to to have someone fix him. And Victor was like finished processing a contract, Machaveli had no clue that he was done until I told him, "Hey, Victor's done. He finished the task." You know, now Mercury finishes analyzing a client's message and she posts it in her channel uh right away as soon as it's done. And if something needs an escalation, she mentions Machaveli directly. All right. Now, I also have a QA agent, which his name is Sunny, and he monitors everything and flags issues in a systems health channel. So, when he finds something, he pings Machaveli. Machaveli will go and investigate the situation that Sunny escalated, and then they go back and forth in a threat until it's completely resolved. And I can see the entire conversation. So, there's a full paper trail here. And it's actually amazing, amazing to watch this happen and to watch these agents go back and forth legitimately talk to each other and get this stuff done and actually fix problems in real time.
Here's what makes this even more powerful, and that's sub agents. So, Mchaveli doesn't just coordinate and per like the permanent team. He can also spawn in temporary sub agents for like one-off tasks right through Discord. So he could say go research this competitor or go audit all the cron jobs or go investigate this error sunny flag so that way Machaveli could stay free right the sub agent will then spin up do the work and reports back in the thread and then it just completely goes away and gets dismissed. So we have multiple agents coordinating through Discord channels in real time and that's when it stops feeling like a chatbot and it feels way more like you're running an actual team and when you actually build it it's it's actually quite liberating. I'm not going to lie to you, right? My Discord server has over 30 channels right now, right? Organized by business and by its function. And there's channels for each agent's direct line. That's what we call them, direct line. It's essentially like private just me and that agent and Machavel because he needs to see everything. And then there's channels for reports, for system health, for recursive learning where like agents can log what they've learned. There's business specific channels where like agents that service a specific business that is a client of mine and they post the work and they give me all the reports. It's literally like having a virtual office at this point. And I l I say that very literally. Like you can walk into any channel and you can see exactly what that department has been doing, how it's been doing, what's getting done, what decisions were made, what problems came up, how they were actually resolved, and honestly everything needs to be documented like this because it's well, it's how you can see where gaps in the system are. All right? Has to be done if you want to optimize for maximum effectiveness and efficiency. And the rule I always enforce is every message between agents must include like a mention so they could talk to each other. And without because without it, they literally never see the message and the conversation will just die or never start. And I trust me, I learned that the hard way as well. I'm not going to lie to you because like my agents were, you know, talking past each other for an entire day before I realized like, hey, no one's even responding to each other, right? Like I was busy working and then I check and see that they're not even talking to each other. They're just stating stuff. And, you know, that was an error I had to fix. But like if you're running more than two or three agents at this point, like you need to set up a Discord. Like this is not optional anymore. It's absolute infrastructure that needs to be implemented into your business. I can't stress that enough.
Now, the fifth thing, and I saved this one for last because honestly, this thing completely separates this agent from being a chatbot to a real AI agent. Now, I talk about this inside of my OpenClaw complete setup guide. So again, you should check that out because I show you a lot of actually how I get it to think like me and and all of this stuff. But you know, no one in the open claw space is really talking about this too heavily and putting a lot of emphasis on this right now. Most people, the average person, they give their agent tasks. Do this, send that, process these things, do this for me, right? And the agent does it, which is great. But the thing is is that this is just task execution at this point. Like it's not effective, right? Your agent just follows instructions. It's not actually thinking like you. It's just a chatbot at the end of the day.
So I started to do something different when I actually was getting annoyed with that is I started asking Makaveli to explain his thinking process. Not just what he decided but how he actually came to that decide that decision. Okay? What options did he consider? What did he weigh out? What trade-offs did he evaluate? What would he have done differently? Right? You need to ask your agent these questions and reverse prompt it. And then the key part is is I logged those patterns into a dedicated file for him. Very important. And to be honest with you, like I realized that every time I was correcting my agent, I was really teaching him my decision-making framework. I just wasn't explaining it to him, right? My way of thinking through problems, the questions that I ask, like, okay, what are all the options? Let's play each option out to its entirety. What does this indirectly affect if I make this decision? Am I trading quality for speed or would Andrew have to think about this after I'm done? Right? Very important. You need to you have to think like this. So instead of just correcting the mistakes one time, I I mapped out my entire framework of how I think and put it in a file. My agent reads it every single session. And now when he faces new situations, something he's never even seen or done before, he doesn't just guess, right? He runs it through my exact framework. He thinks exactly like me. He asks the questions I would ask. He weighs out the tradeoffs that I would weigh out if I was making the decision myself. So now I'm training my AI agent, Machaveli, to think like a CEO, not just be a task rabbit who just does what I say and gets it done without having to think. And the results are night and day difference. All right. Mchaveli just used to bring me problems. But now he brings me a problem with a bunch of solutions with the root cause identified all the options mapped out indirect effects an analyzed and and considered and the recommended path with the trade-offs already laid out. So my only job is to be like yeah that sounds good do it or no here's actually what you're missing. You need to look at this because this is where the gaps are and then go ahead and fix it. Completely different. And think about what this means. Okay. Instead of you being the operator who makes every single decision, your agent now becomes the partner who thinks through problems using your judgment you trained him on. It's like cloning your best instincts, your mental framework, and your decision-making tree logic into a system that runs 247 365. And the beautiful part about this is is this is a compounding effect. Okay? The more you do this, the better the agent gets. Every correction refineses the framework, every decision adds to the pattern library that you're building for it. And after three weeks, Machaveli is autonomous at making decisions now, right? He can just make decisions autonomously. And they're better than some of my human employees that I've had, all right, and worked with. I genuinely mean that, right? Like this is the difference between people who use AI as a tool and people like me who use AI to build genuine operating partners that you would consider to legit be a legitimate partner, right? You're not just telling it what to do. You're teaching it how to think. And once it thinks like you, that's when things get really interesting, but also get very, very dangerous because that completely and utterly frees up your time to go do higher leveraged activities that actually generates you more money or whatever it is you're looking for. Okay?
Now, look, again, I've spent over 500 hours and spent two to two and a half thousand, probably more at this point now on API cost learning all of this. Okay? And these five lessons, they would have saved me weeks of work, weeks, hundreds of hours of work, okay? And probably honestly probably a thousand-ish dollars and wasted cost of just, you know, going through this stuff. And the bonus tip that ties all this together is when you're building out these systems, test each step individually before you chain them together. Okay? I learned this the hard way when I set up my entire five agent content team. Okay? I connected them all through Discord. I let them run and then they got like this infinite loop of cheerleading for each other. It was like it was a absolute chaotic mess, okay? Just complimenting each other, going back and forth, just burning tokens and doing absolutely nothing for hours. And if I tested each agent alone first, this would have never happened. And I wanted to tell you that because you're going to have a team at some point, ideally, and you need to know this, okay? So, you need to stage gate it. All right? Build one step, test it, perfect it, and then connect the next piece to it. And if you're serious about building AI agents that actually run your business, not just as a toy, but for like real operations, subscribe because I'm documenting this entire journey at this point. The wins, the failures, all of my numbers from my AI agency, everything is getting documented. So, go ahead and hit subscribe, hit the bell notification, drop a comment down below if you have any questions about this and anything you want to see on this channel, and smash the like button because I hope you gain some value from this. And I will see you guys in the next.