Transcription
It turns out I was using Claude code completely wrong. And I only found out because I actually went and built the thing.
A few weeks back, Tim laid out four steps to build your own Claude Trading Assistant. And at that point, I hadn't really touched it. So, I followed his steps, got my hands dirty, and went a lot deeper than I expected. Deep enough to hit four mistakes. Each one kind of disastrous. The biggest thinking cloud code was a tool for writing code at all. This is everything I wish I'd known going in because the thing that builds your app can actually live inside it and the four course corrections that got me there.
Welcome to the Trading Floor podcast episode 19. Tim, what is up my man? Long time no see.
>> I know. Good to be back.
>> So was in Ireland for a week.
>> That's right. How did Were you trading sheep futures over there or what?
>> There was a lot of sheep. I'd be long. [laughter]
>> All right. So, this is a followup, right? A few weeks ago, you walked everybody through getting started the four steps like soup to nuts, the whole like and it actually helped me because as you know, I've been kind of sitting out of the ball game watching what everyone else was building thinking, okay, like let me let me kind of pounce on this thing once I know exactly what I want to do.
Definitely. And it was just the first beginner steps, you know, talking about the plan mode, how I did my template, and really just kind of making it your own unique trading assistant. So, it's cool to say that you like did them and uh played around with those steps. That's awesome.
>> I mean, they were great because they were so simple that it I think the biggest thing for me is it just removed the intimidation factor. And like I mean for somebody who's done a lot of script building and and writing in Python, it's like I shouldn't be intimidated by something that's supposed to make it easier, but is a whole new world, right? There's a little bit of that feeling of like, wow, there's a lot you can do. Where do I even start? And I think that that's what our conversation a few weeks ago really helped me with. So I actually went ahead and did the thing. I built it. Um, it's still a work in progress, but it's built, right? And this is we're talking about a trading assistant here.
I was about to say what did we build?
>> Yeah. So, I mean, first of all, I want to point out that and Lance talks about this, right, which I think is super relevant like with something like this, you always have to ask yourself like what's the ROI because what I didn't want to go do was spend a bunch of time on cloud code for the sake of cloud code, like just because it exists, right? I wanted to make sure it actually moved the bottom line. So, I spent a lot of time thinking about what actually matters to me and like what would move the needle. And I came up with two things, right? So, I've always had this thing called a trade log. And I actually made a video about this like years ago. It's on YouTube somewhere. Um, and it's it's a review process, but it's all on Google Sheets. It's it's awesome. Like, it totally works. It surfaces process gaps like things like, you know, are you grading incorrectly or are you not sizing based on your grade? Or how about when your idea is working but you're losing money or how about when the idea doesn't work but you break even, right? That's like kind of execution gaps. So, it's designed to kind of log all those things and highlight those things. The problem with it, it was super cumbersome because it's all manual.
I remember that trade log.
>> Yeah. And and like I mean I I get a lot out of it. Like I love it, but it just takes me so long and it's just not it's not efficient. And so I thought, okay, number one, I want to use Claude to kind of recreate the trade log in a way that is interactive and in a way that really enhances my review process. like if I can not only make it more efficient but also get more out of my review process that I already have an idea of what works for me then that would be a win. Okay. The other thing I wanted to do was some sort of synthesis of information because we get hundreds of emails in our inbox, analyst reports, um day summaries, earnings report analysis, highlights from banks and you name it, right? just there's tons of stuff.
>> Oh yeah.
>> And there's tons of information out there. Like no matter who you are, we build a lot of market views and filters and things like that because we've, you know, really kind of done a good job of leveraging the tech side of things before AI came along, right? So we've also have all of these market filters. And so one of the challenges that I find is every morning coming in, it's like a huge priority gap where I'm like, okay, how many of these emails am I going to read? how how am I going to pinpoint which earnings names to really get into and like how in how much detail am I going to scan over all of my market filters, right? It's like it's it's a really it takes a lot to get that process down to kind of prioritize how you're absorbing your information flow in the morning. So I wanted something via cloud to synthesize that in a intelligent way understanding what's important to me and give me something that's concise and really hits things and breaks things down the way that I want.
This is awesome. I mean I love that you took the steps and ran with it and made it your own because that was the whole point. So, I would say also if anyone is still trying to get into Claude um and you don't know how to go back to that beginner four steps video, but this is music to my ears. So, I'm excited to dive into uh your journey through this.
>> Yeah, man. So, I mean, I would just say, too, like your your four steps is like the perfect four steps to just get your feet off the ground because that's what it did to me. And then it propelled me into this into the stratosphere into this totally other realm that I did not expect to be in.
>> I was doing
>> once it's built, you're Yeah. you need to
>> figure out on your own. Now then I was like off to the races and I and I kind of like hit this point where I was like man like like I didn't expect to be doing these things with my project at this point. And what ended up happening was of course you never nail it on the first try. You have to go through trial and error. And I thought about it and I'm like, "Yeah, man." Like, I made I made four mistakes when I first started doing this that I wish I knew when I started because then the whole process would have been much more clean from the get-go instead of me kind of figuring these things out like halfway through and then playing cleanup.
Definitely. And I mean, I I've ran into a few mistakes myself, so
>> I'm sure. Yeah.
>> Yeah. I'm interested to see which ones which ones you found.
>> Yeah. All right. So, let me just get right into it.
>> Would you like to gain the biggest edge a retail trader can get? All of our daily and weekly in-house trader meetings are now available to you. Just head over to smbtrainingfloor.com to find out more.
>> We'll talk about the first mistake. So, most people use Cloud Code to write code, right? And there's nothing wrong with that. That's amazing. We do that. If you're just doing that, great. I think it's great. But that's it, right? It's really a a fast engineer for a lot of people and a lot of traders I talked to. Like that's that's kind of the the ceiling. For months, I did the same thing. And you know, that was my first mistake because the thing building your app can also live inside of it. And so I I don't just have cloud code write the functions. I wire its reasoning into the product as a feature. So my trading operating system, we'll call it an operating system, but that's just kind of what they call it. It's the website that cloud builds for you when you're basically inputting this stuff and saying, "Hey, I want like I want my review process. I want my email synthesis." Like it all lives on this website. And we can call that the operating system.
>> And that's your assistant, just to be clear. Right.
>> Exactly. That's that's what we call the the trading assistant. it's assisting me with my reviews and it's assisting me with my information intake and generation. Um, so the the operating system reads the market. It breaks down the researcher views my own performance live. And that's what I mean by inference architecture. So that's something that I'll talk about throughout these four points. And that's basically like the idea is like stop treating cloud as a factory and start treating it as part of the machine. like let it do its thing. And so there's there really specific things I found to uh that allowed me to leverage this aspect of the project and that's when it when it really took off and started becoming much more powerful than I ever imagined.
>> All right. So, just to make sure I understand basically you're saying that like there's two distinct like use cases for claude and you're understating this inference as a big way of utilizing it.
>> Yeah. Because I think like cloud code it's in the title and it's like the first thing that blows your mind is you can say hey can can you just make X Y and Z happen for me? Like my dad for instance we talked about him on the last episode. um he he watched your your episode and he created his own Trading View, right? Like he was like, "Man, I don't want to pay for Trading View anymore and and I want to create my own thing that just gives me my like win rate and my stats and and everything
>> and my chart and he did it based on your video, but he's just using it to code and that's where like a lot of traders are." And so, yeah, there's a whole other side of this where and I was talking to my cousin who's a sellside analyst on a on a short side fund in Boston and he he he owns the he owns the uh the analyst company in the firm basically with another guy, you know, they're the partners and they've basically like commissioned a an intern to kind of attack the whole like AI project for their firm. And he was talking to me about how it's just not working for them. And he was kind of asking me questions. And what I realized was that like he's basically giving somebody who knows nothing about his actual process, his actual trading process, his analysis process. He's giving it to a computer science major and say like, "Hey, can you just make this work?" And what I've realized is that you don't it's not a computer scientist that you need to give this to. It's someone who a understands the the trading process, understands the the playbooks, the methodologies, the way that you want to break down catalysts, and it's somebody who can write really well and reason really well and think analytically. Like, I'd probably rather have like a philosophy major or an English major work on this project than a computer science major.
>> Yeah. I mean, it makes a ton of sense. Like I found too uh just with the build mode what we're saying is like just using clog code as a builder as a coder like that is a normal first step I think to take because it's sick you know you make an alert you're like all right now I'm going to make another alert
>> or like you're doing your trading assistant and you can just say all right like I want to see all of my trades by grade for May and you have a one-off plot and then you start saying like all right but I want to do this continually like ongoing.
>> Yeah.
>> And that's really I found myself where the magic happens. Like in the beginning though, it's cool because you're getting a ton of stuff done because it is a factory and it's producing a ton of output.
>> Yeah. Totally. And to and to bring it back home like to your to your question. So there are two sides. There's the there's the factory code make this thing. It's like one-dimensional. And then the the other side of it is is the inference of the language model itself. The re the ability for the language model to infer and reason and synthesize and surface things that are not necessarily directly in the data but that are patterns in the data that that mean something else. Like for instance, you might have 10 different emails of analysts breaking down different stocks. And in no single email does any one of those analysts say that today the robotics theme is really active because there are a bunch of analysts breaking this down. But of course, if you read all those emails, you can infer that and you can synthesize and you can start to draw patterns. and these are the things I'm asking Claude to do. So, it's not just regurgitate the email to me. It's like scan everything, digest it, and then surface these specific patterns that I'm looking for.
>> That's really interesting. Like, is that what led you to make the jump or like how did you find this to be a mistake?
Yeah, because when I started out, I was building it as if it was just a coder. And then what happened was I I didn't go into it thinking that this was a big layer of my project. And and like this is what I was telling my cousin. I'm like, you should be doing this because you know what you're looking for. You know what you're looking for in all these filings. like you're the one that probably needs to give Claude like your process and really break it down and make it understand like what you're looking for because it can't be a generic thing. Like one of the things I've learned is like the more specific you are and the more boundaries you get, the more it's able to leverage its inference and synthesize because if there are no boundaries, there's there's no, you know, every everything's a theme, right? But if you if you for instance uh give it a database and an entire ontology around like what are our themes that we track and what are the definitions of these themes it locks it right in to those definition definitions and then starts firing away. Um, so, what was my mistake? Right? So I started using, this is mistake number one. I started using cloud code only as a code factory. And then only once I started seeing the results did I start thinking like, oh man, like I could probably get it to give me some more information here if if Claude has read all of my emails,
>> right? If we're just talking about that that aspect,
>> um, you know, I'd kind of like to know I'd kind of like to have a theme section where it surfaces patterns that it sees throughout the entire data set in case, you know, something's popping off in the memory world today. That might be significant, right? It might not be like the single stock with the biggest gap and the number one catalyst of the day, but man, if if stuff overnight happens in Korea and there's kind of a story there and then like there's like an analyst talking about MU and then because you know how it goes like remember the um, the quantum breakout last year.
>> Yeah. where those things just ripped out of nowhere it seemed and they broke huge bases all at the same time and it wasn't like a major inflection catalyst that like popped them off but if you read if you read all the emails somewhere in the fine print like Iowa and Q had like a decent um I think they had like a presentation or something and there was like a read through there and then there was like something else small with like and so underneath that there it really was like a quantum day in that pre-market. There was there was some quantum stuff brewing and knowing about that could have even just made you pay a lot more attention to those names at the open considering they were ready to break out. But if you were only looking for what's the hot stock of the day or what's the biggest catalyst or or volume in the premarket, it was probably some some other name. and SNDK uh January breakout with the hidden Samsung news overnight that no one knew about until a day later.
>> Exactly. Exactly. So, these are the things that I'm trying to surface, right? Because we we care about these and and it's really hard to to go over every piece of information in detail. So, I started building it as just code. And then I started to kind of ask it to do these things. And that's when I realized, oh man, there's like a whole other layer here that I did not account for. And the reason why this is a mistake in my opinion is because this is not something to be taken lightly. This is not an afterthought where you're you're basically giving it, you know, what to build and vibe coding away and and then you say, "Oh, by the way, like I want you to like synthesize these things." Like you have to in order to do for it to do it right, you have to really tell it how you want it to think and you have to teach it your own process and methodology and and what is the definition of a theme to you and what are the types of things that you want it to surface. So I realize that's a whole process. So we'll get into that, but that's why it was a mistake because I would have started out the project a lot differently and not had this as an afterthought.
>> Makes a ton of sense. So is the solution just not underestimating Claude and knowing that these are very different these two use cases.
>> I'm glad you asked. So the the solution is to build an inference architecture. So what do I mean by that? An inference architecture is wiring cloud's reasoning into your app as a live feature, not just using it to write code, right? So you wirecloud API calls into the the operating system. So the app reasons over this data, right? So that's the goal. And so the next the next three mistakes that we're going to talk about were mistakes I made in trying to do that. So the the the clear directives in terms of how we're how we're going to achieve this come they surface in the next three mistakes, right? And so, and and by the way, mistake number four is the same engine that breaks down a research email also breaks down like my trades against my own methodology. And so, I'll I'll show you like how I do that um later.
>> I'm excited to get into that one because I haven't done a ton on that front either. So,
>> yeah,
>> I'll save my questions.
>> It's really fun. It surprised me like how good it is. So, I try to run my whole project. Do you want to sit in on all of our daily and weekly in-house trader meetings? There's no bigger edge retail traders can get. Visit smbtradingfloor.com to learn more
>> inside cloud code session. Right in that one session, you open up cloud code and you just start ripping. You just tell it what to do. That's I mean for simple UI stuff and like just you know stuff like what my dad was doing like I think that's totally fine. But if you're trying to add that inference architecture and really tell it how to think and add that layer, this is a really really bad idea, right? The longer a session runs, the more junk it's waiting through. The model gets dumber as it goes. Like I call that like context hygiene. So now I brainstorm in separate chat threads, one per inference layer. And I'll get to like how I think about these layers, but I'm I'm opening up a chat thread that's not cloud code at all. And I'm basically saying, "Hey, I'm building this thing. I'm over here. I've been talking to Cloud Code. We've got a thing going on, but I want your job to be design me the architecture for the inference layer for, for instance, the morning report that we're talking about, this email synthesis." Like I'll just pick one inference layer and say, "All right, we're just going to break down how we're going to do this." And if I need to brainstorm or get feedback or if we need to go back and forth and get really messy about it, I'd rather do it there than do it with cloud code because cloud code is maintaining a code base and you really don't want to contaminate the process there with tons of edits and confusing language.
Yeah, this kind of reminds me like almost in middle school where you had like all of your classes, all the papers, homework, and everything in one binder and then whenever you tried to like find something, it was just messy and like you couldn't find anything. Is that like a pretty good analogy when you're doing it all in like one session?
>> Well, yeah, 100%. Because you're telling Cloud Code the you're telling Cloud Code the structure of this trading assistant, right? the operating system, the website, you're giving it UI instructions and to then also tell it how to think and then think about inputting how you think and give it certain data like our theme database and all the definitions and then answer the question, how is this theme database, this theme tracker, how does that relate to how you break down the morning emails? How does this relate to how you think? Like that's that needs an organization. It can't just be like whatever you feel like telling it on Tuesday and then whatever you then you kind of close the session and then whatever you feel like maybe telling it on Wednesday. Like over time that's going to get lost and you're not going to remember what cloud has inputed into the code. Of course you can ask that back and that's like a good practice but it's it gets very messy like your binder. This is why this is a good followup too because like in that last video we were talking about using the chat claude chat as your plan mode and then you build it but then I think it is like a misconception that then you never go back in chat
>> because you already did your plan but is that yeah
>> that was that was my mistake right so I like ran the whole project inside one cloud code session and then after a while I was realizing that like and like this is the this is probably like a good conceptual best practice for this is like I started realizing that I was I was auditing this morning report and thinking about how it's doing and okay I wanted to like synthesize these things or think about the theme our theme tracker like differently and I was like patching cloud code's mistakes with these like oneoff like additions like don't do this are like, "Oh, don't do." And then you realize that maybe yesterday you told it to synthesize aggressively. And then maybe maybe you realize there's an issue and then you come back like Thursday and you tell it to like don't hallucinate and only surface things that are that are defined in my my theme ontology and you're real like you don't remember but you're telling it contradictions
>> and session in or in like two different sessions like you know different days right over time it's it's very easy to like
>> contradict yourself And the AI has to pick one, right? Or or it just doesn't do anything. And so you have to reconcile those contradictions actively and write exactly what it is you want cloud to do in the architecture. In that way, and I'll get to the solution here. That way, you are keeping everything contained and you have your own you you're you have your own intellectual property of what this thing is that you're building.
>> So, what does this like look like for you? I guess just to play a little devil's advocate,
>> like I think it's common like everyone knows that you want to try to debug and find errors and it will save you time putting in more work being like clean with your chats and your prompts.
>> Yeah.
>> So you don't have to do it down the road. Yeah. But at the same point, you know, this is sounding a little tedious, which I'm I'm interested to see if like you've seen [snorts] that it has saved you a ton of time and energy and just what it looks like.
>> Well, what it's saved me is like to I mean, put it bluntly like a project. Like that's what So yeah, maybe some of this does take more time upfront,
>> but I don't think there's any other way to do it. So like, yeah, I mean it it's a big job to describe how you want an AI to think and how you want it to take your own process into account and use these things. Like yeah, it's I mean that might be harder than vibe coding up some stuff, but you're not going to get you're just simply not going to get the results. Like it's going to start hallucinating. It's going to start making stuff up. It's going to ignore your directives. It's not going to do what you want it to do unless you organize your thoughts and and go through these steps. At least from my experience, because I was doing these wrong and I wasn't getting anywhere. So, talk about wasting time. Wasn't getting it.
>> Yeah.
>> Um so, so yeah, I mean mistake number two is like running the whole project inside one session, one cloud code session. you want to use separate chats for different layers within that same project, right? So, I'll have one that's dedicated to that morning report and we're just talking about how to synthesize emails and themes and what what are we drawing from my own process? How do I want you to think? What's the structure of this report? How does each section think differently? Right? Like the the environment section is thinking like what's important. What does the market care about? The catalyst section is thinking like bottom up like what what catalysts are actually most significant for like the individual name regardless of whether the market cares because think about it like it's reading all the emails. There might be an Nvidia report in there that's like actually you know from a trading perspective a non-event for Nvidia,
>> right? It it just it's not an inflection. It's you know it's in a range. But 80 88% of the emails are going to have deep Nvidia report commentary in there because it matters to the entire market. And so you're going to find that like Claude is going to put Nvidia as like the most significant catalyst. And I'm like no, like this needs to be in top stories in the the overall market tone in the environment because that's what matters to the market. I want the top catalyst to be I don't care if it's a stock no one's ever heard of, but if it hasn't
>> exactly Inod Fastly from last quarter, like the stuff that we trade catalyst plays on, like I want it to surface the names that have inflections based on the catalyst and it might not be talked a lot about in the email. So, you have to teach it like what you're looking for. And so these are all the things that I'm dealing with in on the chat side when in the chat that is dedicated to designing this AM report so that I'm not confusing cloud code with all of these like okay how do we do this hand up this mistake is great for me because I'm even thinking about my own trading assistant and I've like started to do this but not in a clean way but separating ating each session for the different tabs and sections of your dashboard is genius cuz like it is one project. It's one big dashboard, but at the same time like when you're looking at the tendencies of my trading that's so different than reviewing trades or grading accuracy like they're just very different. So I could do a better job of this I know myself.
>> Yeah. I mean, it's basically compartmentalizing but then there's there's kind of an endgame to this. like I'm kind of leading you through where I got and there's like a real serious um like it all comes together in the next mistake. What What are we on? Mistake number two. So So mistake number three now. [sighs]
>> All right. So I vibe coded a lot.
>> Here's a mindbending stat for you. SMB traders have 20x to 40x higher odds of success than independent traders. Find out why and how to greatly increase your odds by visiting smbtradingfloor.com.
>> And I coded a system that I intended to to trust with my real money. Right? So, even even with the chat side, um, where I'm just trying to work through this stuff, it's like, okay, it's great that you're working through this, but then you come back to cloud code and you're like, all right, now that I know what I want to do, can you do this? And then, oh, you didn't quite do it right. Can you change this? And you kind of fall into the same trap where you might have figured out on the chat side what you're how you want to say it and what you want to do, but then you still get into this thing with Claude that where you're trying to audit and refine and the same thing starts to happen, right? So, I I mean I think vibe coding when it comes to um I guess adding inference layers to your project is is super dangerous because it just muddies up the whole system, right? It's like a continuation of the theme that we're talking about, right? Um, what I did is I kept an architecture document for each chat. So for each inference layer that I'm designing within my trading assistant. So let's just define those like there's an inference layer that handles the morning report. So this is the brain that synthesizes all the information and writes the report. There's an inference layer for my daily logs. So that's like if I'm inputting a trade and I want to do a daily review and I want it to ask me questions and I want it to surface like push me on like being more specific and then maybe surface some things that are like you said like tendencies or or things I need to pay attention to. The third inference layer is just my my performance section and that's like the master like if the if the daily log is like the manager, the performance section is the CEO. This is the one that kind of looks down on everything and says, "Okay, over time this is these are the patterns showing up in your trading." Like you're really good when this and this and this are present. Like these are the things you need to watch for. These are the things that are coming up in your reviews. And then um you can slice that through a week, through a month. You could do a monthly review, a weekly review. And you're basically asking us to surface these things. So when I say an inference layer, I just mean there's a separate chat dedicated to each one of these. And then within the chat, and this is what prevents me from vibe coding. This is what this is my workaround. Each chat creates a document that is a architecture document for that inference layer. So as we're working on it, we're basically just coming up with a PDF that outlines everything about what I want cloud code to do. And so now, not only do I have a record of something that's organized and that's mine. So like if cloud code blows up or I want to go to a different language model, like I've got my project, I've got the architecture for my project. It's not just all lost in vibe coding. And now when I want to make a change, I'll take that document, the updated version, I'll send the updated version to cloud code and I'll I'll give it a prompt that says like here are the changes I made to the architecture. these are the things I want you to implement like please read the architecture and then also please give me a report on what is in the codebase that it does not align with this architecture and that's what we call like closing the loop and then if anything needs to be updated in the architecture you go back and you do that and you keep them so that they mirror each other that's great that's interesting so you're saying like vibe coding basically is it's now been such a common term and everyone like kind of says it where it's a productive thing, but you're saying it's just coding, building everything out without actually documenting it and updating a document.
>> Yeah. Like, and it's it's it's kind of just you're flying by the seat of your pants and you're just messing around, which I I mean, it's fun. It definitely works for like UI changes and things that are contained, but I think that if you want to, it's my opinion or from my experience that if you want to ar like really add an inference layer on top of what you're designing, if you really want to tell cloud how to think and how to operate within different parts of your project, I I think you would really benefit from having an architecture that you're building rather than just vibe coding because like that just it's it's like breadcrumbs like you're just dropping breadcrumbs everywhere and maybe eventually contradicting yourself. So I would rather have what we call this in my my cloud chats is like the single source of truth. So we tell cloud code like this architecture document is the single source of truth and it knows that.
>> You know you're in deep when you start saying we as it you
>> I know I know and I look I might be losing my mind. I might be completely losing my mind, but I've made so much progress with this thing. And it doesn't it's not like super time consuming. And in fact, like I was wasting a lot of time before I figured this stuff out because I was going around in circles being like, "Oh, you didn't do this quite right. How, you know, you didn't do this, you didn't do this." And it was like dropping these breadcrumbs and they just weren't fixing each other. But once I really sat down and said, and I'll show you this in a second. I'll share my screen. I'll show you the architecture. Once I sat down and like wrote out like, "Okay, this is how I want you to think about like each section and like really make sure the language is consistent, it nailed it. It like absolutely nailed the the morning report. It was like magic."
Yeah, that's that's awesome. What I love about these mistakes so far, one, I haven't heard anyone really talk about them, and two, like it it just will make everything so much more streamlined. Like no one talks about these things, but this is how you get more done. Like everyone so far, what I've seen on videos, it's like how to build more, you know, build more, more, more. And it's not like building efficiently and well.
>> Yeah. Uh maybe you'll show this is on the example, but I'm curious too like you're updating every single change like no matter how small like or what's a small change that would like does anything not get updated to the architecture doc even if it's super tiny.
So like I would say the really small UI stuff like user interface stuff like so if I say like hey can you know can you make the the font bigger on my my tab >> like that's not going to be in the architecture that's something that's very contained cloud can do that but like what is well it's the like the structure of the report Right. The fact that we're starting with like environment and then we're doing themes and then we're doing catalyst and then like how's each one going to be displayed. What kind of colors are we using for like you know up percentages and down percentages and if are we color coding the the themes and and the calendar and all of that is in the architecture because it's meaningful. But any little like I would just say the only stuff that isn't is just you know those little like tweaks you make to the UI which is super contained.
>> Yeah, that makes sense. Something like font you don't really need to throw in there.
>> Yeah. Yeah, exactly. I will um let me share my screen. How do I do that Kurt? Here is here's my architecture and this is for the morning report and yeah, it's not short but the chat side will write it for you. So again, like it's not it's not too scary. Like I I'm I'm going to want to make sure everything's right that it's saying, but um, but yeah, I didn't have to write this from scratch or or put together like how it looks because the chat side knows exactly what you're trying to do. In fact, like I went back to cloud code and I said, "Hey, like am I crazy? like I've kind of come up with this process and like these writing these architectures for like each inference layer of this project like am I going about this right or am I nuts and it was like no you're this is exactly how you want to be doing it >> and I'm assuming this was a lot smaller at one point and then you keep updating it >> and that's kind of some of the length too.
>> Yeah, 100%. And so, um, just to just to kind of like just to give you a little color. So, on this mistake number three where we're talking about like how not what's the opposite of vibe coding? Well, it's obviously creating a like an architecture document. Um, so the directive here is like I keep a per layer system architecture document, right, as the single source of truth. I update it first with every refinement. And then I prompt clo prompt cloud code with the updated version attached and let it feed the cloud.md file. Right? And this is the only technical thing I'm going to mention this whole time because this is not like a this is not a computer science conversation. But but I would say like the one technical thing to be aware of is the fact that there's this cloud MD file that exists on some of these bigger projects that cloud's maintaining. What is that? Well, it's a file cloud code reads automatically at the start of every session. So your project's like standing instructions basically. So it has like conventions, architecture, vocabulary, how you want it to behave, it has all that stuff in there. So, a lot of what you're giving it here in your architecture is getting translated to how that um cloud.md file is being updated because that's kind of like Claude's version of like how can I keep track of this stuff. Does that make sense?
>> Makes a ton of sense. I don't know how much you want to get into it. I'm kind of curious like what was one like update and what that looks like in the doc.
There are a lot of subsections that you can ignore, but there are there's an A, B, C, and D. And then there's an appendix, which is E. So there's really only five sections. One of them is an appendix. So there's four main sections. So what's the first one? The system's purpose. So you're just, and this is the shortest section. It's like a header. It's like basically saying what is this? What are we doing? What is this? And this is the the architecture for the AM report. So, I'm describing like what this AM report is. Um, and one of the things I've learned with a little bit of research because I'm an idiot and I don't know this stuff automatically. I had to look it up. But it's like how to prompt what are the best ways to prompt cloud and the AIS in general, especially as they get smarter. Apparently, the the optimal way to prompt them kind of evolves over time. And it didn't always be it didn't always used to be this way according to some of the white papers that I read. Apparently, it used to be more optimal if you gave it the steps to take. Now, it's more optimal if you start with the finished product product. So, if you start with what's the definition of success, what does success look like? It's going to do a much better job of getting there if you lead with that.
>> It's a really good point. I just want to emphasize because yeah, that's a totally different thing than saying do five steps XYZ hopefully it looks good,
>> right? And it it's great because that's that's easier for a human too because like we we just want to describe what we want and so it does a really good job of getting there. So part A is just the systems purpose. So I just take it through what what we're doing, what's the actual purpose of the system, um, some of the key things like rules that I wanted to know right up front. So this is kind of like an overview. And then we get to part B, which is how the system thinks. So this is the inference core, right? So this is how I'm basically establishing the synthesis, the inference, like the the cognitive level of of this report. And I'm breaking it down through like for instance, I'll just give you a little tidbit like the four-part inference framework. And so I'm talking about like every section of the report has a question, a lens, a selection, and a synthesis. And so what does that look like? Because this is how I get each section of the report to to basically like give me different information based on telling claude to think differently for each section. So for tone, this is like okay, what kind of day is this? That's the question I wanted to to answer, right? The lens is through the market environment and the broader trading background selection. Well, I want the information that be best characterizes the tape character and risk appetite of the overall market to be in this section. And then the synthesis is just like defining the characteristic of the market environment in the session. So we can do this for each section. And if we go all the way to like catalyst, you'll see how different it is. It's like catalyst. Well, the question is what names are most in play? The lens is an individual company or directly affected basket and we select the names most in play driven by the most significant company specific catalyst and the synthesis is why the catalyst matters to the selected name and what it reveals about the opportunity or risk surrounding it. And so it might kind of sound stupid but you do have to explain how to think if each of these sections are supposed to be different.
>> I mean to me this sounds genius but maybe there's someone out there calling it stupid. I mean, because I'm even thinking catalyst, like how do you expect it to find significant catalysts if all you say is find significant catalysts without being as specific as you are?
>> Yeah. And and so if we jump to um, we're still in B, but if we go to like B3, which is a subgroup, there's a sub section of this architecture that's called catalyst methodology. And we've talked about this on the podcast. We talk about in our mentorship program and we do this with our catalyst plays, but I'm basically saying like, okay, like first of all, don't make stuff up. Like if the information's not there, like don't force it. It can be short. Don't don't hallucinate, right? But yep, like here's our sevenstep catalyst breakdown. Here's how we think of catalysts. And then, you know, you could even go to an earnings catalyst and classify the buckets.
>> Those look familiar.
>> Exactly. And you can explain, hey, um, look, like they don't all fit in these buckets, but sometimes they do, and when they do, we want to know about them. And you can also, because if you don't specify this stuff, ex, you're exactly right. Like, it's just going to say like whatever the big this the mega cap that had earnings that all the emails are talking about is just going to show up first on like biggest catalyst, right? Right. And that's yeah,
>> that's not what we
>> And they always are going to cover Nvidia. So it's like that's not the point
>> exactly. So we're trying to get it to surface this stuff. And I'm starting with just our email basket um because
It has tons of information in there. But then I'm going to branch out to like external sources like Edgar and like macro trends and give it some of our firm data like you, you know, connecting it to our market views and letting it know some of the market-generated information like in the pre-market, so we can start leaning on more information. But first, I want to get it like at least just doing the email correctly. And this might bleed into when you wanted to talk about settings.
>> But this is awesome, just seeing it and seeing the level of detail in this architecture doc.
>> Yeah. And like I mean,
>> very impressive.
>> I mean, Tim, the great part about it though is like I didn't have to make this stuff up because we already like,
>> we already have this stuff. This is just our process. Like we have this,
>> the earnings names are the exact same, which I love.
>> Yeah. Yeah, exactly. Well, because I'm, you know, sharing this with you, so I want it to be universal. You know, this particular report, you can spit out a PDF and have it emailed in the morning. So, when this is ready, I'm going to be like pushing this to you. Um, so of course, I'm going to keep the same names. Um,
>> yeah, I'm solving. I'm pumped.
>> And you know, so, all right, so we're in section B. So, section C, what the system produces, right? So, so far we did what the system, the system's purpose, how the system thinks, and then what it, what the system produces. So, this is where we break down the sections and actually give it the structure. Super simple, right? We're just taking it through like environment. There's a theme section, there's a catalyst section, insights are like analyst insights and stuff like that. And then there's a watch list section that is basically you can input tickers if you wanted to surface stuff about like your watch list. Um, all right, so D, how it's rendered and run, and this is all technical specification stuff that the cloud chat will, will get together for you. So, this is a beauty of using the chat side because cloud chat knows exactly what cloud code needs. So, a lot of the stuff in here is like stuff that cloud chat got together to basically say like, these are, this is, these are the things that we need to like, really explicitly explain about some of the technical renderings of this report to cloud code. So, I actually didn't write any of this. All I did was like audit it and make sure it was right.
Did you have to ask it something to even get it thinking about that, or was it automatically like, "We're going to need to deal with this," or were you like, "Is there something that we need to deal with with the rendering?" That's a bad question. You get what I'm saying?
>> Yeah. No, that's a great question. So, no, I, I didn't have to tell it to do this because like all I, you know, and obviously this came from trial and error, and so sometimes when that happens, like it, it happens naturally and you just, you end up starting doing this architecture and you're like, "Wow, okay, I think this is the answer." But basically, you're talking to the chat and if you tell it, "Okay, the whole point of this chat is to, is to draft this architecture document, and we're going to work through it. We're going to edit it. We're going to take it step by step, and in the end, it's going to be done, and we're going to give it to cloud code." And if we have to make changes, which, you know, you always do, then we can just keep updating this architecture, but that's what this chat's for, maintaining this architecture. Okay, now it knows that and you, and it's going to know a lot just by saying that because this is a best practice, as I learned from talking to cloud code, who said, "Yeah, know, you're on the right track." So, it's going to automatically now know like, and you could even say like, "I just want you to be an expert on how to express stuff to cloud code. Like, I want everything in here to be like optimally explained the way that cloud code would like totally understand it. And if there are things in here that I'm not covering that like should be in here, like I want you to,"
>> cloud code needs.
>> Exactly. And so like, what ends up happening, and this is really the answer to your question, is as long as the chat side knows what the objective is, now you're just talking to the chat side like the way you would as a trader being like, "Okay, my ideal report is this, and like I want like a catalyst section that like reads like this and surfaces these types of names, and this is how I'm breaking it down." And then over time, the chat site is like keeping a log and keeping notes and and then it'll spit back to you like basically organized summaries of like everything that you've described and said, like, "These are the key points that we need to like work into the architecture. Like, do you want me to um put these in right now?"
>> Gotcha.
>> And it'll, it'll just start writing it for you, and then you just, and I would advise like to go back and and read it and make sure everything is right. Um, but it will start writing this for you. And so what it ended up doing was it wrote this entire section. Um, some of the stuff like I knew needed to be said, like there's different additions, and I knew we were going to have to outline that. The visual stuff, like how stuff is like rendering, um, visually, like there's, there's stuff that I, that I knew were going to have to be like addressed, but the chat side did just did a great job of like organizing all that and putting it in here.
>> Yeah, that's great. Like when it's going to send. I saw that with the time stamps.
>> Yeah, exactly. So, all you have to. And sometimes it asks you questions. Sometimes it'll be like, "Hey, um, you explain this, but like that kind of leaves this question, like when's it going to send or whatever?" And it's like, "I'd say it'll be like, I have two questions for you." And then you answer them, and they'll say, "Okay, good. Locked in. We can add this to the architecture. Now, on to the next thing." And it kind of walks you through building this thing. Yeah.
>> Were you using the ask me questions on that? Or
>> Yeah, sometimes I do, or sometimes it just spits questions just anyway. And then the final thing is the appendix, which is literally just two things, like how to build it, and this is super from cloud, like how to build it. This is all like,
>> very technical stuff that it added that I didn't even know it needed this section, but it added it. And then what's next, which is phase two roadmap. So, there are going to be stuff, there going to be things that are going to be ideas that you have that you're going to express to the chat side that are going to be too ambitious to start with. Like it'll be like, "That's a great idea, but I think that's like a phase two item where we need a phase one that's going to include everything that we want cloud code to like build into it right now. And then we want in the appendix like a phase two roadmap that gives cloud code like a heads up for these are the things like coming down the pipeline. We don't want you to build them now, but they're going to be there." And apparently that helps cloud code like leave space for it in the codebase. Like it's better if it knows than if it doesn't.
>> That makes a lot of sense. This is, this is great and something that I need or I want to implement more.
>> Yeah. I mean, it looks like a lot, but it's very easy when you have a chat like basically putting this together for you. That's the architecture, and I'm going to have one of those for, I, I do have one of those for the report, I've got one of those for the performance layer, and one of those for the daily log layer. And each one just has like a very specific um task, and each one then can communicate with each other. Like, for instance, like the performance layer can pull information from the daily log layer to generate its, you know, weekly reports of of what the hell I'm doing wrong. Right.
>> Yeah. What I think is the genius aspect of it is continually updating it too, so you never get it behind and lagging.
>> Yeah, and that, that's a great. So, the chat side taught me this. It's called closing the loop, and it's like a best practice when you're doing this, where every time you update something, especially on the codebase, like you want to close the loop. So rather than just giving cloud code a prompt and saying, "Do this stuff," and then it just kind of says like, "Cool, done,"
>> you want, you want it to give a report back to you of like everything it changed and how it did it.
>> And then what you do is you give that report back to the chat side and say, "Okay, this is cloud code's response." And then the chat side will be like, "Great, closing the loop." and it'll be like, "Okay, it was really smart the way it did this thing, so we need to kind of like update that in the architecture and reflect that." And then it, it, it kind of left these two things like unopened, so we need to actually give it like one more prompt to like check." So you're kind of like leveraging the chat side and and to talk to the code side, and you're kind of staying out of it a little bit.
>> and those things,
>> honestly, you go,
>> just those things should mirror each other by the end. That's closing the loop. The architecture and the codebase should be, should be locked by the end of a prompt. Go ahead.
>> I was going to say, so I haven't been doing this nearly as well. Um, and there was one change, basically I made to my assistant where there was a point where it was like dealing with open trades perfectly fine.
>> Yeah.
>> And then it got tripped up where it started entering them as separate new trades rather than adding my notes and the cumulative P&L to the existing open trade.
>> Yeah.
>> And then all I did, I fixed it, um, but I didn't update any report, and I'm curious if that's going to come back to bite me at some point.
>> Right. Yeah. I mean, those are, those are,
>> but that's the example. Yeah. It's amazing how many artifacts that I'm still finding in the codebase left over from before I started doing this.
>> So like, as I start to make changes, cloud would be like, "Oh, no, wait. That already exists, but it's like under these names."
>> And there's like four of them instead of three of them. So it's going to be like weird to to change. Do you want to do that now?" And I'm realizing, "Oh my god, that was from when like maybe the first like couple days of when I was vibe coding with this thing, kind of describing like what I want." And it's like, there are these, they those things stay in the codebase and they can become toxic after a while. So that's why it's like, start from the beginning. Yeah.
>> Yeah. And that, so that's keep starting this way, keep closing the loop, using the architecture, using the chat side, are all things that I'll be doing like from the get-go when I do anything that's more involved than simply, you know, UI stuff or just simply like code, code-based stuff. Um, mistake number four.
>> Move on to this fourth one.
>> Yeah.
>> Yeah. All right. So, I never taught my assistant who I am.
>> Have you heard this study of tens of millions of traders which shows that only 1% of retail traders actually make it? Don't be a statistic. Visit smbtradingfloor.com to greatly increase your odds.
>> That's mistake number four. So, I started out expecting it to generate and synthesize in like an intelligent way. And that earnings was an perfect example that we I just showed you where um, I thought it was just going to break down the earnings and be great, but it wasn't. It was trash. Like, it was surfacing names that were not significant reports. And I was like wondering why for a while until I finally crossed that threshold and started to find ways to teach it about my process and the way I think. Right? So like my risk tolerance, my methodology, like how I think about trades, how we break down catalyst, our theme tracker, our entire like notion theme database, like give it that understand like, this is what we, when we say theme, this is what we mean. So I built a settings tab to hold all of that and I feed it into the app's reasoning live. So the settings tab lives in the operating system, right? So on my tool. So I'll show you that. So here's my trading OS. So this, this what you're looking at right now is my trading assistant.
>> And we're just, we're in the settings tab. So this is, this is what I'm talking about right now. This was the, the answer to this mistake of of never teaching my assistant who I am. So, this is where I hold a record of all of the things that that outline parts of my process, things I want cloud to know, so that depending on like what we're doing, it can reference these things and respect them. So, you might not reference all of these for every single thing. Like, for instance, the morning report is not going to necessarily reference like my playbooks and like my review process, but it certainly is going to reference how we break down catalyst and the theme ontology and stuff like that. So, just to give you a little rundown, Tim, like under system context, like, and and I could probably change this title because it's a little bit too generic, but it doesn't matter. I've got all our playbooks in here. And like the great part about keeping it in settings is it's like owned by me. So like I can update this stuff at any point. Like if I want to add a playbook, if I want to change some of the variables or or how I'm writing it, like I can do that, but it's all in here. All our playbooks are in here. And then you've got like the grade table risk. And then you've got like catalyst methodology. So this is everything to do with like how we break that stuff down. And then review process, right? So stuff about the trade log, the review methodology, what I use that for, all those different gaps I talked about that I want to be able to to see in my trading and surface. And then of course, like the accounts, which just explain the different accounts. And then here is our our themetology. So this field is going to break down like what is compute? What is robotics? Like here's our taxonomy and how we look at everything. Here's the definitions for every single theme, just so that you know, so that we're on the same page when you're writing this report, that like this is how we look at this thing. And then here's the theme database, which you just upload a CSV and it pops right in here, and this is just straight from notion, but this has like every ticker and every theme tag associated with it in here. And then, you know, I've connected it to Gmail through here and settings as well, but that's kind of besides the point.
>> Yeah, this is sick. Share it. And I love that it's in settings, too. Like, it's a perfect best practice for how to do it. And it really goes along with like, if you give it nothing, like you can't expect like a masterpiece or really good information or like you give it, it's going to give you out.
>> Well, that's the thing that I realized, and that's probably a great place to like land on after all this is like, we know from the quant side doing analysis and and using code that like it's all about the data. Like the be, the better your data is, the better you're going to be. And so I kind of, that's what I think of all of this as is like, okay, if I'm trying to build something and it all it is is code, like I need to give cloud code like really quality data. So what does that mean? Well, it means like everything in the architecture that explains how to think and and how these sections should be, and it also means like everything about me and my methodology. And so the more specific I can be and the more I can give it that's meaningful, and and we also know the more you can not give it that is meaningless, meaning like you don't want to give it noise, um, the better it's going to be. And so it's really no different than the same philosophy that we kind of take to the to the quant trading and the analysis.
>> Yeah, 100%. So, I'm curious. I have a couple questions like on this idea of this whole mistake of you're not teaching it how you think. So, now you are teaching it how you think. You're giving it the settings. Like, how do you know Claude is actually like learning? Or like, are you giving it like pop quizzes? Almost? Are you just kind of seeing every day with the output getting better?
>> Yeah. So, I mean, the number one thing is the output. So there are certain things that you can tell very clearly that it's surfacing directly because it understands what's in that settings. So that's that's the key because of course, what you care about is the product that you're creating. So like, if I'm reading the report and it's surfacing things about the exact themes that we have in our database because that's what I want it to do, like I know it's working, right? So like one of the big changes that I made was, and this is a good example of like when you give it nothing, then kind of like nothing exists like like if the more boundaries you give it, the more it can leverage inference. That's like one of the sort of things that keeps running in my mind. So a good example of this is like, if you just say, "I want a section in my morning report that breaks down the themes that are in play." Claude's gonna make up themes. It's gonna be like, "The global like the global rates, you know,"
>> bond like, you know, the European bond theme this morning and the the Trump taco theme." It's just going to say whatever's in the emails and it's going to be like, "This is a theme." So, you want to teach it what is what is a theme to you, and then even more specifically, what are your themes? Like, if you have a database or watch list that are very specific, which we do, I'm like, "All right, I'm putting this in settings and I'm going to tell you I only want you tracking my themes." And so now, and I, and it knows the definitions and everything. So then I run the report, it's in settings. Now all of a sudden it's like, it's like digital infrastructure, like data centers, memory, robotics, uh perception, whatever. Like, it's exactly our themes, and it's breaking down those things. It's understanding exactly what that label means too because it has the definition. So you can tell that it makes like a huge difference. But the other thing you can do is just talk to cloud code, and you'll start to realize like it knows this stuff because I would maybe ask it some questions about how I'm going to kind of do something with the on the process side, and it would be like, "Yeah, like, so like just like in your trade log, you know, it's going to like surface these like four things." I'm like, "Okay, cool. You know my trade log. Great."
>> Yeah, that's pretty cool. Maybe this is maybe my last question, but are you at a point now where you pretty much trust it? Where like sometimes I feel like with AI, it's like, can you actually take it at like full face value? So if your operating system, your trading assistant like tells you to look at a specific stock for a catalyst or a specific theme, are you double-checking its work anymore, or are you at a point where
>> Yeah.
>> Yeah. No. So that's a great question because, you know, first of all, I'm not, I'm not a developer. I'm not a computer scientist. Like, I'm a trader. Like, I don't know what the heck I'm doing, right? So I'm figuring it this out, trial and error, and noticing the things that are making like the hugest improvements and efficiencies in what I'm doing. And and that's what I'm presenting here today. Um, it comes with a lot of skepticism because obviously this is a giant experiment. Like I told you in the beginning in the last episode, like I hadn't even done anything yet because I was like, "Yeah, I don't, I don't know."
>> I'm kind of skeptical and I've kind of have what I want and I don't want to like waste my time because I want to focus on trading. Right. So,
>> it's, it's a very good question. And so I, I did take a pretty um deliberate order in the architecture documents to address the the sources issue of like, where do you get your information? What can you infer, right? Like giving it boundaries, and then every sing in the report, every single section that like comes with sources at the bottom. So I can go to the emails that then it got the the data from, and I can check it. And so like, I've been doing that, and it's been spot on.
>> Um, it hasn't like hallucinated or made anything up. I would just say like, where it does go wrong tends to be in the structure. So not so much the data that it's giving, but like it's, you know, it's talking about some name that it like shouldn't be talking about in this section because it thinks it's important.
>> I see.
>> And yeah, the reason I ask is that, yeah, this is a huge jump. Like, honestly, this is all sick.
>> And I'm excited to uh get the email in my inbox every,
>> you know, and then it'll be like us like just auditing it. I mean, I think it's going to, I think it's going to help. I think it's going to save time. I think it's going to be efficient. I think it's going to direct our focus to where we want to be, right? Which is like, we want to be able to take 10 or 15 minutes to read something and then know exactly like what themes are in play and then boom, back to the screen. Like, so all of this work really is to in the end of the day make like every morning and every after hours way more efficient. And if, so if I can get that time back, then this is all going to be worth it. And it is an experiment, and I, I'm well aware that these AIs make mistakes at this point in time. And so I'm trying not to give it too complex of tasks, but I am tracking that stuff as we go. And I have been like very surprised at how accurate it's been so far. So, we'll see. It'll be, I, I'm actually looking forward to giving it to you so that then I, I can get another perspective because like, like I said, I might just be like going like off the the reservation completely.
>> Well, you guys are best friends.
>> Yeah. [laughter] Exactly.
>> No, but even go back. I mean, after seeing it and talking to you, like in terms of, is this a high ROI activity for the reasons you just listed? Yeah.
>> Yeah.
>> So, I'm pumped.
>> Yeah. Well, we'll see what, maybe there's a follow-up to this where it all comes crashing down. But I hope, I hope that people listening, you know, can can just gain a few simple process rules for themselves as they do this. If you want to take on like a bigger task in building stuff with cloud code, like especially something that builds the inference into the project, I think that these are really, really beneficial steps to take. Um, and and then maybe we'll find, maybe [clears throat] we'll find some more the deeper we get, but hopefully not. Hopefully we're just, we're just done with this and we can we can move on.
>> Yeah, these were four really good mistakes.
>> Everybody have a great [music] week of trading. Thank you for joining us. We're on Spotify now, as always, if you're watching us on YouTube, and you know, put stuff in the comments or not, right? Like whether I say it or not is not going to make you do [music] it. So, just, just do it if you want, and we'll see you next week.