22:03Jake Van Clief · Talking Head
Folders Over Agents: The AI Layer Nobody Teaches
A 22-minute argument that context architecture — not agents, not prompts, not frameworks — is the only AI layer that survives the next model update.
May 29thThe argument in one line.
A three-layer folder system using CLAUDE.md, workspace CONTEXT.md files, and staged skills does everything a custom AI agent does, in plain English, with no code and no framework overhead.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code regularly and keep hitting token limits or losing context between sessions.
- You want structured AI workflows without writing Python, building agents, or managing databases.
- You switch between multiple types of work — writing, production, distribution — and want the AI to only read what is relevant each time.
- You are a freelancer or solo creator who wants a repeatable system that adapts to any project type.
SKIP IF…
- You are running a mature multi-agent system with durable state and need architectural guidance beyond folder organization.
- Your workflow depends on real-time data retrieval — this system uses static markdown files, not live databases.
The full version, fast.
Most people dump everything into one giant prompt and wonder why the AI loses the thread. This video introduces a workspace blueprint built from folders and markdown files alone: a top-level CLAUDE.md acts as a floor plan the AI always reads; workspace-level CONTEXT.md files tell it what to load for each task type; and skills or MCP servers are wired in per stage rather than loaded globally. The result is token-efficient, editable by hand, and adaptable to any profession without writing a single line of code.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:15
01 · What this video covers
Intro: folder-as-workspace concept, template walkthrough promised
01:15 – 03:30
02 · The problem with how most people use AI
Chat loop, token walls, no persistence, not scalable
03:30 – 05:45
03 · Tokens and context windows
Token explainer, context window as finite resource, burning tokens on irrelevant files
05:45 – 07:00
04 · Workspace blueprint: three areas
Community, Production, Writing Room — each handles a different kind of work
07:00 – 08:30
05 · VS Code and markdown explained
IDE as a cleaner folder navigator; markdown is just readable plain text
08:30 – 10:45
06 · The CLAUDE.md file: Layer 1
Always-loaded floor plan; AI reads it automatically every session
10:45 – 12:30
07 · The three-layer routing system
Map (CLAUDE.md) to Rooms (CONTEXT.md) to Workspace — each narrows context further
12:30 – 14:00
08 · The routing table: the most important pattern
Markdown table specifying read/skip/skills per task — eliminates token waste and wrong guesses
14:00 – 15:30
09 · Production workspace deep-dive
Four-stage pipeline: brief to spec to build to output; each stage loads only its own context
15:30 – 17:00
10 · Skills vs. a system
Skills wired into a layered pipeline outperform skills loaded in isolation
17:00 – 19:00
11 · Naming conventions as a database
Consistent file naming lets AI locate outputs by description — no SQL or vector DB needed
19:00 – 20:30
12 · The folder IS the app
Zero code, zero framework; the folder is the UI; voice control is the next layer
20:30 – 22:00
13 · How to make this yours
Template adapts to any role — creator, freelancer, developer — by swapping layer content
22:00 – 23:18
14 · Research paper and CTA
System grounded in 200 years of software engineering history; template available to paid community members
Lines worth screenshotting.
- Dumping everything into one prompt wastes tokens because AI reads files irrelevant to the current task.
- A CLAUDE.md file at the top level acts as a persistent floor plan — the AI reads it every session without being asked.
- The most important pattern is a markdown routing table: for each task, specify which files to read, skip, and which skills to activate.
- Skills work best wired into a layered architecture, not loaded in isolation — their position inside a thought process is what makes them effective.
- Consistent file naming conventions can fully replace vector databases or SQL lookups for most individual and small-team workflows.
- Three-layer routing mirrors function-call routing from traditional software engineering, now written in plain English anyone can edit.
- Running multiple Claude Code instances in parallel against different workspaces is more practical than building separate agents for each task type.
- The folder itself becomes the UI — no apps, no dashboards, no frameworks required.
- Context windows are finite; routing AI to only the relevant files is the primary lever for keeping sessions under the token limit.
- Skills and MCP servers are more effective when wired into a specific pipeline stage than when loaded globally at all times.
- Workspace templates adapt to any profession by swapping the content of each layer while keeping the three-layer routing structure constant.
- About 84% of people are not yet using AI regularly, meaning the competitive window for workflow fluency is still wide open.
Three routing layers make a folder smarter than a chat window.
WHAT TO LEARN
The reason AI loses context is token waste — reading irrelevant files — and a three-layer folder architecture is the structural fix.
- Separating work into discrete folders with their own CONTEXT.md files means the AI only loads what the current task actually requires, preserving context window capacity.
- A CLAUDE.md file at the workspace root acts as a persistent floor plan — it loads automatically every session and tells the AI where everything lives and how files are named.
- The most important pattern in the system is a markdown routing table: for each task type, specify which files to read, which to skip, and which skills to activate.
- Skills and MCP servers are more effective when wired into a specific stage of a pipeline than when loaded globally — adding them per stage cuts token overhead at every other stage.
- Consistent file naming conventions (draft vs. v2 vs. final, with date prefixes) let the AI locate and retrieve files by plain-language description alone, replacing vector databases for most personal workflows.
- Running multiple Claude Code instances against different workspaces simultaneously is a practical alternative to building separate agents — the folder does the routing that an agent framework would otherwise handle.
- The three-layer structure adapts to any profession by swapping the content of each layer: a content creator relabels Writing Room as script lab, Production as edit bay, Community as distribution hub, without changing how the routing works.
Terms worth knowing.
- CLAUDE.md
- A markdown file placed at the root of a workspace that Claude reads automatically at the start of every session. It functions as a floor plan — describing folder structure, naming conventions, and where different types of work live.
- Context window
- The maximum number of tokens an AI model can process in a single session. Once exceeded, the model begins losing earlier content, producing degraded or incomplete outputs.
- Token
- The smallest unit AI models use to process text — roughly three-quarters of a word on average. Long documents consume many tokens, which is why token-efficient routing matters.
- MCP (Model Context Protocol)
- A protocol that lets AI models talk to external apps and services through standardized plug-and-play integrations, without requiring custom API code for each connection.
- Skill (in Claude Code context)
- A package of markdown files, instructions, or scripts that teaches Claude how to perform a specific task. Skills can be downloaded from GitHub or created custom and wired into a workspace at specific pipeline stages.
- Routing table
- A plain markdown table inside a CONTEXT.md file that specifies, for each task type, which files to read, which to skip, and which skills to activate — the core mechanism of the three-layer system.
- IDE (Integrated Development Environment)
- A code editor like VS Code that displays a folder tree, opens multiple files in tabs, and lets you navigate a workspace without double-clicking through Explorer windows. The video argues you do not need one — you can use Notepad or any text editor.
Things they pointed at.
10:50toolHumanizer skill (GitHub)
10:52toolDoc coauthoring skill (GitHub)
16:12toolMCP (Model Context Protocol)
Lines you could clip.
03:53
“Instead of building one big file, you wanna separate your thoughts, your ideas, your work into separate areas.”
12:42
“This right here is the most important pattern in the whole system. It's just a simple table that tells the AI, for this task, read these files, skip those files, you might need these skills.”
18:35
“The folder becomes your app. This is your UI. What simpler UI than a folder?”
19:09
“Within six months, everyone's gonna be doing this — just talking to your folder setup.”
21:03
“It's folders and markdown files with the understanding of advanced software engineering.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00Hi, everyone. In this video, we're gonna dive into my concepts around folder as a workspace and architecture, what it means to use AI in a way that will actually probably last to the next decade and what I think the industry is actually moving towards.
00:15Now if you've seen any of my other videos, I'm sure you've seen screenshots or snippets of crazy folder structure I'm doing where I'm routing Claude to different folders, my markdown files, using skills at certain times, creating animations or code with it. I have all sorts of different ways I'm doing it, but I don't think I have a video out there that really dives into how I structure it and teaching you to do that.
00:39And I decided to go ahead and create a kind of template folder to walk you through all of this. This is something you could use to be able to apply to any of your workflows, build through it, understand it, and kind of explore, you know, building your own version of this.
00:55Now throughout this video, I'm actually gonna teach some concepts along the way for people who are not familiar with maybe Versus Code, which is the workspace I'm working in. Don't worry. You don't need Versus Code to do the things I'm showing you.
01:07You can do it within Claude. But I just wanna describe, like, what markdown is, what it means to have an IDE, all of these things. And if you are familiar with these, you're a software engineer, you probably are, don't worry.
01:18I'm still gonna be diving into the specifics, and I think you're gonna get a lot of out of this either way. However, sometimes it's also nice to just dive back into the fundamentals. So let me paint a little scene for you.
01:28Right now, most people, if they are even using AI, which fun fact, about 84% of the nation is still not fully using it, but log in and use Claude or ChachiPT or Gemini, and they're typing things. They're diving in. Maybe they get something back.
01:43They start another conversation, start over, and sometimes they're able to share context between conversations. Right? Like ChatGPT and Claude can do that.
01:51I can say, you know, tell me a software stack for Vigilor, which is a software that I was working on for another company, and it's able to go and look through our older conversations, which is nice. It's a little bit of extra there, but it's still not, you know, it's not able to look at files, have persistence, things like that.
02:08And you're constantly having to make these huge prompts. Maybe you have some sort of prompts that you save and throw in or documents, and you're throwing it in, and then you have to start a new conversation.
02:18Throw it in, start a new conversation. You hit a wall. There's too many tokens.
02:21Right? All sorts of problems there. And don't get me wrong.
02:24There's some really good prompt engineering techniques out there. But at the end of the day, AI can only hold so much in a single information or chat area. And even further, that's not your workflow.
02:36Having to have it create a whole bunch of stuff and then reedit it again, it's it's not sustainable. It's not scalable. Now for those of you who don't understand why they struggle so much, essentially, AI reads everything, all of your sentence, and measures it by something called tokens.
02:53A token is roughly three quarters of a rule word or a single word sometimes. The word hamburger could be three tokens. Hamburger.
03:02Hamburger. The term comes from NLP research in the nineties. A bunch of researchers needed a unit smaller than a word because language doesn't all break the same way.
03:14Right? So they borrow token from linguistics, which borrowed it from Old English, which means a sign or a symbol.
03:21A token is just the smallest meaningful chunk of a sentence or a word. There's only so much tokens an AI can have in its context window before it starts failing.
03:34When people say context window, they mean how many tokens the AI can see at once, and that window is in fact finite. If So you dump everything into one file, an AI writing a blog post is also reading your video production notes, you're burning tokens on stuff that doesn't matter. So instead of building one big file, you wanna separate your thoughts, your ideas, your work into separate areas.
03:55This is something I created for all of us. It's called a workspace blueprint. Here, you have three workspaces.
04:01And, again, this is an example. You don't always have to do this. But each one handles a different kind of work, one for the community.
04:07Right? Maybe I have this one for working on content and docs in the community. Production.
04:11Right? What am I building? What are scripts?
04:14Maybe I'm creating animations. Writing room. Maybe I need to have some sort of process of thinking, or I have a client list, or insert as many things as I have there.
04:24We're gonna dive deeper into it. This is a space that does the job well because you can circumnavigate an AI seeing everything and only direct it to what you want.
04:34Let me explain how that actually works, though. So I'm gonna show you inside of Versus Code, which is an IDE, basically a developer environment that allows people to kind of look at folders in a different way. So instead of having to click into the folders like you just saw here, I can just open the folders and see everything and open the files without having to double click the files.
04:57Right? So instead of having to double click this text document and opens another window, I can just bounce between them. It's much cleaner, much easier.
05:04Even if you're not into code, it looks overwhelming. I promise you, all of this is just natural language. This literally reads like a document.
05:10So this is a markdown file. If you haven't seen them before, markdown is just a text file with some lightweight formatting. Right?
05:17You have dashes for bullets. You have hashtags or pound symbols, if everyone else remembers when they were just called that, for headers.
05:26You have all sorts of stuff like asterisks for bolding or doing things in that way. And there's a lot of programs that can actually run this to look a certain way.
05:35In fact, your Claude does exactly that. If you look, when you're talking to an AI, it's writing in markdown already. These boldings, these lines, all of these things, watch what happens when I copy this.
05:48All of that formatting disappears when I paste it into here, and it turns into markdown because that's how it's making it look like that.
05:56Markdown is just a good way to format text. If you're curious, there's a man named John Gruber. He actually created this in 2004.
06:03The whole idea was write something that's readable, is plain text, but can also render into a formatted document.
06:11He named it markdown as a play on markup language, which is the same stuff that HTML, right, the stuff that builds websites, hypertext markdown language, and all it does is mark stuff down with tags.
06:24Markdown strips all the tags and absurdity away and makes it look like something simple like this. But, again, you're probably not here for this. You're here for the file system.
06:34So let's move on to that next step. So in this specific folder, which is an example, it runs on essentially three layers, and there's a reason for each one.
06:44If you look at my Claude MD, my Claude markdown file, this is something that my AI will read every time it's in any one of these folders. So this is something that the AI will always have and always reads.
06:57Imagine it's you're just copying and pasting this into Claude code or into Claude every time you open it. Now you can actually just type in to Claude, read the Claude dot m d.
07:08In this case, it's Claude code. You could be working inside of Claude Cowork as well, which is, again, a video I have on how to install, and it can operate inside of folders in case the Versus Code is too much for you. But just read the Claude.
07:22M and tell me what this is. Before, you had to copy and paste, do all these things. It had to read the in every file that's in here.
07:29In this case, it reads the clob dot m d and immediately, without having to read everything else, understands the product, the process, what's going on, my writing room, my production, my community. It knows where to find it, what the file names are, all from just a single text prompt that allows it to understand where to go, what to do, what are these areas.
07:50But let me describe this a little simpler for those of you who might, you know, feel like this is a bit overwhelming. Layer one in this is the map. This is what loads automatically.
08:01Right? It's looking at it. So you put the stuff the agent always needs to think about, folder structure, naming conventions, where files go.
08:10Think of this as the . You walk into any room, the floor plan is on the wall, and the agent knows where to go. Now layer two is where the floor plan tells you to go.
08:23It's the actual rooms. What is your task?
08:27Go here. I want to write a blog post. Well, then you need to go to here and read this context or this markdown file.
08:35If you want to build a demo or a video, you need to go here and read this context or markdown file. And this could be one that you wrote by hand, or it could be one that you told Claude to wrote. And we're gonna dive into that here shortly.
08:48Layer three is the actual workspace itself. Where do you want your files going? What content are you doing?
08:55If you're writing stuff, where do you want the events to be? Where do you want newsletters to be? Where do you want social to be?
09:01And it's just a file system. Again, if you don't wanna work inside of here, you can actually just go straight into the folder and just create new folders. New text document.
09:12That can be a prompt. That can be a context. Right?
09:15It's that simple, and you can just you can just edit it without any of this. Look. My myclaude.md, this is what it looks like in if you open up Notepad.
09:23Same thing. And nothing breaks when you edit it. You can type whatever you want in here.
09:27It's just English. Now it's good to have it uniform and well, but that's the idea here. Most people are only doing one of these layers, maybe two.
09:37The reason you want to actually have these three layers is it stops the narrow funnel of AI doing too much all at once without allowing you to edit every single part, but still give you the ability to automate the entire process. So, again, the router, the initial clauda m d or whatever you're naming it, is loaded when you start any task.
10:00The workspace is loaded when you're in the workspace. When you want to do production, it's only reading stuff that's in production as it needs to. When you're doing stuff in the writing room, it's only writing it when you're in the writing room.
10:12For example, go to writing room. Let's start making something.
10:20Very little prompting, almost terrible prompting, yet the agent, without wasting a whole bunch of tokens and going through everything, immediately goes to the context file that I have in writing room that describes what it is, describes what to load and what not to load, and just describes the folder structure and what the process is.
10:39First, I understand the topic, then I find the angle, then I write it, then I catch problems. You can also incorporate skills into this. Right?
10:47So you can download the humanizer skill, which is an actual GitHub I recommend you all check out, or, like, doc coauthoring skill, which is another set of markdown files or even Python scripts and tutorials that someone else build to do a certain task. And this is where this whole process is different than just running skills.
11:06You're putting skills inside of a thought process. And as you can see right here, we're in the writing room. Clean slate.
11:12No drafts in progress. Voice is loaded. Style is blog post.
11:15What do you want me to make? From one single prompt, we've gotten it in there.
11:19But while that's going on, I can open up another Claude, and I can say, hey. I want to do some work in production. And it's gonna go in there, and I can do whatever work I wanna do in production.
11:31Right? I wanna maybe make some designs. I wanna create some sort of code for workflows in there.
11:37But the real fun happens is when you're building stuff in production with one of your Claude code instances, you're writing a script, and you can say, hey. Take the script from Writing Room, and let's make an animation out of it in production.
11:54It's moving that file. It's gonna go there. Now it's gonna notice that I don't have any scripts in there when I send this out.
12:00But if I did have a final in there, it's gonna go look for it. Right? There's no scripts.
12:04Boom. It didn't waste a whole bunch of tokens. It didn't do anything, but it immediately knew, oh, well, we have to write a script first.
12:09Or if you have a script somewhere else, you can upload it. You see, most apps or frameworks or agentic things require you to build an agent for each of these. I need a writing room agent.
12:18I need this agent. I need this agent. But the way in which you approach each task is always different.
12:24Why not just have Claude code become the agent you need when you're working in the workspace? And you see from there, you get the most important part of this process is just good routing in English language.
12:38Again, this is all just English. Right? File folder names, titles.
12:42It's describing what you want. Right? This right here is the most important pattern in the whole system.
12:48It's just a simple table that tells the AI, for this task, read these files, skip those files, you might need these skills.
12:57Without this, the AI either reads everything and runs out of the room and just does all sorts of stuff you don't want it to do using way too many tokens, or it guesses wrong about what matters or just doesn't hit what you need, or you can't edit what it creates along the process.
13:12This table eliminates all of those problems. This this system here gets rid of all of that. That.
13:17Now let's go ahead and zoom in a little bit here and actually really look at this kind of folder structure, walk through this pipeline. Imagine you're sitting here, and you open up production, and you go to workflows.
13:28Right? So you know you're doing some sort of animation production or insert whatever it is that this folder is as a separate workspace as part of a larger task flow that you're doing. Production has a pipeline in itself as well.
13:43Four stages. You have to do a brief. You need a spec with a specification, a build, and an output.
13:49I have a brief, some sort of script that I wanna do. I have a spec that's generated from that brief, and then it goes into the builds, and it builds out the animations. And then finally, you have the output.
13:58Output. More importantly, this allows me to have one MD file.
14:03Right? So for my production, I can have a context for this file system that is generating different types of sub agents or ways to look at it. Again, I'm not even worried about agents.
14:12I'm just worried about what the workspace is, what I wanna do in this workspace. If I wanna understand, look up technical standards, look up design rules, I can find that because I might have that in a doc.
14:25Right? So I have components or maybe some way I like to design these systems with my color designs, my headers, my quality.
14:33And, again, these are all just generated from English super short docs. These are visual philosophy or what type of tech you wanna use. And it doesn't always have to read that, but maybe during the brief stage, it does.
14:45Right? When you're sitting there and you're going through and you're loading the brief, well, I need to make sure you look at this text standards when you're making the brief. If it's loading the spec, I need to make sure it looks at the design system and our component library, and then maybe it does need to load the deck as well.
15:00And you can swap this around super easily just by looking at it. This is traditional function calling software routing.
15:07This has existed for decades and decades, but now it gets to be natural language, English.
15:14Now at this point, many of you are like, oh, well, you're just making a bunch of skills. Technically, yes. Now for those of you who don't know what skills are, again, I mentioned them earlier.
15:22I I talked about this idea that you can download them from everywhere. There's PowerPoint skills, and I have other videos on this. But at the end of the day, skills are a process that someone else figured out and designed a set of packages or folders just like I'm doing here to tell Claude how to do something.
15:41The thing is skills aren't just markdown files with instructions. Some are just that, but skills work best when they're wired into a system. One important note too is this is where the difference between it's just skills we're creating here and it's a system.
15:56You're actually putting skills inside of your MD. So in this case, I have in my context for production, I have the fire outlook, what I wanna do, but also, right, this is where you can tell to call skills or MCP servers.
16:13If you don't know what an MCP model context protocol, I think that deserves an entire video in itself, but just think of it as a way that the AI can talk to other apps and services easier.
16:25It's designed to just kind of plug and play it in rather than you have to create all these custom integrations. At certain points, we might want the front end design skill or a web app testing skill or a PDF skill, or I might wanna give Claude the opportunity to look up a skill, find a new skill, or even possibly create one.
16:45You can wire up to 15 skills, 20 skills, a 100 skills into a workspace, or you can perfectly add the skills where you would need them inside the workspace rather than having them loaded at all times.
17:00That's the whole idea here is about plug and play and routing. One other sneaky thing I do to, like, completely ignore, like, databases or anything like that, in my ClaudeMD, the main file at the beginning that shows, you know, every AI or every agent that comes into this workspace can see my entire folder structure navigation, I just add naming conventions.
17:22Right? So if a file's gonna be outputted a certain way, it needs to name it. For blog drafts, it needs to be, like, file name, where it's at.
17:31Is it draft? Is it v two? Is it v three?
17:33Example, API auth guide draft. Right?
17:36Or same for newsletters. Here's the year and day, and then here's what it's kind of is. Right?
17:41Twenty twenty six zero three launch week.md. So the AI knows to organize and move stuff, which comes in handy when instead of having to navigate through these files or having agents that's connected to some sort of, you know, SQL database or vector database or query or Postgres or anything like that, I could just say, hey.
18:01Pull my auth demo demo v two and build a spec from it.
18:11It immediately knows without me doing anything to look where that v two demo script would be because it knows how to find it. It knows to pull it, then it knows to read the docs associated with specs, and then start building it.
18:27I have zero code, technically speaking, running any sort of Python injection or framework or database.
18:35This is tools that people are building right now. They're building apps and crazy Python stuff, which in some very bespoke cases might be useful. But most of the time, for most people, you don't need all that extra stuff to get the process and the job done.
18:50The job to be done is more important than this kind of rigid architecture that so many people are building. You see, . This is your UI.
19:01What simpler UI than a folder? And the best part is I don't even need to technically click on anything. I could just talk with my voice to AI, have it do all the text work for me.
19:11The next stage of this, I promise you, within six months, everyone's gonna be doing this, is just talking to your folder setup. It's gonna be designed and set up to be this way.
19:22It's gonna be around yours, which leads into a good final point. How do you make this yours? This template uses a fake idea with fake process.
19:32Right? Fake blog posts and demos. If you're a content creator, Writing Room might become your script lab.
19:39Production becomes your edit bay, uh, whatever. Community becomes a distribution hub, and you're gonna remove and change these rules to edit your platform, maybe your tone and voice inside of these.
19:53Right? So what is your audience? That's what you wanna hear.
19:56It's working developers, two to eight years of experience, technical decision makers, developer advocates. It might be something completely different. You might be in construction or real estate, but this is roughly what you would be doing across all of them.
20:10And the best part is all you need is one subscription to Claude Code, and you can generate a 100, quote, unquote, apps that are just folders creating what you need.
20:20Obviously, it's much more complicated once you get into breaking down your workflow. But if you're a developer, if you're a freelancer, right, just swap design for engineering and docs or intake and production and delivery.
20:34Right? This workspace changes lightly, but the three layer routing system, the idea that you go from, hey.
20:43Look at this area. This is what you're going to, To lower level context ones, to lower level skills is the idea here.
20:53It's just layered. This isn't a prompt trick. This isn't some sort of crazy infrastructure.
20:59It's folders and markdown files with the understanding of advanced software engineering. Every conversation after that, the AI knows where it is, what to load, what tools to use, and where to put the work in.
21:13Now there is a lot of history behind my thinking. I didn't just randomly come up with this, and there's a lot of people who are already doing this. And the reason they're doing this is because it works.
21:23I'm writing a very large research paper right now that goes into the history of programming. Rules of transparency, rules of composition, um, all the way back down to 1972, and then I'm bringing it forward and applying all this stuff to modern day AI, what it means to have humans in it.
21:39And I specifically talk about the layers that we could actually have, and I actually go into a five layer architecture in the paper. But, realistically, most of you just need to understand the three main layers that we talked about here. I will be making videos on this.
21:51However, it is in the main chat. If you wanna download this and give it to Claude so that it can tell you about it rather than having to read through it, I highly recommend that. In fact, I urge you to do it because some of this is kind of technical information.
22:05I'm being nerdy. I'm being structured in it. But this is layering out what the next decade is looking like.
22:11And this isn't because I'm predicting the future, it's because I'm learning from the past, from the last two hundred years of software engineering, and I mean two hundred when I say that, and applying it to AI. I wanna teach the concepts that last, not the concepts that are replaced next month.
22:30I understand that some of this might be in a little fast. It might be a little confusing. I'll keep making deeper dives.
22:35Give me feedback on what you didn't understand. How? Did I move too fast?
22:40I wanna make these better every day. Again, I'm making these on my own. So, hopefully, this all gave you a good idea.
22:46If you do want access to any of these files or worker templates, I am giving them already to my VIP and my premium accounts.
22:56It's my one way of, like, the work, right, to be able to support this. So if you're able to subscribe, amazing.
23:02If it actually is that much of a financial challenge, please reach out to me. I can try to see if I can get you something, you know, quick and easy for you. But at the end of the day, go go check it out.
23:12Go check out all my other courses that I'll be doing. And again, as always, happy learning.
The Hook
The bait, then the rug-pull.
Most AI workflows die in the chat window. Jake Van Clief spent time studying 50 years of software engineering history to understand why — and built a folder-based architecture that routes Claude to exactly the right context for any task, without a single agent, framework, or line of code.
Named ideas worth stealing.
07:55model
The Three-Layer Routing System
- Layer 1: Map (CLAUDE.md) — always loaded, folder structure and naming conventions
- Layer 2: Rooms (CONTEXT.md per workspace) — loaded only when entering that workspace
- Layer 3: Workspace — output destinations, skills, MCP servers wired in per pipeline stage
Three nested layers of markdown files replace custom agents by routing AI to exactly the right context for each task.
Steal forany AI workflow that spans multiple task types or team members
12:42model
The Routing Table
- Column 1: Task name
- Column 2: Files to read
- Column 3: Files to skip
- Column 4: Skills to use
A plain markdown table inside each CONTEXT.md that tells the AI exactly what to load and what to ignore for each task type.
Steal forany CLAUDE.md or CONTEXT.md file to prevent token waste
14:07list
The 4-Stage Production Pipeline
- Brief
- Spec
- Build
- Output
Each stage loads only its own context — design system for spec, component library for build, etc.
Steal forany creative or technical production workflow
How they asked for the click.
VERBAL ASK
22:51product
“I am giving them already to my VIP and my premium accounts. It's my one way of, like, the work, right, to be able to support this. So if you're able to subscribe, amazing.”
Soft and low-pressure. Full value delivered before any mention of payment. Template access is the only incentive offered. No hard sell, no pop-up link.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
