Stop Making Your AI Dumber With Context

We're stepping into a new era of software development where the latest models are no longer relevant. And how we get ahead is the way we orchestrate AI. And it all begins with the unlikely Ralph Wigum, the famously clueless character from the Simpsons. It all starts with Jeffrey Huntley here in this very, very trippy, very strange video where he, which by the way, neither of us is going to do this concept any justice right now. We're going to do our best, but if you want to dig into it, you really need to go in and watch the source or at least learn about it from the source. What is Ralph? So, Ralph is a new coding technique, an infinite loop until it gets the job done. It solves the problem of like context. So, the problem is is that when you're in claw code and you're building your feature, builds up its to-dos. You know, you might use taskmaster mcp or whatever, but it's building up that context. With every message, you're adding the previous messages to the bloat of that context. What Ralph does, it sits outside and around Claude code and actually gives Claude code the prompt um with a um fresh context. Every everything is brand new. and then it writes the results of that into an external file to keep track of what's been done. And so with every single prompt, it's brand new because AIS will just get dumber the more and more context you add to it. Now, I wanted to show you the actual code. Let's pull out some highlights here. So, here's the PRD. So, you you you create your PRD. What is the overall task at hand? What are we trying to achieve? Um it looks like he's got a progress script here. Um archives and some various things like that. So he's checking for some things. You can see this is Python probably uh bash script and I think he's doing some checks here. Dates checking folders. I want to get to the the juicy bit. So here's a for loop here. Everything before that is just checking for files whether they exist and things like that. So you don't have to worry about that too much. This will this bit of code will run infinitely or at least in the max iterations [music] and it will take the prompt into the AI and take the output and put it back into the external file and once it's done it will call this kind of thing Ralph completed all the tasks. AMP is probably the tool he's using. So I would probably change that with Claude. I would probably add Claude here, not um I don't I don't know what AMP is to [music] be honest. And you run this on your your your Mac will run this code and the script or the prompt will be just stored in a in an MD file in a separate file and that will it what they've done they've piped that into so cat gives you the output of this file. You pipe that into this amp tool or whatever. I don't know what T is, but I think it's I think that might be just returning the response. So, you're getting the response from the prompt. Yeah. And then you're putting that into an MD file and then pumping it back into a fresh, brand new prompt for that AI. Every single time. Every single time. And it And the beauty of this as well, you're giving Ralph autonomy here to decide what the next task should be, or the next appropriate task should be. So it orchestrates itself, outputs that to an MD file and then the next agent will take that and then run with it and and keep doing that and [snorts] it will it will do things one after the other based on a fresh context, a fresh perspective and all that sort of lot. So there will never be crossover and you in theory could just run this infinitely. No task is big or too big or too small for this. Maybe if you know exactly what you want to do and you could get it done in like a quick little bit of context, then whatever. But if if it's a challenging bit of code or a challenging feature or something like that, get Ralph on it so it doesn't get lost cuz that's the problem. There are some agents that get out there in a bit of a loop. And a word on the model and reason why I say models matter less is that people are having great luck using traditionally not so great models on this because the the context is fresh and it just redefineses how we use how we use AI. Like Google will give you the million context window but truthfully that doesn't really solve any problems cuz the more you add to a context the dumber your AI gets. is this has just got fresh context, bit of prompt, bit of output from the previous agent that's run along with it and it's given to you. You mentioned that you're already kind of doing something like this. Yeah, this is interesting because the the moment you mentioned it, I was like so what I do is in cursor I have multi and I I might be dumb for doing it this way. I haven't watched any tutorials on it. So the way I do it, I have multiple chats. So let's say one chat is specific to either a file or something in the same context. So let's say in one chat I change the styles and in another chat I add uh animations and in another chat I have like the base the core like often uh let's say animation engine. So I create like a gap complex function and I have one chat specific to that one and then another one kind of like taps into whatever we are creating in that file to to to do like other things and that's that's how I manage my chats and context because I notice like if I'm if I'm like one second building like the the the whole thing and then I'm asking about a style change and then I'm asking about pushing it to git or something or then I'm having a question to learn something. It it mixes them and sometimes I try to do something and it does one of the previous things because it's mixing them. That's why I decided to separate them and I by the way I do typically open a chat about learning like things that I want to learn. So it's just the chat mode agent mode. Now I take all of those and also mix it with an MD file. just a file where sometimes I ask the the the model to to to say like based on the mistakes you did, this is what I say. Based on the mistakes that you've you did and based on like my preferences or like the things we've been through, based on the the battle we've been through, create a file uh and add all your learnings so I can give it to another AI model. And it typically does it in a way that another AI model um can benefit from it which is also sometimes fun to read because yeah it it tries to not hallucinate. It tries to you know not sound like AI if it if it has copyrightiting in it and things like this. So that's how I do it. So the the MD file for me is kind of like memory. It it's exactly well the chat is memory but the chat is messy and big and it has like a bunch of stuff and it has issues like it has the mistakes that it did but that file is kind of like um the the best memory like the best learnings and the best things that it did and like the notto-dos. I guess I could add these also on the cursor level or like on the project level, but I [music] come up with these as I go, you know, it it's not something that I that I have prior to the project. This is very interesting and I I totally recommend that everyone plays around with Ralph to see if it actually has an impact on getting your features built. It's there's nothing, like I said, there's nothing big or there's nothing too big or too small for for Ralph to run. The important thing here which underpins this whole conversation isn't Ralph. It's the fact that we are now entering this new age where we were once fussing around should it be cla what model should we use? All of this conversation is now behind us. We're now learning what it takes or what it means to be a software developer in 2026 using AI and it's becoming less and less important about what about the code that you write. You will no longer be hired for the code that you write because the code that you write is is a commodity. Now the important aspect is how you scrub up and how you be a better orchestrator, a better engineer to understand how all of these puzzle pieces fit together. And this is so crucial and it's something I love because I get to entertain my UX hat as well as my um uh DevOps hat, how I architect these. I'm not the best. I'm not going to claim to be the best, but I love playing in this world where it's not about whether I can or cannot write a function to do something. It's whether I know that this function fits or sorry, this feature fits in the grand scheme of this project or this tool that fits the users and the demands of the client. And that's where that's where software development is. Now, I mentioned last week that I don't think we know what we're building. don't know what we will be building in in a couple of months. And this further emphasizes the fact that you know you as an engineer again just will not be hired as a as a coder. Your role will exist. You'll be paid to know how everything fits together. Um so gone are the days of of counting lines of code written because it's just not important anymore. Um it's still it's still it's still a troubling time because this does remove the need of potentially several several developers costing a lot of money because if you've got one developer who can just do all this because you don't you know instead of writing six feature each individual developer writing six features you've got one developer writing all six on a Ralph Wigum loop which by the way can iterate through different tasks different feature sets one after the other instead of parallel like you would traditionally do it. So, this is why this is important. Who knows what the next tool might be? You mentioned a character from Rick and Morty. Oh, yeah. Mr. M6. Anyone know about that? Like that could be the next one because you said go on explain. Yeah. So, with Mr. Misix, it's so fun. Like, um Rick does this like it's it's a character to help. It's kind of an assistant. And you you you ask Mr. me seeks to do something, it comes to existence, it does it, and the moment the task is done, poof, he's gone. He doesn't exist anymore. Uh, now that the [snorts] whole joke and thing is about what if you give it a task that it can't do. And that is where it starts to ask for another Mr. M6 to help him and another Mr. M6 to help him. And that is kind of like the sub agents that we do. Um but it is very interesting because it reminded me of that because it it will also try it try until the job is done and then you know next Mr. Mi6 also is kind of like fresh context. Exactly. So anyone who knows it comment. [laughter] Yeah. Yeah. Like a loser. [laughter] But you know, jokes aside, like that could be the next iteration of this. So, I'm digging massively into claw code right now and I'm looking into some of the features and I've been learning that uh slash commands which are shortcuts um can be used by sub agents. So you can tell you can say once you've done this use the slash prit you know get commit slash command or review slash whatever you can tell it to do the slash command. So there's this um there's this inner task that's running that it can run it has an agent has access to things like that. I'm also seeing here that he's got skills. Now, skills are another Claude code uh thing. It could be that the next iteration of a W Ralph Wigum thing is someone's figured out a way that it's a it does spin off sub agents and it does ask for help or something like that. I don't know. I'm just I'm just spitballing. But we're we're this is the new era that we're going in and Ralph just might be the start of it. And I compared it similarly to Deepseek when Deepseek first came out. Deepseek was not the smartest best model in the world. That's not why it made waves. The reason why it made waves is because China proved that you can train an AI for a fraction of the cost that all the big companies were doing. Um, and it be a pretty good pretty okay model. And that's why it shook things up. Um, we're seeing seismic shifts in some of the undertones that are happening, not not in um not in the most obvious ways. So, that's Ralph Wigum in a nutshell. We'll leave links for everything, you know, down below once the video is out and stuff. But yeah, pretty cool. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.