Anti-AI Developer Goes Agentic

Added: 2026-06-01

[00:00:00]Okay, so we have default, which is Sonnet 4.6, which was another error that I made in my last video. I believe I was just using 4.6 for everything.

[00:00:10]So, it's come to my attention after posting my last video on reviewing shitty agentic code that people had some legitimate suggestions for me to actually fix my AI development flow.

[00:00:27]So, instead of resisting that, which we should all do as humans in general, and especially software engineers, is to take feedback.

[00:00:38]Don't take it as an Don't take offense to feedback, but actually try to apply that. If someone is saying something and you and you believe that they know what they're doing, you should take that feedback, go give it a shot, or at least investigate what they're saying, and it could really help you. It could be something that could get you to the next level of a subject that maybe you were weaker in, or something that you didn't know about. So, let me just read a uh comment a couple comments here. I was reading through this morning, actually, and one of them had some great advice, and I I wanted to sit down for an hour session just to see how much knowledge I can pack in, and just reassess this whole food truck pro- project that I was building with the shitty agentic code.

[00:01:31]I'm going to try to redo the project, but I'm going to try to redo it with the knowledge that I gained from the recommendations that other engineers have given me in my comment section.

[00:01:43]So, we're going to start with an hour today just to see how much we can learn for this video.

[00:01:49]And we're going to learn a whole lot more and spend a whole whole lot more time in future trying to get this thing right, because if this is the future, we need to know how to use this technology.

[00:02:02]So, I don't even know how to say this guy's name. I'll include it in the video up here.

[00:02:08]But, this guy says, "I don't intend on sounding rude or insulting, just my opinion."

[00:02:14]After watching through, I heard mostly nitpicks. This is true. I was very opinionated, which engineers can be very opinionated about code and system architecture, design, everything. If you've ever worked in a a role with other engineers, you know that some engineers have hills that they die on, and they will argue and fight to the death to get their point across. So, this is something that's definitely true.

[00:02:43]Except for localized type, I also sense frustration, whether that be job market expectations not being met met. But, seriously, please ignore the internet hype. LLMs are tool, and that's all. I really would like to believe this. I would like to believe that they're a tool that maybe I'm just using improperly.

[00:03:00]The Tailwind point you made, you don't have to use it. On second thought, it's likely used the Next.js script to bootstrap the project. What he's What he's saying is essentially since it was using Next, and Next by default essentially uses Tailwind CSS. So, that's why it wasn't really the agent's fault that it chose to use Tailwind. And I was upset that it had chosen Tailwind as its default and didn't choose something else that I liked more in CSS when in reality, it was never the agent's fault. So, that was human error on my end.

[00:03:36]The generics versus casting point, I question your experience here. You're opting in unsafe JS because it reads better.

[00:03:43]Often times, I do opt for readability because readability for me is especially important for somebody who really really kind of uh to connect syntax together and has to reference document documentation a lot. I I I like things to be readable almost in human just like natural language readable with my code. Now, I understand everything cannot be done that way and there are shorthand tricks in coding that you know, make our job a lot more quick and simpler.

[00:04:18]Uh I think maybe I just need to glance over generic TypeScript knowledge again and get used to using it in different formats. So, I think I'll just re- reassess my whole knowledge base on this topic.

[00:04:34]Try to get better at that.

[00:04:37]Local types point, I agree. Though, that can be solved with context generally.

[00:04:42]You can either scaffold the project before LLMs come in or write a convention/rule section. This is what I really want to spend probably most of my time today on is trying to figure out how to write a a good spec because I I'm starting to understand that the spec is extremely important and spec needs to be extremely detailed for the agent to actually produce a result that is going to be somewhere at least I would hope in in the 80 to 95% range of accuracy.

[00:05:19]We're going to try to figure that out and you're going to watch me literally just investigating Google through AI, whatever it is to try to figure out the best way to start making better specs and plans because this is probably the most important part of agentic coding.

[00:05:36]You made a brief comment about trusting the generated code, which I was hoping you would expand and give your take. I think this is the biggest issue for us alongside of Maybe I'll expand more on that later.

[00:05:49]I'd recommend trying spec-driven.

[00:05:52]I personally found success using open spec and cloud code. I will definitely be looking at those today.

[00:06:00]It's possible your understanding of transformer models is surface at best.

[00:06:04]AI engineering by Chip Huen is a solid overview. This is a book, I believe, so maybe something that I can look into, not today, but later.

[00:06:15]Or when I'm actually having some downtime to read and learn better about how these models work. So, this was a great comment here. And I trust this person because they actually know what they're talking about, and they're giving me concrete answers. They're not giving me surface level understanding. They're actually giving me a deeper knowledge, and actually every point they made was was true. And I was not correct in a lot of areas, and nor am I correct most of the time.

[00:06:46]Um but sometimes I am correct a lot of the time, and that's just the nature of being human. So, let's take what is called constructive criticism, and not get angry about it, but actually try to learn something. So, this was just one comment that was super helpful. Let me just read one more so we can get an idea of what else we can go over.

[00:07:10]Um someone says the about the ternary operators, "I don't see a problem with it. It's a preference. Ternary operators are even recommended in React documentation."

[00:07:20]Fair. I just hate the way they look, and I've never been able to mentally map them quickly, which is why I hate them.

[00:07:27]Um let's see what else we can find.

[00:07:33]The time you waste crying about everything could be used to design your context. We all know what AI can and cannot do yet at this point, but one thing is clear, the output you get is only as valuable as how you organize your ideas in your mind.

[00:07:48]I I agree. I agree that I think my planning is a bit weak.

[00:07:56]A lot of people suggesting Codex.

[00:08:00]Um, somebody I can't find the comment was suggesting some kind of I think it was Codex and I think it was a store.

[00:08:08]Microsoft store.

[00:08:10]And let's see if we can just find Codex.

[00:08:15]Codex app.

[00:08:18]Open AI. Let's get that.

[00:08:20]I don't know if this is paid product or not, but let's see what we can do.

[00:08:30]Could have went with Codex. Claude code is not even close. Speaking to both the Codex app as a harness and GPT-5 as a model, 5.5 as a model. Literally all the issues you ran into would have never happened in Codex.

[00:08:42]Please download it or PS download the Codex Windows app.

[00:08:47]So, I'm getting that. I'm going to investigate that later. Right now, I'm going to focus on Claude code CLI because that's what I've been working with and I want to see if I can make that whole model and planning situation work better so that when I start to build this app, maybe I get a better result. So, anyways, um without waiting around too much longer, I'm going to set a time for 1 hour and I'm just going to put my headphones on here and start researching, practicing things, and maybe if I find something interesting, I'll pop up every once in a while, but a lot of this is probably going to be sped up to the extent where you can see a little bit of what I'm doing and then uh you can go through my learning process. So, let's go ahead and get started.

[00:09:46]>> [music] >> There's something that I really think AI excels in is whenever you have good context of what you want, especially grunt work.

[00:09:57]For example, I just copied this comment and I'm pasting it into Gemini chat and I want it to set me and make me a set of bullet points. Normally, I would maybe type this out in the past, pre-AI, and figure out which points are valuable, but I find the AI really extracts good information out of context that we provide it. So, it's going to make me a list, I hope. Let's see.

[00:10:19]So, I can study.

[00:10:38]>> [music] [music] >> Essentially, I'm just deleting anything that I had related to cloud cuz I kind of whenever I'm learning, I want a fresh start on working with a technology to understand exactly how it works without getting any errors from previous investigations and whatnot. So, this is personal personal preference. So.

[00:11:18]Also, what I'm noticing I've just read a little bit through these cloud code docs is that context is extremely important. And I'm reading about this agentic loop model and how cloud code works and essentially how Claude code is a harness around all the Claude models which includes the actual model doing the research and also tools to allow Claude to do different things like fix bugs or refactor or these kind of things. So this is kind of a simple little graphic here that I was looking at and the first step in actually using [music] these models and these these agents is gathering context for them. So after they gather context, that's when they take action and verify. So if you your plan is essentially your foundation and if your plan is then your the output of the code is probably also going to be really too.

[00:12:26]a Claude Yes, does this generate a Claude directory within here?

[00:12:38]I'm not sure, but let's look at model.

[00:12:42]Okay, so we have default which is Sonnet 4.6.

[00:12:47]Um Opus, which was another error that I made in my last video. I believe I was just using 4.6 for everything, which is probably why it turned out how it did.

[00:13:00]So I'm really curious to see how good I can get the quality of the code to be um after doing this. So Opus and Haiku. So what did it say was actually the best for code?

[00:13:14]Uh was that here?

[00:13:41]>> [music] >> Those are just little nuggets that this documentation is actually great. It takes you on kind of like one of those web adventures almost where one link takes you to the next to the next to the next. So, I started at at this doc right here and then I ended up going to these other ones. So, good stuff.

[00:15:15]>> [music]