0:06 All right, I think we're live. I'm going to wait for it to show up on my other
0:10 monitor here before I get started just But, uh, I think we're good. Okay, nice.
0:19 There we go. Now I got the confirmation. I have to wait for the delay myself to
0:22 make sure we're good. It's been a bit since I've done a live stream now.
0:28 um probably let's see I think August halfway through August was the first or
0:31 the last time. So it's been about a month and a half now almost. So yeah,
0:37 welcome everyone to the next live stream here doing a live AI agent build. It's
0:42 something that I've never done before. I did something kind of similar with N8N
0:47 like way back in December of last year. uh built an agent live, but like that
0:52 wasn't nearly as complete as what I hope to do with you all in this stream today.
0:56 So, I'm really looking forward to this. Before we really get into stuff, I'm
0:59 going to give a couple minutes for people to join and then also uh just
1:04 chat with you guys a little bit and uh talk about what's what's up with the
1:08 stream for today. So, yeah, on my right hand monitor, that's where I'll have
1:11 like any support materials I need to like look at as I'm building. Left
1:15 monitor is where I'll have the chat. So, if you see me looking around, that's
1:18 just because I'm looking at my other monitors here. But yeah, I've got the
1:24 the chat open up on um my lefthand window. And I've done this before in all
1:27 my other live streams in the past. At any point, I can click on anything, any
1:34 message, and then I can have it pop up So, if I go to my face here, you guys
1:59 can see me again, but I don't have the live stream thing here. Um, but if I go
2:05 to my scene, so you guys can see my screen now. You can see it. So, I guess
2:08 I'll just swap to this now so you guys can see my screen. I didn't have it
2:12 configured for when it was like just my face. But yeah, so you can see that I
2:16 can select any message here and then I can display it. So for any questions
2:20 that we've got, um I can like have it pop up on the screen as I'm answering it
2:23 so everyone knows what I'm talking about exactly. So yeah. All right, that's
2:27 that's how things are going to roll here. Uh huge Marley said, "Join
2:33 Dynamus. It's amazing." I appreciate that very much. Means a lot. Uh ever
2:39 looked into AI for video analysis? Is it any good? That's a good question. I have
2:45 not actually. Um, and video analysis can mean a lot of different things. If you
2:48 mean like analyzing it frame by frame for something like to give you feedback
2:52 or to uh look for like copyright material, I haven't done that before.
2:55 Um, but I know that YouTube does a lot of that under the hood actually.
3:00 So, yeah. Great. Looks good. Okay. Glad to hear it. Yeah. And I'll go ahead and
3:06 uh clear this now. Okay. Cool. So, yeah. Uh, how many people we got in the stream
3:10 right now? We got 124 watching live. Okay, sweet. I appreciate that. So,
3:13 thank you everyone for being here. I'm really excited for the stream today. So,
3:17 we're going to build a full AI agent. I want this to be like super practical and
3:21 educational, but at the same time, I want I want it to be rather casual.
3:25 We're just have going to have a good time. And I'm building something from
3:30 scratch, which spoiler, it does mean that I'm going to probably fumble in a
3:33 couple of parts as we're going through the build. It's not going to be
3:37 extremely smooth like uh when I show building an agent in a YouTube video.
3:41 And for anyone that's in the Dynamis community, you know what I'm talking
3:44 about where when I do the live builds there that um it and it's kind of like
3:48 actually educational in its own right to see me mess up on things and have to
3:51 work through issues, but it's the kind of thing that we don't have in a YouTube
3:55 video. So, yeah. And then throughout the stream, I'm going to really try to
3:59 remember to drink some water because otherwise I my voice gets really halfway through. So,
4:08 can't forget to do that. So, yeah. Um, so another thing that I want to say is
4:13 that throughout the process here of building our AI agent, which we'll get
4:17 into in just a second here, and this is the the agent that we're going to be
4:20 building, and maybe you saw this diagram if you uh saw my little teaser
4:24 announcement on Friday for our stream today. I'm going to be going through an
4:31 AI coding process that I have started to cover in a new agentic coding course
4:36 that I've released in Dynamis. And so I'm just going to mention the course a
4:40 couple of times through the live stream. I think you guys will find that fair. I
4:44 just launched the first part of this course on Friday. And so I just want to
4:47 call it out a couple of times, especially how it relates to some of the
4:51 concepts that I'm showing you all um during this stream as we build an agent
4:55 together. So, I'm not going to go into a ton of detail on every part of our
5:00 development workflow with AI coding assistance, but if you want a lot of
5:04 detail and you really want to learn what it looks like to build your own reliable
5:08 and repeatable systems for getting insane results with AI coding
5:12 assistance, then definitely check out dynamis.ai. And the last thing I'll say
5:16 on this before we get right into building the agent, and I'll put this in
5:22 the chat as well, and then I'll also um highlight it here so you can see it. If
5:28 you um sign up for the community during this live stream, this is only going to
5:32 be available while the live stream is going on. I've got a special discount
5:35 for you guys for both the monthly and yearly memberships of Dynamus. So, you
5:39 just go to dynamis.ai. I'll show you the the site right now. You just go to
5:44 dynamus.ai AI and you can join the community. It's currently a 10% discount
5:51 and 25% discount for um anyone joining right now. But for you guys, that's it's
5:55 now going to be 20% for annual or for monthly and then 40% off for annual if
6:00 you join during this live stream. Um, and hopefully as we go through building
6:04 the agent today, uh, and I talk about like how I dive deeper into certain
6:08 things within the new Agentic coding course that's being released, uh,
6:12 that'll peique your interest. That's the it's it's like, you know, kind of a side
6:15 goal that I have for you in the stream here. But, of course, the primary goal
6:20 is to give you guys a ton of value building this agent live. Um, so yeah,
6:25 I'll go ahead and so like this message that I have up right here, I'll share
6:28 that once in a while throughout the stream just to give you guys the
6:31 opportunity if you're interested. Um, but yeah, I'll go ahead and uh take that
6:35 down right now. And uh Oh yeah, we'll just show it later on. So, okay, cool. Joe said, "It's
6:40 great already diving into the new course." Cool. I appreciate it, Joe.
6:43 Yeah. Yeah, we got a lot of people in the community that are already diving
6:46 in, excited about it. So, yeah, that means a lot. Um, yeah, one more question
6:51 I'll answer quick. How is your course for non-programmers? Yeah, that I really
6:55 appreciate you asking this. So, the course is designed, it's made for
7:00 beginners, but designed for mastery. And what I mean by that is like you can jump
7:03 in even if you don't have programming experience. And I talk about how you can
7:07 use AI coding assistance not just to write the code, but also to help you
7:11 understand concepts and even like help you install the things that you need for
7:15 the course like Git for example, um or your IDE. So, yeah. But then also like
7:20 we go pretty deep throughout the course as well, like really advanced concepts
7:23 if you want to, you know, truly be in like the top 5% and beyond of people
7:27 that are building these systems around AI coding assistance. So yeah, cool. I I
7:33 appreciate that a lot, Sebastian. Thank you. Yeah. So with that, uh let's go
7:39 ahead now and get into building our agent. And I'm really excited for this
7:45 because I've honestly wanted to create this AI agent for my channel for a long
7:52 time now. And I haven't gotten around to it just because uh there's, you know, a
7:56 lot that goes into really refining this agent and productionizing it for all of
7:59 you. Like we're not going to like finish it in this stream, but we're going to
8:03 get a really good starting point. Um and I've just been like so busy with other
8:06 content that I'm creating on YouTube and everything I'm doing at Dynamis. Uh, but
8:10 now is the time to build this agent. It is a personal AI coach for both AI
8:16 agents and AI coding. And so the end results and hopefully what we can get to
8:22 in this live stream is to have a chat interface that looks kind of like this.
8:27 So, you know, it's kind of like a a chat GPT style interface that you sign into.
8:31 We'll have superbase authentication under the hood. I want to build that as
8:35 well. And you can talk to this agent like I can say um what is the latest
8:40 news on claude code or I could say claude skills or whatever right like
8:44 asking it questions just like you would chat GBT but under the hood it's going
8:49 to be an advanced rag agent that is trained on all of my YouTube content and
8:54 then later on I want to include a lot of other things as well like open source
8:58 repositories that I'm working on any like articles that I want to share like
9:01 if I'm researching things and I'm like oh this is like so good to include in
9:04 the agent agent, then I can just like dump it in and the rag pipeline will
9:07 pick it up and include it right in the knowledge. So, because like the two
9:12 things I focus on are building agents and using AI coding assistance, that's
9:15 kind of going to be like what the agent is an expert in, right? So, like trained
9:19 on all of my knowledge and using advanced rag techniques to make it
9:23 really accurate um and precise. And so, this chat interface that you're
9:27 looking at right here is the one that I build in the Dynamis AI um agent mastery
9:32 course. And so we'll be using this and building the agent today uh in a way
9:36 where like with the backend for the agent, it can like plug right into this
9:39 front end. So I really do hope to have like a full stack application for the
9:44 agent uh by the end of our stream today. And we'll be using a few different
9:49 technologies to get the job done here. Obviously YouTube, we're going to be
9:53 pulling all of our all of my YouTube videos uh for the rag pipeline. So
9:57 that's going to be our knowledge source for our agent to start things off. We'll
10:02 be using paidantic AI to build our AI agent in Python. Superbase will be our
10:07 database because I also want to use that for authentication.
10:10 And then last but not least, uh I want to leverage Dockling, which I covered on
10:15 my channel recently. It's a a rag library that helps with uh data
10:18 extraction from different file types. And then also a couple of chunking
10:22 strategies that they give us. So, a way to like take all of our, in this case,
10:27 YouTube transcripts and find a way to split them up into different chunks so
10:30 that we can, you know, insert information efficiently into our
10:35 knowledge base. So, we'll be using a strategy for chunking called hybrid
10:40 chunking, which I'm really excited to incorporate dockling for this
10:44 specifically because um hybrid chunking is one of the things that has like
10:48 seriously leveled up my rag game recently. I've been trying a lot of
10:51 other strategies before like even some more like custom chunking ones and
10:55 nothing has worked as well as this. Like it's actually insane. So we'll get to
11:00 that actually towards the start of the stream. Uh because we will be building
11:05 the rag pipeline first. And so this stream pretty pretty simply is going to
11:10 be like let's build the rag pipeline for our YouTube data and using dockling and
11:14 storing things in superbase and then we will build our agent with padantic AI.
11:20 So, it's going to be kind of two loops of of uh using the AI coding assistant
11:25 process that I'll cover with you guys here. So, this is one of the things that
11:29 I get into in a lot more detail in the agentic coding course. But my typical
11:34 process for working with AI coding assistants is I go through this mental
11:41 model of planning then implementing then validating and then kind of doing that
11:44 in a loop for all the features that I want in my codebase. And so like for
11:49 example when we build the rag pipeline to bring in YouTube content into our
11:54 superbase knowledge base that's going to be one iteration of planning then
11:57 implementing and validating. That's what we're going to do together in this
12:01 stream is I'll go through my process of exploring my tech stack and architecture
12:05 really defining things at a high level with just a very unstructured
12:10 conversation with ourselves and then the AI coding assistant as well. And then
12:14 I'll show you what that looks like. Transitioning into a very structured
12:19 plan where we're outlining a complete plan of attack for our coding assistant
12:23 with things like goals and success criteria, the task list that we're going
12:27 to be integrating with archon. A lot of fun stuff in store for that, right? And
12:32 so that's our planning phase. And then once we have our structured plan, we'll
12:36 implement that. And this is where we really delegate all the coding to the AI
12:39 coding assistant. And then we just verify things ourself in the last phase,
12:44 the validation phase. So that really is the loop that we have here that I'm
12:47 covering in this stream. And then when we iterate that would be like for us
12:50 like once we're done with the rag pipeline, then we'll go through
12:53 planning, implementing, and validating again this time for the AI agent, right?
12:58 So like there's the two core parts to the application that we're building
13:01 together. And we're going to do the piv loop as I like to call it, right? Like
13:05 piv loop for each of those parts. So that's the plan. And um yeah, I'm
13:13 going to try to make this like kind of interactive as well. So maybe I'll like
13:17 ask you guys some questions uh just as I am going through things. Um like for
13:21 example, as we're going through the planning here, there's definitely like a
13:25 lot of considerations we need to think through up front for our rag pipeline.
13:29 Like how do we want to ingest our YouTube information? Like do we care
13:33 about getting timestamps or do we not want that? Right? There's those little
13:36 decisions that it's actually really important to make up front because um
13:40 really the goal of the planning phase and this is like one of the most
13:43 important things I want you guys to get out of this live stream and because this
13:46 applies to building this rag agent, applies to building really anything. One
13:50 of the primary goals of the planning phase is to in your structure plan, hear
13:55 me out on this. In your structure plan, you want to reduce the number of
14:01 assumptions your AI coding assistant is making in implementation as much as
14:06 possible. And so that's why it's so important for us to spend time upfront
14:10 planning and thinking through even all the little things. And so that's what I
14:14 want to cover with you guys first. And so before we even get into our AI coding
14:18 assistant, which we'll do that in a sec, I want to just do like a little bit of
14:23 scratchpad planning with you guys. And I think this is like one of those things
14:26 that I don't cover enough on my channel because it definitely fits a live stream
14:32 format to do this kind of thing much better. And so we'll just plan some
14:35 things together and I'll ask you guys some questions about um you know like
14:40 the things that we should consider as we're building this agent and making it
14:44 interactive in that way as well. And yeah, I saw the question um what uh tool
14:49 is this? Thanks Mike for giving the answer as well. This is Excaladraw that
14:54 I have as a plug-in integrated directly within Obsidian. So, Obsidian is a free and local uh
15:00 note-taking like knowledge management platform that I've started to transition
15:03 a lot of things to. I was using a lot of like Google Drive and Aana and stuff
15:08 before. Um, absolutely loving it. And then Excaladraw is a tool that lives
15:12 outside of Obsidian as well, but I have a plugin that brings it directly in
15:17 here. So I can add it add obsidian um or excuse me I can add excaladraw diagrams
15:22 and things like directly within all my other notes and documents and even have
15:27 these diagrams reference all of my notes and knowledge and vice versa. That's
15:35 Excuse me. So yeah, and then the laser pointer. I just press K. Like that's my
15:38 hotkey. And then I have this super cool laser pointer um that uh yeah I've
15:43 started to use this like all the time. Like if you go and you take the aenta
15:47 coding course that I have in Dynamus now you'll see me use this a lot just as I
15:52 explain things like the piv loop for example. So all right so we're going to go I'm
15:58 going to go to white text here. Piv loop number one is going to be our rag
16:03 pipeline. All right. And then piv loop number two is going to be our rag agent. Let's kind
16:12 of uh let's zoom out before we even get into the AI coding assistant. I really
16:16 want to zoom out here and plan things with you guys because this is what
16:21 building real software looks like. You you need to spend time doing this kind
16:24 of thing before you get into the actual coding. Um so I'm actually really
16:27 excited to do this with you guys. So all right, piv loop number one. We've got
16:31 our rag pipeline and I'll I'll just kind of like show off what you can do in
16:35 Excal. Just like making things look kind of nice here. So all right, we got our
16:40 two piv loops. And so for our rag pipeline specifically, I just want to
16:44 think through like what are the things that we should consider before we even
16:47 talk to the AI coding assistant about what we want to build. So we are
16:53 building a YouTube video rag pipeline. Okay. And um what's what should our text
16:58 stack be? Right? Like what tool should we even consider using? Uh well, first
17:02 of all, I mean this part I kind of already have defined, right? Like I
17:05 thought of this definitely ahead of time. We're going to be building some
17:07 things from scratch, but like right now I know that I want to use Python um for
17:11 my programming language for the pipeline. Uh specifically, the tool that
17:17 I'm going to be using to get YouTube transcripts, I'm going to use something
17:22 called Super Data. I've used this in the past. It's not free, but it's really
17:26 cheap. Um because if you don't use a service like this to get YouTube
17:31 transcripts, then um you risk hitting a lot of like rate limits and IP blockings
17:35 and things like that. Um, and so yeah, let me go to the homepage here just so
17:39 you can see. Oh, I spelled it wrong. All right, super data.ai. There we go. And
17:43 this works for more than just YouTube, by the way. And you get your first 100
17:47 requests for free. So, I debated a lot like should I use super data or should I
17:53 just use the YouTube API directly because the YouTube API would be free.
17:58 Uh, but the problem is, and like this is really annoying because like I hate
18:01 having to pay for something just to get YouTube transcripts, but like in the
18:05 end, I think this is so worth it because the YouTube API, you're only able to get
18:10 the transcripts for videos that you have edit permission on. So basically your
18:14 own videos, right? And like that's fine for what I'm building here with this
18:20 agent. Um, but like what if I wanted to extend this so that like we could take
18:25 in anyone else's YouTube channel? Like I could go grab Network Chuck's videos and
18:28 then we got an agent that's trained on Network Chuck or I go grab like Reendor
18:34 and then we got uh like a AI coach that's trained on Reendorses YouTube
18:37 content. Like we could do that and I think that'd be so cool. So I don't want
18:42 to limit myself to my own videos. And so that's why I'm using Super Data here.
18:46 And you just like grab an API key. I'm paying like I think like $10 a month for
18:49 Super Data. It's super cheap. Uh and so that I mean really like a thousand
18:53 videos being pulled a month is all I need. Um, so yeah, that's going to be our tech
18:59 stack here. U and then like other considerations that we need to think
19:03 through other considerations. And this is where I want you guys to uh give some
19:07 suggestions here. So I'll be looking in the chat at the same time that I'm
19:11 typing stuff out. Um, but this is where we need to think about like how should
19:15 we pull data from YouTube and then how do we want to store that, right? For
19:19 example, I'll just give you guys an example here. Uh we don't want
19:27 to pull videos older than a week. Let's just like consider that one of our
19:31 constraints because this rag pipeline I want it to run on a scheduled basis,
19:36 right? Like every day or every hour. It's going to run and check for new
19:39 YouTube videos in a channel that I specify. And maybe let's say we'll um
19:46 specify the channel ID uh in an environment variable, right? So like
19:49 we'll make us a like this instance of the rag pipeline is monitoring a single
19:54 YouTube channel and then if we wanted to like duplicate this AI coach for another
19:57 YouTube channel we just have another instance of the application running for
20:01 example. I think that's the best way to keep it simple right now.
20:07 Uh and then let's see um Joseph said using langraph as a text stack. We could
20:11 incorporate langraph uh though I think to keep things simple right now I'll
20:15 stick with just pideantic AI. My general philosophy is that um once an
20:20 application or like an agent gets too complex for a single AI agent, that's
20:24 when I'll expand to multi- aent and then at that point I'm using Langraph very
20:28 extensively. So I I appreciate the thought. Uh but I definitely want to
20:33 keep it simpler here. Okay. Uh don't want to pull the same video twice. 100%. Yep. So check the uh
20:42 database to make sure video isn't processed already. Very good. Yeah, I
20:47 appreciate that idea. Um, and really one of the beautiful things that, and let me
20:50 move this over a little bit so we got room. One of the beautiful things with
20:55 uh YouTube videos is that you can't actually change the content of a video
21:00 once it's posted. And so, we are guaranteed that if the video is already
21:04 in the database, we don't ever have a need to update it, right? Because like
21:09 if you want to change a video, you literally have to delete it and
21:12 re-upload it. And no one really does that. I don't even think that's an edge
21:15 case that I have to really cover in this application unless I like really want to
21:20 make it robust later on. Uh filter conditions are needed and a
21:24 trigger. Yeah, exactly. So, I have to think about uh think about how to
21:31 to trigger this application. Um filtering for videos. Um well, I mean that kind of is what I
21:39 have here. We don't want to pull videos older than a week. I think that's really
21:42 the only filter that we need. And then obviously making sure that they aren't
21:46 um already processed. So those those are kind of like the two filters that we
21:49 have here. So that's good. And like every single one of the things that
21:52 we're defining here is removing an assumption of the AI coding assistant.
21:57 If we didn't specify a week, then maybe you would think two months and then we'd
22:00 pull in way too many transcripts to process at once and that would just be a
22:03 waste of our super data credits for example, right? Like these are the
22:06 things that we don't want our coding assistant to assume. If the transcript fails, retry. That's a
22:15 good thought. If the transcript fails, retry. And we'll say like retry once.
22:19 Like just give it one more shot. Otherwise, um otherwise uh give an
22:24 error. I don't know exactly what that would look like. We'll probably have
22:28 some monitoring set up at some point for this application. Why Super Data over
22:33 Amplify? Um that's a good question. I I just Super Data works really well for
22:38 me. It's super convenient and cheap. I don't really know like Appify pricing. I
22:42 haven't used Ampify for a while. I'm going to be honest. It probably is a
22:46 solution that would uh work just as well as Super Data. YouTube scraper.
22:54 Honestly, this probably would work well, too. I'm just going to use Super Data
22:58 because that's what I am familiar with and I just I know I'm going to be able
23:01 to validate the code that's outputed a lot better using Super Data. But yeah,
23:05 Ampify would be another good option. I think that like in the end the exact
23:09 tool that we're using to get the YouTube data is not extremely important for what
23:13 we're building. And honestly, I think it'd be very easy to swap this out at
23:16 any point because it's only going to be a few simple API calls to get the actual
23:21 YouTube data. Most of the work is going to go into building the rest of the rag
23:26 pipeline like chunking and storing things in our knowledge base and then
23:29 also building the agent itself. Um, but yeah, Apify is probably fantastic as
23:33 well. So yeah, going back here, Joseph said, uh, textjack sentry. It's a good
23:41 thought. Monitoring for the agent. Um, I would say though that, uh, Joseph,
23:49 the text stack for observability is going to come for the second piv loop.
23:53 So I'm going to actually hold off here. Well, I guess that's fair. We might want
23:58 to do sentry for the rag pipeline and then maybe like an agent specific one
24:02 like Langfuse for the agent. I'm gonna say like for the rag pipeline though I
24:06 just want to keep it simple right now. Uh but I do definitely want to add in uh
24:10 a monitoring solution for the agent itself and then maybe later for the
24:15 pipeline as well. Is it guaranteed to have a transcript with this API? I believe it is. Yep.
24:23 Uh will you add web search capabilities as well and then expand the knowledge
24:26 base to enhance the YouTube knowledge base? Um yeah. So later on I do want to
24:30 do that. Uh, right now though, I want to stick with like just YouTube to make it
24:34 simple at first because it's always important to build in iterations, right?
24:39 Like you could just try to include all of the different data sources, right?
24:43 Like I could right now try to build in articles and I could try to build in
24:46 YouTube content like we're doing right now. I could do open source repos. I
24:49 could give it web search and give it the ability to do like deep research and
24:53 build its own knowledge base. Like I could try to do all of those things, but
24:56 that's going to get so messy for the coding assistant up front that I would
25:01 100% want to do that in another piv loop. Like maybe later on I'll go ahead
25:06 and copy this just to make this nice and visual for you guys. Like maybe like a
25:11 piv loop number three is going to be um deep research. I don't know, maybe
25:14 something like that or just like some other knowledge source. And then maybe
25:20 uh piv loop number four is we're going to do uh let's say like agent
25:26 observability. Agent observability. All right. Some for some reason that's a
25:29 hard word for me to type out sometimes. All right. Right. Like we can keep
25:33 iterating on the agent just as we want to keep adding more features. And so we
25:38 want to start pretty simple. Like in this case it's a rag pipeline just
25:42 YouTube videos. That's it. Make it nice and simple. something that is kind of
25:46 more of a an MVP like a proof of concept. We get the pipeline working.
25:50 Hopefully, we design it in a way where it's easy to add other sources. So,
25:54 actually that could be another consideration here. Make it easy to add
25:59 other data sources in the future besides YouTube, right? So, like building it intelligently like
26:06 in a very considerable considerate way upfront is very important. So it's it's
26:11 good that we're having this conversation just so that we can think about these
26:15 these assumptions that we want to take away or the instructions that we want to
26:20 give to our coding assistant. Uh timestamps, yes or no. What metadata
26:26 chunking size? Very good, Toby. I agree 100%. So yeah, let's say that um we
26:32 don't care about timestamps. Um well, actually, you know what? I
26:34 think we do because if we want to site a source then we could also like even site
26:38 to like the exact time in the YouTube video to go to or like the chunk of time
26:42 that that text comes in. So we can say uh metadata including um video URL uh
26:49 timestamp for the chunk. I hope that is available to us. I don't
26:53 actually know for sure but I think those would be the main things. Um and then
26:57 like we could just say like channel ID as well. I suppose that would be useful.
27:02 Um let's see. Chunking size. chunking size doesn't really apply here actually
27:09 because we are using using hybrid chunking. So really what we're defining
27:15 is going to be the minimum and maximum uh chunk size versus like an explicit
27:19 size. And so we'll say that uh min chunk is going to be um let's do 400
27:23 characters. max chunk or let's just say min uh 400 characters
27:32 max we'll do 1,000 characters I think that'll be a pretty good
27:37 number and then we could always uh set those environ as environment variables
27:43 so set the min and max chunk as environment variables as well. So
27:47 there we can make that one of the things that's configurable as well as the
27:50 channel ID. So we we want to like think about the things that we want to be able
27:53 to define and config versus the things that are just like sitting there in
27:56 code. Like I generally don't want something like this to be hardcoded in
28:01 the code somewhere. So yeah, uh Sean said, "Okay." Yeah. So uh Sean, let me go ahead and uh put up
28:10 your message here because this is good to to know. Um yep. So yeah, Dynamus 20
28:19 is 20% off monthly. Dynamus 40 is 40% off yearly. And then Sean said that
28:24 Dynamus 40 only shows us 20% off. The reason for that, Sean, is because
28:27 there's already a discount for the early adopters like outside of this live
28:31 stream. So, it's a discount on a discount. I know it's a little bit
28:35 confusing, but it is um for the yearly, it's going to be 40% off the the full
28:40 annual price for joining the community. I hope that clears that up. I just
28:44 wanted to address that really quick um before we go on to the the next part of
28:49 uh building this agent here cuz yeah, I think at this point I'm pretty confident
28:54 with the planning that I've done here with you guys and I I appreciate all of
28:57 your suggestions by the way. Like thank you for helping me think about
29:00 everything that we want because now with this we can dive right into our
29:04 implementation. And so I've got the repository open in my editor here and I'm working from a
29:12 blank slate today. The only thing that I have done, and I'll kind of show you
29:15 guys this really briefly, is I've set up what I like to call my AI layer. So, I
29:20 have things like my commands, my reusable prompts. I have some examples that I wanted to
29:26 reference from previous work that I've done. And u I would highly highly recommend
29:33 doing this. This is one of the things that I actually call out in the PIV
29:37 loop. When you are working on a brand new project, generally the things that you're
29:42 researching with the help of your AI coding assistant is online resources
29:45 like how did other people build a similar application and what does their
29:49 tech stack look like? But then also previous projects that you've worked on.
29:53 It's kind of it's like one of the biggest cheat codes when using AI coding
29:57 assistants. When you have been building AI agents for a while, like I have
30:03 obviously, you start to build this big collection of past projects that are at
30:08 least one of them is going to probably be pretty similar to what you're
30:13 creating right now. And so it's a huge cheat code because during the planning
30:17 phase and even the implementation phase, you'll have your AI coding assistant
30:21 constantly reference to the examples that you have prepared for it so that it
30:25 knows how to do things like set up the agent, how to define tools, well, how to
30:31 configure observability with Langfuse for example, and like I literally have
30:36 all of these things ready for it to go um so that I can like tell it as a part
30:39 of the structure plan that I create like here's where you go um so that you know
30:43 how to create an API endpoint around the agent so we can use it in the the front
30:47 end that we have defined right here. Like I said, we're going to like
30:50 integrate it directly into this front end. And the way that the coding
30:52 assistant is going to be able to do that very easily is it can look at the
30:56 example that I have pasted in here. This is my cheap code. I made an API endpoint
31:01 around my agent before. I'm going to do it exactly the same way again. As long
31:06 as this implementation is good and I made it good, then we can just reuse the
31:10 heck out of it for all the agents that we want to integrate into a full stack
31:15 application going forward. And by the way, everything that I have in the
31:19 examples folder here, I also have in the repo um that I will share with you guys
31:24 right now. So this like initial AI layer that has my commands and examples and
31:28 everything. Um I have this pushed to a public repo that I will uh share in the
31:34 chat right now. So there's a link to it if you guys are interested in pulling this. And I mean
31:40 you could follow along right now if you want to. You could just use this as a
31:44 resource to see kind of how I've set up my AI layer like taking a look at the
31:48 different commands that I have. the three core commands that are going to be
31:54 guiding my piv loop as I'm calling it. I have uh one called the primer and the
32:01 primer is um not I'm not going to use this right away. It's more for when
32:04 you're working on an existing codebase. You use this command to instruct the
32:08 coding assistant on how to analyze your codebase to kind of like catch it up to
32:11 speed. So you start a brand new conversation with your coding assistant
32:15 and then it can prime itself on your codebase so it now understands how your
32:19 code works so that you can go and implement a new feature. Right? So we'll
32:23 use this later on. This is a very power like I I whenever I'm working on an
32:26 existing codebase I will literally always start a conversation with a
32:33 priming command like every single time. And then we'll use the create plan uh to
32:39 create that structured plan once we kind of have the initial details set with the
32:44 AI coding assistant. And then when we go into the implementation plan, we will
32:48 use the execute command slash command. So I will uh show you all of this in
32:52 action to make it really concrete for you. And again, these are some of the
32:56 things that I dive into in a lot more detail in the Aentic coding course. like
33:02 not only how these commands work, but literally how you can build your own to
33:07 start to define your own system for working with coding assistants. Cuz in
33:09 the end, there are a lot of things that you can just like take as templates for
33:13 me or other things like the BMAD method for example. But the true power comes in
33:16 when you can design a system that's optimized for your codebase and that's
33:21 what I teach you how to do. So yeah, uh with that, let's open up a
33:27 conversation here. Let's chat with our coding assistant. So, I'm going to be
33:31 using Cloud Code in the stream for today, but everything that I cover here,
33:35 man, you could use any AI coding assistant you want because literally
33:39 like all the commands that I have here, like to create a plan, for example. I
33:43 know that in cloud code is very nice that I can just do like create- plan,
33:48 right? But if you're using something like um I don't know, GitHub copilot
33:52 that doesn't support slashcomands. All you have to do is say, you know, look
33:56 at.cloudcommandcate plan. MD and use this as your instructions right
34:04 now. Right? Like you can because all a a command is is a prompt, right? So you
34:08 just tell it to look at the prompt and execute the prompt. And then it doesn't
34:11 even matter that you don't have the idea of a, you know, slash command. You can
34:18 still use the instruction set here. And so to kick things off, I'm going to
34:25 go into my Obsidian diagram here and I'm going to copy this. So all the planning
34:31 that we did, I'm not just doing this for theoretical reasons like we listed out
34:35 all these considerations together. And this is going to be part of the context
34:40 initially when I uh build the rag pipeline. And so I can paste this in as
34:45 some extra context here. So, we're building a YouTube video rag pipeline.
34:49 Here's our text stack, Python and Super Data, all of our other considerations.
34:55 And then what I want to do before I get into any actual coding is just have a
35:00 conversation with the coding assistant. Like at this point, we can keep it super
35:05 casual. I like to call it vibe planning. This is the first stage to planning. At
35:10 this point, we're having a unstructured conversation. All I want to do, this is what I call my
35:16 layer 2 planning. All I want to do is figure out for this rag pipeline.
35:22 How do I build it, right? Like let's just explore the text stack. Let's
35:26 explore the architecture. We'll have I'll have cloud code like search the web
35:29 so we can understand supera data, how to use supera data, right? Like we're
35:33 establishing all that context so that we're then going to use the conversation
35:38 as a part of the context to build that structure plan. And so I'm going to go
35:43 into uh my coding assistant here. I'm using Aqua Voice for text to speech. A
35:47 lot of you guys have seen me use it before. It's $10 a month. There's also
35:52 another uh free version of or like a similar free thing called um Epicenter
35:58 whispering. So I highly recommend if you're using coding assistance like
36:01 don't type out your prompts. Like please it just saves you so much time if you
36:06 speak them out instead. And so Aqua Voice is what I use every single day.
36:13 Like literally I um I've spoken 300,000 words into Aqua Voice and uh my words
36:19 per minute is 236 when I'm using Aqua Voice. And um I know it's $10 a month.
36:23 That's why I want to suggest a free and local version as well. Um so check this
36:27 out if you're interested. Uh I'll put these in the chat as well. So, uh, free
36:32 text to speech. Um, and then my text to speech. If you guys just want like another
36:40 golden nugget of recommendations here, like you you guys, I'm telling you,
36:44 don't speak or don't type your prompts into AI coding assistance. It is so much
36:49 better to to talk it because watch this. I'll say, all right, so I opened up my
36:54 Aqua voice here by just double pressing alt. Um, well, here I should probably
36:59 delete that. I don't actually care to have that text in there. Um, and I will
37:04 say I'm building a an AI coach. It's going to be an agent that is trained on my
37:12 YouTube data. Now, the first thing that I want to build is a rag pipeline where
37:17 I can take all of my YouTube videos, pull the transcripts from them,
37:21 including timestamps, and then I want to chunk those up and store them in a
37:26 knowledge base. And so my text stack, I'll be using Python for the rag
37:30 pipeline. I want to use supa data to get the YouTube transcripts with timestamps.
37:34 So I want you to search the web for super data as well. I also have the
37:39 documentation for super data pulled into archon. So I want you to incorporate
37:45 archon. Um and then for the database, I'm going to be using superbase. So I
37:49 want you to also like help me think about the database tables that I should
37:52 create to store the YouTube videos and the embeddings and everything. And then
37:57 I also want to use Dockling for the hybrid chunking strategy. So I want you
38:01 to search the web so that you understand hybrid chunking with dockling as well.
38:07 And then also within my PRPs/examples folder, I have a file called uh dockling
38:12 hybrid chunking.py. So I want you to take a look at that. That's a working
38:17 example that I already have for using hybrid chunking with Dockling. And so
38:22 given that tech stack and also all the considerations that I've listed above, I
38:28 want you to um not build like don't code anything yet, but just help me like
38:31 think through if there's any other considerations that I'm missing. And
38:34 then I want you to give me some different options for how I can
38:38 implement this rag pipeline. Boom. There we go. All right. So I'll send this in. And
38:43 look at how fast that is. Like it's blazing fast and extremely accurate. Uh
38:49 the only thing that it didn't do very well is it said uh PRPS and then it like
38:53 explicitly wrote out slash when I wanted to reference the file. So I can just go
38:57 and like edit that. So like sometimes I'll go and edit things very briefly but
39:01 otherwise I mean I trust this. I think this is really good. And again I am
39:07 following no structure in this conversation at this point. Like this is
39:10 kind of a messy prompt but that is totally okay. In fact, um, one thing I'm
39:15 going to do really quickly is I'm going to copy this. I forgot to, uh, add in
39:23 Archon as my MCP server. So, I want to do that really quickly because one thing
39:26 that I did to prep for this, um, stream is I pulled the documentation for Super
39:32 Data into Archon. So, it has access to like all the code examples and things
39:36 for how to pull transcripts from the Super Data API. It's really neat. So, I
39:39 have I have the documentation for Archon already, which I'm not going to cover
39:43 Archon. uh for the sake of time too much in this live stream, but it is going to
39:48 be part of my coding process here. So, I've got the documentation for super
39:54 beta supa data, excuse me. And uh I literally just pulled their llm.ext. So,
39:57 like that's the source that I'm using in archon. So, if you want to follow along,
40:03 I'll actually post this here. Super data llm.ext. I'll put this in the the
40:07 YouTube chat. So if you guys want to include this documentation in your um in
40:13 your archon, you can crawl that. It will take like just a couple of minutes. It's
40:17 pretty short. So I got the documentation for super data. I just need to actually
40:22 hook archon into my cloud code really quick. So I just didn't have it set in
40:27 this terminal here. So I'll add it. There we go. Cloud MCP list. Make sure
40:31 that our connection to archon is good. Looking good. Okay, perfect. So now I
40:35 can go back in uh to Claude and then I've got the prompt that I can paste
40:39 back in here. So okay, a little bit messier now, but that's okay. So I'm
40:42 going to go ahead and send this in. And so yeah, we're not coding anything yet,
40:47 but it's going to search for super data. It's going to understand dockling and
40:51 look at my example. So we're already using my cheat code, right? Like this is
40:54 a something else that I already did. I actually showed this on my channel
40:59 already, uh using hybrid chunking within Dockling. And so now it'll understand
41:04 how to use that as well. So I'm just having it I'm kind of just like
41:08 gathering as much context as I possibly can to make sure that the coding
41:12 assistant is on the same page with me for what needs to be implemented. And
41:15 take a look at that. It's searching archon as well doing all the web search.
41:20 It's spelled dockling wrong, but that's okay. LLMs love to do this all of the
41:24 time. They they will literally spell dockling incorrectly. So yeah. Uh, but I think it'll still get
41:32 what it needs. So, okay. Uh, in the meantime here, don't you need the Dockling docs and
41:40 archon as well? It probably would help, Toby. The reason that I'm not including
41:46 the dockling documentation is because uh the hybrid chunking is the only thing
41:50 that I'm using and this example just like lays it out so clearly that I don't really need it
41:57 to search anything else within dockling. Um but 100% if I was using a lot more uh
42:02 aspects of dockling like if I was working with PDF and Excel extraction
42:06 for example u which I might do later for this agent. So I could just like dump
42:09 PDF documents for research that I found and include that in the knowledge base
42:13 as well. At that point I would probably include the dockling docs in archon too.
42:19 But um yeah I think that's another like good lesson is um don't overengineer
42:25 things right your AI layer which includes your commands. It includes your
42:29 your global rules that I also have set up for this project already. This is one
42:32 of the things I created ahead of time. uh your AI layer includes all the
42:37 knowledge that you have in Archon. Like don't overengineer it. If you're only
42:43 using a tiny part of a library, it might be enough to just include a single
42:48 example like I did here or even just to ask it to search the web on something
42:54 really basic. Now, archon is better than web search because you get to pick where
42:57 it search what it searches through, right? Uh but for something really basic
43:00 like that is going to be sufficient. So, I would just keep that in mind because
43:04 then it's just nice that I don't have to like go crawl dockling and wait for
43:08 that. I mean, maybe it'd be nice to have it anyway, but yeah, I think you get the
43:13 point. Um, looks like claw code failed to load the dockling example. Will it find it?
43:21 Oh, is that really true? Um, oh yeah, you're right. PR Oh, because it spelled
43:26 it incorrectly. Interesting. That's actually a problem. I really appreciate
43:31 you catching that because did you read the and then I'll I'll call
43:36 it out explicitly here. So that is one of the the warnings that I have with um
43:39 speech to text is like sometimes that'll happen where it's not spelled correctly
43:42 so it misses it. Um but this was actually really surprising though
43:46 because generally when claude code u fails to find a file it'll try to search
43:51 for it in a couple of different ways like using reax or blob search or
43:55 something. And so I'm actually surprised that it didn't retry like that. That is
44:00 uh quite unfortunate. So yeah, one of those things that you just have to watch
44:03 out for as you're having it look through a bunch of files, making sure that it
44:06 actually spelled things right and uh read the file and then retried if it
44:10 failed for whatever reason. So now I'm having it read it. So there we go. Okay.
44:15 Yes. Oh, it said it did read it already or unless it's saying that. Um Okay.
44:20 Yeah. So it just read it for the first time because now it's correcting
44:23 something. character based versus tokenbased chunking. It works with tokens. Okay. Yeah, sure.
44:29 All right. So, it did correct itself a little bit. Okay. So, good catch. I
44:32 appreciate that, Tony, because otherwise it wouldn't have been as good a
44:40 Okay. So, here we go. Um, stick with tokenbased chunking. Uh, yes, let's stick with Docklings
44:51 approach. um also give me uh some options for building this rag
44:59 pipeline. Okay, so I kind of distracted it a little bit on something really
45:02 specific. So I'm just going to answer that quickly and then move on to um the
45:06 rest of the questions that it has for me, but at a high level, this is really
45:11 nice. It did a ton of planning for me already. So text stack validation, it
45:17 says it's solid. Uh it says Dockling and Superbase is great. That's good. I mean,
45:21 I could ask it more questions to like really have it validate if I want, but
45:24 for the sake of brevity here, I'll just say that like, okay, good. We're on the
45:27 same page that the text stack is solid here. And then creating a a table for
45:31 the specific YouTube channels, which I think is nice because it's just going to
45:34 be my channel to start, but then at some point I might want to go so that like
45:38 this agent could actually pull videos from other channels, too. and then uh
45:44 pulling YouTube videos in which I assume that it verified with super data that
45:50 all of these fields are accessible when it pulls the transcript from super data
45:54 and maybe that's a question that I could ask as well to make sure that it uh
45:58 really created this excuse me based off of what it can get from super data
46:03 and then we have the chunks where we store our um individual chunks for the
46:09 YouTube transcripts including when the chunk stops and ends. So we can cite
46:13 that as a part of the source like you know this piece of information came from
46:17 uh my video on dockling and you can go jump to 9 minutes in and then that's
46:20 where I'll talk about it right that that that's going to be so cool if our AI
46:24 coach can actually reference like timestamps and videos. That's the goal
46:28 that I've got here. Additional considerations. So some of these things I'll probably go through
46:33 rather quickly just for the sake of time here. But in a real planning, like when
46:37 you're going through the PIV loop and you want to spend a considerable amount
46:40 of time making sure you're on the same page with the coding assistant, like
46:43 this conversation can go for a solid half hour. It could go for it could go
46:47 for even an hour. Like if you really want to make sure that like you have the
46:51 right decisions made around the tech stack, that you're considering
46:55 everything like um what's one thing I called out here? It says that I'm not
46:59 considering API rate limiting and cost management. Um so yeah, I mean that's
47:04 important. um incremental processing track the last process at per channel to
47:08 only fetch new videos. That actually makes a lot of sense. So that's one of
47:11 the things I listed initially that it's getting more specific for me. And so I
47:17 actually appreciate that a lot. And so let's see what it says at the bottom
47:21 here. Um questions to finalize. Which option do you prefer? I prefer option
47:25 number four. Option number four, detailed recommended pipeline architecture.
47:37 YouTube service calls a super data API handles rate limiting implementing retry
47:41 logic. Okay, blah blah blah. Like I said, I would spend a lot more time with
47:45 this normally, but uh I'm going to just say this is good. Like I'm going to act
47:49 like I read through this entire thing and I would recommend doing that.
47:55 Um yeah, I don't want it to be considered a tool because at this point this is just
48:01 the rag pipeline. So, okay, good. I'm going to say that option number four is
48:06 good. So, questions to finalize. Uh, option four is good. And that's where
48:10 it's going to create a separate folder called YouTube_ra. And it'll build the
48:17 pipeline there. Um, token limit 512 max tokens or different. I'm going to say um
48:23 min should be 400, max should be uh 1,000 like we established in our
48:30 diagram. embedding model. Uh stick with all mini LM or upgrade to a larger
48:36 model. Uh let's use let's allow to specify the embedding model as an environment
48:49 So you specify the OpenAI base URL and embedding model. And then
48:55 again this is my cheat code. I already have aenv.ample in my example folder. And so this
49:04 literally has what I'm talking about here where we can set our provider like
49:08 we want to use openAI or Olama or Gemini. Then we set the embedding base
49:12 URL and the API key and the model choice. And so I wanted to look at this.
49:21 to see how this is set up. And so I'm going to build this AI agent. This is
49:23 one of the things that I promised you guys. uh when I was uh you know
49:29 announcing this is that it's going to work with any large language model. So I
49:32 want to make it so that through just configuration I can go between Gemini,
49:37 open router, open AAI for both the uh embedding model and also the large
49:40 language model that we're using for the agent itself once we get to Pantic AI.
49:46 So let's go ahead and do that. And then uh trigger mechanism CLI only for now or
49:50 do you want an API endpoint? Uh CLI is good. So, I just want to be able to to
49:54 um call the script that like pulls my YouTube videos. I just need to do that
49:57 from the terminal. Like, I don't need an API endpoint to do this right now. Um,
50:01 and so, yeah, I'm going to give it my options here. And so, now we're really
50:05 starting to get on the same page. Like, it's asking me questions to fill in its
50:09 knowledge gaps. It's telling me other considerations I need to fill my
50:13 knowledge gaps. Again, we're just reducing the assumptions overall. That's
50:17 100% the goal that we got here. And so it's going to create a a detailed
50:21 implementation plan here, which I'm going to actually interrupt it and say
50:24 create plan. I'm going to call um based on our conversation create a detailed plan um for the
50:34 YouTube rag pipeline implementation. And so what I'm doing right here is now
50:39 calling my command that is going to systematically leverage the conversation
50:44 to do additional research and analyze our tech stack and architecture and
50:47 other decisions. And then it's going to create that structured plan that I'm
50:52 talking about uh within the piv loop here. And so now at this point we are
50:58 going from vibe planning just a more unstructured conversation into our
51:03 structured plan. And so we're leveraging the four pillars of context engineering.
51:08 Memory is our conversation that we just had. Rag is going to be external sources
51:13 like looking at the super beta documentation, super data documentation
51:18 in archon. We're going to define a task list that we're also going to leverage
51:22 in archon. And so we want to split the task up task by t or split the
51:26 implementation up task by task so our coding assistant can knock it out in a
51:31 more systematic approach. And then prompt engineering is really all the
51:35 techniques we have for outlining the different sections of our structured
51:40 plan. And so in the command that I have here, one of the things that you'll see
51:46 is that I outline what the plan looks like, right? Like we have an overview,
51:49 what's going to be implemented, a requirement summary, all the
51:53 documentation we want to reference. Here is our task list, right? Like this is
51:57 prompt engineering. This is defining the core sections we want in our structured
52:02 plan so that we can create a repeatable system where whenever I create the plan,
52:07 I know exactly what my plan is going to look like. So, it's very easy for me to
52:11 validate it as well. And so, there we go. I'm just going to kick this off and
52:16 uh have it run. And while it does run, this is another good time for uh me to
52:19 just share really quickly the Agentic coding course because again, all of
52:22 these things that I'm covering right now, I cover it in a lot more detail in
52:28 the new Agentic coding course that just came out. I haven't Okay, I need to do
52:31 this. I'm I'm so excited for this. I didn't even show the the page for the
52:36 course yet really. Um the first batch, so this this course is being released in
52:39 batches into like straight into the Dynamus community over the next month
52:44 here. And the first batch was just released on Friday, which I I'm just so excited for.
52:51 I like it it's literally been like hundreds of hours prepping this like up
52:56 until this point. Like so much work has gone into this. All the things that I'm
52:59 showing here in this stream, building this agent, like I go into a lot more
53:03 detail, things around the pivot loop, how we can build reliable and repeatable
53:08 systems so that we can create our own process to build our own software
53:13 optimally. Because in the end, most developers have what I like to call the
53:18 system gap where they're using AI coding assistance more willy-nilly, if you
53:21 will, right? And like that's what I'm trying to teach you how to not do
53:25 because people who don't get good results with coding assistance and like
53:30 that that'll happen. Like if you don't have a good system in place, it's not
53:34 going to work that well. Your code's just going to break down, especially
53:37 when you go into production. And so you need to have a system in place. Reusable
53:41 prompts and commands, defining your rules and workflows, how you integrate
53:45 different tools with your coding assistant, how you optimize your code
53:49 base to work with AI, like all these things that are not trivial but very
53:53 important so that you can, you know, actually get good results realistically.
53:57 It takes a lot of work up front and that's what I show you how to do. And
54:02 so, yeah, right now that I have the the link for Dynamus highlighted and also
54:07 the discount code that's only going to be available for this live stream. like
54:11 I've got to make sure you know that so that you don't expect to be able to use
54:15 it after 20% off monthly and 40% off the yearly membership right now um for this
54:20 course and then also everything that we have in Dynamis the community the weekly
54:24 workshops the AI agent mastery course like it it it is um it is really the
54:29 whole shebang um that you're you're getting within Dynamis and so I'm just
54:32 continuing to build more and more into the community so it's just more valuable
54:36 over time and this course is like the big thing that I'm doing right now for
54:41 that. So, yeah. All right. I'll go ahead and uh tear that down here. And let's go
54:47 back over to our coding assistant and see where we're at. Cool. So, it is failing to read some
54:55 files, which it's fine because it looks like it's retrying. Um not really sure
54:58 what it's trying to do. There's probably like something that I called out on
55:03 accident for a pi project tol, but that's all fine. So, okay, we have our
55:08 codebase analyst sub agent that I have running right now to analyze what we
55:12 have set up so far, which is ma basically just going to be analyzing the
55:16 examples that we have at this point. And so, it's going through the examples to
55:20 understand how to build the rag pipeline. I mean, this is the main file
55:23 that I needed to reference right now. So, as long as it does that, I think
55:26 we're good. And then after that, it's going to create our structured plan. And
55:32 so maybe while it's continuing to go here, I'll just uh answer some questions that
55:38 we've got in the chat. So let me take a sip break and I'll open up the chat as
55:47 All right. Can I switch from monthly to yearly? I just joined for a month to see if it's
55:54 worth it. Uh yes, that is an option. Um so yeah, appreciate you asking. If you
55:59 go and message me in the community, I Um, let's see. Can Cloud Code give us a
56:08 warning when we get close to our token limit? H, that's a good question. I
56:13 believe you can with uh Claude hooks. I've never done that before. Uh, but
56:17 also like when you're working in Claude and you get close to your context limit,
56:21 you'll see an error or not an error, you'll see like a little warning right
56:25 here kind of in like the bottom right of your terminal that'll say that you have
56:29 like 5% or 8% left of your context window. And so yeah, I mean that's kind of like
56:35 the warning that you'll get in the UI. If you wanted some kind of alert, you
56:38 you could probably do something with Claude hooks, I would say.
56:42 Uh, what do you think about Claude skills? Yeah, actually I just did a
56:47 workshop in Dynamus on claude skills. Um, which yeah, here I'll go show that
56:52 really quick. So cloud skills for those of you who don't know, it's a kind of
56:57 something that might honestly replace MCP servers because it's a way to load instructions
57:04 into a coding assistant for how to use different tools uh only when it actually needs to use
57:10 those tools. So like for example with archon as an MCP server the context for
57:16 archon is always loaded in to the coding assistant and so it takes a lot of the
57:20 context window but what if like you had a little description of like here's what
57:25 archon does now only use archon as a skill and like load in the context when
57:29 you want to do task management or search the knowledge base for example. So it'll
57:33 load in the instruction set for archon and then like all these scripts to
57:37 basically like use archon with code. That's what skills give us. And I have
57:41 been very impressed with skills. Like it solves a big problem that we have right
57:47 now with um just like a lot of tools like in MCP servers right now where like
57:51 it just takes so much context. And with our AI coding assistants especially, we
57:55 hit those context limits so fast. And so we need a way to like basically only
58:00 dynamically reference different capabilities when we actually need them.
58:05 That's kind of the core of what uh Claude Skills gives us. So yeah, I'm
58:09 very bullish on skills. I think that they could entirely replace MCP servers.
58:14 Um, not to get too technical, the one thing I feel like MCP servers do better
58:17 is environment variables. There's not really a way to like configure
58:22 environment variables for skills like you can do with MCP, like when you, you
58:26 know, connect a MCP client to a server. So yeah, but otherwise I think they're
58:36 So, it's like a branch expansion context updator. You could think of it like
58:39 that. Yeah, because you kind of have different layers of context that you
58:44 load when you need it. Like I can load a skill and that has like the core
58:47 instructions and then the core instructions can also reference other
58:52 documentation to bring in if I need that part of the skill specifically. And so
58:56 that way you can have like a really massive instruction set, but you're not
58:59 like loading it all into the LM at the same time. only when it actually needs
59:03 it. Which by the way is similar to how I set up my global rules. Um, so the
59:07 global rules for this project, just as the coding assistant continues to go
59:12 here, the global rules for this project, I I say like when you are building
59:17 tools, then look at my tool guide, right? So like I have a part of my rules
59:23 for like how to build effective tools. So the this is like part of my global
59:27 rules, but I'm only telling it to look at it if it is actually actively
59:30 building tools. Otherwise, it doesn't make sense to dilute my context with
59:34 this information because like right now as I'm working on the the rag pipeline,
59:38 I'm not doing anything with tools. Like why why should I have that in my context
59:42 window? But then when I'm building the agent and I'm building the rag tools for
59:46 the agent, then I definitely want this, right? So I hope that makes sense.
59:53 Uh Toby said, "The coding course is great. Just finished the two-layer
59:56 planning video today." Awesome. Appreciate that, Toby. Cool. Thanks for
60:01 going through it. Can I watch the stream later on? Yeah, 100%. So, this live
60:04 stream is going to be available on my channel um like pretty much just a
60:09 couple of hours after it's done. It just takes time for YouTube to process it, I
60:14 believe. But if you go to my channel and uh Oh, that's cool. I've never seen
60:17 I've never looked at my own channel from a different account before. Um but yeah
60:21 uh well while it's while it's live I mean but yeah so if you go to the live
60:27 tab then this will be uh a like a full recording afterwards like all my past
60:30 live streams you can view those just like a YouTube video if you just go to
60:38 Uh is the rag agent tables SQL file something you configured before for this
60:42 specific agent or is it from an old agent to show the architecture on how to
60:46 do the tables? Good question Toby. Yeah. So the uh file that he's calling out
60:52 explicitly is this one right here. This is from a an a previous agent. So I just
60:58 want it to reference this SQL file so that it understands uh my preferences
61:02 around building scripts to like set up tables for my agent. And then also I
61:07 wanted to reuse some of these tables like the requests and conversations and
61:11 messages because those are the tables that I needed to set up exactly like
61:16 that for it to be compatible with the uh front-end application that I have right
61:20 here. So I hope that makes sense. So yeah, it's just an example from a
61:23 previous agent. Everything I have in the examples folder is from a previous
61:29 implementation including the SQL. So okay. All right. We have our YouTube
61:34 rag pipeline request. So I'm going to go into my request folder. I typically like to store all of
61:42 my requests in a special folder in my codebase. So I could if I want to even
61:46 like check this into source control so that I have a record of my plans that
61:50 I've used to build everything in my different piv loops. And so I'll go ahead and open this up.
61:57 And so this follows exactly the structure that I have defined in my
62:02 create plan command. And so we've got the overview. We have the requirements
62:08 summary, the architecture requirements, um our research findings. So it did a
62:11 ton of research as a part of that planning to create this structure plan.
62:17 We've got the dockling hybrid chunking, which it keeps spelling it wrong, but
62:20 I'm just going to leave it. I I think that's fine. If I was really nitpicky
62:23 right now and I wanted to spend some time with it, I would correct little
62:26 things like that. But right now, I'm just going to um call that good. We're
62:33 using Superbase with PG Vector. Um and it's referencing things based on
62:38 uh what it saw in the examples, which is looking really good. Uh codebase
62:43 integration points. So, it's it's calling out existing patterns to follow. So, uh, I want to
62:49 make sure as long as it's not going to try to build things into the examples
62:53 folder. I hope that it doesn't do that. Um, so maybe I'll call it out
62:56 explicitly, but otherwise, this is looking really good. Um, yeah. Okay. No, this is actually
63:02 perfect. So, one of the things that I have it give me is the desired codebase
63:08 structure. So, it actually outlines the rag pipeline and it's got my chunking service, my
63:14 embedding service. This looks really, really good. And it's got the tests for
63:18 the rag pipeline as well. One thing that I feel like it doesn't need to do right
63:21 now is create the rag tools. And so I'm just going to tell it here like uh don't
63:25 create the rag tools right now. We're just focusing on creating the uh rag
63:29 pipeline. So update the plan for that. And then also make sure you're very
63:33 explicit here that we're not to actually create any code in the examples folder.
63:38 All right, there we go. So just a little bit of iteration. And I would spend a
63:41 lot more time iterating on this generally, but overall this looks pretty
63:45 good. So, I'm going to jump into an implementation with this in just a
63:48 second here because I want to get the rag pipeline knocked out so that we can
63:53 test this with the YouTube content and then also um go on to making the agent
63:58 and I want to be able to see it in action by the end of the stream here.
64:03 So, okay, let me scroll back down uh implementation tasks. So, it this is
64:06 really good too like we have our task list like starting with environment
64:09 configuration. um going into defining our schemas. Um what else we got here? Database
64:16 migration. Let's see. Creating a our first database schema script.
64:23 Very nice. Um well, here I actually don't like I'm going to say one more thing. Uh don't
64:29 call it a database migration. Just do the initial setup SQL. Uh because I'm
64:34 starting from a fresh database here. Okay, there we go. I'm just kind of
64:38 correcting a couple of assumptions that it has overall. U but yeah, like this is
64:44 looking really really good. The way that it's doing the chunking,
64:47 the embedding service. I'm just glancing over this like really really quickly.
64:53 This is kind of a long plan overall. Um it's 2,000 characters, which is that is
65:00 way too long. So I don't like that because the problem is it like basically
65:06 wrote all of the code already. So, in an ideal plan, and like sometimes Claude
65:11 just decides to do this where it'll um include like all the code even when I'm
65:14 just only in the planning phase right now. Um when it does this, like a lot of
65:19 times I'll just ask it to make it way more concise. I don't need all the code
65:25 right away, right? Um and so for the sake of brevity, I'm not going to ask it
65:28 to do that right now. Typically though, I like my plans to be more like uh 500
65:35 to a,000 lines, not 2,000. like it definitely went too far here,
65:41 but I I think this will still work pretty well because the only reason it's
65:45 long is because it kind of did a lot of the coding already, which it already did
65:51 research on super data and dockling and things like that. So like it knows how
65:53 to use those libraries, right? So I think that the code that is in the
65:57 planning here is probably more than even just pseudo code. Like I think this is
66:01 like good code that it can build right into the file structure as it creates it
66:06 here. So, I I'm just going to let it rip and uh once this is done here, we'll see
66:11 how it goes. Cloud 4.5 loves to just charge ahead with things. Yep, indeed. That is what
66:19 it is doing here. Um yeah, uh in Dynamis, there already exists an
66:25 internal agent as we build on this master class. So, um I'm not sure
66:30 exactly what you're referring to, but in the AI agent mastery course that we've
66:35 got in Dynamis, there is a complete AI agent that I build as a part of that
66:38 that's a template for you. It's got a full rag pipeline that works with local
66:42 files or Google Drive. It's got an AI agent with web search and image analysis
66:47 and rag tools and aentic rag. Um, and I have it both within N8N and in Python.
66:51 And then I have the full front end set up um as well. So like this what what
66:55 you're looking at right here like this is actually the Dynamus agent that I was
67:00 chatting with as that brief example. So yeah uh there is like and there's a lot
67:04 of other like complete agents built out from like workshops and stuff I've done
67:07 in the community as well. So there's a ton of value there 100%.
67:13 So yeah. All right. Um what are we doing here? Okay. So now it's just updating
67:16 the database stuff for me. So good. Okay. We're almost ready for the
67:25 Uh, do you ever find that it hallucinates when you add too much
67:30 context or content? It, yes, 100%. That's why I'm a little uncomfortable
67:34 with how long the plan is right now. Um, but because the plan is mostly just long
67:38 because of code, I think I'll be okay. But yeah, if you were to if this plan
67:43 was 2,000 lines because it was like it literally had to be to like outline all
67:47 of the features I was trying to build at once, then 100% it will hallucinate,
67:52 right? Like as long as you're not asking it to code too much at once, like I
67:55 think this is a pretty reasonable request overall, just based on my
67:59 experience with coding assistance, then you're fine. But if I was trying to ask
68:03 it to like build the YouTube rag pipeline and build the the agent and
68:07 build in observability with language, like all these things, um, and like
68:11 support other sources like GitHub repos, if I was trying to do that all at once,
68:14 then it would hallucinate because this document would be 3,000 lines long
68:18 because I'm just asking for too much at the same time. So yeah.
68:25 All right. How close are we? All right, we are good here. So, uh, we have our
68:29 plan. I'm going to call this good. I'm just going to go and let it rip now. And
68:33 so, generally, I do quite a bit more validation here and shrink it. But I'm
68:37 going to get into the implementation now. So, I'm going to clear my
68:40 conversation. I'm going to say Claude is doing fine because I I wasn't super impressed with
68:46 the plan, but that's okay. So, now I'll go into execute plan.
68:52 And so this command is similar to the create plan, right? It's just it's a
68:55 standard operating procedure. It's a step by step. I want you to read my plan
69:01 that I have here and then I want you to execute it task by task. And another
69:05 thing that I have as a part of this plan here is I uh instruct it on how to
69:13 create the tasks in archon. So all of the tasks that we have in that
69:17 implementation plan part of my structured plan, it's going to create
69:20 these tasks in archon and then it's going to go through this implement
69:24 implementation cycle for each of them. And so going back to our diagram here,
69:29 we just finished our structure plan. Now we move on to implementation where we
69:33 are delegating all of the code to the coding assistant. And so we're going to
69:37 let it go for quite a while here and then we will validate everything at the
69:41 end. Right? So I'm I'm delegating the first step of planning and the last step
69:45 of validating. That's where I want to be very involved. But when it comes to the
69:49 actual implementation, I am trusting but verifying giving it the plan, trusting
69:54 it to do a good job and then I just do the validation at the end. So let's go
69:58 ahead and kick that off here. So I'm going to go in execute plan and then the
70:03 path is PRP request YouTube ragpipeline.md. Just calling out the path to this file
70:08 right here. And we're doing this in a brand new conversation because our plan
70:13 is literally all of the context it needs. There's nothing else it needs
70:17 from my previous conversation. It doesn't need to do any more research
70:20 right now. Like it just needs to look here because we also call out the
70:25 different files and and examples to look at for the implementation as well. So
70:29 I'm going to go ahead and and let it rip So yeah, uh while we do that, I'll just
70:35 uh go ahead and answer some more questions in the chat here. Like I said,
70:39 casual live stream. We just get to watch Claude go Claude code go and uh chat
70:44 about whatever you guys want to. So, let me go back to the chat.
70:48 Oh, I appreciate that. Well done on a great stream. Thank you very much. Yeah,
70:51 and and I've never built an agent live before, so I've never done something
70:55 like this. Um it's a little bit slower than I wanted, but I I also think it's
70:59 important to like explain things as I'm going through it. um instead of just
71:03 like blazing through stuff like executing the plan or creating the plan,
71:07 I'm going as fast as I can um reasonably. I think also I'm getting the
71:13 So Claude code loves to like stutter a bunch when it's doing things in
71:16 parallel. So I'm going to like hide that for a second because that is really
71:19 ugly. Okay, let's see. Um do you try to run as much of this locally? If so, what hardware do
71:28 you have in terms of CPU and GPU? Thank you. Yeah, good question. So um I do want to
71:34 build this agent in a way where I can run things locally. So in the examples
71:39 here, this AI agent that I have as my previous project that is an example, I
71:44 do actually have it set up so that you can change the LLM and embedding model
71:48 to be OLAM if you want to run everything locally. Now, I don't try to run as many
71:55 agents locally as I possibly can because local large language models are always
72:01 going to be not as powerful as the cloud the big cloud ones right now at least
72:08 right now. So like um Cloud Sonnet 4.5 uh GPT5 like those models are better
72:12 than anything you can run locally. uh but if your application has sensitive
72:16 information or you just need it to be private for whatever reason then you
72:21 have to use local AI like it is a super important concept and so I do create a
72:24 lot of agents that are local because I'm working with things like you know like
72:28 my internal knowledge base that I have as I'm planning YouTube videos for
72:31 example and I want an agent to help me with that ideation but I don't want it
72:34 to be something that's just like sending all of my content ideas to OpenAI just
72:38 as an example or like I've worked with companies before where um they do
72:42 actually have like local LLMs running on codebase because it's very proprietary
72:46 technology that they're working on. Um, and so yeah, it it is important to to
72:52 understand how to use local AI and like setting it up in an agent like this,
72:55 it's pretty simple overall. And then also like when you should use local AI
73:00 because it's it's not always a solution. Generally for any kind of like internal
73:02 application, it's it's a good consideration for anything that's like
73:06 an external facing agent like a you know a voice agent that handles inbound calls
73:11 or if you are you know building like a lead handling agent like that kind of
73:15 thing should probably not be using local AI because it's frontward facing like
73:19 it's not sensitive data anyway and then you want it to be faster and more
73:26 As far as hardware, um, let me pull up something really quickly because I have my exact build
73:34 um, saved in pcparpicker.com. So, I can I can load that up. Uh, I'll
73:38 take the time to do this because I feel like you guys would actually appreciate
73:42 this, but I have a saved build. Um, that has my part list. Save part list. Boom. Boom.
73:50 Uh, let me make sure this works. I'm going to make sure this link is not a
73:54 Okay. Yeah. Okay. Let me go and paste this here. So, this is a link to my
73:59 exact PC build if you're curious. I have two 3090 graphics cards.
74:05 I um I got them used off of eBay, so my my computer's not actually as expensive as
74:12 it says here. Uh not nearly as expensive, but yeah, that that's my
74:17 build. So, I hope that helps. And it does do really well for running local
74:25 Um, sounds like the create plan command needs some more guardrails. Um,
74:29 honestly, it usually does a lot more better than this. Like I Wow, that was
74:33 terrible English. Um, usually it does a lot better than what it did now in terms
74:37 of making the plan more concise. But yes, Tony, I I should probably improve
74:42 my create plan. This This isn't like the plan the create plan I typically use. I
74:46 just kind of put something together to make it a bit more concise for this
74:50 stream. Um, but yeah, usually I would have something that's more explicit
74:53 about like being more simplistic and just outlining the the plan versus the
74:58 code. Uh, that's the main guardrail that I should have incorporated better into
75:02 here. So, yeah, definitely should be improving that. Um, I appreciate this bow. The community
75:07 is well worth the money if you're on the fence. It may seem expensive, uh, but
75:11 you get a ton of resources and access to Cole and other very responsive members.
75:14 That means a lot. Yeah. And I'm I'm active in the Dinos community literally
75:18 every single day. And there is an event going on in Dynamis every single day. We
75:23 got workshops, office hours. I started this thing recently called the AI
75:27 exploration hour where um it's just like a more casual event for us to chat about
75:31 the latest AI tech and then just dive into it together. And so if you're
75:35 you're ever curious like what are the things that I should pay attention to
75:38 right now? Like is Claude Skills actually worth it? Like come to the hour
75:42 and we'll talk about the things that are big right now and and explore and see if
75:46 it's actually worth our time. Right? And that's a big thing, like just trying to
75:49 fight through the fluff. Like that's a one of the goals that I have for you in
75:53 Dynamis is to just help you fight through the fluff. Like learn the skills
75:56 that are important to learn. Avoid the the shiny objects that aren't worth your
76:01 time. Uh I think that's that's just like a big problem that we have in the AI
76:04 space right now is there's just so much that is coming out every single week and
76:08 like half of it or more just like doesn't even matter. But it's hard to
76:12 tell what doesn't matter versus like what should we try to incorporate into
76:16 our coding workflows or build into our agents or you know use as a tool to
76:21 replace XYZ. It's hard to figure that Um Cole does a 60 to 90 minute working
76:29 session nearly every Friday at lunchtime Eastern. Recorded has office hours hang.
76:34 Okay. Uh founding member love it. Cool. Yeah, I appreciate that. Yeah, that's
76:37 like just what I said basically. But yeah, I appreciate you calling that out.
76:40 Sorry I didn't see your other message before I basically said that myself.
76:46 Um, which plan are you using right now in Claude? What about rate limits? Yeah,
76:51 so I'm using the um $200 a month max plan right now and I like never hit rate
76:55 limits with Sonnet 4.5. I know people that are using um the $100 a month plan
77:02 and they generally don't hit rate limits either unless they're using Opus 4.1. I
77:06 there's been a lot of complaints about rate limits in general with claude code
77:11 recently and so I know that it is considered like a pricier tool overall
77:14 because you have to have the higher tiers to not hit rate limits. It's well
77:18 worth it to me though because it is the best in my opinion. I think that Codeex
77:23 is the second best coding assistant and I even did a video recently on my
77:27 channel um showcasing using Codeex actually like programmatically through
77:31 the Codeex SDK. That was pretty cool. So I I'm impressed with Codeex as well. I
77:34 think that there are a lot of um very very viable alternatives to claude code.
77:39 Um codeex and the Gemini CLI being the main two that I would uh prefer if I
77:44 couldn't use cloud code. Um so yeah, you're definitely not tied to it. And
77:47 like everything that I cover that I'm covering here, everything I cover in the
77:52 agentic coding course, it is um AI coding assistant agnostic, right? Like
77:56 that the principles that I teach and like what we're doing right here with
78:00 the PIV loop that that works no matter the coding assistant that you're using.
78:04 because I think it's really important to not get tied to one tool. The mantra
78:09 that I always say um and like this is another gold nugget for you guys is is
78:13 uh capabilities over tools, right? Like focus on the high lever skills that are
78:17 going to apply no matter the problem you're solving instead of getting really
78:21 in the weeds about trying to master specific tools. Like I I honestly don't
78:26 care how comfortable I am with cloud code. like it matters to an extent, but
78:31 the thing that really matters is like my skills around how I apply AI coding
78:35 assistance and like how I build my system in a way that works no matter the
78:39 coding assistant. Like that's the real high lever skill there and and what I
78:43 teach. So I hope that makes sense. Um it's always a good idea to break down
78:51 large tasks and create individual archon tasks for each step. 100%. Yeah. And
78:55 actually, let me show my Archon because I don't know if you guys were watching
78:59 the logs here as I was talking, but um we're using the Archon MCP server to
79:05 create a project and tasks for our pipeline here. So, let me go into
79:08 projects. We have our YouTube rag pipeline and take a look at that. It created how
79:16 many? Um, yeah, okay, 12 tasks in total. And it's actually already almost done.
79:19 So, we're just to the point right now we're doing validation. And so one of
79:22 the things that I have built into my structured plan is I give the AI coding
79:27 assistant instructions on how to validate its own work. Right? So we have
79:33 our structure plan. We're in the middle of implementing. And in a second here,
79:38 it's almost done. We're going to do validation, but also validation is more than just a
79:45 human effort. It is also an AI coding assistant effort. We want to be very
79:49 explicit to the coding assistant on how it can check its own work. And that's
79:53 one of the things that I have defined in my structured plan. So I tell it how it
79:56 writes unit tests, how it writes integration tests, and then it'll run
79:59 those and iterate on them until everything is passing. And then at that
80:04 point, it gives control back over to us to do our validations like code review,
80:08 which I'm not going to do much of actual code review in this live stream because
80:12 I think that's more boring and I need to be pretty brief here. Uh, but like this
80:16 is super important. So like don't skip that. Um, and then even if you're not
80:20 very comfortable with using or like if you're newer to coding. Um, and like you
80:24 don't really know exactly how to do a code review, that's okay because you can
80:27 just instead of like looking at the code yourself, you can just ask the coding
80:32 assistant questions like did you address this? Did you address that? How does
80:36 this API work? How is our database being set up? Right? like you can kind of
80:39 review uh just by asking it questions like a you know typical non-technical
80:45 project manager might and then obviously the fun part doing the manual tests like
80:49 we will run the rag pipeline watch it pull my videos and put those into a
80:53 knowledge base that is what we're going to get to in just a little bit here as
80:56 the final layer of validation so validation strategy is one of the things
81:01 that I have in the structure plan and I can show that really quick as we're
81:04 wrapping up the implementation here like if I go to my create plan command and I
81:10 scroll all the way down. One of the things that we have is the validation,
81:14 right? So, like before finalizing the plan um or no, this isn't the right
81:20 place. Sorry, that that's uh that's validation for the the uh plan itself.
81:24 What I meant to refer to is this right here, the testing strategy, right? Like
81:28 create the unit tests, create the integration tests, here are the edge
81:31 cases that we need to cover. And so then with within our plan explicitly if I go
81:41 Oh well I guess I can't search a preview. All right let me scroll down.
81:45 Uh here we go. So phase four testing like unit tests. Here are the tests that
81:50 it should be creating right and then like here are our integration tests and
81:54 then here's our validation around linting. Right? like instructing how to
81:58 do those things task by task and it created that all as a separate task in
82:02 archon as well. So that's what it's doing right now. So it actually is to
82:05 the end of the testing where it's uh creating the or it's doing the linting
82:15 or I guess my pi is for type checking. Yeah. So rough is for linting. Anyway,
82:22 that's that's too technical. Uh, you should try Native Cursor again.
82:30 Ross Mick did a video and said it's ripping right now. I did actually see
82:33 his video. Well, I didn't watch his video, but I I saw it was like
82:38 recommended to me. Um, so yeah, I do want to check it out cuz yeah, I mean
82:43 like part of my my job for all of you guys is to constantly be on top of, like
82:46 I said, cutting through the fluff, like what are the tools that are actually
82:50 worth paying attention to? Um, and like that's what Rasmix is doing as well.
82:53 Like he's trying out Cursor again. He's figuring out like, oh shoot, like this
82:59 is good. And and um yeah, I think I'm allowed to say this. I I got like a a
83:02 little bit of pre-release information that Cursor 2.0 is coming out rather
83:07 soon here. And I am going to be getting um like a preview access to it. So, I
83:11 might be sharing that with you all in a video. Like, we'll see. It just depends
83:15 on how impressed I am with it because I'm not going to show you guys something
83:18 unless I I really think it is worth paying attention to. And like even for
83:22 sponsors, I'll do that. Like I turn down sponsors all the time that I think are
83:27 just stupid. Um or not well that's mean and not not worth your time. Like it's
83:31 part of the fluff. Um just being blunt. And so like yeah I'll cover cursor 2.0
83:35 if I if I really am impressed with it. But like I will try it out 100% Chris
83:39 and I think there is a good chance that it will impress me because I do know
83:43 I've heard recently that it's pretty Um Archon works like a charm. First time
83:52 using it in such a green field project. Cool. Glad to hear it. Nice. Yeah. Yeah.
83:57 If you guys want to try out Archon, it is a completely free and open source
84:01 tool that I've just been building uh on my channel recently or like kind of like
84:04 for my channel recently. Got a team behind it as well. Um and yeah, it is
84:11 all about giving yourself a command center for working with AI coding
84:14 assistants. And so kind of like what we are using it for right now, we have our
84:17 knowledge base where we're pulling knowledge like super data for example.
84:21 So it can like look at how to leverage super data in Python code or using
84:25 pantic AI to build our agent. And then we have the projects as well. So it
84:30 gives a canban board for us to manage manage projects and the coding assistant
84:34 gets to through the MCP server as well. So it's an interface for human and AI
84:38 collaboration, right? Like that's the beautiful thing that we're we're trying
84:42 to do with Archon is like building this single application that has all of the
84:46 human AI collaboration points that you need for um coding and even extending
84:51 beyond coding for just like general agents at some point as well. Uh not to
84:55 spoil too much of what we have in store for Archon, but definitely we got big
84:59 things coming for it. And it's it's a blast to like use this as a part of any
85:03 development workflow like I'm doing right now. So yeah, I appreciate that a lot, Yandi. Thank
85:11 you. Hoping for an Obsidian setup walkthrough. Ah, yeah. Uh, I could do
85:17 that. You know, it's kind of interesting because if I if I cover like my Obsidian
85:22 setup, I mean, sure, I have some AI things in it, but it'd kind of be like
85:26 my first video ever where it's like the main topic isn't actually AI. So, I
85:31 don't know. We'll see. I I might do that, though. I think it would be a
85:34 useful video to just like show how I use obsidian and like the zetocasting
85:42 Love Archon currently grinding through a project for me. Cool Barry, I love it.
85:45 Yeah, that's always the most satisfying thing when you can just have a coding
85:48 assistant rip through something while you're doing something else,
85:52 multitasking, watching a stream, taking your dog out for a walk, playing video
85:56 games, working with another coding assistant, like whatever it is, that's
86:01 always that's always the best. I need a Twitter bot, my dog. A Twitter
86:10 bot. What do I need a Twitter bot for? Maybe crew AI. Crew AI. So, I'm using pantai
86:19 for our build here because I prefer it overall. I think crew AI is a pretty
86:23 pretty great Python library for building agents. It's a good agent framework,
86:26 especially for multi- aents and making that really simple, but I definitely
86:29 feel like I have a lot more control and customizability when I'm building with
86:33 Pantic AI specifically. It's less of an abstraction. And so, um, it's it's less
86:38 of a black box, right? like when you're using Crew AI, they make a lot of um it
86:43 it's I guess I'd say it's more opinionated of a framework like you get
86:47 to customize things less. I think that's Um good question. Do you have a meta
86:55 prompt generator for creating the best possible prompt? Uh example, a prompt
87:02 for plan creation or it's just fluff. Yeah. So the the metaprompt generator
87:08 that I use for AI coding, it really is the structured plan that I outline in
87:16 the plan, right? Like this uh let me go So this part right here, like this
87:24 entire or wait, sorry, no, I need to go down to here. Okay, there we go. This
87:29 right here, all of this is outlining the core components that I want of my
87:33 structure plan. This kind of is my metaprompt. And so like I inject this as
87:38 a part of some planning command and I say like use this as your template to
87:42 create the structure plan. So it kind of is my metaprompt generator if you will.
87:47 Now for like an AI agent system prompt for example, I don't have something like
87:50 that um because I think that would be pretty rigid. Uh, but generally I have
87:56 like my core um components to a system prompt that I'll follow and I'll use an
88:00 AI coding assistant to help me outline that. But I wouldn't I wouldn't say I
88:03 have like a meta prompt generator for something like a system prompt if that
88:07 makes sense. But good question though. Yeah, learn the technology versus the
88:13 tools. Right. Yeah. Like when I said capabilities not tools, that's exactly
88:23 Uh does pantic AI have multi-agent system or are you using langraph? Yeah,
88:28 so pideantic AI does support uh building pideantic AI multi- aent systems like as
88:33 graphs. They do support that. I still think that langraph is a lot more robust
88:37 though. They also had their 1.0 release by the way if you guys didn't see that.
88:40 Uh lang graph and lang chain has their like official 1.0 release like as of
88:44 just a few days ago which is super cool. I think it's a lot more robust for
88:49 things like um human in the loop and uh state persistence things that pantic AI
88:53 might catch up with at some point for their multi- aent like graph library but
88:57 I don't think it's like a huge focus for them. I mean like lane graph is like all
89:01 about multi- aent and then like that's just like a subset of panticai, right?
89:05 So it's just a matter of like langraph is more specialized in that. So I feel
89:08 like they are probably always going to Yeah, let's see where we're at in the
89:15 implementation here. Curious if we can um because I want to move on to the Okay, so we're doing the
89:21 validation now. Got it. Okay, so we're almost there and maybe like we can
89:24 validate things kind of at the same time as AI. So it's creating the unit tests
89:28 for the rag pipeline right now which is cool. So if we go back to archon
89:35 uh we are at the very end here. We still have to do our database setup. Um but
89:39 other than that like we have the validation and linting. We're just
89:42 currently going through all of that. So that's why everything's in the review
89:46 stage right now. And then that'll be moved into the done stage obviously once
89:51 we have the uh testing done probably on our end as well. So, we'll do the manual
89:56 testing. Actually, we can do that right now. Uh, I'm really curious if I could
90:00 just let it rip. Let the rag pipeline rip. So, let's see what we have. And
90:05 again, for the sake of brevity, I'm I'm not going to spend that much time doing
90:09 a code review right now. I would highly recommend not vibe coding. I would
90:12 highly recommend going through and seeing what is what's created, what's
90:17 changed, getting a sense of uh the decisions that the coding assistant made
90:22 when it was writing the actual code. But I just want to look at this really
90:25 quick. For example, the uh chunking service. Let's look at the chunking service. I
90:32 just want to make sure that it's using Let's see. Uh okay. Decisions are okay.
90:47 I don't know. I don't know what this tokenizer map is because that that seems
90:51 a little rigid to me. But anyway, I'll I'll try setting it up in a second here.
90:57 And uh where where is the Dockling stuff? Hello. Let me search for Dockling.
91:08 It better have incorporated Dockling. wait a second. did it not? Where is this? Is this It
91:27 didn't do anything with Dockling. That seems like it it totally dropped the
91:31 ball here. Unless I'm missing something big. Hold on. I think it just made a terrible
91:38 mistake. Where? Let me Let me go back here because do I do I call out Dockling
91:45 in my plan here? I do. But then if I search through my codebase
91:49 for dockling, it uh literally only has the example Yeah, I did try with CK. It didn't find
92:00 it. I think it I think it like You know what? I should have made my plan
92:04 shorter. It's probably my fault for leaving the planet 2,000 lines because
92:08 it is it I think it totally hallucinated that it doesn't seem to be using
92:13 Dockling for Rag at all. Like that's that's stupid. Come on. Like what is it
92:20 doing for chunking here? Um, wow. There. Lesson learned. I I tried
92:24 because I mean obviously I wouldn't keep it as a 2000 line plan for real, but I
92:28 was just trying to go fast for you guys so we can get through everything. But
92:32 that's that is like a live lesson for you guys why it is so important to make
92:36 sure that you're not overengineering your plans or your code because it just
92:39 leads to these problems where it completely misses things that that is
92:43 like stated extremely explicitly in the plan. Like if I search for Dockling in
92:47 the plan, let me open up the markdown again because I I guess I you can't search in
92:51 an overview. Like if I search for Dockling here, like boom, it's it's in
92:57 my dependencies. Um, and then here like we're doing dockling hybrid chunking
93:00 based on the example. Like we we literally have that and it just ignored
93:05 it completely. So yeah, make sure that your plans are not 2,000 lines long
93:08 because I'm going to have to go and correct it now. And like in the end,
93:11 this is going to actually take more time than if I just went and and had the plan
93:17 correct and concise from the get-go. So yeah, but uh I'll still try to run it
93:23 right now and then correct that later. But it is a very good lesson.
93:27 So, let me open up the readme because it will have documented how to run things
93:34 as well. So, I'll go ahead and do a UV sync. So, I'll just follow its
93:43 All right. And then I need to let me start the environment. BNB scripts
93:48 activate. Clear it. All right. And then I need to create my environment variables. So,
93:54 what do I got here? It wants my super data API channel ID and then it wants my uh things for my
94:02 embedding model which looks good my superb basease configuration and my
94:06 agent configuration which I don't need this right now so I don't really know
94:09 why it created that either but that's all good and then I can specify a
94:12 minimum maximum chunks which defaults to the 400 1,000 that I specified. So, like
94:17 clearly it got a lot of my instructions, right? Like it didn't totally botch
94:20 everything. Like even the little details like this it got, but it's just the
94:24 dockling stuff specifically that it must have for some reason made a decision
94:29 somewhere in here. And I wasn't looking cuz I was busy talking to you guys,
94:33 which I mean I it's worth it, right? Like we we'll correct it. We'll we got
94:36 it. But somewhere I made a decision here to not use Dockling for some reason even
94:40 though it was in the plan because it read the whole plan. Like we know it
94:43 did. It clearly did. But yeah, let me go ahead and copy this,
94:51 paste it here, and um rename it to env. So for my YouTube channel ID,
94:58 I just need to go to my uh channel on another browser here. Or actually, hold on. I kind of forgot
95:05 how to get a channel ID because if I just go like this, it is um just like at
95:12 Cole Medine. So, let me search how to get channel ID. YouTube.
95:19 Go to settings on your YouTube account. Navigate to advanced. Okay. Well, I'll
95:25 try to do this really quick. Go to settings on my account. Customize
95:31 channel. Oh, here we go. Okay. All right. This is my my channel ID. All
95:36 right. And so, we'll go seven days back. One for max retries, five for batch
95:40 size. It looks good. I'll need my super data API key which I will set that off
95:45 camera obviously. And then I will need my um credentials for OpenAI and Superbase as well. So
95:52 okay, for these things I'm going to set them off camera obviously. Don't want to
96:01 expose that. So uh let me pull this up and um I'll have to bring this off
96:06 camera as well because I'll grab my super data API. Well, I guess I can just
96:11 show this. I just grabbed it right here. So, super data API. Set that. Boom. And
96:34 Boom. I'll leave the model as text embedding three small. That works for me. And then
96:43 for my database here, data API project URL. Copy that. And then for my key,
96:54 I will reveal and copy that. Boom. Okay. Good. Okay. Wish I could show you guys that,
97:03 but this is the one thing that's like kind of awkward I have to do off camera.
97:07 But yeah. All right. I got everything set. Close that out. Go back to the read me
97:14 here. And uh so now it'll process videos from the last seven days if I run this
97:19 command right here. So apparently this is the moment of truth. I don't have a
97:23 ton of faith in it right now because I skipped through the the planning phase
97:27 too quickly, but let's see what happens. All right, so back in the terminal here.
97:36 Let's go ahead and kick this off and hopefully I don't just get like an error
97:38 immediately so that there's at least something to show. Oh, I don't have my database tables
97:45 created yet. So, that's going to be a problem. also it says that the superbase URL is
97:58 required, but I do have that set. Superbase URL. Okay. So, it's uh it needs to fix that
98:05 as well. But yeah, it never actually created things in the database yet. So,
98:10 that's another thing where I I guess I can't really validate it yet because
98:13 it's still in the middle of uh creating the database for me. So, all right, let
98:18 me go ahead and paste in this error. Uh even though I have the M&M set, also you
98:25 need to create the SQL for me to build the database tables. All right. Yeah, man. It's missing a lot
98:31 of things here because that's another problem. When your plan is too long, it
98:35 uh picks picks and chooses what it does and and it doesn't do things like define
98:42 the SQL for the database and um like it's not even loading in the
98:45 environment variables environment variables apparently. Like all these
98:50 problems with my typical system for AI coding that I just tried to simplify and
98:54 like go through fast here like it would have knocked those things out. But it's
98:58 just because I'm going quickly that it's messing up. Um, so yeah, again, lesson
99:03 learned. But yeah, while it goes and fixes this here, I'll go back into the
99:06 chat here. Just uh keep having a good time with you guys while Cloud Code
99:13 works hard way for me. What is the best free coding model right
99:17 now? Yeah, using the Gemini CLI with their like really generous rate limits
99:22 is the best right now, I would say. Yep. And then if you like I know it's
99:25 not free but like Codex with like the $20 a month plan it gets you really
99:36 The plan had a dockling service. It did. Yep. Yep. It totally missed it.
99:41 It spelled it wrong in the plan. I mean, it honestly might be. I don't think that
99:45 incorrect spelling is what caused the problem in the end because it would like
99:49 realize that like the spelling is incorrect and it's still like know like,
99:52 okay, I should be using this external library to build in the hybrid chunking.
99:57 So, I I think it's pretty nuts. Like, it's it's actually crazy what a plan
100:01 that's too long can end up leading to as It would be a really it'd be really cool
100:12 if Cole could shed some light on the above mentioned concepts are relevant or
100:16 just fluff. If you tell me what concept you're referring to, I could definitely
100:20 answer. um I don't I'm looking through the chat. I don't see anything else,
100:32 but uh yeah, I've noticed that some MCP servers and skills really chew up
100:36 tokens. Usually those are the code review and testing agents. Pay attention
100:39 to these if you're using the lower limit cloud code plans. Yep. 100%. Yeah. And
100:44 like that's kind of part of like context engineering is um how can you create
100:52 commands like slashcomands and sub aents in a way where they uh don't go for too
100:57 long, right? Like being very explicit in your instruction set to like do simple
101:01 validation or do simple analysis or simple review. um because otherwise they're just going
101:07 to go on for forever, right? So yeah. Okay, cool. We got our migrations with
101:12 our database script here. So we got our channels and videos. So
101:15 like this is this is exactly what we have in the plan. Um so this looks
101:21 really good. So, I'm going to go ahead and copy and yeah, like it's basing it. I can I can
101:27 tell just in the way it's creating policies and things that it is also
101:32 following what I have in my examples here, like in this file right here. So, that
101:37 this is really good. Like, it's disappointed in some ways because my
101:40 plan is too long, but like overall it's decent. So, I'm going to go into my SQL
101:44 editor here in Superbase, create a new snippet, and paste this in.
101:49 Whoops, I didn't actually copy it. So, I just have my error message. There we go.
101:55 Paste this in and run it. And let's see if it works first try. There we go.
102:00 Okay. Nice. So, now going to my tables, I've got channels, transcript chunks,
102:04 and videos. Nice. So, looking really good. So, the database is set up. And
102:09 now we can see a moment of truth if we are able to run this thing.
102:15 I would be pretty happy if it worked first try and then we can always add in
102:19 the dock link stuff later. Okay, so we got our logging. This might
102:24 be a little too verbose. Um, let's see. So, the channel. Oh,
102:30 okay. So, it's expecting the channel to already be here. So, I guess maybe I
102:33 just need to add that. Ah, yeah, that makes sense. So, I need to I need to add
102:38 my channel explicitly. Um, and then it'll work. So, name I'll do Cole Dean URL.
102:49 I'll do it like this. uh active true created at good good save
102:55 okay nice so that makes sense like I don't I shouldn't expect my rag pipeline
102:59 to like create the record for my channel so now let me go ahead and run this and
103:04 then we should be good so we're making okay uh okay there's one error that we got
103:17 looks like that's repeating quite a bit so but This this is good. Like an error
103:21 is an opportunity for me to also like teach you how I handle errors. And so
103:27 it's it's um retrying like over and over and again which is also not good. So two
103:30 things that we want to correct here. So I'm going to interrupt it. I'm going to
103:36 copy this error and I'm going to paste it in. And so um here's kind of like the lesson that I
103:45 want to give with this. Um hold on. My live stream just crashed. Let me I
103:50 got to make sure I'm still live because my my web browser just crashed. I think
103:53 I'm good though. Okay. All right. So, the lesson that I want to give with this
103:57 is that when you encounter a problem with your AI coding assistant, like it
104:02 messes up something in the code, it's your job to figure out how drastically
104:07 incorrect it is for something little like this error right here.
104:12 This is pretty small. At least I'm pretty confident that it's small. And so
104:16 I'm just going to, you know, very simply paste in the error and and ask it to
104:20 correct it. If it was something really big, like honestly it completely
104:25 skipping dockling might honestly warrant this. I would actually like go back to
104:29 the planning phase of the piv loop, right? So I'll I'll pull this up for a
104:33 visual again. So like we're right here in the validation phase and we realize
104:37 that it completely messed up something in the implementation phase. So our two
104:43 options is one we can either just like stay in validation and ask it to fix
104:47 something or we can go all the way back to planning and correct our plan and
104:51 then based on that correction have it try to reimplement things like I could
104:56 literally go into my source control here and I could wipe everything start from
105:01 scratch with the plan and have it retry maybe after I fix something in the plan
105:05 like maybe in my plan I need to make it much more concise and then I also need
105:08 it to be very clear that like you should be using dockling for your hybrid
105:14 chunking. Do not skip dockling for the chunking strategy. Right? Like I I could
105:17 add that into the plan and then go back into validation. I'm not going to show
105:21 that right now for the sake of speed, but sometimes it's worth doing that if
105:26 you realize through your code review and manual testing that it really screwed
105:29 the pooch on this one. And it kind of I mean it did in some ways. We're going to
105:33 make it work. We're going to make it work. U but I think that's a good lesson
105:37 to to show here. And so let me go back to the code and uh just have it fixed this here. So uh
105:44 please fix this error. Also when it was getting this error, it was retrying
105:47 infinitely. It should also just retry a single time. So address both of those
105:51 things. Give me a very concise summary of of what you had to do to fix both.
105:56 Right? And so also adding that sentence in at the end because I want to be a
106:00 part of the validation here because again going back to the diagram here,
106:05 the implementation is where we delegate to AI. But then for both the planning
106:10 and validation, we want to be very in the loop. I want to be very in tune with
106:15 what we're planning and how things are actually going in the implementation
106:19 when we get to validation. And uh these are the kinds of things
106:23 like like I I keep telling you guys here that like I'm I'm going over something
106:27 quickly for just the sake of brevity in the stream. Like I keep saying that over
106:30 and over again because like we are already an hour and 45 minutes into the
106:34 stream. like it takes a while to get to that point where we have like a
106:38 completed code and we have gone through the piv loop. And so again just just
106:42 calling this out again like if you are really interested in going really deep
106:46 into these concepts like that is what Dynamus is for in the new agentic coding
106:50 course because in the agentic coding course that is where I don't skip
106:55 through anything and as we go through the mental models and the piv loop and
107:00 how to start to build our systems that are specific and optimized to your code
107:04 bases how we can automate those systems and integrate them with our applications
107:09 even doing remote AI coding leveraging out of the box solutions like the PRP
107:13 framework and BMAD like that. It's we're covering the whole landscape of AI
107:18 coding here and I'm not glossing over anything like I'm forced to do in in
107:22 something more concise like this. Like it's so it's so much detail that I cover
107:27 here and like making sure that we don't have a plan that's 2,000 lines long. And
107:30 I talk about what it looks like to shrink that. Um and even doing like the
107:33 same thing for global rules because we want global rules to be concise as well.
107:37 And that's one of like the big lessons for you guys to get out of this just in
107:41 general is like everything should be simple, concise because LLMs love to be
107:47 uh over overgineering and they love to just talk talk talk in their plans like
107:51 you guys saw and they just like to charge ahead. And so one of the big
107:54 things about building reliable and repeatable systems is that um you want
108:02 to do it in a way where it is um focused on simplicity first. like it's just so
108:10 important to focus on simplicity first. So yeah, um let's see. I gotta my chat
108:16 didn't pop out. Let me try one thing again. So my my stream like crashed on me. So let me try
108:25 putting this in the chat again and seeing if that shows up. Well, like stream didn't uh crash on me,
108:32 but my browser did. So okay, now I can show this. Yeah. Yeah. So all right.
108:35 Yeah. If you're interested in joining Dynamus, there's the link and the codes
108:38 that you can use to get a discount on the monthly and the yearly. So, hope to
108:43 see you inside for the course and everything else we have in Dynamus. So,
108:48 yeah, with that, let's uh head back over to All right, good. So, fix both issues.
108:54 Type error. Uh, super data returns objects, not dictionaries. Okay, makes
108:59 sense. And then infinite retries. Um, okay, good. So, I made an adjustment there.
109:05 Looks really good. So, we can actually try this again. Hopefully, it'll work better this time.
109:11 My goal here is just to have something show up in the transcript chunks and
109:15 videos tables like that. That's the goal that I've got. So, let's see.
109:23 Okay, it failed again. Well, that's fun. I guess we get to iterate more. I was hoping not to
109:27 iterate too much, but uh that's the nature of of AI coding. Sometimes you
109:32 just have to do it. Um cuz I think what I want to do is just
109:38 ask it to let's see I'm getting this error now. I want to just make the
109:41 validation very simple. I think you're overengineering this right now. Just
109:45 need it to be simple. I need it to to work. That that was maybe a little
109:50 silly. But yeah, and then the uh infinite retries was still happening.
109:55 All right. Not spelling it correctly Uh, all right. Greeting champs. Carrot
110:05 to guide a lost muggle non techie. Does coding specific models help in task
110:09 other than coding? Uh, example obsidian related. Moreover, can someone check how
110:13 legit Lemon AI agent is? I'm not familiar with Lemon AI agent. I' I've
110:18 got no idea. Um then but to your other question though for uh AI coding models
110:23 they do definitely help with other things related to file management
110:28 because like coding LLMs are specialized not just on like writing code but like
110:31 handling files right like moving files creating files editing files and so for
110:35 something like Obsidian where it is just like file management they they
110:38 definitely shine with that and more like instruction following as well. So yeah
110:50 Um, Coleman has a true open source mindset. The community is private and
110:53 paytoplay, but his tools are free and available on GitHub. Yeah, I appreciate
110:59 that. I I hope that helps. Um, yeah. So, yeah, the the answer to your question
111:02 is, uh, does the course use your software? Must we use it? The answer is,
111:07 um, no. So when I talk about my open- source projects like Archon or I've
111:11 covered like my PRP framework a lot before uh that Rasmus built um and like
111:16 we've showcased on my channel a lot like I have my specific strategies for AI
111:21 coding um but the course itself doesn't cover those more opinionated things
111:26 until the end of the course. And so very intentionally I'm starting from the
111:29 ground up to help you define your own system so you're not reliant on my
111:33 tools. I talk about how you can use them, but more importantly, I show you
111:37 like how you can um build your own system regardless of the tools in the AI
111:42 coding assistant that you're using. So, really good question. Not limited to my
111:46 stuff. Like I I never want to just like push my stuff on you. I want it to be a
111:50 resource that you can leverage when you want, but you never need to um to like
111:53 get the most out of like what I'm covering. Like I'm using Archon in this
111:56 live stream because it's a fantastic tool, but also you could definitely just
112:01 like have it search the web for super data instead of uh using the super beta
112:06 docs in Archon. Or you could have it uh manage its own internal tasks instead of
112:10 using archon for that, right? Like all those things. It's never a requirement
112:15 to use archon as an example. So all right, let's try this out here.
112:20 Yes, you're you're very welcome, Stephen. Okay. Um, I think it's actually working
112:32 really well now. Let's go. All right. Finally, our rag pipeline is in
112:38 business. Let's go. Let's check the Let's actually check the tables here.
112:45 Uh, videos. Nice. Yep. Yep. Looking really good. Okay. Nice. So, let's let's validate one
112:56 of these. I'm going to check this video. The only problem I have is it it feels
113:01 like it went it's going longer than one week. So, I think that is bad. Like how
113:07 Okay, this is See, this is two weeks old. So, okay, things are working
113:12 almost. It's just like um not respecting the fact that I only want to go one week
113:16 out. So, I'm going to I'm going to cut this short here. But, it's working now.
113:19 Like, we're making good progress. And a lot of this iteration, I want to be very
113:22 clear, I probably wouldn't have to have done if I was a lot more explicit and
113:27 concise in my planning. I'm going to say that over and over again because I
113:31 honestly regret not spending more time at that with you guys here because we
113:34 probably would have gotten to where we are now, but it's just a better planning
113:37 document in like the same amount of time at this point. But, you know, hindsight
113:41 2020, I've never built an agent live with you guys. So, that's part of what I
113:44 have to learn as well just as I'm doing it here. So, I I appreciate your guys's
113:49 understanding as I'm uh as I'm doing this, but yeah, it looks like it um went
113:54 really far. Like, if I go to this video now, >> uh which one is this?
113:59 >> Archon beta launch live stream. This is from two months ago. Okay, so
114:04 not what I wanted it to do, but it is kind of cool that like that was actually
114:08 pretty quick to get the transcripts pulled and embedded and inserted into my
114:11 knowledge base for the last two months worth of videos. like that is pretty
114:18 neat. So, okay. Um, but I'll just go into the coding assistant and say
114:22 everything is working great now. However, uh, it's it pulled videos from
114:27 like even two months ago, like 20 30 videos, uh, when it should have just
114:30 done the last couple from the last week. And so, it definitely is not respecting
114:34 the um the day limit that we have set in the environment variables right now.
114:40 Boom. All right. So, we're almost there. And then honestly in the meantime um I can and thanks for
114:48 the kind words David and Frosty Frog. I appreciate it guys. Yeah, do my best. I
114:52 mean this this is like super fun. Like I'm having a blast with you guys even
114:54 though it's like a little bit of scrambling in my mind trying to figure
114:57 out like okay how fast should I go versus how comprehensive should I be?
115:00 That's always a challenge with these live streams. But yeah, while this is
115:04 going here, I'm going to actually open up another cloud instance in parallel
115:09 because I want to do the implementation for the agent now. And since I've
115:13 already covered the pivot loop and we already have our considerations here, um
115:19 well, I guess we don't have it for this one. I'm going to go through this one a
115:22 lot faster now, right? Right? Cuz like I already went through the full piv loop
115:26 with you guys and we talked about um the commands, we talked about archon, we
115:29 talked about how we are going through the vibe planning. I'm going to go
115:33 through this one a lot quicker now and it also should be simpler because we
115:36 already have the core part of our codebase set up and like our database
115:40 tables and everything. And so now I'm going to go into a parallel session
115:44 while the other thing is still fixing some stuff for the rag pipeline. And
115:49 let's go ahead and knock out the agent. And so I'm just going to say, well,
115:53 first of all, I'm going to do the slashp primer. So I'm going to prime the coding
115:59 assistant on my codebase. Um, well, you know what? I need to fix up this primer
116:06 here. Read the key files in source. Um, there we go. Cool. Because I had I
116:09 didn't have that customized this codebase yet. So I'm going to go back
116:13 into cloud again. Just going a little bit quicker here so you guys get to see
116:19 see a speedrun. Um, so all right, slashp primer. All right, so I'm going
116:25 to do a primer now. So I talked about the primer just a little bit earlier,
116:30 but it's a command to just instruct it very simply on like here's what I want
116:33 you to read through and understand so that you're caught up to speed on the
116:36 codebase, right? Because we started a brand new conversation, but I wanted to
116:39 understand what I already have in my codebase as I go into planning for the
116:43 next feature. So, I'm going to let the primer run. And then while it runs, I'm
116:47 going to go in and let's just do some planning for our agent here. So, I'm
116:51 going to go ahead and delete these two things because we're not working on that
116:55 right now. I wish I had time to get into other things like observability, but
116:58 definitely won't. So, I'm going to say building a rag agent to search my
117:04 YouTube knowledge base. And so, the text stack here is going to be Python,
117:09 Pyantic AI, and Superbase, right? And then for other considerations here, we
117:14 need to be able to specify the LLM provider um and model inv
117:22 like in the examples folder. Right? So we have that one of those things in the
117:26 examples folder already. And I'll say we want a rag tool to search chunks and a
117:34 rag tool to read an entire transcript. And so transcripts are short enough
117:38 where they can definitely fit in the context for an LLM. And so if we find a
117:42 chunk for a transcript and the agent decides like, oh, I actually want to
117:46 read the full transcript. I want to give it a tool to be able to do that. Like
117:50 that's kind of the agentic rag piece that at least at a in a simple starting
117:55 sense will give our agent the ability to um search our knowledge base in
117:59 different ways and make it more adaptable. So it can like actually think
118:02 to itself, I want to read the entire transcript and not just focus on that
118:06 little segment because maybe the user's question is more overarching. So I'll do
118:14 that and then I'll say um we need to use the match function in the
118:23 SQL file for the rag search tool. What other consideration maybe you guys can
118:29 uh um share with me in the chat what kind of other considerations you think
118:32 we should have for building this podantic AI agent that can search our
118:41 and and Bo the the tool that I use to fetch YouTube transcripts is called
118:44 supera data super data it's this one right here very very affordable and then you could also
118:51 use the YouTube API directly if you wanted to um only work with your own
118:55 videos. I'm using super data because it's a way for me to crawl like any
119:00 video from any channel. It's very nice. Do we need a chunker? So, we don't need
119:03 a chunker here because we already took care of that in the rag pipeline. So,
119:06 right now we're just building the agent with the tools to be able to search. Um, another thing
119:14 that I can think of here is um we need a we need to build the agent API exactly
119:21 like agent API.py pi in the examples folder. Right? So I'm calling out like
119:25 we have the example there. Uh and then I can even say like follow agent.py
119:40 from the example folder. Right? So like keep calling out like I said our cheat
119:44 code existing projects bring it into examples reference it have it look there
119:48 to see how we do things like setting up our database tables and our agents and
119:51 our tools. Like I I I want to keep leaning on that as much as I possibly
119:55 can. It's so important. Maybe a limit to the size of the transcript. I like that idea. Yeah. Uh
120:04 rag tool for pulling transcript. Uh limit to a uh size specified innv.
120:12 Right? So like just in case we try to pull like an eight-hour long YouTube
120:16 video, we could limit it. Um, and then like maybe even the tool would say like
120:19 this has been limited because it's too long or something like that. Uh, rag
120:24 aent citation with timestamps and video ID. I love that. Yeah. Um, rag tool to
120:31 site timestamps and video ID. Well, even better, we could say to site video URL
120:37 and timestamp uh URL with timestamp even because there's a way to do that in in YouTube
120:43 where you can have the the time stamp be a part of the URL so that when someone
120:46 visits that link, it takes them right to that time in the video. So, I think
120:50 that's really good. Cool. Include citations. Yeah. So, Tony said it too.
120:54 Love that. Yeah, I definitely should not be missing that in my considerations.
120:58 So, I think this is good. This is good to start. Um, so I'm going to copy this,
121:05 go back in. Um, first of all, for our other instance of cloud code, we are now
121:11 only going back to 7 days. So that's good. Um, so now within here, it has
121:15 primed itself. So we finished the priming. So now the coding assistant
121:20 understands our current project. And so I'm going to paste in just like I
121:24 did before the other considerations. And I'm going to speak to it again. So right
121:28 now, this project only has the rag pipeline. Now I want to build the
121:31 pidantic AI agent. So take in all the considerations that I have specified
121:35 above. I'm going to be using again Python pyantic AI for the framework and
121:41 superbase obviously to reference uh the knowledge that we're storing from the
121:45 rag pipeline. I also by the way have the documentation for pyantic AI available
121:53 in archon uh for you to search. So take a look at that right here.
122:00 All right. Uh then I also want you to reference quite a few files that we have
122:05 in the PRP's examples folder. And so within backend agent_appi,
122:11 we have agent_appi.py and agent.py as well as tools.py. I want
122:16 you to read all three of these files in entirety because I want to create the
122:22 pyantic AI agent in a very similar way. And the the this agent also leverages
122:26 the environment variables that we have in thev.ex example file there in the
122:31 example. So I want you to take a look at that. This will give you a really
122:34 holistic picture of how to build the agent to work with any LLM. How to build
122:39 tools and the dock string for tools. Uh and then the other really important
122:44 thing here is that uh I want to copy what we have for the agent API.py like
122:50 pretty much exactly like I want to build the API endpoint in exactly the same way
122:55 so I can use it with the same front end that I already have created. And I want
123:00 this API to be um probably a lot simpler than the one that we have in the
123:02 examples. Like I don't need anything with me zero that I have in there right
123:07 now. Um but yeah, I definitely want to have the same structure. And you can
123:12 literally copy the db_youutills that we have in this example folder. You
123:16 can copy that and use that exactly because I have all the functions here to
123:20 manage the conversations and messages in the superbase table already. Um, but
123:24 then just make sure that you add the tables that we have from
123:30 Make sure you add the messages and conversation and requests and all of
123:34 that um to the primary SQL that we have in our migrations folder here.
123:39 Um, and so yeah, right now I want you to explore all these files and help me come
123:42 up with a plan and my different options. Don't implement any code yet.
123:46 Okay, that ended up being way more than I thought it was going to be. But just
123:50 as I was talking, I was realizing all the different things I wanted to
123:54 incorporate from my example. And I'm going fast here. I know there's like a
123:56 lot of different things that I'm referencing from my previous project. Uh
124:01 but yeah, like I I want to like have it work explicitly so that it can um be
124:05 leveraged immediately within the website that I have here. And so yeah, I'll let
124:11 it go and I'll I'll see. And like this time I'm also going to tell it to create
124:15 a lot simpler of a plan. And so hopefully this will give you guys a nice
124:18 demonstration of uh how things look a lot different when we do have a simple
124:22 plan and things aren't overengineered uh because Claude likes to talk talk talk.
124:29 So all right, context window exhausted. Well, this is a brand new context and uh
124:33 even with all the talking that I did here, like it's not that bad overall.
124:36 Like this is only a couple hundred tokens. So I think we're good. um
124:41 because it costs a lot more tokens when Jen said, "I always use big plans, but
124:57 have a step-by-step plan. Do step one, then step two, then no problems with
125:00 forgetting things." That's good. Honestly, that's kind of what I was
125:03 going for, but it still forgot things. So, I don't know. It It could have been
125:07 more of a fluke, too. It is pretty crazy that it uh ignored Dockling. Um or maybe
125:12 there's something that I'm missing and it actually is leveraging it, but I
125:17 think that it missed it, but oh well. All right, cool. So, it's it's searching
125:22 the rag knowledge base for Pantic AI. That's looking really good. So, that'll
125:26 help create its plan as well. And so, once this is done, then I'll I'll run
125:30 the create plan command. So, all right. Yeah. So, uh, while this is going,
125:39 um, one thing I want to say is the the use case specifically. Let me actually
125:44 pull up I'm going to pull up something else here. Don't mind me. I'll go to the uh, course
125:51 introduction that I have for so this is like the Excaladraw diagram that I have
125:56 for the introduction to the Agentic coding course. So, uh, let me put this
126:06 So for the agent coding course, this is the introduction that I got. One of the
126:09 things I want to call out really quickly I think is super neat is when I show the
126:16 principles of AI coding and creating a full system around AI coding. I'm not
126:22 just doing it on like some random project I'm going to scrap after the
126:25 course is done. I wanted to build something freaking awesome. So, not only
126:29 do you like have the course and like everything you learn from it and the
126:31 systems you build from it and the templates there, but you also get a
126:37 complete AI agent built as a part of the course. And the agent is the Obsidian
126:42 second brain assistant. So, it's leveraging obviously Obsidian and then a
126:46 lot of like the same stack that we're covering in this stream right now like
126:49 with Pineant AI for building out our agent like that's the framework and then
126:54 Superbase for our database. And so it's an agent where like you can literally
126:58 have it um brainstorm based on your documents, help you organize your notes
127:01 and knowledge in Obsidian. There's a lot of more tools that we're going to add as
127:05 we go throughout the course like with the very full schedule that we have for
127:09 the course over the next month. Like yeah, so it's a really awesome use case.
127:13 So I just want to call it out explicitly um because this is going to be like
127:15 another really big resource in the Dynamis community. So again, if you're
127:19 interested in taking your AI coding to the next level and you really want to
127:23 dive deep into some of the things I'm covering in this stream, then yeah,
127:28 definitely check out dynamus.ai. And the codes that I have on the screen right
127:33 now, uh those are going to go away right after the live stream. Like this is your
127:37 chance to get a discount that I'm not going to have uh for a long time again.
127:41 So yeah, 20% off the monthly and then 40% off the yearly membership. and
127:44 you're getting the agentic coding course, the AI agent mastery course, the
127:48 whole community access to me every single day and weekly workshops that I'm
127:53 doing every Friday, which those are a blast as well. Like in the last
127:58 workshop, um we covered Claude code or not Claude Code, we covered Claude
128:02 skills and uh plugins like both in one some cool use cases that I built out. So
128:05 like we're always covering like the latest tech and then using the workshops
128:09 as a way to like dive deeper into like really specific things for some of these
128:13 concepts as well. And so I'm actually going to be doing workshops related to
128:17 the agentic coding course throughout the next month as we're releasing the
128:22 batches. Um, and so I'll actually show you the schedule really quick. So like
128:26 we did uh we're doing a live workshop next Friday related to the first block
128:31 that was released just for the course now. And then like for the second
128:35 release block, there's going to be another live workshop in a couple of
128:38 week like in a few weeks from now and then a third one and then a fourth one.
128:41 So, it's it's almost like if you join right now, you kind of get to be a part
128:44 of a cohort going through the course with me for the first time with all
128:47 these live events to make it even more interactive. So, yeah, it's it's pretty
128:51 exciting. Like, this is like the main thing that I'm I'm working on right now.
128:55 Um, and it's been a blast and like a lot of the things that we're we're building
128:58 together here, but just going that much So, yeah. All right. Uh, is your course
129:07 a live course or is it pre-recorded? Can we view at any time? Um, yeah. So, good
129:11 question, Stephen. So, it is it's all going to be pre-recorded and then we
129:16 have the live workshops that are going to uh be a part of each of those
129:20 released batches, but the recordings of those will also be posted after. So, you
129:24 can view at any time. It's meant to be a course that will obviously stand the
129:27 test of time because of the principles that I cover and like will be available
129:31 for you to to view forever and then I'll evolve them over time. Like, if there's
129:34 ever anything that becomes like irrelevant in the material or like I've
129:37 evolved my approach, the course will evolve as well. Well, it's the same
129:40 thing with everything else in Dynamis is I I work my butt off to keep it like
129:44 really up to date and making sure that um like I said, I always help you fight
129:47 the fluff and just like work with or like focus on what works right now,
129:52 right? When is the first live course start? So, the first the the first uh batch of the
129:58 course is released. It came out on Friday. Um and then the uh first
130:03 workshop for the course is going to be this Friday, right? Right. So, like you
130:07 have a week to like go through the first few videos and then there's a couple of
130:10 like exercises that we have there and then you get to go into the workshop
130:13 where I cover my solution for the exercise and then later that'll just be
130:17 like another pre-recorded video basically in the course is the way that
130:20 I'm structuring it. So, it works either way. Um, but it like joining right now
130:23 is when you have the discount during the live stream here. So, yeah. All right. Let's go back here.
130:32 Uh, all right. We have our plan. So, I'm not going to read through all of this
130:37 right now. Typically, I would. So, you've heard me say that for the other
130:42 plan as well, but yeah. Um, so we're copying the exact structure of agent
130:47 API. Uh, looks good. So, we're doing the same thing for authentication and rate
130:50 limiting and conversation management. Like I I this is like a super robust API
130:54 that I created as a part of the Dynamus AI agent mastery um implementation.
130:59 So, it's recommending option one, minimal coach. So, okay, looks good.
131:04 Creating our tables, updating our environment variable. We're creating
131:09 only three tools like this. This is good. everything looks good. I option one is
131:22 definitely what I want. Um, but I don't want a tool for format video citation.
131:25 Like that should just be deterministic code that's included right within the
131:29 rag tool. Okay. Okay. Otherwise, cuz right like I don't really want it to call a separate
131:34 tool just to format things. Like that's stupid. But everything else is looking
131:39 really good. So API layer uh with authentication making sure it's testing
131:43 everything. And then it's giving me a couple of options and like I am just
131:46 going to go with the most simple one right now. U because also that if I
131:50 don't go with the most simple one, it's just going to be overengineered. Like I
131:53 can absolutely guarantee it. So I'm going to send this in and then after it
131:56 acknowledges that then I'm going to go ahead and create the plan. So I'll just
132:01 do create plan and I will say uh create a simple plan for the implementation
132:11 of the agent based on our conversation. Keep it between I'm going to be explicit
132:17 here. 500 and a,000 lines. Um cuz yeah this this is going to be better. And
132:20 honestly I should have done that for the other one also within the create plan.
132:24 This is also an opportunity to evolve my system, right? Like I could go in here
132:29 and be like um you know important. The plan should be between 500 and 1,000
132:36 lines. And maybe there's a lot more that I should do to like truly make this more
132:41 elaborate. But like even a simple addition like this is potentially all it
132:46 takes to learn from what happened previously and then fix it for the next
132:50 implementation. I'm actually going to try this. I'm going to delete this line
132:53 because now I have that as a part of my plan. Um, well, actually that's not true. I I
132:58 would have to reload claude code for this to take effect. So, I'm just going
133:01 to add it in uh right here, but you get the idea. So, keep it between 500 and
133:07 1,000 lines. All right, there we go. So, it's going to create an AI agent with two simple
133:12 tools. Nice and basic. I like that. And then we already have our knowledge base
133:15 set up. And it's actually kind of nice that it has that it went way too far
133:19 because now I have a lot of data to search through with the agent here. So,
133:24 we will see. We'll see if it works. I'm going to let it rip again and I'll
133:29 spend some more time in the chat as uh All right, you're good. All good. I I
133:41 know fat fingers happens. No worries. You're late, it seems. Well, that's all
133:44 good because the whole recording will be posted for the live stream after it's
133:47 done. I think it just takes a couple hours for YouTube to process, but you're
133:52 all good. Glad to have you here. That's Okay. Yep. Um Okay. So, Stephen, another good
134:05 question. What happens if the plan must be longer than 1,000 lines? Say 2,000.
134:09 Do you have to do a two-part plan and you build each project separately and
134:14 merge later? Yeah. So, I honestly I would say that if your plan has to be
134:18 2,000 lines, you're trying to do too much at once. Like I would I would try
134:23 to not like make a two-part plan. I would I would think of it more as like
134:26 splitting your implementation into two separate tasks entirely and then do the
134:32 piv loop for each of those. And so you're not splitting up the plan.
134:36 You're going one level higher to like splitting up the actual work that you
134:40 need to do. Um so yeah, I mean you could try to do 2,000 lines, but you're probably going
134:46 to run into the same thing that I did. Um, maybe there's a way to optimize my,
134:50 you know, execute command, for example, that would have it like handle every
134:55 every task better, but I told it to do things task by task in a loop and it
134:59 still missed the dockling stuff. So, yeah, I think just like you're doing too
135:02 much at once, so if it has to be longer than a thousand lines, and I'm going to
135:05 I'm going to have it do just uh 500 lines to a,000 here. Yep. You're very welcome.
135:14 I've done the this exact thing for a local rag agent. That's cool. Stevan,
135:17 what are you using to extract transcripts? So, yeah, I'm using a a
135:21 tool called Super Data. That is my tool to get the job done here. And so, I like
135:26 Super Data because it works for more than just like my own videos. Like, if I
135:29 use the YouTube data API, it only works for my own videos or ones that I have
135:33 edit access to. And then also, it avoids like rate limits and IP blocking and
135:37 stuff that you can hit if you're using like a just like a Python library for
135:41 example that just like crawls directly. So, they have like proxies and things
135:45 built into the platform. So, it's pretty robust and and production ready, I would
135:50 say. So, I like it a lot. Course amount is very high for poor
135:54 countries where $1 equals 90 in their currency. Yeah, I understand. I I wanted
135:58 to bring this up really quick cuz I saw a couple comments on that. I do
136:02 understand, but uh geographical price parity is something that I um I'll have
136:06 to look into more, but like initially looking at it, like just in the
136:10 platform, I don't think it's supported. Um, and I think that really is the case
136:15 unfortunately for most platforms like school, uh, Circle, which is what I use,
136:19 if you're using like I think it's called, like I don't I don't think they
136:24 support that. Um, so yeah, but I mean I there are people in the community from
136:29 all over the world. Um like we got a ton of people um in uh Germany, people in
136:35 Brazil, we got people um in like Saudi Arabia, like Dubai, um we got people
136:41 from South Africa, Australia. Um actually, so the United States has the
136:45 most members. I probably not a surprise to anyone. Um and then uh Germany is
136:51 second and then Australia is third. So I track metrics for where everyone's at in
136:55 the community. Um, it's pretty cool and like you don't have to share your
136:58 location. So, it's just for the people that do like obviously there there's no
137:01 privacy issue. Like you can be as secretive as you want when you're in the
137:04 community. Like it's totally up to you. But like for the people that share and I
137:07 think it's cool that it'll just share like generally their country or state or
137:11 whatever. Like it's cool to see the stats around that. A lot of people from
137:15 Canada as well of course. Um I think Canada's fourth and then like the UK is
137:21 fifth I believe. Um so yeah it's pretty cool. Um, yeah, I think Belgium sixth or
137:27 something. Um, I mean, I don't remember off the top of my head getting into
137:31 those numbers, but yeah, it's pretty sweet. So, yeah, very global community
137:34 and that that's like super exciting to me cuz there there's people that are
137:37 posting things and sharing things like every hour of the day because it's a
137:42 global audience. So, I always like wake up to like like five things where it's
137:45 like, "Oh, that's super interesting." Like every single day. It's pretty cool.
137:54 There's so much free stuff on YouTube from Cole and others. Yeah, that's
137:58 right. Yeah, I'm I'm like I talk about my community and my course a lot uh
138:01 today because that is like the forefront of my mind. Like it just came out, but
138:05 like 100% like I'm always going to be giving a ton of free stuff on YouTube as
138:08 well and that's never going to go away. Like my ethos is always going to be open
138:13 source um and just like giving a ton of value through the content on my channel,
138:16 but the community is just a place to go deeper. I think that's like a good way
138:19 to put it and to really cover the things that like I don't have the ability to
138:24 like even if I wanted to. I can't go through my whole process in the stream
138:29 today in a lot of detail because I'm going to lose my voice. Like I already
138:31 feel like I'm going to lose it pretty soon and I think it is going to be time
138:35 to wrap up the stream pretty soon here. I just want to try to knock out this
138:39 plan and hopefully see the agent in in action and then um call it good there.
138:44 So yeah, you're doing a good job. Learned a lot from me. I appreciate
138:47 that. Thank you, Stefan. That means a lot. Cool. Stephen, thank you. Favorite AI
138:58 coder to follow online, bar none. That means a lot. I appreciate that. Yeah.
139:02 Part part of the reason I I love these streams is because you guys are like
139:05 really engaged and like I mean you guys are encouraging a lot too. Like it's
139:09 awesome. Like thank you so much for the kind words. Like I don't take that
139:13 lightly. So yeah. And and I I'll keep it up. like everything I'm doing, I just
139:17 want to keep improving myself for for you guys like more and more every single
139:19 day in the way that I'm presenting and the systems that I'm building and the
139:30 Yes, you're very welcome, TS. Appreciate that. Uh, could super data pull menu items
139:36 from a restaurant website or would you use another web scraper? That's a good
139:41 question JP. So, Super Data is more specifically for social content.
139:47 If I go to their main website here. Um, so yeah, like YouTube, Tik Tok,
139:51 Instagram X. If you want something more for web scraping in general, I would
139:58 highly recommend crawl for AI. Also, crawl for AAI is free unlike Superbase.
140:02 So, it's not like as robust for something like YouTube transcripts, but
140:07 for craw crawling general websites, u crawl for AI is fantastic. And there's
140:11 also a couple of videos that I've done on crawl for AI. If I go to uh my like
140:17 literally my second most popular video ever. I'll post this in the chat quick
140:22 for you. Um here I'll title this a crawl for AI intro. It's a little bit of an
140:25 old video. So they've changed some things with their library, but you'll
140:28 still get the general gist of uh how to use crawl for AI. So you run it on your
140:32 own computer. It uses a headless browser to crawl any site that you give it the
140:35 URL for. You can pull the content in markdown which markdown is like the
140:39 perfect format to feed into LLMs or to ingest in a knowledge base for rag like
140:42 what we're doing in this video or in Cool. All right. Uh let's see where we at with
141:03 the plan here. Okay. So plan is done and uh I'm going to go really fast through
141:09 this here and then implement it. So okay are you guys seeing this?
141:24 It just generated a 2500 line um That is terrible. That is insane. I have never I have
141:42 never had this happen before. Even if I didn't specify a limit, a line limit.
141:48 Like I don't know what is going on here. This needs to be Wow. That
141:59 Why in the world is this 2500 lines long? That is actually insane.
142:05 I've never had this happen before. There there's um I think it's Well, it's like
142:09 there's something else that's kind of like Murphy's law, but like for live
142:14 streams like uh if something can go wrong, it's going to go wrong in a live
142:19 demo. and Claude Sonnet 4.5 deciding to completely ignore my instruction
142:26 is uh yeah that is crazy. I have used this exact prompt before. I know this is
142:30 like a more simple version of what I used for my planning um command. I've
142:35 used this exact thing before a dozen times even though it's the simple
142:40 version and I haven't specified how long to make it before like I am now. Like I
142:43 did it in the command and then I also did it explicitly in the prompt because
142:47 I didn't restart cloud code. I've never had to do this before and the agent that
142:54 it's building here is basic. There's two tools that I'm giving it.
143:03 That is so crazy. Repeat at top bottom in all caps. Yeah, I might have to Well, I'm just going to
143:10 ask it here. So, we should be good. That is I mean I'm just going to laugh at it
143:13 like that. That's actually pretty frustrating, but uh shouldn't take long
143:17 to correct it and I'll just have it make it concise here. It's just it's funny.
143:21 Like whatever. All right, Claude, you do what you want to do. Um does Archon use
143:26 crawl for AI natively? Yes, it does. Yeah, thanks for for asking that because
143:29 I should have called that out right away when I was was recommending crawl for
143:34 AI. Archon does actually use it. So when you guys are running Archon locally, you
143:38 are running crawl for AAI locally and it's it's leveraging crawl for AAI to
143:42 crawl the documentation and other knowledge that you're bringing into your
143:50 Uh maybe Claude just had a great need to communicate today apparently. Yeah. I
143:55 mean for some reason Claude like well really this Oh wow. Wow. I made it really short. That's kind
144:01 of uncomfortable. Um, see, look at what the heck. I said 500 to a,000 and it doesn't live because
144:08 now it's too short. That is insane. That is so weird. Um, yeah. Okay. Sorry to your your comment,
144:19 Toby. Um, I was going to say like these models like they act differently each
144:22 day. Sometimes it literally feels like there's something different and and it
144:26 could actually be like there was an update to the system prompt for cloud
144:29 code between times that I've used this command. like it honestly might be. And
144:34 so it is kind of something where like you do have to constantly evolve your
144:38 system to adapt to the tools that you're using. And that is one of the things
144:40 that I'll be covering in the Agenta coding course as well is how you can
144:45 adapt your systems as um better LLMs come out or LLM change like system
144:49 prompts change in your tools, right? Or like you have a just like a new tool
144:55 that you're using. So yeah. All right. Um so we're excluding all
145:04 yeah, cuz I don't know if it's referencing the examples as much as I
145:12 you know what? I'm going to I'm going to just go for it. I'm going to just
145:15 implement this because I I'm running out of time here. So, slashexecute plan PRPS
145:22 requests YouTube re or AI coach agent. This is actually kind of a
145:26 nice test. It It's not ideal, but it's kind of cool for this live stream to be
145:29 able to show like here's what happens when your plan is way too long and then
145:32 here's what happens when it's a little bit short. I'm curious. I I think it
145:36 might actually be able to knock things out even with a simpler plan of attack
145:41 here because in the end, simple is always better than overengineered.
145:47 This is too short to my liking probably because I think that if I had a bit more
145:50 text, I would be able to reduce some more of the assumptions the coding
145:53 assistant has to make during implementation. But I still think this
145:57 is going to be way better than having a 2000 line uh liner because then it's not
146:01 going to miss stuff, right? Like it's going to knock out these tasks one by
146:04 one in Archon and it's it's going to do a good job. So yeah, let's uh watch
146:10 these populate in real time here. Go back over to Archon YouTube AI coach agent.
146:18 All right. So, it's going to create all these in parallel here. So, we'll see
146:26 Try it with GBT codeex again. Frankie says, "Yeah, you know what? This live
146:31 stream is making me want to try codeex again. I'm not going to lie." Uh, yeah,
146:37 cuz this has been frustrating. In the end, it's fine. like we got the
146:41 rag pipeline to work. It's not ideal yet. And I'll probably have to do some
146:45 work to incorporate dockling correctly and all of that, but I mean it's it's
146:50 working well, but yeah, I mean like what the heck. I don't know why it's it's so
146:54 stupid in the planning phase. Usually the planning phase is the simple phase.
146:59 So I don't know. I I do want to try codeex again, but okay, here we go. We
147:03 got all our tasks in Archon now. Well, it's creating some of them. Yeah, there
147:06 we go. Okay, so they're coming in one by really Belgium on your short. Well,
147:13 yeah. Yeah, there's a lot of people in Dynamus in Belgium specifically.
147:18 Yeah, I don't know if it's like I I think it is top 10 for countries. Pretty
147:27 Claude is quite chatty. Yep, it is. I know, but I tried to be explicit like
147:31 you have to be short and it still wasn't. Yeah, usually that that is all
147:42 does anyone else notice a slightly different persona with each chat? It
147:47 does kind of mimic your own prompting. Um, like this is really funny, but if
147:51 you like include swear words in your prompts, you know, like you drop an
147:55 Fbomb a couple times, it'll actually start to do it itself. Uh, so if you're curious, you could try
148:00 it. I mean, I don't really want to try it right now, but yeah. So that might be
148:04 why if you're just talking to it in a different way, it kind of emulates that
148:08 even if it's like really subtle things. Otherwise though, it does depend like if
148:12 Claude code actually updates the system prompt or like Codeex updates the system
148:15 prompt under the hood, that'll change its persona as well. Yeah. Using Sonnet 4.5, it's generating
148:22 mistake after mistake. It's like overloaded. Yeah. Um that is kind of
148:27 like a really real problem that people and I honestly myself right now is
148:32 facing is is um these models like they seem to vary in performance and like
148:35 these providers like Enthropic and OpenAI they promise it's not happening
148:39 like we're not quantizing our models. We're not doing extra rate limiting for
148:43 different users and things like that but it really feels like it's happening.
148:46 Like I I don't know. I I don't trust it entirely. Uh so I think you're on to
148:52 something. Yeah, Richard gets it to swear a lot. It's funny. Yep. I know some people that make
149:00 it as well. It's funny when it does, but yeah. Okay, so we're ripping through the
149:04 implementation now. So, the plan is um I'm going to let the implementation rip
149:09 and then we'll test it and then we will uh call the stream there because I am
149:13 losing my voice a little bit. Yeah. But for those of you guys who've like stuck
149:16 around for this entire stream and then we still got more to come. Like I I want
149:19 to see this in action. I think you guys should hold out for the end. We'll see
149:22 it in action. But I was just going to say that like this is um this is like
149:27 been a long stream and and uh I appreciate all you guys for for sticking
149:31 around and hanging out with me here watching me build this agent. I I want
149:34 to keep building this out for the channel, not just as this use case for
149:39 the live stream, but really to make this something incredible for all of my
149:43 YouTube content. you can go here and like ask questions like how do I
149:46 incorporate this into my codebase or like how do I build an agent that is
149:50 similar to how Cole built XYZ right like being able to ask those questions and
149:55 have it guide you I think will be like a really phenomenal resource and then
149:58 maybe like even in the Dynamis community making this agent like be able to
150:02 reference my course material as well like for the people that are in Dynamus
150:06 u that would be super cool so yet another reason to come join the
150:09 community because I want to build this agent into the community as well for the
150:13 things that we got there. Um, so that like someone in the Dynamus community
150:17 could like come into this agent and ask questions and it would reference both my
150:21 YouTube transcripts and the agent to coding course transcripts and the AI
150:25 agent mastery course transcripts and the workshop transcripts. Like imagine that
150:29 like that would be such a powerful agent. And it's a little bit sloppy
150:33 building the foundation right now, but that is what I'm doing. Um, yeah. So
150:39 it'd be pretty cool. I'll probably have to go back and improve some things with
150:43 better planning um off stream, but I just wanted to show you guys what it
150:46 looked like as much as I could Yeah, you're welcome, Toby. I appreciate
150:54 it. Uh yeah, you're welcome as well, Tony. Is this going to be a weekly live
150:57 stream? I'm not planning on doing live streams on YouTube weekly. I want to do
151:01 them monthly, though. Um because what I do in the Dynamis community is weekly
151:04 live streams essentially when I do my the Friday workshops like every single
151:08 Friday. I've been I have done a live event every Friday uh since April and uh
151:14 that is kind of like the weekly stream basically. So yeah, but like still want to do a ton
151:23 Yeah, Frankie doesn't trust him either. The difference in the results are
151:26 sometimes too great for the same question. Yeah, 100%. Uh Raymond said,
151:31 "I've noticed that or here, let me uh highlight these again. I've noticed that
151:36 um based on the time of day, effort that the AI puts into the work varies. Yep.
151:40 It feels like when the batch processing begins to get congested, it's token
151:45 effort is adjusted. Yes, it does. And they say they don't do it, but like I
151:50 really think you're right. I I do. It's unfortunate. Like I get why they have to because a
151:57 lot of these companies are spending more money they're making because they're
152:00 just trying to like, you know, get the next round of funding. um you know get
152:06 as much of the market as they can and almost you know like creating a monopoly
152:09 and then jacking up prices right like kind of like what Nvidia has done in the
152:14 past. Um so it's like losing money now in the hopes of like getting the market
152:16 and then just like forcing prices on them later. I don't want to be all doom
152:20 and glue and I'm not saying that like is exactly what's happening for sure, but I
152:24 think that's like a viable theory. I I'll just leave it at that. Like I'm not
152:28 trying to to get super opinionated or political or anything, but I think it is
152:32 an interesting conversation. Uh definitely worth considering. Uh like
152:36 just watching out for that as you're coding. Like it is kind of crazy that my
152:42 my uh create plan command that I have here that's worked so well for me before
152:47 has just led to these like kind of poor planning documents today. And the
152:53 documents were good. Like it it followed the steps here. It was just specifically
152:57 the length which it kind of feels like it's just like a big problem with claw
153:00 that would just happen to be amplified today. honestly is what it feels like.
153:08 Yeah. Lost leader. Yep. Yep. Thank you so much. Initial plan has
153:13 created a very nice UV run rag pipeline with Dockling. Feel quite powered up
153:17 right now with your instructions and guidance. That's cool. I appreciate you
153:20 following along and and building like that. That's that's really cool. So, I'm
153:23 glad that you have created something awesome. Sweet. Yeah. And I never expect
153:28 people to follow along cuz I know that it's a lot just to like digest what we
153:31 got going on here. But that's super cool that you did that you uh built something
153:35 fantastic. So congrats. Seems like uh your coding assistant didn't ignore
153:39 Dockling like it did for me. So kudos to you. I wish mine used Dockling, but I'll
153:47 I'll make it later. Don't worry. cool. I appreciate that. It's been an
153:54 incredible stream. Seeing this build bit by bit makes us feel a part of the
153:56 process. can't ask for anything more. That means a lot. I'm glad that that you
154:00 like that you guys are liking it, even considering um how it's uh yeah, just I
154:04 mean it takes time to go through this process and that there's a reason why my
154:07 YouTube videos don't cover certain things just to make it more concise. So,
154:12 I'm glad you guys are liking that. Um yeah, why you stopped using BMAD? I
154:15 haven't actually stopped using BMAD. just um I don't use it a lot I would say
154:19 but I am going to cover it in the agent coding course as one of those uh primary
154:23 options for like a you know an out-of-the-box solution for an AI coding
154:27 system. The one thing I will say on the BMAD method and honestly a lot of different
154:33 strategies for context engineering is I think that they are very prone to
154:36 overengineering because when you have so many steps in a process like here are all the different
154:42 sub aents and like each one of these sub aents is contributing some work or like
154:47 some planning to the project it just leads a lot of times to more getting
154:50 done than actually needs to get done. I think that would be my main critique for
154:53 something like BMAD. I'm not saying I'm I still do think it's a great framework
154:58 and the guy who created it is like seriously a genius. The ideas are really
155:03 solid and profound in some ways. Uh but you just have to be careful. Like some
155:07 of the issues of Claude overengineering are just amplified if you're not careful
155:13 how you use them, I would say. So yeah, like I I had a 2000 line plan created
155:16 the first time I used BMAD, right? Like just like what happened here. Uh but
155:20 like here I'm using a process that I know like generally doesn't do that,
155:23 right? And like be ad might not either, but yeah, just it happens. So you have
155:26 to be careful. So yeah. All right, let's um Oops. Didn't mean to highlight that one again.
155:34 Let me go back over to Archon and see where we're at. Just do a refresh here. All right, cool.
155:42 So we got 10 in review and we are basically to the testing phase now.
155:47 Okay, so here's what I'm going to do. Be very fast with the testing. I just I
155:54 want to get to the point where I can validate things with you guys here. So,
155:57 I know it's a little silly, but I'm just going to ask it to blitz through the
156:02 testing here. So, let me take a look at the agent. So, we got agent.py.
156:07 We got our system prompt, which this is probably one of those things I would
156:11 iterate on a lot um before I just, you know, deploy this to production, but it
156:17 looks pretty good overall. Oops. Like, I like how it even has like an example
156:20 interaction in the system prompt. That's pretty neat. So, we define our agent.
156:25 That looks good. Uh, we have our rag tool and then we have our get full transcript
156:30 tool and it's just calling a single function that we're importing from our
156:35 uh tools folder here. So, yep, looks really good. We got our
156:41 service where we Yeah. Okay. So, again, really fast, really high level, but this
156:45 is looking great. So, looking good. Uh, the other thing I want to look at was
156:48 the API endpoint. So, we have our main.py up high where we are yeah
156:54 building our API very similarly to how we did in the example. So it it really
156:58 does seem like the more concise plan did a much better job. It's not skipping
157:01 through anything that I can tell here. And that dockling thing I picked out
157:05 like almost immediately. So I I've worked that muscle over time
157:08 or how to validate things really quickly with AI coding assistance. So I found
157:12 that it was missing dockling right away looking at this initially at least and
157:16 in a realistic setting I would want to dive a lot more deep into this but like
157:19 right now it doesn't seem like it's missing stuff. This API endpoint looks
157:24 really solid and so it's running uh pantic agent endpoint. Um yeah this looks really good.
157:36 This is kind of weird. Okay. Well, I'll Yeah, I having imports at the bottom is
157:41 stupid. Maybe I'll move this quick. So, that seems like a bit of a iffy thing,
157:49 but I'll ignore that for now. Those little smells like that that you like
157:51 code smells that you want to look out for. U but let me see how I start the
158:03 I guess I have to do setup guide. configure environment variables,
158:12 verify installation, run the pipeline. Okay, so it hasn't done documentation
158:15 for the agent yet. So this is just everything for the rag pipeline right
158:22 now. So that's fine. Um, what I want to do Okay, good. It created the tables for
158:29 nice. Okay. Yep. Okay. So I'll run the table creation here. So, I'm just trying
158:34 to like do as much in parallel as I can to get this to be quick. Um, create a
158:38 new snippet. So, I'll create all the tables it told me to create for the rag
158:47 Okay, there we go. Channels, conversations, messages, requests. Looks
158:54 good. Nice. All right. And then go back to the coding assistant here.
159:02 How far is it? Okay, so it's running the validation, which I really hope this
159:05 doesn't take that long. Uh, make sure you update the readme after with instructions on how to run
159:15 the agent and how to specify the port the agent runs on. All right, just want
159:19 to make sure it does that as well. It I think it already has it as a task to
159:23 update the documentation, but well, yeah. All right, that's fine. It already
159:27 has it, but I'm just making sure it's explicit in doing that here. So okay.
159:31 Yeah, things are looking a lot better. This this shorter plan like definitely
159:41 All right. So yeah, as validator is going here. I think uh this is the the
159:46 last time that I want to uh talk about the uh agentic coding course really
159:50 quickly because this is like very we're almost done with the stream here and this is
159:57 the the last call for the discount for the Dynamist community. So, I'm going to
160:00 pop it up one more time, one last time. Got to do it because the course first
160:06 batch just came out on Friday. And we're going to be doing this live like
160:09 workshops at the end of every release batch kind of coming in as like the
160:13 first cohort for all all of us together to master building reliable and
160:17 repeatable systems for working with AI coding assistants. And um yeah, let me
160:23 pop up the page one more time here. So in joining the Dynamus community like
160:27 you get like this course is a part of the community included directly and uh
160:32 it is all about building reliable and repeatable systems to get the most out
160:36 of AI coding assistance. This you're not going to learn this anywhere else es
160:39 especially not like with a community of support coming alongside you as well.
160:43 Like I'm here every single day to not just like help you go through the course
160:47 but also to answer any questions as you build your own system that's optimized
160:51 for your codebase. Like that that's the system guy. That's what I help you
160:56 through in the course. And um the primary goal like this is one other
161:00 thing that I haven't talked about in this live stream yet that I think that
161:02 you guys will find like really interesting. The main mental model that
161:08 I have in the course is that every mistake is an opportunity to evolve your
161:14 system. It's kind of like what I did when the plan that was created here was
161:18 way too long. Like I went in and this is like a really basic example obviously
161:23 but like I made an addition to my my prompt. This is like you know kind of
161:26 part of my system. It's a very simple system but it's a part of my system. I I
161:31 saw a mistake and I corrected it. And now every time going forward I have this
161:36 you know very slightly evolved version of my system. And that's like the main
161:40 mental model that I have for you. So when we when we see a mistake with the
161:45 AI coding assistant instead of doing a one-off fix, we fix the system and then
161:51 we retry, right? Like that's the high lever skill that I show you how to
161:55 develop in the course, making every mistake and opportunity and really like
162:00 that is your competitive advantage, right? And so like that's that's the
162:03 main thing that I'm covering here. Um, and along with everything else you get
162:06 in the Dynamus community, like even the Agenta coding course itself, I've
162:10 designed it to be well worth your investment in the community every single
162:13 month or, you know, every year if you go with that route. So, yeah, 20% off with
162:18 the code that's going to expire after the live stream for the monthly
162:22 membership and then 40% off the annual. So, jump on it now. Like, I'm I'm taking
162:26 it down like right after a live stream. So, this is your chance here. And if
162:29 you're already in the Dynamus community, then like go at it right now. Like, this
162:32 is available for you. It's super exciting. I'm I'm here to answer any
162:36 questions as you go through it as well. let's call it at closing time. Yep,
162:47 man. Your uh live streams are always instructive, entertaining. Love to be a
162:49 part of the Dynamus community. Highly recommend anyone interested in
162:52 developing AI to join. Keep the great job. I appreciate a lot, Mr. Nato. Thank
162:57 you very much. Um, does it make sense to sub to Perplexity Pro? Does it have any API
163:07 with utility like other models? I do use Perplexity myself. I use it plus OpenAI
163:11 deep research typically like at the same time when I want to do a lot of
163:15 extensive research on something. So, I like Perplexity. I think that it's
163:18 actually pretty impressive that they've stood the test of time just as other
163:22 tools have come out like uh you know anthropics deep research and opening
163:26 eyes deep research and lang chains deep research and like there there's still
163:28 like I feel like it's very really valuable especially the speed at which
163:32 it is able to analyze sources and things I'm in says Mark looking forward to
163:42 learning awesome Mark welcome to Dynamus I appreciate it a lot unless you're
163:45 saying you're already in but I'm I assume you mean that you're in now. So,
163:49 welcome. Welcome. Super excited to have you in. Cool. Uh I don't know what Claude is
163:56 doing here. Wait, this is actually bonkers. It Okay. Oh, okay. That was just a UI
164:04 glitch. I all of a sudden it's just I I don't know. Do you guys see that? It was
164:07 just like validator. It was like repeating like over and over and over
164:11 again. That was like really weird. Um Okay. But it it like corrected itself.
164:15 It was just a UI thing. It wasn't actually calling my validator a million
164:18 times, so that's good. Okay, cool. Stephen just joined. Welcome, Stephen. I
164:28 appreciate it a lot. Cool. Glad to have you in. Appreciate it, Chanton. Thank
164:32 you. Gemini is amazing at deep research. Yeah, 100%. Mr. Nato and Mark, looking forward to
164:40 learning with all of you. Cool. Cool. Welcome as well. Um, just use Reflexi MCP tool and a
164:47 workspace directive to consult Replexi on all problems that require more than
164:52 one attempt. Nice. That's That's smart. Okay, let's see here. Karine said, uh,
165:08 it is the 80% 20% rule. BMAD shows you the rest of the 80% up front when you
165:12 have an MCP with 20% of the effort. just the happy path. Maybe you could elaborate a bit more on
165:17 what you're getting at. I think what you're saying is that BMAD is it gets
165:21 you 80% of the way there, but then it's up to you with your process and your
165:25 tooling to finish, which I would agree with that. Maybe that's not exactly what
165:28 you're saying, but yeah, a lot of these frameworks like PRP and BMAD, like
165:32 they're not meant to like fully replace whatever kind of process you have for
165:36 using AI coding assistance, but they do augment, right? It's just another thing
165:40 in your tool belt to augment your Okay. So we're updating the readme here.
165:52 I think I will have to uh set my Okay. I'll have to set my environment
165:56 variables. So I will do that off camera Um because I just have to Yeah. So I
166:04 have to configure things for the AI agent. Let me do that really fast. I'll I'll
166:11 just use GPT4.1 Mini for my model right now. We'll just keep it simple.
166:19 All right. API port 8040, right? That's what I have running in
166:22 Docker. I'm going to try I'm going to try my best to get a demo for you guys
166:27 working right before I close the stream here. And then I'll upload everything to
166:33 GitHub later as well. Um API is running on 801. Okay, let me correct that. So, I'm going
166:41 to close the API that I have running in my Dynamus agent already because I'm
166:45 going to spin it up with what the code we just created. And fingers crossed it
166:53 nails it. We'll see. Okay. And if if it fails right now, I probably
166:57 won't iterate that much because I'm about to lose my voice. But this has
167:01 been a blast. Uh oh, you reached a super data rate limit. Yeah, that'll happen. Yeah, you
167:06 probably don't want to pull like a million videos in within a minute or an
167:11 hour or whatever it is. So, okay, let me go. Okay, so we we updated
167:23 and then uh run the AI coach. Okay. Okay. The instructions are a little
167:29 jank, but this is going to work, so we're good. All right. Run this. And I
167:39 And I don't know why. Why is the host Well, I'll try this. I'll just try this.
167:47 Okay. Uh we got an error right away. Okay. So, I'm going to I'm going to try to iterate
167:53 a couple of times. So, uh why is the host 0.0.0.0? Shouldn't it be local host? Unless I'm
167:58 running in a container, because I'm not running in a container right now. And
168:02 then also I got this attribute app not found in the module error. So address
168:05 this as well. Okay, cool. Because it doesn't seem like a drastic issue, right? So like I
168:11 shouldn't have to go through the piv loop again. So send in the error, ask it
168:15 to think through and find the implementation or find the fix and then
168:20 just let me know what went wrong. Like that's my standard process for working
168:23 with these little problems. Hopefully it's a little problem. I don't know
168:26 exactly. I'm not taking the time to dive into it right now, but hopefully it is a
168:31 a smaller problem. So, let me restart this. Um, oh, it's being stupid. It won't let
168:36 me close out of it. There we go. All right. All right. I'll be ready to run
168:40 it again once it uh corrects itself here. Yeah, Raymond, that is true. For some
168:56 reason though, whenever I when I have it run on um o some and if it's not in a container like
169:02 sometimes it like a local host connection won't work. May maybe I was
169:06 doing something wrong and I just like messed up my own knowledge previously.
169:13 I don't I think this will be fine. So I'll try this. You're probably right,
169:18 though. Um I'll just run this. I think I think this will be good.
169:27 Okay. Come here. There we go. I'll just delete this and uh re-enter
169:39 Oh, thank you for the $10 donation. I appreciate it a lot. Thanks for all your
169:42 videos. They are very helpful. You are quite welcome. And also, I think I
169:47 missed a donation earlier um from uh Dax. I think like at the start of the
169:51 stream. I just like that triggered my memory. I'm really sorry, Dax. I missed
169:54 yours. So, thank you as well if you're still here for your $10 donation and
169:58 yeah, for the one right now. I appreciate it a lot. That means a lot.
170:01 Thank you for your support. And yeah, let's uh I'll do it for you. Let's see
170:06 if the API endpoint works for you. Uh okay, looks good. Cool.
170:15 So, moment of truth. Um well, hold on. I Or maybe I'll just start the front end.
170:30 Yeah, hold on. I got to do this off camera, but I'm going to set it up so
170:35 that I can run my agent. so I have to set up environment variable. Well, actually, I don't have
170:49 to do this off camera because it's a front end. So, these aren't these aren't
170:53 actually secret. So, I'm going to go back to my superb basease here. And I'm
171:00 going to grab my anonymous key. And I'm going to paste this in here. I
171:04 mean, this looks kind of messy, but I don't care. All right. Anonymous key and
171:16 Nice. And then I'll open up a new terminal So, I'm starting my front end, which I
171:25 my front end is something that I've already built. So, I'm not like showing
171:29 how I built my front end from scratch here, but it's running on port 8080. And
171:33 I have it uh connecting to the agent on localhost 80001 API/patantic agent,
171:38 which aligns with the port that I have it running with the code that we just
171:43 built together right now, port 801. I I have absolutely zero guarantee that
171:47 this is going to work right now, but that would actually be insane if it
171:53 worked. Let's try this. So, I have to remember I think I I use
171:59 like a throwaway email for this. Um Oh, man. I don't remember the login
172:09 Oh, it worked. Okay. Okay. Here we go. Moment of truth. I'm in my
172:13 chat interface. So, I'm signed in. I'm going to send a message. I'm going to
172:16 see what happens. I'm going to just say Okay. Um, oh, it's just the user is not present in
172:28 Okay. Okay. So, that's fine. I forgot to create my user in the user table. So,
172:32 when the when a user first signs up for the platform, it automatically creates
172:35 this record, but I made this table after. So, that's why I'm getting this.
172:41 Um, let's see. There we go. Boom. Boom. And I am an admin. Yes. Okay. Save.
172:48 Good. Reload. Look at that. Okay, it actually worked. Now, now the real
173:01 test is the rag tools here. But okay, let's just take a second and appreciate
173:06 that this actually worked. I I told it to reference this API file like copy the
173:11 exact API endpoint to work with this front end and it actually did like this
173:15 agent is working in the front end. Like we can see that the these are the logs
173:19 that we got from the agent that we just built right now. That is super cool. Um
173:25 so now what I can say is like oh I don't even know like uh search the
173:33 knowledge base for um what's someone give me a good question.
173:38 um for like what what I could ask it based on one of my recent YouTube videos
173:45 like just someone throw out something search the knowledge base for whatever
173:48 like I'm going to ask it some question here I just want some example like from
173:51 like my dockling video or like an archon video trying to think
174:00 search knowledge base for um the the goals of archon I don't know
174:06 something like this. What is Yeah, what is Archon? Thanks, Richard. That's good.
174:10 Search for what is archon. Okay, this might fail. There might be an
174:17 error. We'll watch for an error here. Okay, that actually worked it. So, it
174:24 used the rag rag search and it embedded No way. Archon appears to be a platform
174:33 or tool related to AI or database management. Okay, that's kind of a dumb
174:39 answer. Um, but I think it actually Yeah, so it's currently in beta with
174:42 experimental features ongoing development. Um, run database migration.
174:47 Okay, so the answer is not ideal, but that's probably just in the system
174:49 prompt that I'd have to like tweak on like what it says. Um, it uses Superbase
174:54 for local data storage, which that's that's hilarious that it says Superbase
174:59 instead of Superbase. Um, and then it links to a video. Okay, here's the real
175:02 moment of truth. If I open this video, is this actually going to be Archon? Add
175:06 a timestamp. That is Hey, look at that. It linked me to the beta launch live stream
175:14 >> specifically to the quick start. >> Take a look at that. That is so cool.
175:20 So, okay, man. In the end, this worked beautifully. It took a little bit to get
175:23 there, >> but wow, what a big payoff. Like, it actually worked.
175:29 >> Definitely problems that we have to address here. This is a really awkward answer, right?
175:35 Like I'm not going to try to say that like this is perfect, but it it cited
175:39 its source with a link that I can click on that brought me right to the video.
175:43 It clearly understands everything from the transcript. There's some
175:46 misspellings here. Maybe we want to like integrate some strategy in the rag
175:50 pipeline where it'll like analyze the chunks to fix misspellings. We'd
175:53 probably want to adjust the system prompt so the answers are less awkward.
175:57 like quite a few things that we can correct here, but like this is a really
176:01 solid proof of concept and we actually did it. We got the big payoff and we did
176:05 it live. I didn't know if I was going to be able to do this. 3 hours is a lot.
176:10 Like my voice is horse now, but like this was so worth it. So cool. Let's ask
176:16 another question. Let's say um uh search for dockling hybrid chunking
176:24 and uh link me to the part of the dockling video where I talk about that.
176:28 Let's try it. Let's try another one All right. And this is like blazing
176:35 fast. Like holy cow. Um okay. At this point discussion rag strategies. Um
176:42 okay. All right. Is this dockling? like you're absolutely right. I
176:46 apologize for so this this answer is not ideal. >> I mean it's still working. It pulled
176:50 something. It just didn't pull it correctly here. Um if I wanted to like
176:53 really dig into this, that's when I had set up like agent obserability and like
176:56 figure out the chunks that it pulled and everything and it uh didn't find my
177:02 dockling video. Uh how about my video where I talk about dockling? The main
177:09 topic of the video is dockling. Another thing that I I I should probably do is
177:13 give the agent a tool to list out the videos that I have here. Um, like one
177:17 thing that Super Data apparently didn't do correctly is it didn't pull the
177:20 titles. Or maybe that's something I have to fix in my rag pipeline. So again,
177:24 things are really not perfect here. Like it probably can't actually find the
177:30 exact video because of this. Um, okay. Well, oh, it worked. Okay, cool.
177:36 It found the video. Nice. Okay, so that's really good. So now can I say
177:41 okay uh search this video specifically for when I cover hybrid chunking. Now
177:46 another thing that I haven't given this agent is the ability to limit its search
177:50 to a specific video. I don't think it actually is going to be able to do that.
177:55 Um yeah. So like >> all okay it actually did find it. Okay. So
178:01 we're good. So it's it's like not perfect but it's still kind of
178:04 impressing me here even given like how much I know I have to iterate on it. So
178:08 yeah, this is uh the part of the video where I'm covering hybrid chunking. Like
178:11 this is my example where I'm showing uh the output of the hybrid chunking script
178:16 that I actually use as the example in our codebase for today. So that's pretty
178:20 cool. >> Awesome. Okay, that that's enough of our testing here. And and by the way, what
178:27 I'm doing here, going back to our diagram, um that is the manual testing part. Uh
178:34 hold on, let me go back. Okay. So, what what we're doing right here is the the
178:38 manual testing, right? So, like the code review I did very quickly. The AI coding
178:42 assistant took care of unit testing, integration testing. We're just doing
178:45 the manual testing, making sure things work. And yeah, we found quite a few
178:48 things to improve, which that's fantastic. We'll do that. I'll do that
178:52 as well. But like, man, we got a really good starting point. So, all right,
178:56 let's go. Skipper Chunk rocked it today. Cole, just joined the diamond community.
179:01 Looking forward to all your excellent content. I appreciate a lot. Thank you
179:06 very very much and uh welcome to Dynamus. Really excited to have you in.
179:11 Cool. And and Craig, it is it is handling citations. If I go back um
179:15 yeah, like it's it's got the links right here with the time stamps and
179:18 everything. So yeah, we got we got our uh citations. Pretty sweet.
179:25 Nice one. Yeah, appreciate it, Chad. Cool. All right. Uh search for the most viewed
179:34 video. I'll try uh search for my most viewed video. I don't think that I have
179:38 that metadata though, so I don't think it's going to be able to pull it.
179:42 Um, yeah, I don't think it has access to that, but we'll see. Yeah, well, it's
179:48 Yeah, because it it doesn't even understand that it's like a separate
179:53 thing than a rag search. So, it just tried to like search for that text in my
179:56 transcripts. But that would be another cool thing to add 100%. uh Matias would be to add in
180:03 the ability to like search metadata on videos as well like what's my most
180:06 recent video, what's my most viewed video, what's my video that has the most
180:10 likes or you know whatever it is. So yeah, there's so many more ways that I
180:13 can expand this agent and I want to keep doing it and it's actually freaking
180:16 awesome that we got this far in only a few hours and I didn't even do that much
180:19 prep before the stream because I really wanted to build it from scratch with you
180:23 guys and show what that looks like realistically. So this is super cool.
180:29 Um yeah, what are the three most co important coding principles for AI coding? That's
180:36 a good question. Maybe I'll actually end the stream with uh this question here.
180:39 Um but one more thing I want to say after. So hold on, but I'll I'll um
180:43 answer this here. So yeah, I would say the first one is avoid overengineering.
180:47 Kind of like what we covered like I'm actually glad that that came up so much
180:50 alive today. Avoid overengineering through the way that you're prompting
180:53 and asking for simplicity and and validating for simplicity. Um the the
180:58 second most important coding principle uh well actually this is probably the
181:00 most important is you want to have a system for working with your coding
181:04 assistants like you want to define a process for I mean really everything that I cover in
181:11 the piv loop like with uh you want your process for planning your process for
181:14 implementing and your process for validating and you want to create it in
181:17 a way that's reusable right like you have your rules your global rules you
181:21 have your commands you have your hooks like whatever you want as like the
181:25 package of what I like to call your AI layer, you want to define that in a way
181:29 where you can use it for any project and create that as a a reliable and
181:33 repeatable system. And so never just like prompt like any anytime you create
181:39 a prompt two times that's really similar that screams that you should turn it
181:44 into a command and reuse it later. And then the third principle that I
181:49 would uh say and like really these are all principles that I've covered in this
181:52 stream today is that um you should delegate as much code as possible to the
181:56 AI coding assistant, right? Like planning and validating is where you are
182:01 in the loop. That's where you want to inject yourself as much as possible so
182:04 you're not just vibe coding. But then assuming you have a really well
182:07 structured plan that gives you the freedom to delegate all the code to the
182:10 AI and then you're just validating at the end, right? So delegate code.
182:16 Really good question though. Reference good examples says Bo. Yeah, referencing
182:20 good examples like I did for our implementation today. Clearly because it
182:23 worked super well. The examples I had in here are coming directly from the
182:27 Dynamis AI agent mastery course and the resources I have for you guys there. Um
182:32 and so yeah that uh clearly worked well giving it the ability to build that API
182:36 endpoint perfectly first try. full conversation management, token
182:39 streaming, message handling, everything um in the database in the front end like
182:45 first try. So, super cool. But yeah, all right, I'm going to go
182:49 ahead and uh wrap up the live stream here. So, thank you again for everyone
182:54 who was here. Uh especially those of you guys who were here the whole time, like
182:59 got 170 people on the stream here, uh which is an incredible number given um
183:04 how long this stream was and how much I'm losing my voice right now. But, uh,
183:08 yeah, it's it's fantastic. So, yeah, last thing I'll say just as I wrap
183:11 things up is, um, yeah, I'm going to have this, uh, this discount for the
183:16 Dynamus membership up for just a tiny bit longer and then I'm tearing it down.
183:19 This, this really is something special for the live stream because I really am
183:22 doing this live stream, this whole thing, building this agent with you as a
183:26 celebration of the first release batch of the Agentic Coding course that's
183:30 available to everyone in Dynamis now. So, if this seems interesting to you,
183:35 like get in on this and uh 20% off monthly, 40% off yearly membership for
183:39 Dynamus right now um for this course that I'm just going to be pumping out
183:42 more and more for over the next month and then obviously more and more coming
183:46 to Dynamus well beyond this later as well. So, yeah, even the course itself
183:49 I'm designing to be worth your investment in the community and just
183:53 going so much deeper on all these things that I covered in the stream today. And
183:57 u yeah, always will be doing things like this. I always will be giving away uh so
184:02 much on YouTube, but um there there's there's more special things in in
184:07 Dynamis, and the the level that we go into there is um unmatched for sure. So,
184:14 all right, sweet. This was fantastic, guys. So, yeah, thank you everyone again
184:19 for for being a part of this and I'm super stoked that we have the agent
184:22 working at the end. Like, that was so cool. Like, big payoff. We had it. And
184:26 thank you everyone for your engagement, your questions as well. And uh with
184:32 that, I will hopefully see a lot of you guys in the community. And uh I'll keep
184:37 seeing you guys on YouTube. More content coming out next week around building AI
184:41 agents. I got a really interesting video on like how to just like keep things
184:44 simple when building agents for Wednesday. So you guys will see that on
184:47 Wednesday. Hopefully you're looking forward to that. So with that, hope you
184:52 guys have a fantastic rest of your