I'm in Danger - Ralph Wiggum in Claude Code

Speaker 0: So, the idea was that I've kind of I've done a lot of work with kinda trying to use agents and stuff, and cloud code to actually build out products and AI. And I've decided to go, like, AI, seeing all of the issues with just running it myself, providing prompts to the actual agents, AI of the base ones. What happened was it was like, okay. This is good, but it's also really bad. And then I took the next step, which was like, okay. Speaker 0: Let's try b mod. So you actually give it very structured inputs, and you have very structured, kind of agents executing things. And that was great, but you spend around 1 whole unit of, like, cloud code context just to set up, your PRD. And then AI that time, you've spent a bunch of money. So I was like, okay. Speaker 0: What's the next step of this? And then I saw that there is the so if people have seen there we go. So if people have seen the Ralph Wiggum kinda new new approach where you can use where you can guess Claude code to AI of break down a general PRD into small AI steps and keep running while it executes these by creating new agents and then killing them in between these things. Now, what I was trying to do is I was trying to take this 1 step further, emphasis on AI, because it kinda worked, and then it kinda didn't. And then I tried to kind of improve it a little bit, and then it completely failed. Speaker 0: So I have a feeling that the best way to do this is then 1 shot it again. So just clean everything out and just go do it from scratch. But do it AI. I probably could, but, AI, you can already see or MRF, the whole thing. But the thing is, it's AI, there were a number of learnings that I used in this over the space of the actual, kind of to implement this. Speaker 0: So starting off, AI Claude has a couple of really cool features. I don't know if people have tried them already, but plugins. I would highly recommend people to just go, if they use Claude code, do slash plugins and install all of these. If you're using it for anything work based and if you're, like, doing any sort of UI front end, get the AI kinda, like, plugin and just tell it, no. You're doing it wrong. Speaker 0: Keep, like, trying until you get it right. Now with this, what it does is you can see that it starts up the actual BMAD agent. So I tried to actually put on the BMAD approach, so breakthrough method for agile AI driven development. So it gives it a bit more, like, information on what each agent must do. And then on top of it, I tried to use the tune, so, token oriented object notation. Speaker 0: So it's a a little bit of an optimization on the usual, like, dot m d approach that you would kind of feed your agents. There I did a slight analysis, of how Claude actually expected the optimization would be, And it was only, like, $20 worth of, IDENTIFICATION, so there is further improvement to be done, or there's actual testing. Because whenever if anybody has done it and has asked it to estimate work, it gives tens of hours, and then you get it to run and it's, like, 30 like, not even 30 minutes, like, maybe 5 minute max, and it's done. And you're AI, okay. Take these kinds of things with a pinch of salt. Speaker 0: But, essentially, I tried to get it to kinda give these kinds of things and do all this. AI, tried, and got nowhere, but a lot of these learnings so the main ones are use the plugins. Also, a really cool plugin, actually, if you go into the AI, find there's, like, a marketplace of plugins that really there's there's, like, 1 1 provider that has, like, tens of thousands of stars. Click into their get their AI of like AWS slash cloud plug in. Really good, really helpful if you're, like, trying to deploy stuff or migrate stuff. Speaker 0: I've been doing it for my brother recently. He has a he has a start up, and he's like, yeah. I'm moving to this. And then I'm like, yeah. Let me just blast my €200 a month cloud code at it. Speaker 0: But yeah. It was 100 until yesterday when I was like, shit. I'm not getting anything done. But, yeah, that's it. Essentially, the moral of the story is, like, I called it I called it AI in danger because you AI of are in financial danger. Speaker 0: You will just increase those Tiernan until Tiernan being clogged tiers and also human tiers. But yeah. It's it's a cool experience. I'm gonna keep tinkering with it. Hopefully, I'll get it to work. Speaker 0: But if I don't, I don't. If I do, thank you. Thank you, Andre. Any questions for Andre, guys? Yes. Speaker 0: There you go. Speaker 1: Hello. Oh, it's so weird to hear your voice like this. Iman here. I just wanted to say, first of all, thank you, Alessandro, everyone that recognizes this. This is amazing. Speaker 1: I feel so, I don't know, in awe when I see you guys. Anyway, amazing stuff. So my question is actually very basic, but I was just curious when you start testing these agents, In which areas do you experiment these agents on and why? Speaker 0: Like, in which, sorry. Yeah. Speaker 1: So Sorry. I I build agents, and I work with agents not in this level of, AI not a developer. And whenever someone Callan Speaker 0: delinquency. Speaker 1: Yeah. Yeah. And when someone, talks about, AI, I build this agent, in which area which for what purpose and why would you test it on that first? Because if I get to start doing something like this, where would you recommend me starting? Speaker 0: I would so great question, actually. With agents, I wouldn't start from scratch. I would actually start on top of something that already creates the foundation. So something like BMA or something, I think the Clod have their own, like, agents.md. Like, starting off on that, you then just add skills. Speaker 0: AI, like, something that I've noticed in my own work is you have your QA agent, let's just say. So, like, you've written a bunch of code and you wanted to test it. You tell that agent that, okay, Run the QA. It does a load of stuff. It uses up a bunch of tokens, but it actually doesn't really do anything. Speaker 0: And it'll create a lot of, like, smoke tests and a lot of, like, liar tests. And then you're AI, no. Don't do that. Make sure that these tests are, like, true tests. And then it's like, okay. Speaker 0: Sorry. That is a great idea. Essentially lines. Yeah? You're absolutely right. Speaker 0: Yeah. You're absolutely right. I completely so starting with that, like, having those and then telling it to do stuff. So, like, you have your dev agent, something that, like, when I started my job, made a huge mess that was, like, doing a bunch of different like, dealing with a bunch of different bugs. And then once AI was like, okay. Speaker 0: Push this into push this into the repo and, create a pull request. It pushed the thing that I was working on, deleted the other task that I was working on, and then I'm like, well, class. This is amazing. So then what I did was, okay. How do I solve this? Speaker 0: Git work trees is a great way to actually overcome this, where you, like, you copy the code base every single time you're doing something new, and that way, like, the push is unique. So you kind of by messing around with the agent, you give it little, like, additional skills. It's like teaching anybody. But, like, to your original question, I'd say start with something like, you're not going to build out, like, if you're going with, like, development and stuff. You're not going to build out a website, like, just purely in Angular from the very start. Speaker 0: Like, go with React. And then, like, once you have React, then you can be like, okay. Does this actually do the job for me? If it does, perfect. If it doesn't, like, then you can have a look what do I actually need. Speaker 0: You can then look at these other agent AI of, like, frameworks, methodologies, but yeah.