// transcript — 831 segments
0:00 Why This is So Powerful
0:01 Last night, I was laying in bed and I texted Klouse to build me a YouTube
0:04 dashboard. And I told her that I was about to go to sleep so that I wouldn't
0:07 be around. So, it spun up the initial page and then it gave itself six more
0:10 to-dos that it would take care of while I was sleeping. And then when I woke up
0:14 and I opened up my computer, I had this. I've got a YouTube analytics dashboard
0:17 where I can track my 7-day stats, my 30-day stats. I can look at my videos
0:20 and see how they're performing relative to other videos. I've got a hub to look
0:23 at all my comments, whether that's hot topics, common questions, pain points,
0:26 or just looking at all the recent comments on here. It's able to pull in
0:29 competitor videos and see which of these are breaking out. And then it uses all
0:32 of that data to help me come up with ideas. So, it comes up with titles for
0:36 YouTube videos that I could make based on my audience and based on the trends.
0:39 This is just V1. It did this while I was sleeping. So, this is my dashboard that my
0:45 ultimate assistant Klouse built for me. In here, I can see all the tasks that
0:47 he's working on, what's in the backlog, and I can even see up here when he's
0:51 actually working or when he's not. So, last night before I go to bed, I
0:54 basically say, "Hey, all night when I'm sleeping, I want you to build me a
0:57 YouTube dashboard." And so we put together a speclist. He threw all of his
1:01 tasks in the to-do section as you can see. So of course I go to bed. I wake up
1:04 feeling like a kid on Christmas. And then I check in on the morning and he
1:08 says, "The YouTube dashboard MVP is done." And so I come into the dashboard
1:11 this morning and I can see that everything has now been moved over here
1:15 to done. And this is what we got. So hopping back into my Cloud dashboard, if
1:18 I now go to the log, I can see everything that he does. So if I scroll
1:22 down to where it would have been this morning. So last night I went to bed at
1:26 like midnight. So right here is where I started this. And then after we kicked
1:29 off all the spec docs and I added all of these tasks to his backlog. Look what he
1:35 did. 12:31 did something. 12:48, 1:15, 142, 208, 234, like all night up until
1:40 like 7 when I checked in on him and he started doing some daily pull stuff in
1:43 the morning. He was just working on the YouTube dashboard all night long. So I
1:46 mean, if that doesn't at least get you a little bit excited, then I don't know
1:48 what will. So today I'm going to be going over this dashboard, how I built
1:52 it, how you can create your own version of Clouse. So, pretty much everything
1:55 that I've done with my Clawbot here and then I'm going to talk about five hacks
1:59 for a smarter assistant because it is a really smart tool, but there's a lot of
2:01 things that you have to overcome and I've had a lot of frustration when I was
2:05 building out this dashboard and getting clouds to work in a way that actually
2:07 makes me more efficient rather than less. Now, the reason that I'm actually
2:10 so excited by this and I've been having so much fun with it is because it's like
2:14 the end ultimate assistant except for it is 15 times more powerful. This is one
2:18 of the videos I made about Enidend that went viral and it's the ultimate
2:21 assistant which I've configured to have an email agent, a calendar agent,
2:25 contact agent, content creator agent, a search tool. If you think about it, the
2:29 functionality is really not that much. We're able to talk to it through
2:32 Telegram and it responds to us in Telegram. And so all of the sub aents I
2:35 obviously had to build. So for example, the email agent looks like this with all
2:38 the email tools. And if I wanted to add a different type of agent, I would have
2:41 to go into Niten and build that out. But what happens with Claus is really cool
2:44 because it's like let's say you have five functions and if you want to figure
2:47 out if you can do something else like hey Claus can you send voice notes? Sure
2:51 let me just research how that works. Okay cool here's how I could do that. Do
2:54 you want me to actually build this in? So basically any feature that you could
2:57 describe Klouse will figure out how to do it and if it's possible it will just
3:01 do it for you. What's on your end of course is actually maybe going and
3:05 creating an API and putting that in your or maybe going and paying for a plan and
3:08 then giving him that username or something like that. So yeah, here's
3:11 where I put in my Claus master guide access and capabilities. Right now, this
3:15 is what it's got access to. GitHub, Brave Search, OpenAI, 11 Labs, PDF
3:19 Shift, Appify, ClickUp, Google Workspace, and then some YouTube stuff
3:22 like that. We've got understanding of how the memory system works. It
3:25 understands when to send voice notes, which is only when I send him a voice
3:28 note. Here we have some information on what it should do freely and what would
3:30 require permission. And then you can see what we've built together. So the first
3:33 one is obviously the cloud dashboard, and here's what it does. We've got the
3:36 YouTube intelligence dashboard, and here's what it does. I'm still impressed
3:38 that it built this while I was sleeping. Super cool. It can do branded PDF
3:42 reports and then it has scheduled triggers to do things like daily pulses,
3:46 audits, email monitoring, checking tasks, and obviously this isn't
3:50 completely updated because I added some more stuff in after this guide. All
3:54 right, so before we actually break down exactly what I've done and how you can
3:57 replicate this, let's just go over a few more features in this dashboard so you
4:01 can get a feeling of what I've actually built here with Klouse. So, of course,
4:03 we are talking to Klaus through Telegram. And you can see the last thing
4:06 I just had him make a change. I said, "This is perfect. You're the best." and
4:09 he reacted to my text, which I thought was pretty cool that they can react. And
4:12 then the reason why it says model claopus 4x5 is because I wanted to have
4:16 visibility. So every time I messaged him, I told him to tell me what model
4:20 he's using. So this started as a pretty simple dashboard just so I could track
4:24 what was going on. But I told him, "Hey, your job is to save me time. I want you
4:27 to be proactive, understand my workflows, and suggest things all the
4:31 time." So specifically in this dashboard while we were building it together, I
4:34 said, "Suggest features, whatever you think might work and whatever you think
4:37 will help you work better and help me work better. Just tell me that you're
4:40 going to implement it." And so one of the first things that he did was he
4:43 built this little thing up in the top left so that I can actually see him,
4:46 which was a little bit scary, but also kind of cool because now I can look over
4:50 and I can see if it's idle, then he's not working on anything. But if he's
4:53 thinking or working, then I know that there is something going on. So let me
4:56 go ahead and shoot something off and we will see him spin up and start working.
4:59 All right. So, just to start off here, I'm saying, "Hey, Klaus, I need you to
5:02 spin up a sub agent to do some in-depth research about what's going on today
5:05 with Naden. There was some news about a security thing, and I want you to create
5:09 a report for me." So, obviously, we can see up here that it is thinking. It also
5:12 has spun up a sub agent, which we are able to visually see, which is pretty
5:17 cool. And I will get back to you once we have that report. And of course, it was
5:20 able to add an inrogress task right here. And then when it's done, hopefully
5:23 it should move it over to done. But that's a really cool thing about this
5:26 Cananban board is that if I have a task and I come in here and I just add a new
5:30 to-do, then Claus will automatically pick that up and start working on it. Or
5:34 similarly, if I give him a note down here, I can send something off just for
5:37 him to log it in his memory. Or even if I wanted to take action on it. So like
5:41 here, I said testing if this works. When you see this, tell me a joke. And then
5:44 it shot me a joke in my telegram. And every time it actually acknowledges
5:47 these, it marks them as seen by Clouse. So I know which ones it's processed or
5:51 not. So here's a great example of how it knows about me and my business because
5:55 it came back and said the TLDDR is anyone running any needs to upgrade
5:58 because of this vulnerability. So do you want me to turn this into some content
6:00 for you? Because it could be a quick YouTube short or community post which is
6:04 the kind of thing that your audience needs to hear about fast. But anyways, I
6:07 just said no post for now. I just want the report. So it's throwing that
6:10 together now. All right. So it said that the report is done. If I close out of
6:13 Telegram, we should see that this got moved to done. And the reason why it's a
6:16 different color is because these are actually a priority switch, not like a a
6:20 status. As you can see, if I move this over here, it's just a priority thing.
6:23 So anyways, if I go ahead and click on the docs tab right here in our
6:26 dashboard, you can see that we have a new doc about security, which is end
6:30 security vulnerability report. So now I'm able to actually read through this
6:33 entire report that it made. We've got the date, classification, the exact
6:37 summary, vulnerability details, and we can continue to scroll down and see this
6:41 entire markdown format report that it was able to make for us. And if I wanted
6:43 to change anything about this, I could go ahead and edit it, and then it would
6:47 bring up this editor for us to change something in the doc, save it. So now we
6:50 have a really cool place where every single time that we ask Claus to create
6:53 some sort of file for us, it's able to just drop it in here, and then we can
6:56 look at it and save it. So I thought I'd show you guys off one of those heartbeat
6:59 functionalities. And I'll explain exactly what a heartbeat is and how that
7:02 works in a bit. But I'm going to go ahead and add a new task right here that
7:07 just says, um, make Nate laugh. And I'm just gonna say, "Usually in the
7:11 afternoon, Nate gets sad, so tell him a joke about AI." So, while we're waiting
7:15 for Claus to actually pick that up and shoot me off a joke, let's see what we
7:18 have done here, which are some of our scheduled deliverables, which are
7:21 workflows that happen either daily or weekly. So, for example, with the weekly
7:25 YouTube audits, every single week on I think it's like Monday night or Sunday
7:28 night, I'm going to get one of these YouTube audits. And obviously here, I
7:31 was playing around with some different formats, but this is kind of what they
7:35 look like now. They have my brand information, although this one didn't
7:38 have the logo. Let's try this one. It'll have my logo. It'll have my brand
7:42 guidelines. And it sends me a SWAT analysis and a YouTube channel audit
7:45 about strengths, weaknesses, metrics, and some other things that we need to
7:48 do. Obviously, there's a bit of a spacing issue here, but we get some
7:52 recent content analysis, strategic recommendations. And what you'll notice
7:55 here is that we're signed in here as Klouse. So, Klouse has its own Gmail,
8:00 Drive, Docs, Sheets, its own accounts for pretty much everything because I
8:03 didn't want it to be in my environment. I wanted to treat this as a person that
8:07 I could talk to, forward emails to and have it respond like that. So, it
8:10 created a drive called cloud deliverables and that's where it was
8:12 able to store all of these things like the daily pulses, the SWAT analysis, the
8:17 logos, audits, and YouTube audits. So, obviously security is a huge topic when
8:21 we're talking about Cloudbot. So, the security audit that I had it run for me
8:24 here, this was actually just a Google doc rather than a PDF, but it pointed
8:28 out all these issues. It ranked them for different risk levels, and then I was
8:31 able to have Claus just go ahead and fix all these things. And I'm constantly
8:34 checking with Klouse to make sure that we are completely fine and that our
8:37 port's not exposed and stuff like that. And obviously I'm not like a security
8:40 expert. So I'm doing the best that I can working with Klouse with cloud code with
8:44 perplexity figuring out how I can be safe. But everyone out there that wants
8:47 to experiment with cloudbot, just make sure that you are thinking about
8:50 security and not giving access to tons of different random things if you don't
8:53 really know what's going on. Okay, so what I did is I said, "How often does
8:56 your heartbeat run and what does it actually do?" And it said that my
8:59 heartbeat runs every 30 minutes. And every time that my heartbeat runs, it
9:03 basically wakes up Klaus. And Claus will make sure that both the dashboards are
9:05 running. It will do a sync. It will check for notes, which you can see it
9:08 actually just moved the make Nate laugh from in progress to done, I believe. As
9:13 you can see, it will monitor X and Twitter. It will do a daily pulse. And
9:16 the idea is that I wake up with no memory. So when I have my heartbeats, I
9:19 stay proactive and I check in on things and do things in the background for you.
9:23 So anyways, then it said, "Yep, just saw that task Q. Make Nate laugh high
9:26 priority." And the joke is, what do you call an AI that finally passes the
9:30 touring test? Unemployed, the humans just move the goalpost again. I don't
9:32 really get that one. So, let's actually move on to the next part, which is how
9:35 can you create your own version of Klouse? So, the first thing, of course,
9:38 you want to do is set up Cloudbot. Whether you're going to do that on a Mac
9:40 Mini or whether you're going to do that on VPS, I will tag a full setup guide
9:45 video right up here that I made a few days ago and you can get that set up on
9:49 a VPS. So what I actually did is I came into Klaus and I was like, "Hey, I want
9:53 to teach my YouTube channel how they can basically build exactly what I built for
9:58 you and how we configure you." So I had it create this guide called build your
10:01 own Klouse, a step-by-step guide to creating an AI executive assistant.
10:05 Clouse handles daily briefings, monitors my business, runs security audits, and
10:08 proactively services opportunities, and you can build your own version. So time
10:11 investment, 4 to 6 hours over the weekend to get the core system running,
10:15 and cost about 50 bucks a month. I'm going to say that it's going to be
10:17 higher than that. So, I've obviously spent lots of hours playing around with
10:20 Cloudbot, doing different things, and I spun up this one to turn this into an
10:23 executive assistant. And you can see with just this specific Cloudbot that
10:27 I'm using, Klouse, I've used almost a quarter of a billion tokens in the past
10:32 just 3 days, which translate to about $223 in actual money. But keep in mind, all
10:38 of this has been done with Opus 4.5. So, if you went to Sonnet 4.5 or if you use
10:42 a locally hosted model, your cost would be less. And the reason why I'm not
10:46 using like my max plan, the $200 a month plan for which I would get way more
10:50 tokens with is because Anthropic has been banning accounts left and right for
10:53 violating terms of service. So if you are currently using Cloudbot with a max
10:56 plan or any Cloud plan at all, you have to switch it over to API. So anyways,
11:00 what happens when you actually get this set up? Well, you have to get onboarded.
11:03 So the first thing it's going to start to do is it's going to start to ask you
11:06 questions. So you have to give it an identity. You have to talk about who are
11:09 you, what its role is, what you want it to do, how it should behave, and you
11:12 should also just brain dump everything about you and your business. So, I spent
11:15 like almost an hour just talking to Claudebot, talking to Klouse, letting
11:18 him get to know me. I'm telling him, ask me questions. What do you not know about
11:21 me? What do you need to know more about? And this ends up creating a soul file.
11:26 So, the soul.md is basically who is your AI and what's its job? And then also
11:29 creates a user.md file, which is basically who am I and what do I do? So,
11:32 right here, I just asked Claus to summarize the soul and user files. And
11:37 it came back with, "I'm Klouse, your executive assistant. My job is to log
11:40 everything, track all my work, communicate casual but concise, be
11:43 resourceful, and the user summary is that you are Nate based in Chicago. Your
11:47 background is nontechnical, graduated May 24, blah blah blah." So, it knows
11:50 all of this kind of stuff about my business, about the platform, even know
11:54 some people on my team and stuff like that. So, here's where I talk about
11:56 having the getting to know you conversation, of course, covering
11:59 everything that you want it to know about your business. And then, like I
12:02 said, prompt the AI to ask you questions and be proactive about it. and the AI
12:06 will basically interview you which makes everything better. So step two is to
12:10 create dedicated accounts for your AI. So I don't think that you should be
12:13 giving Klouse or whatever you call it access to your accounts directly. I
12:17 didn't want Claus to be logged in as you know my main email or my main calendar
12:21 or my main ClickUp cuz you just don't really know what could happen. So
12:24 instead if I give it an email account and I wanted to help me manage my email,
12:27 well I can just forward at things and I can CC him on things and that way he's
12:31 kind of looped in as I would treat an actual like executive assistant. The way
12:35 that I think about it is if you had like a VA or someone that you brought onto
12:38 the team and you wanted to take on a lot of administrative roles, would you give
12:41 them right away like credit card information, bank account information,
12:45 passwords, access to everything? I I probably wouldn't. And on top of that,
12:50 Claus has its own Drive, Docs, Calendar, so that it can see my calendar and it
12:54 can, you know, share files with me and request access, but it's not actually
12:57 going to be deleting things from my environment just in case. Same thing for
13:00 task management. It's got a ClickUp account. It can view my list. It can see
13:03 what's going on in the business, but it's not able to really delete things
13:06 and change anything unless I give it specific access to do so. And I put it
13:11 in its own environment with its own list that I manage my tasks on now that it
13:14 can actually like change things and move them around just to see how it's going
13:17 to work. And then of course, when it comes to credentials, you want to be
13:21 storing this in av file. You don't want to give it straight to Klaus in a
13:27 conversation history in Telegram in the, you know, hub. So you can say, "Hey
13:30 Claus, I don't want you to ever mention an API key in our conversation. I just
13:34 want you to use placeholders and just help me get into myv file and I can drop
13:38 them in there." So you can either do that through your terminal or you can
13:41 just open up in your file explorer and go to thev file and add in all of your
13:46 API keys. So then of course on top of this I just gave it read access to
13:49 pretty much everything. All the APIs that it needs to use for you know
13:52 analytics or appy and things like that. I gave it pretty much read access unless
13:56 it needs to do something. And if it needs to do something, I would set up
13:59 like a spending limit or I would just basically make sure that there's very
14:02 strict permissions. So it can look at my calendar. I can forward it emails. It
14:06 has visibility into my tasks. And it has access to certain things that I have as
14:10 far as like socials, but it can only read and extract information. So it
14:14 can't post on my Twitter or reply to things on YouTube or Twitter, but it can
14:17 grab a bunch of data, which is super helpful. So phase four is about
14:21 developing the proactive mindset. This is where it gets really, really powerful
14:23 and where it starts to feel different than just like a cloud code or, you
14:28 know, an NN AI agent. So, I'm consistently telling it, you know, I'm
14:31 running a lean business. I want to save time. I want you to figure out
14:35 opportunities to be proactive and make my life as easy as possible. Only loop
14:38 me in when you really need something from me. So, you can say something like,
14:41 "Based on everything you know about me, my business, and my goals, what are all
14:44 the ways that you could proactively help me? Don't wait for me to ask, but what
14:48 would you do if your job was to save me time every single day?" And after that,
14:51 it just started brainstorming and it suggested a ton of ideas. So, at the
14:55 moment, the proactive things that I have it doing are this. I've got a morning AI
14:59 news briefing. And it's not just general, it's also like super specific
15:02 to me. I've got ClickUp task summaries every day. So, every morning it looks at
15:05 my ClickUp and it says, "Hey, here's what you've got today." And also, it
15:09 will try to give me things proactively if it can help with any of those. So,
15:12 sometimes it'll do extra research for me. If it's like a content flow or if
15:15 I've got a really busy day, it can maybe suggest moving some of my, you know,
15:18 events on my calendar, things like that. It does email monitoring every single 10
15:22 minutes. So, in its inbox, if I ever send it something, it should respond
15:25 within 10 minutes or it should, you know, kick off research or do something
15:28 else about it in 10 minutes. And with the dashboard notes, it's supposed to
15:31 check this every 5 minutes, but we just saw an example where it thought it was
15:34 every 30. So, there is a bit of a back and forth. I will say the persistent
15:38 memory is cool, but it's not perfect at all because it chooses what memories to
15:42 store and then sometimes it just forgets things. And I'm going to talk about this
15:44 later. You can see that I've got a memory system explainer doc that I had
15:47 to create. But the memory can sometimes be a little frustrating. But honestly,
15:50 that's to be expected. So, you just have to figure out, okay, how can I be
15:54 smarter than this thing and prompt it and use it in the right way because yes,
15:57 it's really smart, but you have to use it the right way because it can also be
16:00 super super dumb. I will admit last night I got pretty frustrated at it and
16:03 I started calling it dumb and these other names because I was just like I
16:07 had had it. Anyways, weekly it does these things as we already talked about
16:10 and now we have the save me time framework. So ask your AI, "What
16:14 currently takes me 20 plus minutes that you could turn into a two-minute
16:17 review?" And asking it questions like that really helps you spot these
16:20 opportunities. So when you actually want to set up those automated workflows,
16:24 it's so so simple. You just ask it to do something. So once it started analyzing
16:27 my YouTube and Twitter, I said, "Oh, that's actually really cool. Can you
16:30 just set this up so that every morning at 7:00 a.m. you do that?" And of
16:34 course, I could come back right here to the actual log and see right here, 7
16:37 a.m. daily pulse has been delivered. And then if I go up to 8 a.m. we have the
16:41 heartbeat dashboard notes none Gmail none clickup summary sent to telegram.
16:45 So this just shows that it is working on a cron and all you had to do was ask it
16:49 in natural language. So it's super cool. Anyways then we got to the point where I
16:52 was like okay I want to build a dashboard and I know that a lot of you
16:55 guys in the last video I showed the dashboard asked about how I built it and
16:58 really it's just a matter of going back and forth but there are some hacks that
17:02 I want to talk about that really make it easier to actually be able to build the
17:04 dashboard out. So first of all you want to have a really clear goal of what
17:07 you're actually building. So my components are a status panel. So I can
17:11 see if the AI is working idle or offline. I can see if there's sub aents
17:14 and I can see what task it's currently working on. Of course, we've got the
17:18 actual cambban board. So the to-do, the in progress, the done. The AI will
17:22 update it as it works. The activity log is like a non-negotiable. You have to
17:26 see every single time you do any action, you have to log it. That's just the
17:30 ability for us to come here, drop in a note, and then have it be able to store
17:33 that as memory. And honestly, this was more so for me testing if it would
17:36 actually work. Because realistically, if I wanted to drop in a note, I would
17:39 probably rather just come to Telegram and shoot off a message. But the tough
17:42 part can be actually getting it to remember to do things specifically like
17:46 updating the Camban board or logging things. And so what I realize here is it
17:50 has to do with the memory. And a lot of times it won't actually store things as
17:54 long-term memory or daily memory. So to contextualize this a little bit, let me
17:55 Memory Explained
17:57 go to the memory system explainer doc. So this basically explains how
18:00 Cloudbot's memory works because it's not super intuitive. We obviously have the
18:04 soul file, the user file, the identity file. So, it's got these that it is able
18:07 to check in on, but then what it starts to do is it creates these different
18:11 logs. So, it creates a daily log, which are raw notes from each day. And that's
18:14 basically like what happened, decisions made, context. It's a journal. But
18:20 again, the tough part is Claus chooses what to put in the daily log journal.
18:23 And so, even sometimes I'm like, hey, log that because you have to remember
18:27 it. Sometimes it won't. The long-term memory are more of like the curated
18:30 highlights about your facts, your lessons learned, your business, things
18:33 like that. And then when you're working on specific projects, it can create
18:37 project specific memories. And so it really is a matter of prompting it like
18:40 make sure you save this to that project memory or save this to your daily log.
18:43 And if you're not doing that, then you're going to be really confused as
18:46 far as like where is it getting into information and why is it confused.
18:50 Okay, I just searched our chat history for the word frustrating because I knew
18:53 I've said it to Klaus many times. And so this is exactly what I'm talking about.
18:57 I said, "This is a problem. I need to be able to know that you have memory that's
18:59 persistent. So if we're having a conversation, you can see what I'm
19:02 talking about because you just sent me a message about 4 minutes ago and all of a
19:06 sudden you forgot that context. So what can I do in your configuration so that
19:08 this doesn't happen again? It's really frustrating. I need to be able to trust
19:11 that I can have an ongoing conversation with you. And then I would basically
19:15 just have to repaste in messages so it could reand context of what we were just
19:19 talking about. And I'm not even exaggerating. It would be like I would
19:22 say hello and then I would say my name is Nate and it would be like okay cool.
19:26 And then I'd say like how's how are you doing? And it would be like what's your
19:28 name? It's like that quick. it would forget what we were talking about. So
19:31 really the key thing to remember is that it's always basically going to wake up
19:35 with no memory. So those files are the memory. And so that's why I'm telling it
19:39 log everything you do. Check these files every time. Put that in the right spot.
19:42 The 5 Hacks
19:42 And so this is a really nice segue into the hacks that I wanted to talk about
19:46 because as I have been obviously playing around with Cloudbot for days now, I've
19:50 learned a lot of things about the way you can communicate with it to actually
19:53 make it do what you want to do. So the first one right here is plan first. This
19:57 one is so important. If you've been using cloud code for a while, you know
20:00 how important it is to use plan mode first and then go ahead and execute. But
20:04 the tough part about planning first is sometimes it'll make an amazing plan.
20:07 I'll spin up three sub agents. It'll build me a beautiful plan and then I'll
20:11 say nice, execute that and it'll be like execute what? And that just makes me
20:15 want to throw my monitor against the wall. So what you could do there is you
20:18 could copy and paste the plan into the next message and shoot it off. Or maybe
20:21 you could just say build a really good plan and then execute. Cuz that still
20:25 increases the quality of the actual solution itself. But another thing I
20:28 started doing which actually helped a lot and I'm I'm going out of order a bit
20:31 here but I was just creating files all the time. And so that's why I really
20:35 wanted to make this dashboard where I could search through all my documents.
20:38 So if I go to like a you know PDF generation I can search through and we
20:41 can see every doc that I've created and I knew that this was going to be huge
20:45 because I ask it to create docs all the time. Basically my hypothesis was if it
20:49 creates me this plan and then it creates a document out of it. It can obviously
20:52 read those documents. So, I create a plan doc and then I say, "Cool, execute
20:56 that plan doc." And then it's like it's it's caught up on all the context right
20:59 there and I don't have to get so frustrated. So, plan first and create
21:02 files for basically everything you're doing. Just tell it to create a doc and
21:05 put it into your dashboard. So, the next one is about proactivity. And so, we
21:08 already talked about this one a little bit, but it's just the idea of thinking
21:11 about how can you make this thing really, really awesome because if you
21:16 take away the proactivity, it's very, very similar to a lot of the other AI
21:19 tools that you probably already use. So proactivity not just meaning doing
21:22 things on a schedule because any automation can check Twitter every 30
21:25 minutes and give you a breakdown. Productivity in the way of like what do
21:28 you know about me and my business? Where do you see what you're doing and where
21:32 I'm doing and figure out how you can save me time and solve problems before I
21:35 even know that those problems exist? So like reading through emails and reading
21:38 through ClickUp. And if you realize that there might be an action item, just go
21:41 ahead and take that action. or when you're running a SWAT analysis for me
21:44 and you have action items, why don't you just take those action items, put them
21:47 in your own to-do list, and then start like banging them out, which is
21:50 basically exactly how we were able to actually have Claus work all night long
21:54 is I said, "Okay, let's brainstorm here. Let's figure out what it would take to
21:58 build a YouTube dashboard and let's basically make like six tasks to do
22:02 this." And they're chronological. And what I want you to do is on your
22:05 heartbeat every 30 minutes or so, pick up a task, contextualize. So read the
22:09 GitHub repository, read all of the past stuff that's going on, and then build.
22:13 And then before you actually shut down on that task, recommit it to GitHub and
22:17 re-update like all the information so that next time you wake up with fresh
22:20 context and you pick up this task again, you can actually understand where you
22:24 just left off. So that is the really important part that takes a little bit
22:27 of more time to orchestrate. And like I said, I got a little frustrated trying
22:30 to figure out how to do it. But it just takes being logical and talking clearly.
22:35 Now, the next thing is discipline. So, it's going to make a ton of mistakes.
22:38 Like, it just will. I've seen it make so many more mistakes than I thought it
22:42 would. But the idea is when it makes a mistake, that's an opportunity to learn,
22:47 both you and Klouse. So, when it makes a mistake, I basically say, "hm, what I
22:51 want you to do is spin up some agents and analyze and audit what you just did.
22:55 Why did that break? Why did it not work? Analyze all the other options. And most
22:59 importantly, tell me which option will make sure that this doesn't ever happen
23:02 again." And then turn all of that into a doc and store that somewhere. And so
23:07 it's like how can I give it how can I let it learn from its mistakes rather
23:11 than just saying you suck, you're doing this bad, try again. And then the fifth
23:15 one was about memory. So all of these hacks kind of, you know, flow into
23:17 memory. And we've talked about memory a little bit, but that's like, like I
23:20 said, that's something that's really frustrating. The reason I threw this on
23:23 here is because it's not just something that's super intuitive. It just takes
23:27 actual time and repetition talking to this thing and understanding how it
23:31 learns and how it sees things. So you'll get better at understanding how to tweak
23:34 your prompts a little bit. so that it's all in one go rather than multiple goes.
23:39 Because what's cool about Klouse is you can shoot off like five messages in a
23:42 row and it will cue them. So it will go ahead and do the first one and then it
23:44 will do the second one and it'll do the third one. But it's doing each of them
23:48 individually. So if I schedule off like three different things, it will do them
23:51 in order. So go ahead and tell me a riddle. Find me a YouTube video today
23:56 about Claudebot that's performing really well. Do some research on the difference
24:00 between pancakes and waffles. find me some exp posts today about the end
24:04 vulnerability thing. But anyways, my point being, if you wanted to send off
24:07 all of these three, but you wanted them to be together, you have to put it in
24:11 the same box and the same message as you send it off. So, it actually treats it
24:14 as one message. Now, I do have one final bonus hack for you guys, which is cloud
24:18 code. If you get to a point where you're confused and you just don't really
24:20 understand what you should be doing with your maybe your configuration or your
24:24 keys or security, then definitely utilize cloud code. And I will say I'm
24:28 not ashamed to admit that I did use cloud code. So I set up a project called
24:33 Klouse and I basically gave cloud code access to this project, right? And it
24:37 can see my backups, my configs, my docs, my scripts, the cloud.mmd file, my
24:41 credentials and it understands how to help me set up the environment. And
24:44 sometimes it actually helps to be able to plan in here first and then take that
24:48 plan to Klouse because if you were going to use like a claude or a perplexity to
24:52 help you set up this stuff anyways, you might as well do it in a claude code
24:54 project that can actually help you create files and navigate through a
24:57 project like this. It just felt like there was way more context, especially
25:01 when I could give it a special system prompt to help me with this project. So,
25:05 if you hit a roadblock, maybe just try throwing some stuff into cloud code and
25:08 Want to Master AI Automations?
3:09 Klaus Showcase
3:11 where I put in my Claus master guide access and capabilities. Right now, this
3:15 is what it's got access to. GitHub, Brave Search, OpenAI, 11 Labs, PDF
3:19 Shift, Appify, ClickUp, Google Workspace, and then some YouTube stuff
3:22 like that. We've got understanding of how the memory system works. It
3:25 understands when to send voice notes, which is only when I send him a voice
3:28 note. Here we have some information on what it should do freely and what would
3:30 require permission. And then you can see what we've built together. So the first
3:33 one is obviously the cloud dashboard, and here's what it does. We've got the
3:36 YouTube intelligence dashboard, and here's what it does. I'm still impressed
3:38 that it built this while I was sleeping. Super cool. It can do branded PDF
3:42 reports and then it has scheduled triggers to do things like daily pulses,
3:46 audits, email monitoring, checking tasks, and obviously this isn't
3:50 completely updated because I added some more stuff in after this guide. All
3:54 right, so before we actually break down exactly what I've done and how you can
3:57 replicate this, let's just go over a few more features in this dashboard so you
4:01 can get a feeling of what I've actually built here with Klouse. So, of course,
4:03 we are talking to Klaus through Telegram. And you can see the last thing
4:06 I just had him make a change. I said, "This is perfect. You're the best." and
4:09 he reacted to my text, which I thought was pretty cool that they can react. And
4:12 then the reason why it says model claopus 4x5 is because I wanted to have
4:16 visibility. So every time I messaged him, I told him to tell me what model
4:20 he's using. So this started as a pretty simple dashboard just so I could track
4:24 what was going on. But I told him, "Hey, your job is to save me time. I want you
4:27 to be proactive, understand my workflows, and suggest things all the
4:31 time." So specifically in this dashboard while we were building it together, I
4:34 said, "Suggest features, whatever you think might work and whatever you think
4:37 will help you work better and help me work better. Just tell me that you're
4:40 going to implement it." And so one of the first things that he did was he
4:43 built this little thing up in the top left so that I can actually see him,
4:46 which was a little bit scary, but also kind of cool because now I can look over
4:50 and I can see if it's idle, then he's not working on anything. But if he's
4:53 thinking or working, then I know that there is something going on. So let me
4:56 go ahead and shoot something off and we will see him spin up and start working.
4:59 All right. So, just to start off here, I'm saying, "Hey, Klaus, I need you to
5:02 spin up a sub agent to do some in-depth research about what's going on today
5:05 with Naden. There was some news about a security thing, and I want you to create
5:09 a report for me." So, obviously, we can see up here that it is thinking. It also
5:12 has spun up a sub agent, which we are able to visually see, which is pretty
5:17 cool. And I will get back to you once we have that report. And of course, it was
5:20 able to add an inrogress task right here. And then when it's done, hopefully
5:23 it should move it over to done. But that's a really cool thing about this
5:26 Cananban board is that if I have a task and I come in here and I just add a new
5:30 to-do, then Claus will automatically pick that up and start working on it. Or
5:34 similarly, if I give him a note down here, I can send something off just for
5:37 him to log it in his memory. Or even if I wanted to take action on it. So like
5:41 here, I said testing if this works. When you see this, tell me a joke. And then
5:44 it shot me a joke in my telegram. And every time it actually acknowledges
5:47 these, it marks them as seen by Clouse. So I know which ones it's processed or
5:51 not. So here's a great example of how it knows about me and my business because
5:55 it came back and said the TLDDR is anyone running any needs to upgrade
5:58 because of this vulnerability. So do you want me to turn this into some content
6:00 for you? Because it could be a quick YouTube short or community post which is
6:04 the kind of thing that your audience needs to hear about fast. But anyways, I
6:07 just said no post for now. I just want the report. So it's throwing that
6:10 together now. All right. So it said that the report is done. If I close out of
6:13 Telegram, we should see that this got moved to done. And the reason why it's a
6:16 different color is because these are actually a priority switch, not like a a
6:20 status. As you can see, if I move this over here, it's just a priority thing.
6:23 So anyways, if I go ahead and click on the docs tab right here in our
6:26 dashboard, you can see that we have a new doc about security, which is end
6:30 security vulnerability report. So now I'm able to actually read through this
6:33 entire report that it made. We've got the date, classification, the exact
6:37 summary, vulnerability details, and we can continue to scroll down and see this
6:41 entire markdown format report that it was able to make for us. And if I wanted
6:43 to change anything about this, I could go ahead and edit it, and then it would
6:47 bring up this editor for us to change something in the doc, save it. So now we
6:50 have a really cool place where every single time that we ask Claus to create
6:53 some sort of file for us, it's able to just drop it in here, and then we can
6:56 look at it and save it. So I thought I'd show you guys off one of those heartbeat
6:59 functionalities. And I'll explain exactly what a heartbeat is and how that
7:02 works in a bit. But I'm going to go ahead and add a new task right here that
7:07 just says, um, make Nate laugh. And I'm just gonna say, "Usually in the
7:11 afternoon, Nate gets sad, so tell him a joke about AI." So, while we're waiting
7:15 for Claus to actually pick that up and shoot me off a joke, let's see what we
7:18 have done here, which are some of our scheduled deliverables, which are
7:21 workflows that happen either daily or weekly. So, for example, with the weekly
7:25 YouTube audits, every single week on I think it's like Monday night or Sunday
7:28 night, I'm going to get one of these YouTube audits. And obviously here, I
7:31 was playing around with some different formats, but this is kind of what they
7:35 look like now. They have my brand information, although this one didn't
7:38 have the logo. Let's try this one. It'll have my logo. It'll have my brand
7:42 guidelines. And it sends me a SWAT analysis and a YouTube channel audit
7:45 about strengths, weaknesses, metrics, and some other things that we need to
7:48 do. Obviously, there's a bit of a spacing issue here, but we get some
7:52 recent content analysis, strategic recommendations. And what you'll notice
7:55 here is that we're signed in here as Klouse. So, Klouse has its own Gmail,
8:00 Drive, Docs, Sheets, its own accounts for pretty much everything because I
8:03 didn't want it to be in my environment. I wanted to treat this as a person that
8:07 I could talk to, forward emails to and have it respond like that. So, it
8:10 created a drive called cloud deliverables and that's where it was
8:12 able to store all of these things like the daily pulses, the SWAT analysis, the
8:17 logos, audits, and YouTube audits. So, obviously security is a huge topic when
8:21 we're talking about Cloudbot. So, the security audit that I had it run for me
8:24 here, this was actually just a Google doc rather than a PDF, but it pointed
8:28 out all these issues. It ranked them for different risk levels, and then I was
8:31 able to have Claus just go ahead and fix all these things. And I'm constantly
8:34 checking with Klouse to make sure that we are completely fine and that our
8:37 port's not exposed and stuff like that. And obviously I'm not like a security
8:40 expert. So I'm doing the best that I can working with Klouse with cloud code with
8:44 perplexity figuring out how I can be safe. But everyone out there that wants
8:47 to experiment with cloudbot, just make sure that you are thinking about
8:50 security and not giving access to tons of different random things if you don't
8:53 really know what's going on. Okay, so what I did is I said, "How often does
8:56 your heartbeat run and what does it actually do?" And it said that my
8:59 heartbeat runs every 30 minutes. And every time that my heartbeat runs, it
9:03 basically wakes up Klaus. And Claus will make sure that both the dashboards are
9:05 running. It will do a sync. It will check for notes, which you can see it
9:08 actually just moved the make Nate laugh from in progress to done, I believe. As
9:13 you can see, it will monitor X and Twitter. It will do a daily pulse. And
9:16 the idea is that I wake up with no memory. So when I have my heartbeats, I
9:19 stay proactive and I check in on things and do things in the background for you.
9:23 So anyways, then it said, "Yep, just saw that task Q. Make Nate laugh high
9:26 priority." And the joke is, what do you call an AI that finally passes the
9:30 touring test? Unemployed, the humans just move the goalpost again. I don't
9:32 Build Your Own Klaus
9:32 really get that one. So, let's actually move on to the next part, which is how
9:35 can you create your own version of Klouse? So, the first thing, of course,
9:38 you want to do is set up Cloudbot. Whether you're going to do that on a Mac
9:40 Mini or whether you're going to do that on VPS, I will tag a full setup guide
9:45 video right up here that I made a few days ago and you can get that set up on
9:49 a VPS. So what I actually did is I came into Klaus and I was like, "Hey, I want
9:53 to teach my YouTube channel how they can basically build exactly what I built for
9:58 you and how we configure you." So I had it create this guide called build your
10:01 own Klouse, a step-by-step guide to creating an AI executive assistant.
10:05 Clouse handles daily briefings, monitors my business, runs security audits, and
10:08 proactively services opportunities, and you can build your own version. So time
10:11 investment, 4 to 6 hours over the weekend to get the core system running,
10:15 and cost about 50 bucks a month. I'm going to say that it's going to be
10:17 higher than that. So, I've obviously spent lots of hours playing around with
10:20 Cloudbot, doing different things, and I spun up this one to turn this into an
10:23 executive assistant. And you can see with just this specific Cloudbot that
10:27 I'm using, Klouse, I've used almost a quarter of a billion tokens in the past
10:32 just 3 days, which translate to about $223 in actual money. But keep in mind, all
10:38 of this has been done with Opus 4.5. So, if you went to Sonnet 4.5 or if you use
10:42 a locally hosted model, your cost would be less. And the reason why I'm not
10:46 using like my max plan, the $200 a month plan for which I would get way more
10:50 tokens with is because Anthropic has been banning accounts left and right for
10:53 violating terms of service. So if you are currently using Cloudbot with a max
10:56 plan or any Cloud plan at all, you have to switch it over to API. So anyways,
11:00 what happens when you actually get this set up? Well, you have to get onboarded.
11:03 So the first thing it's going to start to do is it's going to start to ask you
11:06 questions. So you have to give it an identity. You have to talk about who are
11:09 you, what its role is, what you want it to do, how it should behave, and you
11:12 should also just brain dump everything about you and your business. So, I spent
11:15 like almost an hour just talking to Claudebot, talking to Klouse, letting
11:18 him get to know me. I'm telling him, ask me questions. What do you not know about
11:21 me? What do you need to know more about? And this ends up creating a soul file.
11:26 So, the soul.md is basically who is your AI and what's its job? And then also
11:29 creates a user.md file, which is basically who am I and what do I do? So,
11:32 right here, I just asked Claus to summarize the soul and user files. And
11:37 it came back with, "I'm Klouse, your executive assistant. My job is to log
11:40 everything, track all my work, communicate casual but concise, be
11:43 resourceful, and the user summary is that you are Nate based in Chicago. Your
11:47 background is nontechnical, graduated May 24, blah blah blah." So, it knows
11:50 all of this kind of stuff about my business, about the platform, even know
11:54 some people on my team and stuff like that. So, here's where I talk about
11:56 having the getting to know you conversation, of course, covering
11:59 everything that you want it to know about your business. And then, like I
12:02 said, prompt the AI to ask you questions and be proactive about it. and the AI
12:06 will basically interview you which makes everything better. So step two is to
12:10 create dedicated accounts for your AI. So I don't think that you should be
12:13 giving Klouse or whatever you call it access to your accounts directly. I
12:17 didn't want Claus to be logged in as you know my main email or my main calendar
12:21 or my main ClickUp cuz you just don't really know what could happen. So
12:24 instead if I give it an email account and I wanted to help me manage my email,
12:27 well I can just forward at things and I can CC him on things and that way he's
12:31 kind of looped in as I would treat an actual like executive assistant. The way
12:35 that I think about it is if you had like a VA or someone that you brought onto
12:38 the team and you wanted to take on a lot of administrative roles, would you give
12:41 them right away like credit card information, bank account information,
12:45 passwords, access to everything? I I probably wouldn't. And on top of that,
12:50 Claus has its own Drive, Docs, Calendar, so that it can see my calendar and it
12:54 can, you know, share files with me and request access, but it's not actually
12:57 going to be deleting things from my environment just in case. Same thing for
13:00 task management. It's got a ClickUp account. It can view my list. It can see
13:03 what's going on in the business, but it's not able to really delete things
13:06 and change anything unless I give it specific access to do so. And I put it
13:11 in its own environment with its own list that I manage my tasks on now that it
13:14 can actually like change things and move them around just to see how it's going
13:17 to work. And then of course, when it comes to credentials, you want to be
13:21 storing this in av file. You don't want to give it straight to Klaus in a
13:27 conversation history in Telegram in the, you know, hub. So you can say, "Hey
13:30 Claus, I don't want you to ever mention an API key in our conversation. I just
13:34 want you to use placeholders and just help me get into myv file and I can drop
13:38 them in there." So you can either do that through your terminal or you can
13:41 just open up in your file explorer and go to thev file and add in all of your
13:46 API keys. So then of course on top of this I just gave it read access to
13:49 pretty much everything. All the APIs that it needs to use for you know
13:52 analytics or appy and things like that. I gave it pretty much read access unless
13:56 it needs to do something. And if it needs to do something, I would set up
13:59 like a spending limit or I would just basically make sure that there's very
14:02 strict permissions. So it can look at my calendar. I can forward it emails. It
14:06 has visibility into my tasks. And it has access to certain things that I have as
14:10 far as like socials, but it can only read and extract information. So it
14:14 can't post on my Twitter or reply to things on YouTube or Twitter, but it can
14:17 grab a bunch of data, which is super helpful. So phase four is about
14:21 developing the proactive mindset. This is where it gets really, really powerful
14:23 and where it starts to feel different than just like a cloud code or, you
14:28 know, an NN AI agent. So, I'm consistently telling it, you know, I'm
14:31 running a lean business. I want to save time. I want you to figure out
14:35 opportunities to be proactive and make my life as easy as possible. Only loop
14:38 me in when you really need something from me. So, you can say something like,
14:41 "Based on everything you know about me, my business, and my goals, what are all
14:44 the ways that you could proactively help me? Don't wait for me to ask, but what
14:48 would you do if your job was to save me time every single day?" And after that,
14:51 it just started brainstorming and it suggested a ton of ideas. So, at the
14:55 moment, the proactive things that I have it doing are this. I've got a morning AI
14:59 news briefing. And it's not just general, it's also like super specific
15:02 to me. I've got ClickUp task summaries every day. So, every morning it looks at
15:05 my ClickUp and it says, "Hey, here's what you've got today." And also, it
15:09 will try to give me things proactively if it can help with any of those. So,
15:12 sometimes it'll do extra research for me. If it's like a content flow or if
15:15 I've got a really busy day, it can maybe suggest moving some of my, you know,
15:18 events on my calendar, things like that. It does email monitoring every single 10
15:22 minutes. So, in its inbox, if I ever send it something, it should respond
15:25 within 10 minutes or it should, you know, kick off research or do something
15:28 else about it in 10 minutes. And with the dashboard notes, it's supposed to
15:31 check this every 5 minutes, but we just saw an example where it thought it was
15:34 every 30. So, there is a bit of a back and forth. I will say the persistent
15:38 memory is cool, but it's not perfect at all because it chooses what memories to
15:42 store and then sometimes it just forgets things. And I'm going to talk about this
15:44 later. You can see that I've got a memory system explainer doc that I had
15:47 to create. But the memory can sometimes be a little frustrating. But honestly,
15:50 that's to be expected. So, you just have to figure out, okay, how can I be
15:54 smarter than this thing and prompt it and use it in the right way because yes,
15:57 it's really smart, but you have to use it the right way because it can also be
16:00 super super dumb. I will admit last night I got pretty frustrated at it and
16:03 I started calling it dumb and these other names because I was just like I
16:07 had had it. Anyways, weekly it does these things as we already talked about
16:10 and now we have the save me time framework. So ask your AI, "What
16:14 currently takes me 20 plus minutes that you could turn into a two-minute
16:17 review?" And asking it questions like that really helps you spot these
16:20 opportunities. So when you actually want to set up those automated workflows,
16:24 it's so so simple. You just ask it to do something. So once it started analyzing
16:27 my YouTube and Twitter, I said, "Oh, that's actually really cool. Can you
16:30 just set this up so that every morning at 7:00 a.m. you do that?" And of
16:34 course, I could come back right here to the actual log and see right here, 7
16:37 a.m. daily pulse has been delivered. And then if I go up to 8 a.m. we have the
16:41 heartbeat dashboard notes none Gmail none clickup summary sent to telegram.
16:45 So this just shows that it is working on a cron and all you had to do was ask it
16:49 in natural language. So it's super cool. Anyways then we got to the point where I
16:52 was like okay I want to build a dashboard and I know that a lot of you
16:55 guys in the last video I showed the dashboard asked about how I built it and
16:58 really it's just a matter of going back and forth but there are some hacks that
17:02 I want to talk about that really make it easier to actually be able to build the
17:04 dashboard out. So first of all you want to have a really clear goal of what
17:07 you're actually building. So my components are a status panel. So I can
17:11 see if the AI is working idle or offline. I can see if there's sub aents
17:14 and I can see what task it's currently working on. Of course, we've got the
17:18 actual cambban board. So the to-do, the in progress, the done. The AI will
17:22 update it as it works. The activity log is like a non-negotiable. You have to
17:26 see every single time you do any action, you have to log it. That's just the
17:30 ability for us to come here, drop in a note, and then have it be able to store
17:33 that as memory. And honestly, this was more so for me testing if it would
17:36 actually work. Because realistically, if I wanted to drop in a note, I would
17:39 probably rather just come to Telegram and shoot off a message. But the tough
17:42 part can be actually getting it to remember to do things specifically like
17:46 updating the Camban board or logging things. And so what I realize here is it
17:50 has to do with the memory. And a lot of times it won't actually store things as
17:54 long-term memory or daily memory. So to contextualize this a little bit, let me
17:57 go to the memory system explainer doc. So this basically explains how
18:00 Cloudbot's memory works because it's not super intuitive. We obviously have the
18:04 soul file, the user file, the identity file. So, it's got these that it is able
18:07 to check in on, but then what it starts to do is it creates these different
18:11 logs. So, it creates a daily log, which are raw notes from each day. And that's
18:14 basically like what happened, decisions made, context. It's a journal. But
18:20 again, the tough part is Claus chooses what to put in the daily log journal.
18:23 And so, even sometimes I'm like, hey, log that because you have to remember
18:27 it. Sometimes it won't. The long-term memory are more of like the curated
18:30 highlights about your facts, your lessons learned, your business, things
18:33 like that. And then when you're working on specific projects, it can create
18:37 project specific memories. And so it really is a matter of prompting it like
18:40 make sure you save this to that project memory or save this to your daily log.
18:43 And if you're not doing that, then you're going to be really confused as
18:46 far as like where is it getting into information and why is it confused.
18:50 Okay, I just searched our chat history for the word frustrating because I knew
18:53 I've said it to Klaus many times. And so this is exactly what I'm talking about.
18:57 I said, "This is a problem. I need to be able to know that you have memory that's
18:59 persistent. So if we're having a conversation, you can see what I'm
19:02 talking about because you just sent me a message about 4 minutes ago and all of a
19:06 sudden you forgot that context. So what can I do in your configuration so that
19:08 this doesn't happen again? It's really frustrating. I need to be able to trust
19:11 that I can have an ongoing conversation with you. And then I would basically
19:15 just have to repaste in messages so it could reand context of what we were just
19:19 talking about. And I'm not even exaggerating. It would be like I would
19:22 say hello and then I would say my name is Nate and it would be like okay cool.
19:26 And then I'd say like how's how are you doing? And it would be like what's your
19:28 name? It's like that quick. it would forget what we were talking about. So
19:31 really the key thing to remember is that it's always basically going to wake up
19:35 with no memory. So those files are the memory. And so that's why I'm telling it
19:39 log everything you do. Check these files every time. Put that in the right spot.
19:42 And so this is a really nice segue into the hacks that I wanted to talk about
19:46 because as I have been obviously playing around with Cloudbot for days now, I've
19:50 learned a lot of things about the way you can communicate with it to actually
19:53 make it do what you want to do. So the first one right here is plan first. This
19:57 one is so important. If you've been using cloud code for a while, you know
20:00 how important it is to use plan mode first and then go ahead and execute. But
20:04 the tough part about planning first is sometimes it'll make an amazing plan.
20:07 I'll spin up three sub agents. It'll build me a beautiful plan and then I'll
20:11 say nice, execute that and it'll be like execute what? And that just makes me
20:15 want to throw my monitor against the wall. So what you could do there is you
20:18 could copy and paste the plan into the next message and shoot it off. Or maybe
20:21 you could just say build a really good plan and then execute. Cuz that still
20:25 increases the quality of the actual solution itself. But another thing I
20:28 started doing which actually helped a lot and I'm I'm going out of order a bit
20:31 here but I was just creating files all the time. And so that's why I really
20:35 wanted to make this dashboard where I could search through all my documents.
20:38 So if I go to like a you know PDF generation I can search through and we
20:41 can see every doc that I've created and I knew that this was going to be huge
20:45 because I ask it to create docs all the time. Basically my hypothesis was if it
20:49 creates me this plan and then it creates a document out of it. It can obviously
20:52 read those documents. So, I create a plan doc and then I say, "Cool, execute
20:56 that plan doc." And then it's like it's it's caught up on all the context right
20:59 there and I don't have to get so frustrated. So, plan first and create
21:02 files for basically everything you're doing. Just tell it to create a doc and
21:05 put it into your dashboard. So, the next one is about proactivity. And so, we
21:08 already talked about this one a little bit, but it's just the idea of thinking
21:11 about how can you make this thing really, really awesome because if you
21:16 take away the proactivity, it's very, very similar to a lot of the other AI
21:19 tools that you probably already use. So proactivity not just meaning doing
21:22 things on a schedule because any automation can check Twitter every 30
21:25 minutes and give you a breakdown. Productivity in the way of like what do
21:28 you know about me and my business? Where do you see what you're doing and where
21:32 I'm doing and figure out how you can save me time and solve problems before I
21:35 even know that those problems exist? So like reading through emails and reading
21:38 through ClickUp. And if you realize that there might be an action item, just go
21:41 ahead and take that action. or when you're running a SWAT analysis for me
21:44 and you have action items, why don't you just take those action items, put them
21:47 in your own to-do list, and then start like banging them out, which is
21:50 basically exactly how we were able to actually have Claus work all night long
21:54 is I said, "Okay, let's brainstorm here. Let's figure out what it would take to
21:58 build a YouTube dashboard and let's basically make like six tasks to do
22:02 this." And they're chronological. And what I want you to do is on your
22:05 heartbeat every 30 minutes or so, pick up a task, contextualize. So read the
22:09 GitHub repository, read all of the past stuff that's going on, and then build.
22:13 And then before you actually shut down on that task, recommit it to GitHub and
22:17 re-update like all the information so that next time you wake up with fresh
22:20 context and you pick up this task again, you can actually understand where you
22:24 just left off. So that is the really important part that takes a little bit
22:27 of more time to orchestrate. And like I said, I got a little frustrated trying
22:30 to figure out how to do it. But it just takes being logical and talking clearly.
22:35 Now, the next thing is discipline. So, it's going to make a ton of mistakes.
22:38 Like, it just will. I've seen it make so many more mistakes than I thought it
22:42 would. But the idea is when it makes a mistake, that's an opportunity to learn,
22:47 both you and Klouse. So, when it makes a mistake, I basically say, "hm, what I
22:51 want you to do is spin up some agents and analyze and audit what you just did.
22:55 Why did that break? Why did it not work? Analyze all the other options. And most
22:59 importantly, tell me which option will make sure that this doesn't ever happen
23:02 again." And then turn all of that into a doc and store that somewhere. And so
23:07 it's like how can I give it how can I let it learn from its mistakes rather
23:11 than just saying you suck, you're doing this bad, try again. And then the fifth
23:15 one was about memory. So all of these hacks kind of, you know, flow into
23:17 memory. And we've talked about memory a little bit, but that's like, like I
23:20 said, that's something that's really frustrating. The reason I threw this on
23:23 here is because it's not just something that's super intuitive. It just takes
23:27 actual time and repetition talking to this thing and understanding how it
23:31 learns and how it sees things. So you'll get better at understanding how to tweak
23:34 your prompts a little bit. so that it's all in one go rather than multiple goes.
23:39 Because what's cool about Klouse is you can shoot off like five messages in a
23:42 row and it will cue them. So it will go ahead and do the first one and then it
23:44 will do the second one and it'll do the third one. But it's doing each of them
23:48 individually. So if I schedule off like three different things, it will do them
23:51 in order. So go ahead and tell me a riddle. Find me a YouTube video today
23:56 about Claudebot that's performing really well. Do some research on the difference
24:00 between pancakes and waffles. find me some exp posts today about the end
24:04 vulnerability thing. But anyways, my point being, if you wanted to send off
24:07 all of these three, but you wanted them to be together, you have to put it in
24:11 the same box and the same message as you send it off. So, it actually treats it
24:14 as one message. Now, I do have one final bonus hack for you guys, which is cloud
24:18 code. If you get to a point where you're confused and you just don't really
24:20 understand what you should be doing with your maybe your configuration or your
24:24 keys or security, then definitely utilize cloud code. And I will say I'm
24:28 not ashamed to admit that I did use cloud code. So I set up a project called
24:33 Klouse and I basically gave cloud code access to this project, right? And it
24:37 can see my backups, my configs, my docs, my scripts, the cloud.mmd file, my
24:41 credentials and it understands how to help me set up the environment. And
24:44 sometimes it actually helps to be able to plan in here first and then take that
24:48 plan to Klouse because if you were going to use like a claude or a perplexity to
24:52 help you set up this stuff anyways, you might as well do it in a claude code
24:54 project that can actually help you create files and navigate through a
24:57 project like this. It just felt like there was way more context, especially
25:01 when I could give it a special system prompt to help me with this project. So,
25:05 if you hit a roadblock, maybe just try throwing some stuff into cloud code and
25:08 seeing what it can do. But anyways, there's nothing on the canban board
25:11 anymore, which means I need to load up Claus with some tasks. So, I appreciate
25:14 you guys watching the video. And if you learned something new or you enjoyed,
25:16 please give it a like. It definitely helps me out a ton. If you guys want to
25:19 dive deeper into this kind of stuff or you appreciate this style of video, then
25:21 definitely check out my plus community. The link for that is down in the
25:24 description. We've got over 3,000 members who are building businesses with
25:26 AI. So, if you want to be surrounded by that energy, then check it out. Anyways,
25:29 that's going to do it for this one. So, if you guys enjoyed or you learned
25:31 something new, please give it a like. It definitely helps me out a ton. And as
25:34 always, I appreciate you guys making it to the end of the video. I'll see you on
