// transcript — 541 segments
0:00 Intro
0:01 I started using Claudebot as an experiment. It has quickly become one of
0:06 the most important pieces of technology in my life. And if you haven't heard of
0:10 Claudebot, it has been going absolutely viral. It is the most personal and
0:15 capable AI assistant that I have ever seen. It allows you to get real tasks
0:19 done by connecting the different services that you use like Gmail,
0:23 Telegram, Asauna, Slack. Everything can be connected together and you just chat
0:27 with your personal assistant with your Claudebot directly through whatever chat
0:31 interface that you want to use. I personally choose Telegram. Cloudbot
0:35 learns about you and continuously gets better. It is truly incredible. But a
0:40 lot of people still don't know how to use it, and that's what I'm going to
0:43 show you today. By the end of this video, you will become a Cladbot expert.
0:47 But before we can actually use Claudebot, we have to get it set up. And
0:51 there are a number of different ways to do that. Installing it on your local
0:55 machine, using external hardware, or what we're going to be doing today,
0:59 using a virtual private server. And we're going to be using Hostinger to do
1:03 it, who is also the sponsor of today's video. So, huge shout out to them. And
1:07 it is basically one click to get it up and running on Hostinger. So, the
1:11 benefits of using a VPS are super simple to set up. It is always on and it is
1:17 secure because it is completely isolated from any of your personal devices. And
1:22 Hostinger is offering 10% off their already cheap VPS service with one-click
1:27 Cloudbot install. All you have to do is use the code MatthewB and go to
1:30 hostinger.com/matthewb. I'll drop everything in the description
1:33 below. All right. And getting it set up with Hostinger is dead simple. You just
1:36 come to this page. I'll drop the link down below. And all you have to do is
1:40 click deploy. On this page, you can select your time duration. And this is
1:42 where you're actually going to purchase the VPS. Right here, you can select 1
1:46 month, 12 month, 24 month. I'll start with 24 months. And right over here
1:50 where it says have a coupon code. Yes, we do. Matthew B. Apply. And there you
1:56 go. You get 10% off. Now continue. And there you go. Payment done. Then it's
1:59 going to bring you to this configuration page. You could put your API key from
2:04 Anthropic, OpenAI, Gemini, XAI, anything else. And then click deploy. And then
2:08 you can connect everything else you need by opening up terminal and going through
2:12 the open claw onboarding. And of course all of this is optional. And here you
2:16 can see we're good to go in Telegram. I said hello and my clawbot says hey
2:21 what's up. All right. So the first thing to really understand are all the
2:25 different MD files that you have that really form what this Cloudbot is for
2:31 you. And there are a bunch heartbeat identity memory soul tools and user. So
2:36 this is the soul.md file. This is basically defining what the personality
2:40 of your Claudebot is going to be. Feel free to come in here and edit as you see
2:44 fit. Then we have the skills. This is really what makes Cladbot so incredibly
2:49 capable. This is where you give it all the skills that help it browse the web,
2:54 check your email, access Twitter, and all of your custom skills as well. So,
2:58 anything that you tell Cloudbot to do that it doesn't already know how to do,
3:02 it's going to create a skill from that and save it here. You can think of a
3:06 skill as a repeatable process. Anything where you have multiple steps and you're
3:11 using different tools along the way. And so that's defined here. So we have the
3:15 humanizer skill for example and a whole bunch more in here. And the skills
3:21 reference tools. And tools are pieces of code that allow it to actually
3:24 accomplish those things that the skills want to do. So here's an example. Here's
3:28 fetch.js. And it's basically a JS file that allows my agent to plug into ASA really easily
3:36 and grab information from my ASA board. And again, you don't need to write any
3:40 of this. It's all done through natural language through Telegram or WhatsApp or
3:45 wherever you're using your Claudebot. Then we have the identity file and it's
3:50 very similar to Soul in a lot of ways. It basically tells Claudebot how it's
3:54 actually going to interact with you from a tone and emojis perspective. So we see
4:01 name, Claude, vibe, emoji, avatar, and so that's what's in the identity file.
4:05 Cloudbot now has a memory folder in which it stores all of its memories
4:09 about you. And you can just let it store everything. You can always go in there
4:13 and prune things that you don't think it needs to know anymore. But for the most
4:16 part, there's no harm in just having it store most of what it learns about you.
4:22 Then we have the heartbeat file, which basically runs every so often. I think
4:26 by default it's every 30 minutes but it's whatever you want to have happen on
4:31 a recurring basis but we also have cron jobs and I'll get to that in a moment
4:34 and again this entire cloudbot project will grow with you so as you figure out
4:39 ways that you want it to run you just tell it and it will start to learn to do
4:42 that. All right so those are all the basics I want to get into a little bit
4:45 more advanced stuff now. All right so next I want to talk about model
4:48 selection because that's probably one of the biggest open questions that people
4:52 have right now is what model should I be using? What should be the primary model?
4:55 What should be the fallback model? Should I use multiple models? Do I use a
5:01 local model? And the answer is yes, all of it. And I've set up my model routing
5:05 to be pretty sophisticated. And I want to show you. And the way I want to show
5:09 you is by asking my cloudbot. So I'm going to say, show me your model
5:12 selection logic hierarchy. So it's going to read from its own files and show me
5:16 right now. So the primary model I'm using right now is Cloud Sonnet 4.5. And
5:21 that is because it is a workhorse. It is much cheaper than Opus. And so for the
5:26 default interaction through chat, that is what I'm using. But I also have a
5:32 fallback chain. I use Gemini 3 Flash, which is incredibly fast and cheap. I
5:37 have Opus 4.5 here. I have Open Router that I'm using to route to other models
5:41 that maybe aren't on my initial list. And I was even running local models, but
5:46 I'm not doing that in this moment. And so you can have local models running for
5:51 things like cron jobs and other really basic tasks that don't require frontier
5:56 models, unless of course you can run one of those frontier open source models
5:59 locally. And there are a few different ways that you can have it select which
6:04 model it's using. You can simply say in natural language, switch to Sonnet 4.5.
6:09 Hit enter and it's going to switch to that model for this chat. It will also
6:14 make a decision about which model it should use at any given time depending
6:18 on the task complexity. So we say done back on sonnet 45. You can also type
6:23 slashmodel and if you ever get disconnected or you run into a rate
6:27 limit on one of your models, use /model to show or set the different models that
6:32 you have. So the way I'm doing it is for anything coding, I'm usually using
6:36 sonnet, but if it's an especially complex coding task, I tell it to route
6:41 to opus 4.5. For most kind of basic interactions, I'm either using Sonnet.
6:45 Sometimes I'll switch to Haiku because it's even faster and much cheaper. And
6:50 of course, I'm experimenting with models all the time. And remember, the model
6:54 you select definitely affects the personality of your Claudebot. Now, if
6:57 you don't care about that, don't even worry about it. Just use any model
7:00 that's best for your use case. But if you do care about the actual
7:04 personality, you're probably going to get familiar with one model, and
7:08 switching is a little bit jarring. Now, Clawbot packages up everything about its
7:12 personality, all of your memories, and will send it to each of these models
7:16 individually, but the way that the models themselves handle all of that and
7:21 respond to you still will be different depending on the model. Okay, now I want
7:25 to get into skills and tools. This is really what makes Claudebot special.
7:29 This is the thing that sets it apart from any other tool that I've ever used.
7:33 It just so easily plugs into any service, any product that you're using.
7:38 I have it plugged into my Gmail, to my Drive, to my calendar, a sauna, Slack,
7:44 HubSpot. A majority of services that I use, I allow Claudebot to plug into.
7:49 Now, I'm fairly technical, so I really have thought about the security
7:53 implications about connecting all of these different services, and I'm doing
7:56 so in a secure way. I'm going to talk about security later in the video, so be
8:00 sure to stick around for that. But to plug it into a new service, it is
8:05 simple. All you have to do is tell it. That's it. It'll look for skills. If it
8:09 doesn't already have the skill, it will write one itself. You might have to grab
8:14 an API key from somewhere, but that's about it. You do it all through chat
8:19 with natural language. And you can even get very sophisticated with your model
8:24 selection. So, I'm a cursor user and I installed cursor agent in my hosting or
8:31 VPS and I'm actually able to control my cursor agent through Claudebot through
8:37 Telegram. So, anytime that I have a complex coding task, I'm telling my
8:42 Claudebot to delegate it out to my cursor agent and it does that and it's
8:47 super impressive. Now, the only thing to remember with that is that cursor agent
8:51 has basically no personality. So sometimes when it responds back to you
8:53 and Telegram that it's done or giving you an update, it basically strips all
8:57 the personality that you're familiar with. And again, it's a little bit
9:01 jarring, but feel free to plug in the tools that you're most comfortable with.
9:05 And I would recommend starting with more lowrisk tools. Now, if you want to
9:10 browse new skills or you want to get inspiration for things that your Cladbot
9:15 can do, go to clawhub.com. This is the official Cloudbot skill repository by
9:20 the Cladbot creator. So up here, click skills, and you can just see a ton of
9:24 different skills that are available. Now, keep in mind, these are written by
9:28 other people. There's always a chance there's some malicious code in there.
9:32 So, what I like to do is tell my Claudebot, go download this, but first,
9:37 scan it, make sure there's nothing malicious in there, and I want to always
9:42 use the best possible model to do that because the best possible model is going
9:46 to be less susceptible to prompt injection. Okay. Next, another core
9:51 mechanic of CloudBot is it being able to do things for you on a scheduled basis.
9:55 You can have it be on a recurring basis. You can do a one-off scheduled task. And
9:59 all of this is going to be done through cron jobs. And if you're not familiar
10:02 with that term, it basically just means a scheduled thing to do. And you simply
10:09 just tell it to do so. So, in 1 hour, remind me to drink water. I'm going to
10:12 hit enter. And then it's going to create a cron job for me. And 1 hour from now,
10:17 it's going to message me back again. and tell me to drink water. And these can
10:22 get as complicated as you want. So, for example, one thing that I did was I have
10:27 kind of a weird trash pickup schedule at my house. It happens every Monday and
10:32 every other week. It's some rotating recycling. Sometimes it's cans,
10:36 sometimes it's bottles, sometimes it's paper. And I could never remember what
10:41 it is. So, I actually just told my Cloudbot, "Here's the schedule." like
10:44 just took a picture of it from the recycling company's website, uploaded
10:48 it, and say, "Look at what the schedule is, and every Sunday, I want you to tell
10:52 me which cans to take out." And now it does that. It tells me exactly which
10:56 Groups (ADVANCED)
10:56 recycling to put out on Sunday. Okay. Now, I'm going to show you one of the
11:01 most powerful unlocks that I have found with Claudebot, and that is using
11:05 Telegram groups. Now, previously what I was doing is just DMing with my Cloudbot
11:10 directly. And that was fine, but we would just have infinite history of
11:14 conversation. But if I wanted to have multiple tasks running in parallel, it
11:18 was very difficult. I would say, "Okay, hold on to that idea. Let me come back
11:21 to that and let me switch back to this other task I had you doing." And it was
11:25 just so complicated. And then I figured out that you can actually use Telegram
11:30 groups. And with a Telegram group, you can set up different topics. And so you
11:33 can have a bunch of different conversations going in parallel with
11:37 your Claudebot. And that is beneficial in multiple ways. One, you can have a
11:44 very seamless conversation in each of the topic channels. And that allows you
11:49 and Claudebot to stay on topic and not get confused about other things going
11:54 on. That also saves on context window and memory because instead of loading up
11:59 the entire conversation history which might include days, weeks, even years in
12:04 the future of conversation history, it loads up only what is in that topic,
12:10 saving you room in the context window. So here's an example. I have one for
12:14 video research. I have Twitter research. I have an ebook I'm putting together. We
12:17 have content analysis. We have a cloud skills topic and so on. And once I find
12:22 that I'm done with a conversation or a topic, I go ahead and just delete it.
12:27 And to do this in Telegram specifically, you just set up a new group. You add
12:31 Claudebot as the only user. You make Claudebot an administrator of that
12:36 group, and it should just work. And now you also have to tell Claudebot to reply
12:42 to every message in the group chat, not just ones that is specifically tagged in
12:46 because it's thinking it's going to be in a literal group chat with maybe
12:50 dozens or hundreds of people. So, it's only going to respond to messages that
12:54 it is specifically tagged in. And you don't want that with this setup. You
12:57 want it to respond to every single message. So, just tell it to do that and
13:00 Self-Improvement (ADVANCED)
13:01 it'll update its settings. All right. Here's another advanced technique that I
13:05 use. As you're using CloudBot a lot, it is storing so many memories about you.
13:10 It is creating skills and just basically creating all of this potential clutter.
13:15 And so what I have it do daily, listen to this. Set up a daily review of your
13:20 main files including agents.mmd memory, anything that you can suggest. Then
13:25 propose changes here and see if I want to make them. Basically, it is a full
13:29 audit of its codebase every single day. So here it says, okay, I'm going to
13:34 review agents.mmd, memory.mmd, tools, soul, identity, user, heartbeat, and
13:39 then it's going to look for things like outdated info, conflicting rules,
13:44 undocumented new workflows, lessons from recent failures, etc. And anytime it
13:49 finds anything that it thinks, hey, I should actually update that or clean
13:52 that up, it's going to recommend it to me, and then I'm going to decide if I
13:55 want to move forward with that or not. I definitely recommend doing this just to
13:58 Image Generation
14:00 keep your Claudebot nice and tidy. And you can also have full multimedia
14:04 capability. So you can have it create videos and images and voice, anything
14:08 that you want. Here's an example. I have it plugged into Nano Banana. And I'm
14:12 just going to say create an image with Nano Banana of a lobster. Okay, so I had
14:16 to go back and forth with it a little bit, but here it is. Picture of a
14:20 lobster. And again, you can give it voice capabilities with something like
14:25 11 Labs. It can read images. Just drag and drop in Telegram or whatever chat
14:30 app you're using. So, the capabilities are truly endless. You just need to tell
14:33 Security
14:35 it what you want to do. But now, here's the really important part. Let's talk
14:39 about security. I have thought a lot about security best practices. I am far
14:44 from a security expert. So, if you have any thoughts about any of these, please
14:47 drop them in the comments below. So, we're going to create a new topic in
14:51 Telegram and we're going to call this security. So I will tell it something
14:56 like never store an API key or token anywhere but A.Env file and never
15:01 include AENV file in your git. Now it should do this automatically, but I like
15:05 to reinforce that. And so you're going to want all of your API keys, all of
15:10 your tokens all in one file. And then you don't want to include it in your git
15:13 because when you push your Git to your GitHub account or you save it somewhere
15:17 else, you don't want that exposed in the Git history. Now, if a lot of this
15:22 sounds very technical, don't worry. Just type this in and that's all you need to
15:28 do. Also, OpenClaw Clawbot comes with a security checkup feature. So, you simply
15:34 type OpenClaw security audit. So, you can do this directly through Telegram,
15:37 but I'm going to go back to Hostinger's terminal and we're going to do it there.
15:40 So, from the root directory, you're going to type openclaw security audit
15:45 and then hit enter. So, here it ran a security audit and we see two warnings.
15:51 trusted proxies missing and chmod 700. So go ahead and fix those. And so you
15:55 can go ahead and fix those by typing openclaw security audit- d-fix. So I'm
16:00 going to go ahead and hit that. And there it fixes it automatically for me.
16:01 Hosting OpenClaw
16:06 Next is just how you're hosting your cloudbot. And that is the beauty of
16:10 hostinger. You put it on hostinger and it is completely isolated from the rest
16:16 of your environment. It is sitting in its own virtual private server all by
16:21 itself. It doesn't have access to other things in your keychain on your local
16:24 computer. It can't just take over your computer and delete a bunch of files. It
16:30 is all isolated there. And that is super important. The next is I want to explain
16:34 the concept of clean versus dirty data. Think about anything coming from the
16:38 internet, anything coming from outside of your completely enclosed system being
16:45 dirty. Meaning somebody, a malicious actor could be trying to write a prompt
16:50 injection. Here is an example. If you have your Cloudbot reading your emails
16:55 and somebody knows that your Cloudbot is reading your emails, they can send you
16:59 an email with a prompt injection in the body of the email. So your Clawbot goes
17:04 to read the email, gets prompt injected, and then reveals sensitive information.
17:07 And so just keep that in mind. I'm going to say it again. Any time that your
17:14 Cloudbot is exposed to dirty data, it has the potential to be malicious. So,
17:19 be very wary. And there's a few things to do about that. One, again, just be
17:24 mindful and try to limit exposure to dirty data. Also, the better model you
17:29 use, the less susceptible it is to prompt injection. So, Opus 4.5 is going
17:34 to be much less susceptible to prompt injection than Haiku would be. Also,
17:40 Claudebot has prompt injection detection built in, but it's not perfect. It never
17:45 will be perfect because that is the nature of AI. Another security tip, make
17:50 sure you are updating frequently. Just a day ago, OpenClaw updated with a bunch
17:56 of new security features. Next, don't trust skills. Claw Hub has already made
18:00 a bunch of changes to have better security around skills that people
18:03 submit, but ultimately that is dirty data. It is somebody submitting a skill
18:08 and there were a lot of crypto scams already in the early days. Try not to
18:12 trust skills for as much as you can and ultimately your Claudebot can write any
18:17 skill that it needs. It doesn't actually have to download a skill from somebody
18:21 else. So if you can write the skill yourself, if you can have your Cloudbot
18:25 write the skill itself, always do that. And be thoughtful about the integrations
18:29 that you're adding. If it's something that has highly sensitive data and you
18:35 really don't want it exposed ever, you want zero chance of that and it's
18:39 connected to external services that might be prompt injected. Just be very
18:44 mindful about it and you can always limit your exposure by limiting the
18:47 integrations that you're using. But again, the trade-off is capability
18:53 versus risk. Last, on the security front, for any complex coding task or
18:58 really any complex task at all, especially if there are going to be
19:02 changes to files or changes to different integrations that I have, I always have
19:07 my Claudebot propose what it's going to do before actually doing it. So, it's
19:11 kind of like plan mode. Think about what you want to do. Think about the best way
19:14 to do it. Tell me what you're going to do, and I'll let you know if I want you
1:32 OpenClaw Setup
1:33 below. All right. And getting it set up with Hostinger is dead simple. You just
1:36 come to this page. I'll drop the link down below. And all you have to do is
1:40 click deploy. On this page, you can select your time duration. And this is
1:42 where you're actually going to purchase the VPS. Right here, you can select 1
1:46 month, 12 month, 24 month. I'll start with 24 months. And right over here
1:50 where it says have a coupon code. Yes, we do. Matthew B. Apply. And there you
1:56 go. You get 10% off. Now continue. And there you go. Payment done. Then it's
1:59 going to bring you to this configuration page. You could put your API key from
2:04 Anthropic, OpenAI, Gemini, XAI, anything else. And then click deploy. And then
2:08 you can connect everything else you need by opening up terminal and going through
2:12 the open claw onboarding. And of course all of this is optional. And here you
2:16 can see we're good to go in Telegram. I said hello and my clawbot says hey
2:20 MD Files (IMPORTANT)
2:21 what's up. All right. So the first thing to really understand are all the
2:25 different MD files that you have that really form what this Cloudbot is for
2:31 you. And there are a bunch heartbeat identity memory soul tools and user. So
2:36 this is the soul.md file. This is basically defining what the personality
2:40 of your Claudebot is going to be. Feel free to come in here and edit as you see
2:44 fit. Then we have the skills. This is really what makes Cladbot so incredibly
2:49 capable. This is where you give it all the skills that help it browse the web,
2:54 check your email, access Twitter, and all of your custom skills as well. So,
2:58 anything that you tell Cloudbot to do that it doesn't already know how to do,
3:02 it's going to create a skill from that and save it here. You can think of a
3:06 skill as a repeatable process. Anything where you have multiple steps and you're
3:11 using different tools along the way. And so that's defined here. So we have the
3:15 humanizer skill for example and a whole bunch more in here. And the skills
3:21 reference tools. And tools are pieces of code that allow it to actually
3:24 accomplish those things that the skills want to do. So here's an example. Here's
3:28 fetch.js. And it's basically a JS file that allows my agent to plug into ASA really easily
3:36 and grab information from my ASA board. And again, you don't need to write any
3:40 of this. It's all done through natural language through Telegram or WhatsApp or
3:45 wherever you're using your Claudebot. Then we have the identity file and it's
3:50 very similar to Soul in a lot of ways. It basically tells Claudebot how it's
3:54 actually going to interact with you from a tone and emojis perspective. So we see
4:01 name, Claude, vibe, emoji, avatar, and so that's what's in the identity file.
4:05 Cloudbot now has a memory folder in which it stores all of its memories
4:09 about you. And you can just let it store everything. You can always go in there
4:13 and prune things that you don't think it needs to know anymore. But for the most
4:16 part, there's no harm in just having it store most of what it learns about you.
4:22 Then we have the heartbeat file, which basically runs every so often. I think
4:26 by default it's every 30 minutes but it's whatever you want to have happen on
4:31 a recurring basis but we also have cron jobs and I'll get to that in a moment
4:34 and again this entire cloudbot project will grow with you so as you figure out
4:39 ways that you want it to run you just tell it and it will start to learn to do
4:42 that. All right so those are all the basics I want to get into a little bit
4:44 Model Selection
4:45 more advanced stuff now. All right so next I want to talk about model
4:48 selection because that's probably one of the biggest open questions that people
4:52 have right now is what model should I be using? What should be the primary model?
4:55 What should be the fallback model? Should I use multiple models? Do I use a
5:01 local model? And the answer is yes, all of it. And I've set up my model routing
5:05 to be pretty sophisticated. And I want to show you. And the way I want to show
5:09 you is by asking my cloudbot. So I'm going to say, show me your model
5:12 selection logic hierarchy. So it's going to read from its own files and show me
5:16 right now. So the primary model I'm using right now is Cloud Sonnet 4.5. And
5:21 that is because it is a workhorse. It is much cheaper than Opus. And so for the
5:26 default interaction through chat, that is what I'm using. But I also have a
5:32 fallback chain. I use Gemini 3 Flash, which is incredibly fast and cheap. I
5:37 have Opus 4.5 here. I have Open Router that I'm using to route to other models
5:41 that maybe aren't on my initial list. And I was even running local models, but
5:46 I'm not doing that in this moment. And so you can have local models running for
5:51 things like cron jobs and other really basic tasks that don't require frontier
5:56 models, unless of course you can run one of those frontier open source models
5:59 locally. And there are a few different ways that you can have it select which
6:04 model it's using. You can simply say in natural language, switch to Sonnet 4.5.
6:09 Hit enter and it's going to switch to that model for this chat. It will also
6:14 make a decision about which model it should use at any given time depending
6:18 on the task complexity. So we say done back on sonnet 45. You can also type
6:23 slashmodel and if you ever get disconnected or you run into a rate
6:27 limit on one of your models, use /model to show or set the different models that
6:32 you have. So the way I'm doing it is for anything coding, I'm usually using
6:36 sonnet, but if it's an especially complex coding task, I tell it to route
6:41 to opus 4.5. For most kind of basic interactions, I'm either using Sonnet.
6:45 Sometimes I'll switch to Haiku because it's even faster and much cheaper. And
6:50 of course, I'm experimenting with models all the time. And remember, the model
6:54 you select definitely affects the personality of your Claudebot. Now, if
6:57 you don't care about that, don't even worry about it. Just use any model
7:00 that's best for your use case. But if you do care about the actual
7:04 personality, you're probably going to get familiar with one model, and
7:08 switching is a little bit jarring. Now, Clawbot packages up everything about its
7:12 personality, all of your memories, and will send it to each of these models
7:16 individually, but the way that the models themselves handle all of that and
7:21 respond to you still will be different depending on the model. Okay, now I want
7:25 to get into skills and tools. This is really what makes Claudebot special.
7:29 This is the thing that sets it apart from any other tool that I've ever used.
7:33 It just so easily plugs into any service, any product that you're using.
7:38 I have it plugged into my Gmail, to my Drive, to my calendar, a sauna, Slack,
7:44 HubSpot. A majority of services that I use, I allow Claudebot to plug into.
7:49 Now, I'm fairly technical, so I really have thought about the security
7:53 implications about connecting all of these different services, and I'm doing
7:56 so in a secure way. I'm going to talk about security later in the video, so be
8:00 sure to stick around for that. But to plug it into a new service, it is
8:05 simple. All you have to do is tell it. That's it. It'll look for skills. If it
8:09 doesn't already have the skill, it will write one itself. You might have to grab
8:14 an API key from somewhere, but that's about it. You do it all through chat
8:19 with natural language. And you can even get very sophisticated with your model
8:24 selection. So, I'm a cursor user and I installed cursor agent in my hosting or
8:31 VPS and I'm actually able to control my cursor agent through Claudebot through
8:37 Telegram. So, anytime that I have a complex coding task, I'm telling my
8:42 Claudebot to delegate it out to my cursor agent and it does that and it's
8:47 super impressive. Now, the only thing to remember with that is that cursor agent
8:51 has basically no personality. So sometimes when it responds back to you
8:53 and Telegram that it's done or giving you an update, it basically strips all
8:57 the personality that you're familiar with. And again, it's a little bit
9:01 jarring, but feel free to plug in the tools that you're most comfortable with.
9:05 Skills
9:05 And I would recommend starting with more lowrisk tools. Now, if you want to
9:10 browse new skills or you want to get inspiration for things that your Cladbot
9:15 can do, go to clawhub.com. This is the official Cloudbot skill repository by
9:20 the Cladbot creator. So up here, click skills, and you can just see a ton of
9:24 different skills that are available. Now, keep in mind, these are written by
9:28 other people. There's always a chance there's some malicious code in there.
9:32 So, what I like to do is tell my Claudebot, go download this, but first,
9:37 scan it, make sure there's nothing malicious in there, and I want to always
9:42 use the best possible model to do that because the best possible model is going
9:46 Scheduling
9:46 to be less susceptible to prompt injection. Okay. Next, another core
9:51 mechanic of CloudBot is it being able to do things for you on a scheduled basis.
9:55 You can have it be on a recurring basis. You can do a one-off scheduled task. And
9:59 all of this is going to be done through cron jobs. And if you're not familiar
10:02 with that term, it basically just means a scheduled thing to do. And you simply
10:09 just tell it to do so. So, in 1 hour, remind me to drink water. I'm going to
10:12 hit enter. And then it's going to create a cron job for me. And 1 hour from now,
10:17 it's going to message me back again. and tell me to drink water. And these can
10:22 get as complicated as you want. So, for example, one thing that I did was I have
10:27 kind of a weird trash pickup schedule at my house. It happens every Monday and
10:32 every other week. It's some rotating recycling. Sometimes it's cans,
10:36 sometimes it's bottles, sometimes it's paper. And I could never remember what
10:41 it is. So, I actually just told my Cloudbot, "Here's the schedule." like
10:44 just took a picture of it from the recycling company's website, uploaded
10:48 it, and say, "Look at what the schedule is, and every Sunday, I want you to tell
10:52 me which cans to take out." And now it does that. It tells me exactly which
10:56 recycling to put out on Sunday. Okay. Now, I'm going to show you one of the
11:01 most powerful unlocks that I have found with Claudebot, and that is using
11:05 Telegram groups. Now, previously what I was doing is just DMing with my Cloudbot
11:10 directly. And that was fine, but we would just have infinite history of
11:14 conversation. But if I wanted to have multiple tasks running in parallel, it
11:18 was very difficult. I would say, "Okay, hold on to that idea. Let me come back
11:21 to that and let me switch back to this other task I had you doing." And it was
11:25 just so complicated. And then I figured out that you can actually use Telegram
11:30 groups. And with a Telegram group, you can set up different topics. And so you
11:33 can have a bunch of different conversations going in parallel with
11:37 your Claudebot. And that is beneficial in multiple ways. One, you can have a
11:44 very seamless conversation in each of the topic channels. And that allows you
11:49 and Claudebot to stay on topic and not get confused about other things going
11:54 on. That also saves on context window and memory because instead of loading up
11:59 the entire conversation history which might include days, weeks, even years in
12:04 the future of conversation history, it loads up only what is in that topic,
12:10 saving you room in the context window. So here's an example. I have one for
12:14 video research. I have Twitter research. I have an ebook I'm putting together. We
12:17 have content analysis. We have a cloud skills topic and so on. And once I find
12:22 that I'm done with a conversation or a topic, I go ahead and just delete it.
12:27 And to do this in Telegram specifically, you just set up a new group. You add
12:31 Claudebot as the only user. You make Claudebot an administrator of that
12:36 group, and it should just work. And now you also have to tell Claudebot to reply
12:42 to every message in the group chat, not just ones that is specifically tagged in
12:46 because it's thinking it's going to be in a literal group chat with maybe
12:50 dozens or hundreds of people. So, it's only going to respond to messages that
12:54 it is specifically tagged in. And you don't want that with this setup. You
12:57 want it to respond to every single message. So, just tell it to do that and
13:01 it'll update its settings. All right. Here's another advanced technique that I
13:05 use. As you're using CloudBot a lot, it is storing so many memories about you.
13:10 It is creating skills and just basically creating all of this potential clutter.
13:15 And so what I have it do daily, listen to this. Set up a daily review of your
13:20 main files including agents.mmd memory, anything that you can suggest. Then
13:25 propose changes here and see if I want to make them. Basically, it is a full
13:29 audit of its codebase every single day. So here it says, okay, I'm going to
13:34 review agents.mmd, memory.mmd, tools, soul, identity, user, heartbeat, and
13:39 then it's going to look for things like outdated info, conflicting rules,
13:44 undocumented new workflows, lessons from recent failures, etc. And anytime it
13:49 finds anything that it thinks, hey, I should actually update that or clean
13:52 that up, it's going to recommend it to me, and then I'm going to decide if I
13:55 want to move forward with that or not. I definitely recommend doing this just to
14:00 keep your Claudebot nice and tidy. And you can also have full multimedia
14:04 capability. So you can have it create videos and images and voice, anything
14:08 that you want. Here's an example. I have it plugged into Nano Banana. And I'm
14:12 just going to say create an image with Nano Banana of a lobster. Okay, so I had
14:16 to go back and forth with it a little bit, but here it is. Picture of a
14:20 lobster. And again, you can give it voice capabilities with something like
14:25 11 Labs. It can read images. Just drag and drop in Telegram or whatever chat
14:30 app you're using. So, the capabilities are truly endless. You just need to tell
14:35 it what you want to do. But now, here's the really important part. Let's talk
14:39 about security. I have thought a lot about security best practices. I am far
14:44 from a security expert. So, if you have any thoughts about any of these, please
14:47 drop them in the comments below. So, we're going to create a new topic in
14:51 Telegram and we're going to call this security. So I will tell it something
14:56 like never store an API key or token anywhere but A.Env file and never
15:01 include AENV file in your git. Now it should do this automatically, but I like
15:05 to reinforce that. And so you're going to want all of your API keys, all of
15:10 your tokens all in one file. And then you don't want to include it in your git
15:13 because when you push your Git to your GitHub account or you save it somewhere
15:17 else, you don't want that exposed in the Git history. Now, if a lot of this
15:22 sounds very technical, don't worry. Just type this in and that's all you need to
15:28 do. Also, OpenClaw Clawbot comes with a security checkup feature. So, you simply
15:34 type OpenClaw security audit. So, you can do this directly through Telegram,
15:37 but I'm going to go back to Hostinger's terminal and we're going to do it there.
15:40 So, from the root directory, you're going to type openclaw security audit
15:45 and then hit enter. So, here it ran a security audit and we see two warnings.
15:51 trusted proxies missing and chmod 700. So go ahead and fix those. And so you
15:55 can go ahead and fix those by typing openclaw security audit- d-fix. So I'm
16:00 going to go ahead and hit that. And there it fixes it automatically for me.
16:06 Next is just how you're hosting your cloudbot. And that is the beauty of
16:10 hostinger. You put it on hostinger and it is completely isolated from the rest
16:16 of your environment. It is sitting in its own virtual private server all by
16:21 itself. It doesn't have access to other things in your keychain on your local
16:24 computer. It can't just take over your computer and delete a bunch of files. It
16:30 is all isolated there. And that is super important. The next is I want to explain
16:34 the concept of clean versus dirty data. Think about anything coming from the
16:38 internet, anything coming from outside of your completely enclosed system being
16:45 dirty. Meaning somebody, a malicious actor could be trying to write a prompt
16:50 injection. Here is an example. If you have your Cloudbot reading your emails
16:55 and somebody knows that your Cloudbot is reading your emails, they can send you
16:59 an email with a prompt injection in the body of the email. So your Clawbot goes
17:04 to read the email, gets prompt injected, and then reveals sensitive information.
17:07 And so just keep that in mind. I'm going to say it again. Any time that your
17:14 Cloudbot is exposed to dirty data, it has the potential to be malicious. So,
17:19 be very wary. And there's a few things to do about that. One, again, just be
17:24 mindful and try to limit exposure to dirty data. Also, the better model you
17:29 use, the less susceptible it is to prompt injection. So, Opus 4.5 is going
17:34 to be much less susceptible to prompt injection than Haiku would be. Also,
17:40 Claudebot has prompt injection detection built in, but it's not perfect. It never
17:45 will be perfect because that is the nature of AI. Another security tip, make
17:50 sure you are updating frequently. Just a day ago, OpenClaw updated with a bunch
17:56 of new security features. Next, don't trust skills. Claw Hub has already made
18:00 a bunch of changes to have better security around skills that people
18:03 submit, but ultimately that is dirty data. It is somebody submitting a skill
18:08 and there were a lot of crypto scams already in the early days. Try not to
18:12 trust skills for as much as you can and ultimately your Claudebot can write any
18:17 skill that it needs. It doesn't actually have to download a skill from somebody
18:21 else. So if you can write the skill yourself, if you can have your Cloudbot
18:25 write the skill itself, always do that. And be thoughtful about the integrations
18:29 that you're adding. If it's something that has highly sensitive data and you
18:35 really don't want it exposed ever, you want zero chance of that and it's
18:39 connected to external services that might be prompt injected. Just be very
18:44 mindful about it and you can always limit your exposure by limiting the
18:47 integrations that you're using. But again, the trade-off is capability
18:53 versus risk. Last, on the security front, for any complex coding task or
18:58 really any complex task at all, especially if there are going to be
19:02 changes to files or changes to different integrations that I have, I always have
19:07 my Claudebot propose what it's going to do before actually doing it. So, it's
19:11 kind of like plan mode. Think about what you want to do. Think about the best way
19:14 to do it. Tell me what you're going to do, and I'll let you know if I want you
19:18 to do that thing or not. All right. So all of this is best practices, how to
19:22 get the most out of Cloudbot, but what am I actually using Cloudbot for? And so
19:27 I wanted to show a few examples of real use cases that I'm doing and how I set
19:30 them up. So check this out. So one, I frequently have ideas for video topics
19:35 and I have an entire pipeline that I use to track video ideas to research them
19:39 and I basically automated the entire thing with Claudebot. And so here's what
19:43 I do. I drop a link in Telegram and that link is something I want to talk about.
19:47 Whether it's like an Xlink about some model release or a website that I wanted
19:51 to talk about that looks interesting, a video, anything at all. Telegram throws
19:56 it to Cladbot. Cladbot does research on that topic using the Brave API. It will
20:02 also check Twitter for trending tweets about that topic using the Grock API.
20:09 Then it will output an ASA task for me with all of that information. so my
20:13 entire team can see it at all times and that ends up in the video ideas bucket
20:19 in ASA. Super useful. One drop of a link and it's all done for me. Next, I'm
20:23 frequently checking my YouTube statistics and I want to know how my
20:26 videos are performing and history of videos. And so I gave my Cloudbot access
20:32 to my YouTube data API and analytics API. So I set up a topic in Telegram
20:37 like I showed you earlier. I ask in Telegram, okay, give me, you know, the
20:41 last three videos, how are they performing? Cloudbot goes fetches via
20:46 the YouTube API and the analytics API and it posts back to me in Telegram and
20:50 I can actually have it post to Slack and tell my team how specific videos are
20:55 doing. Very useful, very easy. It's basically like an analyst at all times
21:00 in Telegram. Next, I have it do meeting prep for me every single day. In the
21:05 morning, I have it look at my calendar, find all of the meetings, filter by
21:09 anybody that I'm meeting with that is external, meaning not part of my
21:12 company, and not just some event that I have to remind myself about. Go look up
21:16 information about that person, how we met, what the context of the meeting is,
21:21 and then give me a summary each morning about everybody I'm meeting. And so, the
21:25 way I do that is I have it with a crown job in Telegram each morning. It fetches
21:30 the Google calendar. I have it check my Gmail and I only have it check emails
21:35 from people that I know so to keep it very safe and not expose myself to
21:39 prompt injections. And then I have it report back to me in Telegram. All
21:43 right, so those are incredibly important automations that I've set up with
21:46 Cloudbot. And so that's everything. I know this was a long video. I know there
21:51 was a lot to it, but this is honestly one of, if not the most useful pieces of
21:56 artificial intelligence that I've ever used, even more so than Chai GPT. And so
22:00 that's how to get the most out of it. That's how to set it up with Hostinger
22:05 and be super secure using a VPS. Thank you again to Hostinger for sponsoring
22:10 this video. Again, go get your HostingerVPS. hostinger.com/matthewb.
22:16 Make sure to use the code matthewb. You get 10% off and it is a one-click
22:20 install. They've been a great partner. I'll drop all the links down below.
22:24 Please check it out. It really helps the channel if you go use them. If you
22:27 enjoyed this video, please consider
