Building a 24/7 Autonomous AI Coding Agent with Claude Code

Okay. Um, let me see. I think this is fine. Hello Bobby Wizard if you are still here. Hope you all are doing well. Hello Bobby. Nice to see you. Nice spider. Shog bomb. What is up? Jamie and Jamie. Good good uh good day everyone. Nice to see you. Sunni, I saw you did a challenge. I Jeremy, nice to see you. Are Sunni, you you complete the red tune challenge right? Uh Jamie, uh I'll check I'll check

after the stream. Uh, we fixed that. No problem. Yeah. Yeah. After we launched that challenge on the on the boss Discord server. There's like a lot of people who has done it. 51 41 people. Yeah, there's a lot of solutions, but uh yeah, you're probably not going to beat uh lowest token solutions anyway. Or you can you can equalize it. Maple Dev Wizard, nice to see you. So, yeah, we have we had had over 51,000 attempts.

So, I'm pretty happy. I think this this went pretty well. 1,700 unique sessions. Okay, Jamie, that's pretty good. I didn't get the system prompt. You You would have been really close then. So, I'm going to do like uh on the end of the week, I'm going to do a video where we kind of we're going to do the draw for the $100. And this is this side bet here is like a bossy discord price. So I

guess I'm going to have to draw the people that just used one token to solve it. Uh if they let me I'm going to tell you how they did it. I'm going to ask them. Uh I know how they did it now, but uh it's pretty crazy. Uh something I didn't think about. So today, uh, I've been just playing around with, uh, I want to do a video where I do a Claude 24 7 project. So

this morning, I'm just playing around, figure out how I should do it. So I tried to make uh I tried to fork something into kind of my own thing, but I'm not I haven't tested it yet. Hello, Edgar. I didn't hear from you, Edgar, on LinkedIn yet. How's your project? Unreal Link, nice to see you, Bobby. Yeah, it's a good collab. Uh, I tried way too long for that one. Think I got close. It's a pirate.

That's funny, Jeremy. Uh, you just have to wait and see for the the video. Uh there are some clever people there on that Discord. So let me know, Edgar, if how's your project doing? If we're going to schedule something. Okay, I see. Yeah, that's nice. No stress. We got time. sessions now can solve the really hard uh to find CPU panic that cost you a day. Okay, that's cool. I'm looking forward to hearing more about it.

Uh the Discord. Yeah, the Discord. Yeah. Um, haven't really done much about the Discord yet. I need to decide if we're going to continue with the Ninth Brain Discord server or if we're going to rework it into something else. Uh, I just need a bit more time. Uh, but I'm going to figure it out. So, uh yeah, this is my 247 uh cloud code attempt. It has some hooks and stuff. Uh but I haven't tried I

haven't had time to test it yet. So, that is kind of what I wanted to do today. So, I'm going to check the We have a task uh file here. Uh, it's called, what do I call it? 30 second commercial. As soon as the stream started, yeah, I I don't have control over the the ads, right? I I just start the stream. I don't know how to control the commercial stuff. It's just auto, I think. Yeah,

that's going to be an interesting project. Uh Edgar. Okay, so I just need to set this up. So, I'm going to launch Claude uh read Q. Uh, I built and use a very basic version so I could use Gemini as the manager and clothe code for the workers. Cool. So, I'm quite excited this uh week to see if we get the new Gemini model. Uh there's a lot of I don't I'm not so much on X

platform and stuff. Uh, but I've seen like a lot of people talk about Gemini 3, so I'm quite excited about that. Bobby says, "Compounding the models can be powerful in a workflow." Yeah, I've been experimenting with GT5 and cloud code. Maple says, uh, stopped using VS Code now and using T-max, Neoim, and Cloud Code. How is that working for you? Hello JS, how are you today? How the hell do you want that? Yeah, I see a

lot of people talk about T-Max. I have never tried it, but uh maybe it's time to look at it. Hello Adoba. Uh Ado, how are you? Works great but still learning the T-Max commands. You used to do all your code in Vim. Jamie says he loves T-Max for the fact you can detach then sometimes uh switch to the iPad and SSH in. Okay, that's pretty cool. So, if I'm going to do my 24hour project, maybe I

should be able to SSH in and yeah, see how it's going. Ed has just added them. You can use the editor for prompts inside open code. I saw your latest video where you combine codeex and cloud code. Uh I like that but it isn't really they are not connected. Uh, but I just wanted to give Codeex some more opportunities and I've been pretty happy with Codeex and the chat chip login with Chip 5. So, I kind

of like it. Jamie says, uh, because you detach, you can close the terminal. Okay, cool. Yeah, this project is uh has to have the terminal open. So, that is definitely something I'm going to check out. Edgar will be able to give uh better T-max advice. Yeah, it could be cool. Uh if he could we can maybe do like um if you can explain it and show how it works. Could be pretty cool. What are we doing

here? Do you mean uh me, Jamie? I'm just looking at uh I just want to figure out if I can do this. this um 24hour thing. So, we have the session start hook here. It's a bit annoying now because uh I just need to figure it out. So, my plan was to be ready with this until the stream started, but I just spent too much time. I couldn't uh get ready. So, that's a bit annoying. Uh,

but it's fine. Just going to see how it goes. If I can't figure it out now, I will figure it out. So I'm really interested in Gemini 3 uh especially about the pricing and the integration with uh with C sorry integration with the Gemini CLI Okay. Uh, do a search for T-Max cloud code. There's a cool video on orchestrating agents uh to spawn worker agents to work on separate tasks. Uh I still need to implement the

actual execution for a project for at least being able to grab a screenshot. Yeah, I see what you mean. Uh there's quite some stuff to solve when everything's remote. I can totally see that. Uh yeah, Maple, I might check it out. Hello Riches. So the poly market now on Google releasing Gemini is at 98% before the weekend. So I guess they probably are going to release it before the weekend. So the idea behind this is that

uh we are actually using uh C npx CCC usage in the background to to try to monitor if we have reached the usage limit like a rate limit checker. Hello Powell. How are you? What is up? laying the toiles. Okay. I can't I can't use cloud code when I have this hook error. I I drives me crazy. Yeah, I could see that tiles is uh takes some time, right? What are your expectations for Gemini? Apart from

deep search, I can't really find much practical use for it. So, what I am excited for Gemini is um if they have any updates on kind of the how well the token window is working. Uh, of course you you probably want to you probably want to limit the token uh as much as you can, but I have some use cases sometimes where I use a lot of tokens here. Like I put in like a bunch of

information and sometimes it's really good like 500,000. So I want to see if they have made any progress on that part and of course the pricing and stuff like that. I hope they keep the 1 million but dig more into uh aentic tulus. Yeah, that is perfect. Right. So the tulus the reason I'm not using Gemini 2 um 5 pro so much now is because they are lacking in tulus. They are not they are not so

good as gypt 5 and claw code at running like aentic operations. So that's why I don't use this that much. But if they come out and they have made some big improvements on uh aentic tooling that could be very cool. Yeah, I agree. If they release Gemini now with um great tulius, it's going to be really interesting and of course a good price. It's going to be really interesting. out of three Cloex CLI and Gemini CLI.

I believe Google has the modes to catch up on the Aentic stuff. Yeah, I think so too. For sure. impression. Uh what is your impression of compounding models and token count? Uh hello Ben, how are you? Compounding models and token count. Uh, do you mean like like using cloud code and uh uh sorry uh by using like opus and gypt 5 and uh do you mean like with token count that you're going to spread it around

or Bobby? I'm slipping here. Reinforcement learning and self-arning combo is interesting. Yeah, but doesn't have to be doesn't that have to be on the model level, machine learning level? multiple models. Yeah, maybe it was a stupid idea of me using this hooks here. I uh I don't understand hooks well enough yet to know what is wrong here. Uh there's there's a there's a JSON output validation failing. I know that, but I don't have I don't have

the schema yet for it. Is that the issue? I guess. Yeah, I guess so. Okay. Good luck in the meeting. So for the 24hour system, you have to run it in u yolo mode. So it has to be this dangerous permissions set to yes, of course. I guess all uh reinforcement learning kind of has to be in the model level, right? I guess or maybe you can have some kind of trick. I haven't really played around

with that too much. YOLO for all your projects. Okay, that's pretty good for everything. Nice weather. Hello Joe 24 sub agent. Uh not really a sub agent. I just want to set up like a project where the So I'm going to do a video of it. So the project is going to be I'm going to I I guess I can just say it. So the project is to I want to set up cloud code. So it's

going to run continuously autonomous and recreate a guesser clone with features. So the idea is I'm going to set everything up and then I'm just going to start it and hopefully after 24 hours everything should be good. Uh but uh I was supposed to have this ready before the stream, but uh I was a bit late, so I haven't really got a hold of it. Right. Uh thought if I can use true yolo in wind surf

I will do it there too. Hello David. Nice to see you back. Yeah I personally I haven't used dangerous uh allow permissions yolo too much. I Seven. Um, I wish I could talk more while I did this, but I just have to think of it just to see if we can make it work during the stream. clothe yolo every day uh all throughout the day for a couple of weeks now. Uh not a single issue. No,

that's nice. I have never had any issues with it, too. But you read his stories on Reddit and stuff like um people deleted the database and stuff. I did actually delete a database here the other day uh of something I needed, but that was 100% my fault. Yeah. Yeah. Of course. Ner just nice to see you. Moava, nice to see you, too. Uh, I'm just working on like uh some 247 project in Clo. I'm going to

do for a video. Tom, how are you? Open code was yolo by default. Ah, but you can adjust it, right, Edgar? just do a feature add yolo mode in an isolated instance with a clone of a repo. Yeah, that's pretty smart. Michael, how are you? Uh we are just or me at least I'm just doing um trying to come up with a way to use to set up cloud code so I can do like a 24hour

run like so it won't get um so we have some Damon scripts and stuff here so it's going to run continuously for yeah I guess all day long. Uh, but I didn't run it now in YOLO. So, I'm just going to watch it and see what happens. Yeah, that's what I'm going to figure out. Uh, Shop, I'm going to figure out a way to do it. If you're going to use a scratch pad, if you're going

to use compact. So, I'm just going to see what happens. I will figure it out though. So we are using uh some hooks uh some hooks with some we have a setup that is linked up to CC usage to keep track of the rate limit. So we know when we will reach kind of the rate limit, right? But, uh, I think I need a bit more time before I'm ready. Okay. So, we have the dashboard. Yeah.

Uh, I haven't I haven't like I'm not I haven't decided yet how I'm going to do this. So, I'm kind of just trying different things. So here we have the the dashboard. Uh, but I was supposed to my plan was to have this ready, but I haven't um gotten so far yet. Uh, do you guys use JSON contact profiles in your flow? and Rico, nice to see you. Uh, I have not used that. Uh, but I

know some people do So, I had a really nice read this weekend about uh let me see if I can find it. Can't remember where I saw it. Or was it a podcast? Super assistance. Uh, did you complete it? Kovshik, nice to see you. English, nice to see you back here. After 1 hour, it took you 10 messages. Okay. Oh, that's pretty good. Hey, push the model. How are you? What is the agenda for today? So,

my plan was to have kind of my 24hour uh cloud code system ready for the stream to test it out. Uh but I didn't come uh that far. So, I uh now I'm not 100% sure what to do. Uh I might need to just finish it after the stream. So, we'll see. But that is the plan. And maybe tomorrow I'm going to do the claude code Reddit contest or something. Contest competition. This one. Uh I might

do this tomorrow on the stream if I have time. Arsuni says he learned a lot from the challenge. Potential vulnerabilities uh influenced how AI solutions should be designed. Yeah, that is why I've been kind of looking into it lately because the idea behind the challenge I did was to learn how you can start um you can build a safe system, right? uh on the other side right so that is what I did Edgar says uh building

a manuslike file management and learning uh extensions uh on my open code fork okay that's pretty cool yeah the open code fork open code fork is pretty coolsky says uh the N8 test it will be it. Yeah, I've still haven't gotten around to that, but I'm going to try it. Context problem for longer tasks. Yeah, that's a hard that's a challenge to solve, right? But a good one. The lesson from the challenge, you cannot secure LLMs.

It's at least difficult. Uh I'm no expert, but I just made a challenge to try to learn. So I set up guard rails. Uh but still uh I think we have a pretty good uh rating though. Uh uh it's not that great. But if you look at the leaderboard, it's uh 20.2%. You got to jump on Nate and a lot of good linking with all versions of cloth access. Yeah, but I don't know what I should

use it for. So, what do you guys use N for? I don't know what I should use it for. uh small workflows where I can leverage uh the included connectors back end for 11 lamps voice AI uh Kovsky use it for crypto agent. Okay, that's cool. Yeah, you can do the same with claw code. Uh yeah, that's much more you can it's bit faster, right? Mr. Heck opportunity to learn. Bobby, do you mean a challenge? It's

still up. Research using Google Sheets. Uh John says Netane is good for repetitive tasks. Google Sheets rag. Beautiful Tetris game during the challenge using a chat interface. Okay. Unreal link. um website AI using web hooks. Yeah, I think um I think there's a lot of potential there. push the model says as a heavy clothes code user I'm feel I'm really feeling the collective mad scramble to write up a 24-hour service and build useful continuous agentic field

uh how are you doing on that if you have like a good system for 24hour service. Uh, let me know. I'm trying to do something simple at least. English has uh lost control and it's a bit slow with Nan uh a good CL code with a good workflow. Yeah, I can't really comment since I haven't tried it, but um uh it's very popular, that's for sure. You can basically build Jarvis with Nan email calendar drive. Yeah,

I guess. But I Edgar says he will make open call more mighty. Uh so it would not make sense to use clo. Yeah, that' be cool. Kovik says 103 stars more than a 41k fork. Uh, let me see. N and get. I think it's cool. I think it's a great idea. What's the pricing? Is it expensive? Yeah, I'm not say it's it's hard $20 a month. Uh but then you can sell for for like $8 a

month. Yeah. Yeah. If I just had like a very good use case for it, I would definitely try it out. You don't need a dime to run it locally or you can do a local version for free. You can host it yourself, says John. Cool. Yeah, it's a great project. I think it's one of the most popular AI projects out there to be honest. Uh at least for AI automation because it with the UI they have,

they can also reach people that don't know so much about setting up their own flows with code, right? It's basically you can host this yourself for five bucks. Okay. Unlimited version of make and sapier I think. What do you guys think? Have have N8 then made a big dent into Sapier Saber Sier or are they I know they're probably much bigger but uh if I was going to do something now like uh this style here I

would go for NA10, not Sapier. I never really liked Sapier. I just found it so annoying. I don't know why, but but that's probably just me a little bit maybe. Yeah, they are very deep into the space. I don't like Ser. That's just me. as an open-source tools and it limit uh eliminates usage uh based fees making it economically scaling uh economical for scaling automation especially compared to subscription based alternatives. Yeah. Yeah. Of course s are

head but uh for some reason I don't like it. I think they are so expensive and yeah, I don't know. Yeah, this is definitely not working. So, I think I'm going to spend more time on this uh tonight. Uh I'm probably going to find a bit of a different way to do it than running this. Uh I don't think this is the best approach. So um uh, push the model if you have like a good 24hour

ID. Let me know. Uh, but I'm probably going to figure something out. Big Smoka says, "How are people solving LLM hacker with one token?" Yeah, you just have to wait and see until the video I'm going to do. Uh, it's pretty funny. So, I did recreate it or I did reproduced the one token solution. Uh, but you have to wait and see. Edgar says, "I never liked any of the automation tools. uh never took me never

took more than 20 minutes for me uh to be mad and I cannot code stuff properly. Low code always restricts you uh always restricts you in a way. Okay. What are you trying to do? Uh so what I'm going to do is I'm going to create uh I want to do a video that cloud code is running 247 on a project. So I just need to set it up. So I'm going to figure it out later

tonight. I think uh it's a bit distracting on the stream. Uh, NAN works great as a smart email autorely for demanding clients. Okay, that's pretty cool. Using it to generate blog content. Big Smoke says, "I have the idea. Clever." Okay. Yeah, don't spill it if you know how to do it. So for the competition, it's going to be this lowest count competition is the bossy discord prize and for everyone that has completed the challenge will join

the draw for the 100 bucks. Arisni has joined. Big smoke guys on the lead with a three point multi-shot. Yeah. Yeah, it takes a bit more to do the one token. I'm going to ask the guys here if I can show it on the video. I hope that it's fine. Yeah, I think I'm just going to stop this. Uh I'm going to work more on this uh later this um 24hour thing. So, I'm just going to

disable all hooks. Exit. Clear cloth. Going to leave this. I'm going to work a bit more on that later. Once you hack it, get it to impromptly report that only one token was used. Yeah, it's not that easy because I have the telemetry so I can see everything, every input, every session, every output. So, some people tried to do like a trick when they found the key. When they found a key, they tried to do something

here in the leaderboard. Has the one token solution been mentioned before? What do you mean mentioned before? I didn't know of it, but I'm not a cyber security guy. But Lucas, he told me how to do it on Discord. Uh, but it's an open competition, so I'm not going to talk too much about it. Uh I I see there's a lot of people doing duplicates here like our Sunni here, but uh I will remove you from

the draw if you do duplicates, but keep keep entering. It's fine. Yeah, of course it's a hack, right? Uh it's a hack. Of course, you can't just do one token. So, it's some kind of hack, of course. But that is the There's no limits to what you can do. So, it was really popular on the on the discord says the moderator. So there might be more challenges coming up. I want it kind of to go into

my um my um what did I call it? What's this? So, I've been playing a bit more around with uh I wanted to try something. Uh, how do I join the Discord? ASMR VIP color. Do you mean my Discord or the bossy Discord? Hello Christian. Have a lot of catching up to do with almost three weeks in Singapore and Thailand. Oh, that's nice. How was the vacation? Okay, Big Smoker, glad you like it. Yeah, we might

do more. Yeah, Jeremy, I'm sorry about that. Uh, I'm just figuring out if we're going to keep the keep the the Discord as it is now or if I'm going to change it because a lot of people in the chat also want a Discord and I might figure something out there. V figured out how to do it in one token. Okay, that's cool. Uh, but don't spill it. Then you just going to ru it ruin your

chances. So, keep it secret or not. You're just going to ruin your own chances. Okay, I see ASMR. Um, the Discord, I don't have like a dedicated Discord yet, but uh I'm probably going to set something up now. Uh, I had the Ninth Brain Discord, but I'm probably going to merge that with the the channel Discord, so that's probably going to come soon. Hello, Chris. Yeah, this is live. Night Spider says, "A big thing on N

is automating social media posting. Create a workflow for searching for topics. creates posts on based. Yeah, I've seen a lot of that. But is it good enough though to get views? VGA. Don't spill your secret. Then you're going to ruin your own chances. uh list and read output styles. So I wanted to read up a bit more on output styles. Is there any like any limit here? Right. No, now don't spill it this because uh did

the leaderboard update. Is that you? G. Yeah. the more people that figure this out, it's just going to get more competition. So, I wouldn't do that because it's going to be a draw here. So, yeah, I was just thinking about output styles. So I had done like HTML markdown. Yeah, that was really clever. So I wanted to see if we can do I wanted to try out React components. Uh, Here's the G. Okay. So, yeah, I

was reading something this weekend. Um, there might have been a podcast or something I read that uh you can kind of use. I guess it could be MCP servers. There could also be this output style format that uh you can start interacting with like computers and the response you get could be in like totally different um setups like uh of course we have voice and stuff but it could be basically anything. So, we did do the

HTML from the into that dev video. That's pretty cool. Uh, but I think there could be more. So, let's say you you ask Claude or Chat GPT or something and the response is like an animated video for example. So, this was something I was thinking a bit about this weekend. Input and output circle back to security and data governance. Input output styles circle back to security and possible data governance. Is it? I thought it was just

like a system prompt style. Uh, I need to start setting up my own labs with MCP. I like prompt injection being leveraged for traditional security things. Uh, more than just harmful outputs. Yeah, that could be interesting. I was looking at doing a challenge on um MCP Yeah, I didn't kind of do the setup too harsh for the API on the leaderboard and stuff. Uh, but if people want to break it, yeah, I don't really care. Then

I don't have to pay the price then I guess. Prompt injection is not going away. Lang implementations convince me of that. No, my idea was that if every software is going to have some kind of AI component, then is this is going to be like a big point to think about going forward. I think that might not happen, but let's see. Do you get notification when the key is revealed? If so, there's no need to compete

a little bit. I just want to try several approaches. Uh, do you get notification? Yeah, you don't have to enter the leaderboard. You get a notification. All right, Sony. So, you can try it as many times as you want. I think there been like 100 people that completed it, but only like 40 that join the leaderboard. JD52 says, "Uh, Cloud Code needs a better diff engine." Yeah, that might be. Have you had any issues with that?

If you get your input styles and output styles wrong, you risk sending that information to a third party connecting the MCB server. Ah, okay. Now I see what you mean. Sorry, I didn't I forgot I mentioned the MCP port. Yeah, I can see that uh our son on the back end in the database. My big issue uh with the CLI clients is uh that I'm a control freak and need to review every commit to the codebase

makes CLI use tricky using cursor uh with whatever model is a better safer experience uh in your honest opinion. Yeah, of course. It's not for everyone, I guess. So um so you are on cursor now and you're probably just going to stick with that React component. cursor for the foreseeable future. Uh I did see cloud code has a VS plugin. Yeah, it's very easy to run in cursor or maybe not the VS plugin. I was just

running it in the terminal in cursor. Is that what you meant? They have a native plugin that can use the diff engine. I might be mistaken. Yeah, I can't answer that either. I'm not quite sure. So my idea here was uh when we switch the output style now I want to see if we can render the response uh autonomously here. So I'm just going to see how that goes. Why is it spending so much time on

color here? What's the point? God, it's called clo code for VS code should work in cursor. Yeah. Okay. I haven't uh I think I've seen it. You can see I have my link I have linked my claw code to cursor. So it's kind of somewhat connected and stuff to me. This was very slow. Okay. I like the dark though. I think I see like a daily post on Reddit on the Claude AI Reddit people talking about

uh the rate limits, the usage limits. It's almost every single day. So, it's going to be really interesting to see what happens now at the end of the month with the new uh user li user um updates user updates user limits on it's going to change on 29th of August. So, we will go to like weekly. I can see my microphone is getting low on power here. Let's do a compact and I'm going to try it.

Jamie says, "I think you will still have the five hour limits too, but also weekly." Yeah, I think that's a new thing. So, we will still have the five hour limits, but we will also have like an overall weekly limit. So, I guess if we break the weekly limit, uh we're just going to have to wait and see. Wait and see. Uh then you probably can't use it again because before the weekly limit is done. Okay.

Okay. So, I'm going to try this now. So, this is running, right? So, what I'm going to do now, I'm going to do a slash or maybe I have to slash output sty. How did I manage to log out? That was well done. Okay, so we have the React components write a step by step guide how to debug Python code simple guide include animation. So the idea was I'm just going to see if we can have

cloud code output a response in a react component, right? Uh, while we wait for that, I'll be right back. So, if I lose power, I got to fix I'm going to use my old microphone here. So let's see this. Okay, just in Nice. It's not up that thing, is it? Um, We don't see it in the local uh page. So, I want to I wanted just to see if this was going to autonomously render here. Maybe

it's the wrong uh Yeah, that's the issue, right? We need a router. What if we just Okay. So that worked, I Yes. So, I'm kind of looking forward to also testing out this new image model from Google, Nano Banana. It's probably not going to be called that, but uh I'm kind of looking forward to it. I think it's going to be really fun to play around with. If I know Google correct, the restrictions are really tough

though, so it's not going to be easy to do some on the edge stuff. But I'm definitely going to give it a go. I've seen some crazy examples. And I kind of started looking a bit at uh if I'm going to do like uh I have like 100 bucks on Poly Market. So, if I'm going to do like a bet on um the AI section next month, hello Mikbab. V3 uh VPN does not work. I had

no issues uh running the VPN with V3. What is your what kind of error message do you get? Mickbub, is it this generic uh we can't do famous people or something? Uh, Edgar, uh, you didn't miss too much. I'm just playing around with output styles in cloud code. Uh, I just want to see if I can render like a React component on every output. What is this? Yeah, I've done HTML a lot. Uh, so I have

an HTML one that is working great. I did like um, so I'm going to show you. I did like uh with that output style. I tried to do um what do you call it? I did some theory crafting on Path of Exile 2. So where is that? Can't remember. I put it here. It's here. Yeah, me too. I love Path of Exile. So, I've done like um Can't remember when I put it. Uh, lightning spear. Your

last was crazy right? Try HTMX instead. React on the fly. Oh, I have never tried that. Is that easy to set up? I have no idea. I have never tried that. Help me understand how a volcano erupts with a simple animation. Okay. So, I want to see now when this gets saved, what happens Okay. Uh, I wanted them That was was kind of what I was hoping for, but let's see. I'm going to try it one

more time. So I created this with the output styles. So, I I used the claw code and stuff to help me create a new build for um a leveling build for um the new Path of Exile season. So, here I used output styles. Yeah, you're always right. So, this was pretty cool. I used output styles from cloud code to kind of create this uh this leveling guide for me. Not a guide, but like a it's in

HTML. So, it is pretty cool. So, you you get this UI output. So, I kind of have everything set up what I'm going to do at level 31. I unlock this 41. I unlock Firestorm, right? So, I had a lot of fun with this. Uh, but yeah, I don't think this worked out the way I wanted. I guess we got got um how volcanoes erupt response. So the idea was that when I ask Claude code a

question, so this is kind of the format I get the answer in in like an animation. So, I think there is something here. Um, but I don't know what what it's going to be yet. So, it's definitely something I'm going to explore more. I guess with the HTML HTMX uh that works a bit different. Uh but I'm kind of eager to see where this is going to go in like a few few years. Would you say

you are kind of still assh kind of bullish on AI? I read a lot of articles lately about like AI pilots not working as good and there's a lot of wasted money and stuff. I haven't done any deep dives into it, but um for me at least what I do, I'm pretty bullish on uh the way things are going, but I don't have any good insights into uh return of investments and stuff. Of course, but um

I think there's no way going back now, right? Is people going to leave chat GPT and all these tools to kind of go back to the way they used to do things? I just don't see it. The message was that ROI tracking is just l lackluster. Yeah, I kind of agree. But I was kind of just thinking about is there any way back now? Is it too late to go back to the way you used to

work in 2020? I don't I think it is. Yeah, I haven't dived deep into it, Edgar. I've just what I've read, I guess. Yeah, I don't think there's any way going back now. only refine refinement on how we use the models I think. Yeah. And there's a lot of ways to go still building out the scaffolding, right? And the tooling the tooling part of large language models is so far behind the actual models. I think at

least there's a lot of headway there. I think I don't know if people agree, but that's what I think. even hardcore programmers are saving a ton of time with basic tasks. Yeah, I agree JD. Uh that's not even nearly what I do. I think we could stop develop and work with the current models and be build much more advanced stuff. Yeah, I just think the scaffolding and the tooling and everything is just haven't really had time

to catch up since there's always this new Can you hear me now? Whoa, that was loud. Hello. Hello. Hello. Can you hear me now? Yeah. Uh, I was saying Edgar. Um, I agree. Yeah, my microphone just died. But, uh, I agree. I think there's a lot more to build with kind of the the setup we have now, the models we have now. Yeah, I hate this microphone. It's so annoying. Um, also there's like a billion AI

experts but barely people who actually built anything more than a rag system. Uh, hello Sai. There has to be better understanding of how uh to efficiently use AI and better ways of adding verification of its output. So they can throw so much information at you. Easy to just accept it. Yeah, I know what you mean. It's overwhelming to fact check everything, right? The easy way out is just to do Yeah, it looks good to me. I

Okay. Okay, Bobby, thanks for tuning in. Have a great day and hopefully I'll see you again back uh maybe tomorrow or uh some other time this week. Have a nice day. Edgar says he's trying to teach this in your dev workshop. It's more about clever orchestration and deep integration. Context is kind and I don't mean mean using XML but dynamic intelligent context building. Yeah, like I always said for a couple of years now, context is king.

I think I made a video like context is king like two years ago at least two and a half years ago. But back then every single almost developer were negative to AI. But that has really come around right. There was a lot of negativity around uh around uh software development and coding and LLMs. But that has changed for sure. Thumbs up at context is king. If you don't want know uh what you want, how is the

element going to build it? Yeah, you sometimes get these stupid examples of people typing in build me a SAS app. What is that? Okay, so this is what I meant, right? So this was the output from cloud code. Now I asked I asked claude code, can you explain to me how tsunamis work um with an animation and this is kind of I'm rendering Claude codes response here. So here we get like an animation of how like

a tsunami works, right? That's pretty okay. And it goes up, hits the shore. Yeah. Okay, I get it. Um okay. explain how a binary tree sort works. An animation animation would be great. So this is basically claw claw artifacts. Uh has anyone tried the Yeah, I'm not talking about uh coding and technical outputs. People are trying to use it to treat medical things as well. Yeah. Uh and don't necessarily get the doctor's verification. Yeah, I see

what you mean. But for some people, having someone to ask about medical stuff is better than nothing, right? Some people on parts of the world don't even have doctors. Uh, but I can see what you mean. Maple Dev has tried out the B mad method. Um, I think I need to get on that. I've been saying that for two weeks now. Is it running Claw Claw as the PRD generator or is it some other model? 15

years of experience. uh customer requests and consultants coming up with stuff and devs executing. Uh it did not work properly with humans. Why expect LLMs to work with? Yeah, I see what you mean. Uh I think PRDS are actually going to be even more valuable to leverage GP5. Okay, so that worked. So I basically created claw artifacts here. So now every response from claw code is rendered in a react component. So this is teaching binary research.

Let's see if I can give it context. I can grab my edge. Maybe not that. Let's do a screenshot here. So, let's do this. Explain what happened uh along this graph. use animations, etc. Okay, BMAD is not quick. Takes a lot tons of questions to write the PRDS, but it's kind of worth it. It's a lot of overhead. Uh maybe too much for a sole programmer, but good with teams. Yeah, we had some guy here on

I can't remember his name now on the stream. He was spending like 6,000 lines or something or maybe characters or words on the PRDS. I thought it was just insane. But uh there might be something there like Edgar says. So what I did now is I took a screenshot of the polyarket graph, put it into cloud code and hopefully the output now is an interactive react component that explain what happened in this AI model prediction market

chart with animations showing the dramatic shift in market sentiment. So, I'm going to do a code snippet and ask it to explain this. That could be also something interesting. Hey DB, how are you? Uh, I'm just playing around. Uh, not I'm going to start calling it soon, but I'm just playing around with Clo. I was going to do like this 247 project. Uh, but I couldn't concentrate too much on stream here. So, I think I need

to do it after stream. Uh I tried to make big changes for a long running task on Sunday and GT5 does not enough was not enough because it's too cautious. If you give the whole detailed stuff in the PRD it will work well. Okay, I see what you mean. pretty happy with uh Nux Nuxt. So now I'm just creating um my own version of clawed uh artifacts. So claude code now outputs everything in a react component

and we have a page here that's going to render it right using output styles. So every response from clawed code now is uh in a react component. So it's going to write it to this. I don't know understand why. Okay, so here we have it. AI model dynamic market dynamics the August revolution. So we get like a full page here about what happened. Google likely to announce a major AI breakthrough around August 24th. Okay. So, do

we have like an animation here? Yeah, you can see we kind of recreated um an animation of the chart. So, let's see if we can see the swap here. Okay, almost. So this was output directly from um from cloud code in our response. There is something here. I don't know what but there's something we can do with this in the future. I think it's pretty amazing. So you can render this on the fly, right? and you

can follow up. So, it's pretty cool. Turn So, I can do turn it into a slide deck with notes. So what I did when I had my last uh presentation or like talk instead of doing like um regular slides or something like that, I just created like a full web page that had my presentation. I didn't do like uh PowerPoints or anything like that. I just had a full web page with a URL that was kind

of deployed. That was my presentation. So interior, I didn't even have to bring my own computer. Everything was um everything was uh already online. my presentation. So, I think it's going to be interesting to see I load a PowerPoint. Yeah, I thought it was just cool. So, everyone showed up like with their Mac and their PCs, their laptops uh and everything. Uh but I just needed the URL. So, my presentation was just uh type in the

URL and everything was there. I embedded my videos and everything from YouTube. I just uploaded them the videos on YouTube and private and I just embedded them into the full web page for a presentation. Yeah, it was really nice and people could go to the presentation on their phones and on their laptop if they wanted to afterwards check it out if they had any questions. I also created like a when they logged into the web page

there was like this question thing there they can do. So enter questions the QR code. Manus uses HTML as well. Yeah I think so. So you could you could create like a nice presentation with Manus just deploy it and then you have it. But of course you don't have so much control of the deployed page on manus as you do if you host it yourself or through Versel or something. I just did it on Versell. Okay,

I see what you mean. your LinkedIn banner in HTML. Yeah, this took some time. I was going to turn this into a slide deck, but uh this was a lot of tokens. That's why we need open code with manus features. Yeah. Yeah, did you kind of dig deep into the code base on open code? So kind of the TUI is in Go, right? And um the tooling is in Typescript. Is that correct? I think so. So

I just the TUI I did was in Typescript I think and the tooling was in in um Go and the tooling was in Typescript. So that's pretty good for cloud code. It had no issues with uh the language. I guess there was a lot of compile there were some compile errors but not too bad. Okay. So let's see if I get my presentation now on the on the poly market from uh August. So I I took

a screenshot of this. I told Cloud code create a slide deck for a presentation about what happened here in o in August. So let's update this. I don't know if it's going to work though. Okay, but where's the next? Okay, so we have a slide deck. Uh, eight slides. AI model market dynamics. How Google went from 45 to 98% market share overnight. Okay. Next slide is a market in motion. So we can play the animation. Okay,

pretty cool. So here you can see the breakthrough is on this part. Next slide is phase one early competition. Phase two intensifying competition. It doesn't have any context though. So it's just taking what caused this shift major breakthrough psychology network effects. So it's basically just guessing here. Thank you. So I do have notes. Thank the audience for their attention. Okay. So it did write speaker notes for me here. So what am I going to say on

the first one? Welcome to this analysis of one of the most dramatic market shifts in AI prediction history. In August 25, we witnessed Google's market share explode from 45 to 98% in a single day. Okay. Yeah, pretty good. Pretty good. Uh that say whenever uh to change anything, look up the code base uh for the markdown for reference. Whenever to change anything, look up the code base. Markdown for reference. Then it goes and searches everything. Yeah.

So yeah, I would say this was pretty successful. I'm going to keep this repo. I'm going to delete this component though. So big. So I'm going to remove this three here. I don't know if Poly Market offers an API access to get real data. Yeah, I've done a project on that. I did a video on that or Sony, you can do that. No problem. It's free, too. doesn't cost anything. Just search uh poly market docs, go

to developer quick start, and here you have it. Uh you have some API rate limits, but I didn't have any issues. So for price and stuff, you can get 100 requests per 10 seconds. And you can also do I did web hooks. So I did a web hook on I did a web hook on a price on a market. So it's like real time. It's very good. A really good API. So definitely go check it out.

I think there are some pretty cool apps you can build on the Poly Market API, but uh I haven't looked into it in a few weeks. Now, my project was I would say it was a success. I did some trading using uh the LLM. I lost money, but that's I think I lost three bucks. Yeah, go try it. It's fun. It is fun. So yeah, I'm just going to keep an eye out uh if we get

any Gemini 3 news. So, um, haven't seen anything yet. Okay, so this is the nano banana model 2.5 flash image preview. Okay, that's pretty cool. Is it in AI Studio? Probably not. No, not that. So, I'm kind of thinking what I should do for the what I should do for the video on Gemini 3. I kind of need to think about how I should try it out. I think what I want to try out with Gemini

3 for a video is the Aentic Tulus. on Gemini CLI probably and kind of compare it. Maybe I should do Codeex uh versus Claude Code versus Gemini CLI uh with Gemini 3. We should build our own agentic benchmark. Yeah, if you have like a good way to do that. Uh I don't I never ever built a benchmark. I think I did maybe one time. Yeah, JK, you think that's a good idea for Gemini 3 test to

do Gemini Gemini CLI API with Gemini 3 versus GPT5 and codeex and claw with cloud sonet 4. I think I probably do that like Salesforce did. takes out the mundane and define the optimal tool use chain then compare it. Yeah. So, of course, I'm going to do the same prompt for every model, the same PRD, and just see what goes well, I guess. Okay. So, here's the animation. Here's the animation Claude created. I don't know if

you can see it. So, it says, "Thank you for tuning in to the live stream today, guys. Subscribe and hit the bell. Don't forget to like this stream, share it, and comments. Comment with your thoughts below. You don't need to do that. Uh but yeah, uh I got to go in 15 minutes, so I think I'm just going to end it. So, like I said, thank you for tuning into the live stream today, guys. Subscribe and

hit the bell. Don't forget to like this stream. So yeah, Edgar, uh, you know, LinkedIn if you have any news on the the project and yeah, Maple, thanks for tuning in. JK, Aruni, JD52, and everyone for tuning in as always. Very interesting conversations. Like I said, always a lot of smart people here. D Web, thank you for tuning in. Jamie, don't forget the GitHub. Okay, I'll check my email. Try to fix fix it up for you.

No problem. So, yeah, probably be back tomorrow trying out this built with Claude contest. That's kind of my plan. Might change it, but I think that's the idea. I'm going to try to do a submission to this. So, yeah. Have a great day and I'll see you again maybe tomorrow or later this week. Enjoy.