The Missing Piece Is Ideal State

0:19S1:Hey. What's up? So Theo's got me triggered right now. He basically put out this video here talking about how. Yeah, turn off cloud code's memory. That's what the title of the video is. I want to let you hear the opening piece that kind of triggered me here. Made me want to do this video. So let's listen here real quick.

0:35S2:Work effectively in a code base. It needs to know what the code base is, where things are, and how to get stuff done. I found that a lot of people want to store this information in some special way that AI agents will operate with. Well, I've seen so many different systems and libraries, plugins and features and stuff that I have built over the last few years to try and automatically encode what they're doing in the code base in some special way. That'll make the agent always do what they want. I've even noticed that in cloud code recently, it seems to use its memory a whole bunch, where it will save things in some magic hidden file somewhere on the computer after you ask it to do something, and now it will keep randomly doing that forever. You can't tell from how I've been spending this so far.

0:57S2:I'm not a fan of these memory systems. I don't think memory is the right place to keep track of how things should be done in your code base. That worked okay.

1:03S1:I don't think memory is a good place to keep track of how things should be done in your code base. That's what he said. Okay, so here is my response to this. The primary cause of AI systems not making us happy, is it not knowing what we want? Okay, now he is objecting to a specific thing in a memory system, which is this automatic writing that cloud code is doing. So basically cloud code has a feature. If you turn it on where it will write to this memory system. I think it's literally memory.md unless they changed it and that memory MD would be consumed every time you start a session. So it's a cool idea, right?

1:47S1:But he's saying it decays over time. He's going to show you how it decays over time. And he's actually right about that. But he is still throwing out sort of this baby with the bathwater, basically still saying that memory systems are bad. I'm going to present an alternative view, which goes against both Theo and Matt Pocock on this, and kind of takes us in this direction of talking about bitter pill engineering, and what are we actually trying to do with a harness? So let me let me show you the architecture that I use, which is open source project. It's called life OS. This is the overall system that I use and I think it is superior to what Matt Pocock is doing and to what Theo are doing. Okay. And to be clear, they are both way, way better developers than me. Ten years ago, like I was not a developer, I was a programmer. And they were both hardcore real developers. So in terms of development skill, I think they are just way better. Okay, no question there. Unfortunately, I think they're still in that world of building development skills, specifically.

2:44S1:Matt Pocock. And I think Theo looked at his skills recently and wasn't a huge fan. But anyway, Matt Pocock is doing amazing stuff on the development side. Theo is obviously doing amazing stuff. He's actually my favorite AI creator right now. I watch like every video that he puts out, and that's why I was specifically triggered by this because I think he's getting this wrong. So the way that I approach this in life OS is I am specifically addressing the problem universally across the entire harness of one problem, the AI, not knowing exactly what would create euphoric surprise in my brain. That is the ultimate mission. It's actually in our system prompt is to pursue what would make euphoric surprise in my brain. Okay, that is the meta goal of all the entire harness. Okay, so in order to do that, guess what you need, you need extremely detailed knowledge of exactly what I want. Okay. And this is where I think Theo is getting confused and he's conflating multiple things because his video is basically saying memory systems are bad. Okay? That's ultimately what he's talking about, but he's kind of throwing out the whole thing with

3:48S1:this specific problem that he has, which is automatic systems that write crappy memory files are bad. Well, of course we agree with that, right? So let me show you an alternative. Okay. Have the goal of the harness be to produce euphoric surprise in you, the user, the principal. Okay. Second, I believe and this is a strong claim, I believe that the way that we've been doing development before, and the way that Matt Pocock is currently doing it with all these extra skills, and you've got to call the extra skill and you're producing these test documents you're producing, these plans and these specs and then a PRD, and then you've got big documents and you turn it into small documents. And then the Ralph Loop actually, where you take one agent and you work on one particular feature.

4:29S1:In my opinion, these fail for one specific reason. They don't have the plot. When you break things apart, you tear them apart. They cannot see the plot. The plot is, in my opinion, for all of AI, the ideal state. Okay. You cannot achieve euphoric surprise. Your harness cannot achieve euphoric surprise for you if it doesn't know what you want. Okay. And here's the trick. I could be asking for a beautiful picture of a watercolor butterfly. Or I could be asking for come up with an amazing business plan and idea that's going to make me tons of money. Or I could be asking for build me a beautiful website that talks about typography. All of those things can be articulated in one sentence or two books. Okay. How much detail is in my mind for what a perfect typography site looks like, how much detail is in my mind, and how much have I given the AI. The gap between these two is the problem with all AI harnesses.

5:26S1:It is the problem with all agentic work. It is the problem with all AI, prompt engineering, intent, engineering, graph, all this stuff. It's all floating around. It's all orbiting this core concept. This is my argument to you. Doesn't mean I'm right, doesn't mean you have to agree with me. This is my argument. I think I am right, okay? It is the ideal state. What we do in life OS is the ideal state artifact one document. Now, if you have some giant system, you're building an entire role playing game. You can have one Isa that links to other ISAs and they're all cross-linked. Like if you've got a full geography system and a language system over here and like creating NPCs over here, like you could break it apart if it's big enough. But for all intents and purposes, that is still one giant Isa document. Now, what a single document allows you to do is what we have in our document. And I'm going to go into detail in another video where I take you through the code version of this, I want to show you the architecture first. So this architecture basically says, here is the stated goal. Here's the stated ideal state. We also

6:26S1:have a section where you can just dump a full text conversation. I can actually call Chi and say, hey, I want to make this thing. I literally call him on the phone and he answers and we talk back and forth. It's absolutely insane. But you could also use the discuss skill, which is similar to Matt Pocock's really excellent Wayfinder and grill me skills. So you're gathering information, but what he's doing is he's kind of putting it in all these various places and he's breaking it into pieces and recombining it. It's like it's not one unified thing that he is enhancing with this skill. And that's the difference. That's the difference. When I do discuss or I call Chi or we text back and forth or whatever, whatever the interface is, we are updating, we are filling in, we are enhancing, we are improving the quality of the description of what perfect state looks like, of what ideal state looks like for this website, or this local application, or this mobile app, or whatever it is, or a picture of a butterfly inside of here. In the ideal state, we have these. These are individual testable claims. Boolean. Is

7:26S1:this true or not? Along with a testing function that can make sure it's real. So this is the beauty of this thing. The ideal state describes what you're trying to build. It also has the specific criteria for building it. Okay, here's exactly how you build this. It must do this. This form must do this. It must be secure. It must submit to the following location. It must only accept these inputs. It then updates this database. Right. So very specific, very clear things just like we've been doing with spec driven development and test driven development for all this time. Right?

7:54S1:We're bringing it over into this larger sort of master concept. So it is full of ideal state criteria which are inside the ideal state artifact, the ideal state criteria. This is insane. They are the build criteria. They are also the testing harness. The ideal state artifact is the testing harness as well. When I have a production app, I have surface the surface.ai. I have human 3.ai. Both two products that I have that have users that are out there doing real things. In the private repo, there is the ISO document. If I get rid of my entire system and come back and say, okay, I've been whatever meditating for six months and I come back and I've got a brand new MacBook, I set up life OS or whatever. I say, pull up the Isa for surface.

8:39S1:I don't say, hey, go find every single session where we talked about surface. I don't say, hey, go find all these different random documents spread all over the place, plan file spec files, every conversation we ever had, all the places where I corrected you and fixed previous errors that we had in the Isa and went back and forth. I don't have to say go and collect all of that. It is in one place. Now you might be thinking, okay, yeah, but this document is going to get really huge. Well not really. That is getting easier and easier to pass.

9:07S1:First of all, for AI, as the context windows get better as we get more efficient, but also the name of the game for life OS is rip grep and JQ right. We find the things we need in the document. We don't have to parse the whole thing every single time we have sections in there, you can go and pull out the appropriate section. And most importantly, the work being done by all these agents. It's all versioned, it's all in GitHub, it's all committed. You can roll back and you're also updating the changelog, the list of changes that have happened to the OS or to the ISO document for this particular task. And it's all being updated.

9:40S1:The history of decisions is all there. It is one single document. So I could just pull this up for a production app, and I have the full thing to continue working on to add new things to. If I want to add a feature, guess what? It gets added as a feature. We add additional ideal state criteria. And those are the things that get worked on. And those are the things that get tested to make sure it's actually functioning. So that is the ideal state system. My argument is that it is the way things are going to go. I said late last year that this is the way things are going to go. It hasn't happened yet.

10:09S1:I thought it was going to happen soon. I thought people like Theo and people like Matt and other people building stuff like me on YouTube would already be talking about this and the fact that we have to unify, and the fact that all of AI's problem is the gap between what's in our brain and what we want versus what the AI knows, right? Another way to think about this. It's called writer's block or writer's blindness or expert blindness, where you try to explain something, you try to write a book or something, and it's just a garbage book because you haven't explained actually the stuff that you know. Right? So that's a good reason to have an editor because they can actually pull that out and say, you need to explain this more. This is not obvious, right? Well, the ideal state document and the discuss skill is designed to fill that in. Right. And it actually triggers now at this point, if you don't have enough detail, it actually triggers discuss and has a conversation with you back and forth, which then fills in the ice more and more.

10:59S1:So central concept, you have to synchronize your brain and your AI harness across the entire system. Okay. And let's talk about skills next because that's also where this is. So I have a skill system. It's a custom skill system. It's based on cloud code, but it's a specific skill system that has three main components. So the skill.md is basically the overview, and it's got a router in there similar to our cloud.md. It's got a router for workflows.

11:26S1:Workflows are specific prompts, specific tasks that get done within the skill. So the workflows are write the blog post, edit the blog post, publish the blog post, scan the blog post for errors. I have one specifically, it's called enrich. So that goes and takes all my previous content, finds out if it's relevant to this area, and it adds a link for where I said those three words. It links to a previous thing. So I probably have like 15 different workflows inside my blogging skill. So the concept is I want to blog workflows are these 15 things.

11:56S1:And then the third piece, which is pretty unique here is as much as possible is turned into CLI tools, deterministic code, direct API calls. I use very few MC because I've reverse engineered the MC into CLI tools. So the workflows call the code. The code executes, hopefully doing as much as possible, and the concept calls the workflows, right? So I just say, hey, I want to blog about this and it goes and builds a blog. Now let me show you something. This is my site here. Sorry about the flash bang. It's not quite a flash bang because it's not exactly white, it's sepia. But this here is my blog and I want to scroll down here.

12:32S1:Look at this latest content. Do you notice anything about these images? Look at that. Not half bad. Okay, notice that it's kind of sepia. Got a little bit of purple here. Got the dark charcoal. Oh, look at this one. Really cool. Do you notice that the concept of the art matches the content? All right. So I just went and edited this. I actually used this device and which is recording. I can just go for a walk and talk to this device. Or I could call Kai on the phone and say, hey, here's an idea. And I just ramble for the whole one hour walk. I'm just rambling about text and I then say, go write a blog post. Well, I didn't make any of these prompts.

13:13S1:Another crazy experience with AI. My AI found what was killing my Mac and helped me replace two apps in about an hour. Okay, that is talking about specifically this app right here, which we made together. Well, really, I made it for me, but I gave the requirements and guess what? It built an Isa. There is an Isa for actually the UL bar app. Anyway. Diversion there. This piece of art was made from the content that I did. All of that is in my blogging skill. My blogging skill only has it in one place, which is in the blogging skill. Every single thing that I love that I. Okay.

13:48S1:My typography. You have no idea how many hours I've been blogging since 99. This system is perfectly optimized to be exactly the perfect blogging system that I like the way this subtitle is done, the way the title is done, this little marker right here. The entire thing is perfectly in my ideal state, which is articulated perfectly in the blogging skill. So this is where I'm disconnecting from. Theo. Okay, he's saying memory systems are bad. I'm saying that the entire harness. This is all text files, right? What's the difference between a system prompt and a cloud.md file and your skill files, which is also a markdown file, the main skill.md and your prompts. Okay. The code is quite different because that's deterministic in code, but ultimately the entire harness is a whole bunch of guidance. It's a whole bunch of stewarding and shaping and adjusting to get things the way that you want them. Right? So all I'm doing is I'm making that extremely explicit. So this entire thing, I literally just talked now and I say, go make a blog about this and it goes and makes a blog about it. And this technical diagram right here is

14:55S1:created by my technical diagram workflow inside my art skill. So look at this. I didn't link this chi link this for me by following the skill which perfectly articulates my ideal state for a blog post. So all of this is all the same idea. Okay? It is. How do I get to the perfect blog? How do I get to the perfect blog post? The perfect design? Okay, same if I'm building an app. Same if I'm building a new platform, a new security tool, I have ways that I want it to be done perfectly. And that is what I articulate inside of the skill system. So I wouldn't put that in the system prompt. I wouldn't put it in the cloud.md file. I put it in the skill system because it is a particular task. It's a specific task. We are blogging. I've got other ones for making outbound voice calls, so I can call and schedule a restaurant for me and put it on my calendar.

15:48S1:So it's got a calendar thing. It's got an email thing, it's got an outbound call, it's got blogging, and I've got about 150 skills, small specific skills that articulate how to do this perfectly according to what I consider to be perfect. So this is why I disagree with what Theo is saying about memory systems. He's thinking of it as a whole bunch of like little tiny mini rules. No, no, the memory system is the cohesive system that merges your brain with your harness. That's what it is. Okay. In my opinion. And the whole centerpiece of the entire thing is this ideal state artifact. It is the fact that for anything we are building, we can always come back.

16:27S1:Let me give you another example. Let's say that we are building as I am with a buddy of mine named Andrew. Building kind of like a gaming system, a system of building systems, actually of building other gaming systems. Well, when I go to do that, I can't remember everything all at once. When I go on my first walk, I'm talking about this other area. And I actually have tons of detail about this other area, about the integrity of the combat system and the simulation system to be able to simulate auto battlers or tank tower defense like type of games or whatever. Right. Well, I can't remember all of that at any given time. In fact, someone asked me, oh, what makes your gaming system better? In fact, people ask me what makes life OS better? I can't articulate all of this, which is why I'm doing this video.

17:09S1:I'm trying to articulate all at once cohesively, right? The point is, at any given time you might not remember. You yourself cannot remember because the idea is so big and it has so many components and there's so much detail. Well, where are you going to put that? Where are you going to put that stuff? Are you going to go and gather prompts and talk to your AI? And it's going to put it in some session file. No, because that will be lost and it will not be unified. Right.

17:32S1:So the purpose of this discussed thing is to keep adding pieces, as you remember, as something breaks, as you have a new idea, as you realize one of your previous ideas is total garbage. You're like, hey, remember, we're going to do that combat integrity system. We're going to use this method. Yeah, I watched a bunch of videos. It's total garbage. Let's do this instead. Boom goes and cleans it up. The ideal state is now improved, not in some random location in the actual ideal state artifact, right.

17:57S1:The authoritative artifact. All right, so that's the skill system. That's the ideal state artifact system. This is basically the Kanban board that shows the work that's actually happening. And the algorithm is really just this hill climbing system. It's also got a model router that switches between like fable GPT five six, soul GB. Five six Cyber or Daybreak for all my offensive security stuff. It actually uses a whole separate agent. Actually, I could show you that real quick. Yeah.

18:24S1:So down here, if you look at this haiku, sonnet, opus, soul, Gemini, grok, cyber and Max and Forge and Helios. So basically Helios is his own agent that does only cyber stuff anyway. Not important. Not sure why I showed you that. All right. So going back to this. Oh yeah. Evals. That's the other really cool part of skills. But we can talk about that later. This basically shows all the different content that's being executed. I think I showed you that earlier or it's in another video. You should watch that as well.

18:50S1:But the algorithm is just how you're actually moving through the ideal state. Which models are you using, which intelligence level are you using? And it actually pivots up and down based on the work task. So if you shouldn't be using fable for this, switch out to soul or switch out to haiku in some cases, although not very often, or sonnet or opus or whatever. So ultimately the system is euphoric, surprise and ideal state. The system prompt is just like super minimalist. This is just a context routing system, and the skill system is quite unique and I think really helps with the determinism and basically making it more dependable to actually pursue your ideal state as things are being executed. And the algorithm is kind of just like managing the whole thing, managing the general hill, climbing from current state to ideal state for anything you're trying to accomplish.

19:37S1:So this is what I want to say, just responding to Theo here again. I think me and Theo and Matt also are going to agree about this, I think, in the future. And I could be wrong about certain pieces about this as well, right? I've incorporated tons of stuff, for example from Wayfinder or from Grill me or some of his other skills. I've incorporated tons of stuff into life OS from Theo, from all the videos I've watched. Mad respect for both of them. I just think that this concept is so fundamental and first principles based, that I think everyone's going to be moving in this direction.

20:07S1:And if you like it, if you try it, if you hate it, if you absolutely love it, either way, just comment and we'll see you in the next one.