0:02 OpenClaw is the most powerful AI agent
0:04 framework in the world right now and
0:06 it's about to replace your entire
0:08 workflow. While everyone else is still
0:10 prompting chat GPT one message at a
0:12 time, you could be running a specialized
0:14 team of AI agents that work for you
0:17 24/7, remember everything, and cost
0:19 significantly less than a full-time
0:21 employee to operate. The problem is
0:23 there are a lot of tutorials out there
0:25 showing you how to set up Open Claw on a
0:27 Mac Mini or VPS, but nobody's showing
0:29 you how to actually get it to do
0:32 anything. I spent over $200 in the last
0:34 48 hours stress testing the system so
0:36 you don't have to. In this video, I'm
0:37 going to show you how to set up the chat
0:39 model so you don't break the bank, how
0:41 to deploy a self-healing team of AI
0:43 agents that communicate with each other,
0:44 and how to set up the infrastructure
0:47 that lets your AI run on autopilot. And
0:48 if we haven't met yet, my name is Duncan
0:50 Rogoff. I'm a former art director for
0:52 brands like Apple, PlayStation, and
0:54 Nissan. And I now run a six-figure AI
0:56 agency. In under 12 months, I've grown a
0:58 following of over 110,000 people, and I
1:00 lead a community called the Buildroom
1:02 where more than 2,000 are already using
1:04 AI to grow their audience and generate
1:06 leads for their business in under 3
1:08 hours per week. So, focus in, close all
1:10 your tabs, and let's build. So, here we
1:12 are inside of OpenClaw, and it's not
1:15 always the sexiest to look at, but today
1:17 I really want to cover like the most
1:19 impactful things that I've set up that
1:20 actually help me run my business. These
1:22 are a couple things like communicating
1:24 with it, using Telegram, setting up
1:26 different LLM models to handle different
1:28 tasks so I can save on cost, like using
1:30 cron job to just tackle tasks like a
1:32 daily morning briefing or saving this to
1:34 GitHub or things like that. And then a
1:36 breakdown of how I've set up my AI agent orchestras
1:38 orchestras
1:39 we have the orchestrator agent that
1:41 communicates with all the other agents
1:42 which really allows us to be as
1:44 efficient as we possibly can be. I'll
1:46 probably bounce back and forth between
1:48 openclaw and some slides I have prepared
1:50 because like I said openclaw isn't the
1:51 prettiest to look at, but I do want you
1:52 to understand where everything lives.
1:55 Now when you first get into openclaw, it
1:57 is essentially a baby. It knows nothing.
1:59 So it is really up to you to configure
2:00 this the right way and there's not a lot
2:02 of information out there on how to do
2:03 it. So there are a couple things that
2:05 I've learned that have actually been
2:07 really really impactful for me. So the
2:09 very first thing that is really helpful
2:11 is actually setting up different LLM
2:14 models for different tasks. So let me
2:15 just show you what I mean. Can you tell
2:18 me which LLM models we are using and for
2:20 what tasks? So now I can just
2:23 communicate with Open Claw directly here
2:25 in the chat in the browser and basically
2:27 it's showing me that now I've actually
2:29 switched between a couple of different
2:31 models depending on the complexity.
2:33 Right? So for here we can say the
2:34 default for everyday task we're using
2:37 Gemini 3 or Flash preview which like
2:39 gives me a 90% cost reduction. So when
2:41 I'm just chatting back and forth with
2:43 the agent itself it's using Gemini which
2:45 is super super cheap to use. Then from
2:48 there for any of the content sub agents
2:50 we're using like a step up model Claude
2:53 Sonnet 4.5 which is like a really great
2:55 model we're using for lead magnet and
2:56 writing and our LinkedIn post because
2:59 Claude is super super good at writing.
3:00 And then for the most expert layer,
3:02 we're switching dynamically to Claude
3:04 Opus 4.6, which is the latest anthropic
3:06 model. It's the most expensive to use,
3:07 but it's the best at complex reasoning
3:10 or complex tasks like coding or
3:11 automation builds or any sort of
3:13 multi-step reasoning. And so basically,
3:16 you can just talk to OpenClaw like it's
3:18 a smart assistant and tell it to figure
3:20 it out. Like you can just say to it
3:21 things like, "Hey, I want you to set up
3:23 different LLM models for different
3:25 tasks. Use Gemini for this, Sonnet for
3:26 that, and Claude Opus for this." And
3:28 then that way you can actually save on
3:30 cost instead of relying on heavier and
3:31 expensive models for everything. One
3:33 other super valuable setup hack is
3:35 switching between your Claude Max plan
3:38 and the API like dynamically and
3:40 intelligently. Can you tell me how we
3:41 are intelligently switching between the
3:44 Cloud Max and the API? So basically
3:45 there has been like a lot of chatter
3:48 right now about how Quad or Anthropic is
3:50 banning people who are using the Claude
3:52 Max subscription in OpenClaw because
3:54 it's just kind of like racking up the
3:56 credit so to speak. But we basically are
3:58 able to integrate a system that will
4:00 dynamically switch between the two. So
4:02 as you start to hit rate limits on your
4:04 Quad Max plan, it'll dynamically switch
4:06 over to the API when needed. So this is
4:07 essentially how it's doing it.
4:09 Basically, it just uses the Quad Max
4:11 plan first. Right now, I'm just paying
4:13 $100 a month cuz I'm not ramping this up
4:14 like crazy. And then once you start to
4:16 kind of like hit those rate limits,
4:17 it'll switch over to the API because the
4:19 API does end up actually being really
4:21 expensive if you're only using that. If
4:23 you need to know how to get access to
4:25 your max plan on OpenClaw, literally
4:27 just ask. It walked me through the setup
4:29 and how to get my Quad Max API key in
4:31 like two seconds. The next is a
4:33 directive that I actually got from a
4:35 friend. Can you tell me about the figure
4:37 it out directive? So basically, this is
4:40 just like an instruction to OpenClaw to
4:42 basically just like be smart, figure it
4:44 out. If I ask you to do something, just
4:46 like handle it. So I don't have to have
4:48 this constant like back and forth with
4:50 OpenClaw to actually execute on
4:52 something. You can see here it's the
4:54 core operator philosophy. Basically, I
4:56 can't is not vocabulary. If I don't know
4:58 something, learn it now. So, this way
5:01 I'm not constantly having to like baby
5:02 open claw. It's just going to go out if
5:04 it doesn't know how to solve a problem.
5:06 It's going to research the internet and
5:07 figure out how to solve something. It's
5:09 going to search for docs or tutorials.
5:10 It's going to reverse engineer stuff.
5:12 It's going to look at APIs, right? It's
5:14 going to try multiple approaches to
5:16 actually solve the problem first before
5:18 coming to me and asking me questions.
5:20 That way I don't have to hold his hand
5:21 through everything. So if you want just
5:23 pause this, take a screenshot here, drop
5:25 it into openclaw and say, "Hey, like
5:27 this is how I want you to behave." This
5:29 has actually saved me a lot of time. So
5:30 now the next thing to understand is
5:32 basically the agents infrastructure.
5:35 This is kind of like having OpenClaw act
5:37 as like a different team of sub agents.
5:38 And this will allow you to save on costs
5:40 and be as efficient as possible. Cuz
5:41 what I was finding is that if you're
5:43 just doing everything with a single
5:45 agent in the main chat, it gets really
5:47 bloated really quickly. it tends to
5:49 drift or hallucinate. You're constantly
5:51 having to like restart the chat new. Oh,
5:53 one thing that's also helpful in the
5:55 chat. If you just type slash new, you
5:57 can basically start a new session. It'll
5:58 clear the memory or the cache and
6:00 basically gets rid of all the previous
6:02 context. So that way you don't have kind
6:03 of that bloat and that drift happening.
6:05 There is a really clear sort of
6:06 orchestrator pattern that's been
6:08 developed with all of these AI agents in
6:09 general where you have kind of like the
6:11 orchestrator at the top level and then
6:13 it speaks to all of the other agents
6:14 beneath it and tells them how to
6:17 communicate with each other. I put
6:19 together this little guide about the
6:20 ultimate guide to just multi- aent
6:22 systems like how to set these up, how to
6:24 design like a sub agent, especially for
6:27 OpenClaw using the soul.md file. I'll
6:28 leave a link in the description where
6:30 you can just get that totally free so
6:32 you can have access to this. The most
6:34 important thing inside of these agents
6:36 to pay attention to is these files. So,
6:38 there's an agent soul tools file. I'm
6:39 going to walk you through what each one
6:41 of those is so you understand. And
6:43 basically, I found that not every agent
6:46 needs all of these files. Basically, you
6:48 have me. I just have my Duncan agent as
6:50 the main agent. And this has access to
6:52 everything. And then it just essentially
6:55 provides all of the sub agents with the
6:56 context that they need and nothing else.
6:58 So again, you're not sending extra
7:01 tokens to the LLMs to process. You're
7:03 only sending exactly what it needs. So
7:06 first is the soul.md file. This is the
7:07 brain. This is the agents identity and
7:09 the core instructions. It defines who it
7:11 is, how it behaves, and crucial
7:13 behavioral boundaries. So like what to
7:15 do and what not to do. So if we come in
7:17 here to the solemn MD file here, this is
7:19 just for my main agent, right? You're a
7:21 Duncan's orchestrator agent. You do not
7:23 execute tasks yourself. You route them.
7:26 So you're a task router and coordinator.
7:27 You basically need to identify which sub
7:29 agent should handle the task. You should
7:31 spawn that sub agent with a clear
7:32 complete task description and then
7:34 report back the results when they
7:36 arrive. So again, it's this two-way
7:37 communication between the orchestrator
7:39 and then the agents that live beneath
7:40 it. These are the tasks that we expect
7:42 it to delegate. And then these are the
7:43 kind of tasks that we expect the agent
7:45 just to handle itself without having to
7:47 call a sub agent. Inside the community,
7:50 I do have this AI agent dream team stack
7:51 which comes set with all of these
7:53 prompts for an orchestrator, a
7:54 researcher, writer, chief of staff, and
7:56 a builder to do code and things for you.
7:58 It comes with a prompt and so basically
7:59 you could just come down here and copy
8:02 this entire prompt into openclaw and
8:03 it'll set up this entire agent
8:05 architecture for you. And so the next
8:07 file to understand is this agents.md
8:09 file. This is the directory of all of
8:11 the sub aents that we create. And so it
8:13 lists each of the sub aents and the
8:15 specific roles and their strength. So
8:17 this is an example of the type of file
8:19 that only your orchestrator agent needs.
8:21 It basically goes ahead and it just
8:23 describes each of the sub aents. And so
8:25 basically the orchestrator needs to
8:27 understand what each of the sub aents
8:28 are like something that creates
8:30 Instagram and Tik Tok carousels,
8:31 something that writes Twitter threads,
8:33 something that generates images or like
8:35 basically creates lead magnets for me,
8:36 right? So it's really clear what each of
8:38 the sub aents do. So the orchestrator
8:40 understands which agents to call for
8:42 what task. But this is a really good
8:44 example like inside of like my carousel
8:47 creator, there is no agents.md file
8:48 because the carousel creator doesn't
8:51 need to understand the agents around it
8:52 because that's a job for the
8:54 orchestrator. So we're not sending
8:56 additional context that the sub agent
8:57 doesn't need. The next file you need to
8:59 understand is this tools file. And this
9:01 is just basically anything that your
9:04 agent needs to execute on its task. So
9:06 it doesn't need access to things like
9:08 GitHub or NADN. For me, I gave the
9:10 carousel creator access to like images
9:12 of me so that it knows what I look like.
9:14 So therefore can pass my image to Nano
9:16 Banana to generate images. And then any
9:18 information about the process itself,
9:20 right? Do we have a specific structure
9:22 for in my case like this set of slides,
9:24 right? And then here I give it access to
9:26 notion so that it can connect to notion.
9:27 And once it's done creating a carousel,
9:29 it can basically save it inside of
9:31 notion for me so that I can have access
9:33 to it and review it. There is a file in
9:34 there called identity MD. I don't really
9:36 use it too much, but if you want your
9:38 agent to have like a specific identity
9:40 or behave a certain way or have certain
9:42 properties or characteristics, this is
9:44 where you would create that information.
9:45 And again, you can always tell OpenClaw
9:48 to create it for itself. The user.md
9:49 file is everything about you, the
9:52 operator, the founder, whatever it is,
9:53 right? So, this is everything about you.
9:55 So, the agent will read this to
9:56 understand like how it should
9:57 communicate or how it should behave or
9:59 how it should write or things like that.
10:00 And the last one that is arguably one of
10:03 the most important is memory.md. This is
10:05 like the long-term knowledge. So this
10:07 will actually get updated automatically
10:09 by openclaw the more you interact with
10:11 the system and the more information it
10:12 has about you and your processes and the
10:14 way that you like to work. So we can see
10:16 here that this has a ton and ton of
10:18 information about me like the technical
10:19 stack that I like to use some
10:21 information about my philosophy and
10:22 things like that who I collaborate with
10:24 all of my social proof how I like to
10:26 operate all of these things. These are
10:28 super super detailed. But if we come
10:29 down here like into the carousel, you
10:32 can see I don't have the memory in here
10:33 because I'm leaving it up to the
10:36 orchestrator to decide which pieces of
10:38 information should get passed to any of
10:40 the sub agents. So again, we're not
10:41 overloading the system. If you're trying
10:43 to figure out how you can create all of
10:45 that information easily, I have this
10:47 great prompt which is to create a data
10:49 packet of information about yourself. If
10:51 you've been talking with Claude or if
10:52 you've been talking with Chatbt, you
10:54 would just come over to one of those
10:56 LLMs and paste this in and you're just
10:59 going to get basically a giant PDF full
11:00 of information about your business, the
11:02 projects you've worked on, your goals,
11:03 your communication style, and things
11:05 like that. And you can just copy and
11:07 paste that information into OpenCL and
11:08 say, "Hey, here's everything you need to
11:10 know about me. What should we do with
11:12 this?" So setting up these agents the
11:14 right way in the beginning is extremely
11:16 impactful to the efficiency of your
11:18 system and it's going to save you so
11:20 much headache in the long run. Before I
11:21 show you the jobs that I have OpenClaw
11:23 doing for me automatically, one of the
11:25 first things you should do is set it up
11:26 to communicate with you via Telegram or
11:28 WhatsApp. So that way you can take this
11:30 on the go. So I was actually out this
11:33 morning and I was texting with OpenClaw
11:34 and I was asking it to create that kind
11:37 of AI dream team starter pack for all of
11:39 you guys. So basically, you can see here
11:41 I just had it say, can you create one
11:42 thorough master prompt that would
11:44 instantly deploy and set up a team of
11:46 powerful agents for anyone? And then
11:48 said, hey, I love this idea. A starter
11:49 team can deploy. Here are the four sub
11:51 agents, the researcher, the writer, the
11:53 chief of staff, and the builder. They do
11:54 all of this. And basically what it went
11:56 ahead and did is it created this entire
11:58 document for me in notion. I didn't
12:00 write this. I didn't type this in. It
12:02 figured it out. It organized this. It
12:04 did all the design. It created this
12:06 prompt for me and for you guys. It's
12:08 just a massive, massive timesaver. So,
12:10 the other way I like to use this for my
12:11 business is there are a couple of what
12:13 are called cron jobs and they're called
12:14 cron because it has to do with, you
12:16 know, chronological or time. These are
12:18 basically just jobs that will run on a
12:19 schedule. And this is generally like
12:21 pretty ugly to look at, but you can just
12:23 talk to openclaw and say, "Hey, I need
12:25 something that does xyz." One of the
12:27 first things I had it build was this
12:29 idea of a morning briefing. So, every
12:30 day when I wake up, I actually have a
12:33 message inside of Telegram. It looks
12:35 like this. It basically gives me
12:36 yesterday's progress like what we
12:38 accomplished yesterday, what we have to
12:40 prioritize today, anything like any
12:42 pending decisions, like any feedback
12:43 that it needs from me and then it
12:45 basically just sends this to Telegram.
12:46 So you can see here that it says here's
12:48 the top three today. I should post a new
12:50 carousel to Tik Tok or Instagram. I
12:52 should write a new Twitter thread and I
12:53 should basically just test the setup
12:55 guide end to end. Here are the tasks.
12:57 These are open clause tasks. It's going
12:59 to monitor Twitter for any auto replies.
13:01 And so basically if I have lead magnet
13:02 posts on Twitter, it's going to
13:04 recognize when somebody like comments
13:06 the word AI and it's automatically going
13:08 to DM them for me and send them whatever
13:10 the lead magnet is. And so here the
13:11 quick decision section is basically
13:13 feedback that it needs for me like yes
13:15 or no. Do I want to add a hosting or
13:16 affiliate link to my setup guide? Like
13:18 yeah, for sure let's do that. Do I want
13:20 to ship a Twitter thread today or do I
13:21 want to wait for something else to be
13:23 done? Right? And then I also had it just
13:25 like give me ideas like based off the
13:26 conversations we've been having, what
13:28 we've been building, like here are some
13:29 growth ideas to help the business grow,
13:31 help the community grow, all of those
13:33 good things. So, do I want to create a
13:35 quality control miniourse for the build
13:36 room? Do I want to turn the open claw
13:38 setup into a YouTube tutorial series
13:40 like we're doing right now? Do I want to
13:41 cross-promote the comment content
13:43 engine, which is basically like a little
13:45 micro SAS app I built the other day?
13:46 Pretty cool. So, this will just run
13:48 every day at 7 a.m. So, if there's like
13:50 a structure that you want for this, just
13:52 tell Open Call, hey, I want a daily
13:54 morning briefing at 7 a.m. Send me these
13:56 things. On the flip side, I had it
13:58 actually create this midnight daily
13:59 tracker. It's basically just going to
14:01 review what we did for that day, and
14:02 it's just going to log everything in
14:04 notion for me. So, that way I just have
14:06 something to reflect back on. I can use
14:07 this for content. I can use this for
14:08 marketing. It's just for my kind of
14:10 keeping to understand what we've done.
14:12 So you can see here we can see like
14:13 here's what we accomplished like cost
14:15 optimization strategy for Tik Tok
14:17 carousels, content brief, here are the
14:18 key decisions that we made about like
14:20 aesthetics and my brand, right? And
14:22 here's what's kind of in progress or
14:23 what's blocked. Here are all the skills
14:25 and files that we created. And so again,
14:26 this is just a way for me to keep track
14:28 of everything that we've been building.
14:29 This one's honestly a lifesaver. You can
14:31 just create a repository on GitHub and
14:33 you can just say, "Hey, OpenCaw, I need
14:35 you to back up all of our code into
14:37 GitHub every day." And so this way all
14:39 of your configuration files, your memory
14:41 scripts, everything is just pushed to
14:42 GitHub. So that way if anything breaks
14:44 or you're just running into problems,
14:46 you can instantly roll it back to the
14:48 repository. So to a working state. If
14:50 you want access to my AI Dream Team
14:52 prompt library and autopilot stack, it's
14:54 inside the buildroom. All you have to do
14:55 is come into the lesson, open this up,
14:57 and I give you ready to deploy, copy,
14:59 and paste prompts so you can set up all
15:01 of these agents and cron jobs just like
15:03 I have. All right, back to it. And
15:05 arguably the most impactful cron job I
15:08 have set to run every week. It is this
15:10 weekly trends analysis report. It's
15:12 basically going to look at Reddit,
15:14 YouTube, and X to figure out what people
15:15 are actually talking about based off of
15:17 kind of some keywords, my audience, my
15:19 ICP, their pain points, and things like
15:21 that. And so gives me like a top pick
15:23 like content idea number one. I gave
15:25 Open Claw full access to my consulting
15:27 business. Basically, OpenClaw is the
15:29 viral moment right now, which is why I'm
15:31 talking about it in this video. It gives
15:33 me a full execution strategy. It gives
15:35 me more content ideas based off of kind
15:37 of the findings from the different
15:39 channels. Gives me some opportunities to
15:41 create different lead magnets for my
15:42 audience. So again, I can send this
15:45 information to my lead magnet sub agent
15:46 to create them for me. And it's going to
15:48 go ahead and rank all this content for
15:50 me. It's going to break it down by
15:51 different platforms like different
15:53 platforms, LinkedIn posts, YouTube
15:55 videos, short form videos, any like
15:57 audience language that I can like kind
15:59 of pull in, right? any pain points and
16:01 objections that I can really hit on so
16:02 that I'm speaking directly to my
16:04 audience's frustrations. And so this is
16:06 probably one of the most impactful
16:08 things that I have created. And then at
16:09 the bottom, it just kind of links to all
16:11 the sources that it pulled from. If you
16:12 want to get access to any of the prompts
16:14 that we use today or learn how to build
16:16 a highly profitable personal brand using
16:18 AI, just check the link in the
16:19 description to join thousands of other
16:21 people inside the buildroom. If you
16:23 think open claw is cool, just check out
16:26 this video up here of 63 insane use
