Do You Actually Need AI Agents? (Why I Use Folders Instead)
A 43-minute whiteboard argument that the agent you were about to build is already a folder of text files.
Posted
3 days ago
Duration
Format
Essay
educational
Views
27.9K
598 likes
57 · 43
Big Idea
The argument in one line.
Most AI agents are custom plumbing built to deliver context that a folder of plain text files already carries, so the work shifts from building connectors to writing down your instructions.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You are about to build an AI agent with a vector database, a set of API connectors and custom tools, and you want to know whether that architecture is still necessary.
You already use Claude Code, Codex or a similar tool and keep adding markdown context files without a clear model of why that works better than integrations.
You lead or train a team on AI tooling and need a plain-English explanation of what a transformer is doing that does not reduce it to autocomplete.
You want the history under the tooling: Jacquard, Lovelace, Turing, Shannon, Hopper, and why each one was the same move repeated.
You evaluate AI vendors and want a way to tell when a product is selling you orchestration you could replace with a directory and three tools.
SKIP IF…
You want a step-by-step build walkthrough. There is no screen share, no code and no file listing, only a whiteboard and an argument.
You need the case for genuinely multi-agent systems at scale. The video concedes specific scenarios exist but never explores them.
You already hold a working mental model of transformers, context windows and tool use. The middle 15 minutes will be review.
You dislike history digressions. Roughly half the runtime is computing history before the argument closes.
TL;DR
The full version, fast.
Asking an AI to build a PDF for your investors requires three things the model does not have: how you like PDFs built, who you are, and who your investors are. The old answer was an agent, meaning a database, an API connector written in Python, a document tool, a web scraper and error handling, rebuilt per output format. The argument here is that all three inputs are plain text, so they can live in nested folders, and the only code you need is a harness of three general tools: read a file, write a file, search the web. Every frontier model ships with that harness. Adding a capability then means adding a folder, not building another agent. The back half traces the same move through Jacquard, Lovelace, Turing, Shannon, Hopper and Python to show that every layer of computing collapsed a pattern into a word, and English is now that word.
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.
The prompt is boxed on the board, then decomposed into three unknowns: how you build the artifact, who you are, and who the investors are. Each one becomes a folder of markdown.
03:56 – 10:05
02 · How agents used to get built
The traditional architecture drawn out in full: SQL or Drive for your data, a Python connector past the API gateway, a separate tool for PowerPoint, a scraper for the investors, error handling on top. Then the admission that Copilot made the worst decks he has seen, and that the fix was never integration but having the model emit HTML or images instead.
10:05 – 11:39
03 · Unix was right in 1970
Thompson and Ritchie separated data into a navigable file tree, and that choice still carries every workflow in every company. A machine doing the navigating does not retire the fundamental.
11:39 – 13:01
04 · The harness: read, write, search
The minimum tool set a model needs to work a folder, drawn as three words next to a circle. Google, Anthropic and OpenAI are all converging on it, and it ships attached to every serious model already.
13:01 – 17:10
05 · What the model actually is
Attention Is All You Need compressed to a whiteboard: the sky is blue at sixty percent, beautiful and grey behind it, and a preceding phrase reshapes the whole distribution. The point is that your context is a probability layer, not a knowledge upload.
17:10 – 20:47
06 · The folder as a context layer
Writing down opinions, instructions and data makes the user the one choosing the model's context instead of the reverse. Adding a folder adds a capability, so one model becomes the PDF agent, the deck agent or the marketing agent on request.
20:47 – 24:01
07 · Hello world, and the loom
Print hello world is one visible line on top of forty years of invisible code, the same stack any vibe-coded script lands on. Then back to Jacquard, whose punched cards translated a woven pattern into levers up and down.
24:01 – 29:43
08 · Lovelace, Descartes and thinking machines
The Analytical Engine as gears numbered zero through nine, Lovelace's 470 pages of notes against a 120-page book, and her Note G claim that a machine can never originate. Then back a century further to Descartes, who set two tests for a thinking machine before anyone had built one.
29:43 – 32:54
09 · Turing, binary and Enigma
The imitation game, the Lovelace objection and Turing's reply that nothing is original. Then binary as assigned patterns of off and on, and Enigma as a machine that augmented a human rather than replacing him.
32:54 – 36:22
10 · Shannon and Hopper's compiler
Claude Shannon separating meaning from the engineering problem, then Grace Hopper mapping English verbs onto machine instructions with Flow-Matic, and being told by working engineers that she had given up control of the code.
36:22 – 39:58
11 · Fortran to C to Python
Each generation collapsed a known pattern into a shorter word. Fortran still flies NASA hardware, C added English verbs like print, Python hid whole C programs behind objects, and all of it compiles down to electrical switching.
39:58 – 42:49
12 · English became code
The closing claim: build me a PDF for my investors is now a line of code, English is the newest abstraction layer, and the unsolved part is standardizing how a person hands over intent. The folder method is offered as one attempt, not the answer.
Atomic Insights
Lines worth screenshotting.
The agent you are about to build exists to deliver three things the model lacks: your standards, your identity and your data. All three are plain text.
Adding a capability should mean adding a folder, not building a second agent with its own connectors and error handling.
A harness is three tools, read a file, write a file, search the web, and every frontier model already ships with it.
Folders were invented between 1969 and 1970 and nothing about a machine reading them instead of a human retires that fundamental.
Microsoft Copilot produced the worst AI PowerPoints on record because the model could not control the software, not because it could not design slides.
The workaround for unusable software control was never better integration, it was having the model write HTML or render an image and paste the result in.
Context does not improve the model, it changes which answer is most probable. The UK sky is grey until the model knows who is asking.
Reducing a transformer to a word predictor erases the decades of failed research that made next-token prediction land on correct answers.
In a world of cheap answers, the questions and the instructions are the scarce input, which puts the person back at the center of the decision.
Ada Lovelace's notes on the Analytical Engine ran 470 pages against a 120-page book, so the appendix was the real work.
Descartes set two tests for a thinking machine in the 1640s: hold a conversation rather than answer commands, and do something outside its design.
Turing's answer to Lovelace was that nothing originates, because every human is running programs installed by school and environment.
Grace Hopper had a working compiler that nobody would touch, because engineers said you lose control of the code. That objection has been recycled for every abstraction since.
Typing a single character runs hundreds of lines of code, which means every developer alive is already trusting abstractions they cannot inspect.
Build me a PDF for my investors is now literally a line of code, so English is the newest layer of a 200-year abstraction stack.
Takeaway
Instructions in folders beat one agent per job.
CONTEXT OVER PLUMBING
Most agent architecture exists to deliver three pieces of plain text the model is missing, so writing those down in folders replaces the connectors, the tools and the error handling.
01Build me a PDF for my investors
A request like build me a slide deck for my investors needs three things the model lacks: your build standards, your identity, and who the investors are.
All three are expressible as plain markdown and text files in nested folders, with no database, no API connector and no retrieval code.
02How agents used to get built
The classic agent build wires a database, a code connector past the API gateway, a document-generation tool and a web scraper before it answers once.
Copilot produced unusable decks because the model could not control PowerPoint, so the real fix was emitting HTML or images and pasting the result in.
Changing the output from a slide deck to a PDF breaks that architecture entirely, because there is no tool for the new format and the whole chain gets rebuilt.
03Unix was right in 1970
The 1969 Unix idea of separating data into a navigable file tree still carries every workflow, and a machine doing the navigating does not retire it.
Take the fundamentals you consider too simple to matter seriously, because the ones you take for granted are usually the ones holding everything up.
04The harness: read, write, search
A harness is the minimum general tool set: read a file, write a file, search. Nothing task-specific belongs in it.
Every frontier model already ships with that harness, which removes most of the reason to build a custom connector in the first place.
05What the model actually is
Transformers assign probabilities to the next word, and context placed ahead of a phrase reshapes those probabilities without changing the model.
The UK sky is grey is the most likely answer, but the right answer depends on who is asking, which is exactly what your context supplies.
06The folder as a context layer
Writing down your opinions, instructions and data makes you the one choosing the model's context rather than inheriting whatever it assumes.
Adding a folder adds a capability, so one model becomes the PDF agent or the deck agent on request instead of requiring a separate build.
07Hello world, and the loom
Print hello world is one visible line sitting on forty years of invisible code, which is the same stack any AI-written script lands on.
Jacquard's punched cards in the 1800s turned a woven pattern into levers up and down, which made physical logic the first program.
08Lovelace, Descartes and thinking machines
Ada Lovelace's notes ran 470 pages against a 120-page book, so the thing filed as commentary was the actual contribution.
Descartes proposed two tests for a thinking machine in the 1640s: converse rather than answer commands, and act outside its design.
09Turing, binary and Enigma
Turing argued imitation of thought matters more than thought, since nobody can define thought precisely enough to test for it.
His reply to the Lovelace objection was that nothing originates, because knowing all the facts of a system never means knowing all its consequences.
10Shannon and Hopper's compiler
Shannon separated meaning from the engineering problem: anything describable in logic can be built into hardware.
Grace Hopper's working compiler was rejected because engineers said you lose control of the code, which is the same objection AI tooling hears now.
11Fortran to C to Python
Each language generation collapsed a repeating pattern into a single word, so print stands in for assembly, bytecode, binary and electrical switching.
Python repeated the move at a higher level, hiding whole reusable C programs behind readable statements and object definitions.
12English became code
Build me a PDF for my investors is now literally a line of code, which makes English the newest layer of a 200-year abstraction stack.
The model still has no access to your intent, so the open problem is standardizing how you hand it over, not building more agents.
Glossary
Terms worth knowing.
Harness
The small set of general-purpose tools wrapped around a language model, typically the ability to read files, write files and search. It is what turns a model that only predicts text into something that can act on a folder.
Skill
A downloadable folder of instructions that teaches a model how to produce a specific artifact, such as a slide deck. Despite the name it is usually just markdown files describing a process and some assets.
Folder architecture
Organizing your standards, your identity and your reference data as nested directories of plain markdown and text files so a model can navigate only the parts a given request needs.
API gateway
The access layer in front of a database or service that requires requests in a precise format. It is the reason traditional agent builds need a code connector for every data source.
Transformer
The model architecture introduced in the 2017 paper Attention Is All You Need. It assigns probabilities to possible next words and lets surrounding context reshape those probabilities.
Analytical engine
Charles Babbage's proposed mechanical computer, driven by gears representing digits zero through nine and programmed with punched cards. Ada Lovelace wrote the notes that described how to program it.
Imitation game
Alan Turing's proposed test, later called the Turing test, in which a machine passes if a person conversing with it cannot tell whether the other side is human.
Compiler
A program that translates human-readable statements into machine instructions. Grace Hopper built one of the first, Flow-Matic, by mapping English-like verbs onto long sequences of assembly code.
Binary
Representing information as patterns of zeros and ones, which map onto electrical switches being off or on. Four switches give sixteen distinct values, which is enough to start doing arithmetic.
Bytecode
The intermediate form Python compiles to before anything reaches machine code. It is one more layer in the chain running between a typed statement and electrical signals.
Object-oriented language
A style of programming that groups related behavior behind a single named thing, so a block of reusable code can be invoked as one word instead of rewritten.
“Because in a world full of cheap answers, questions become valuable. Instructions become valuable. This places the person at the center of decision-making. This is an augmentation rather than just an automation.”
complete thesis in four sentences, no setup required→ IG reel cold open↗ Tweet quote
20:10
“Instead of having to build agents for every process, you're building instructions for each process and letting one agent, one AI, become what it needs when you need it.”
the entire argument of the video compressed into one sentence→ TikTok hook↗ Tweet quote
10:50
“If you didn't have folders, wouldn't your entire workflow collapse?”
rhetorical jab that lands instantly on anyone who works on a computer→ TikTok hook↗ Tweet quote
39:35
“Every single time you type a key, the letter K, the letter J, into a keyboard, that is hundreds of lines of code being executed. You are living on the top of giants right now.”
concrete image followed by a clean emotional landing→ newsletter pull-quote↗ Tweet quote
35:50
“They said it's too unknown. You can't control the code. It's going to make too many errors. You're not thinking about the code anymore. Does that sound familiar?”
historical quote that turns into an indictment of present-day AI skepticism in one beat→ IG reel cold open↗ Tweet quote
16:50
“To sit there and diminish it as simply a word predictor is to diminish the years of effort and research that people have put into it.”
direct swing at the most common dismissal of LLMs→ TikTok hook↗ Tweet quote
17:40
“This is the opposite of what a lot of people are worried about with cognitive thought and the models controlling you. We can flip it on its head and we can decide what its context is.”
reframes AI anxiety into an agency argument→ newsletter pull-quote↗ Tweet quote
28:10
“If you can learn the history of the why, the how becomes much easier to understand. I could sit here and try to teach you how to program in Python. But if you don't understand why, how doesn't help you at all.”
justifies the entire second half of the video and works standalone as a learning principle→ newsletter pull-quote↗ Tweet quote
25:40
“The book was about 120 pages long. Her notes were 470 pages long. So the real book was the actual notes.”
clean fact plus punchline, no context needed→ TikTok hook↗ Tweet quote
41:10
“The best engineering in the world happening with AI is happening in English language.”
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.
17px
metaphorstory
When I put this into an AI, there's a lot of different things that happen. Now, if you've been following me for a while, you know I use a lot of folder architecture. So I'm actually going to start at that level of kind of, hey, I've designed a whole bunch of instructions into a folder, and my AI is going to follow those instructions every time I ask a question.
So in this case, let's say I have a folder, and it has instructions on how I like to build things, right? And that's just a folder with some markdown files, maybe a couple text documents that literally is just me describing, this is how I like slide decks, maybe there's some assets in there, right, for different images, things like that.
And that's a nice subfolder there. And I describe how I like to build things in it. Then I have another folder that describes me, right?
And we'll have that be maybe my brand voice, where I'm from, my past, things like that. And again, just text files just all organized neatly into an easy to read navigable file but then i might have certain artifacts i like to build with and write with so a slide deck might have a skill, which if you go online and use Claude or Codex or Gemini, there's things called skills that you can download that do certain tasks or create certain things.
And there is one that you can get right now that says slide deck builder. And some of them usually build HTML, so big old code pages, and then convert that into PDF or convert it into a PowerPoint. Let's just say that's someone else's little folder process, because if you open them up, it's just marked down in folders, that says how to build a slide deck specifically.
And the reason we had to describe that is because if I had asked this a year ago or two years ago, it has no idea what a slide deck looks like, or if it does, it kind of doesn't know the best way to put it, the best practices. Someone figured that out, put it into a box, which, again, is just a folder, and then gave that to you.
right and if you've watched my other videos we describe how you need data context and tools to make an ai not stupid in other words to make it actually create good stuff this was one of the solutions to doing so right we've gotten our build folder maybe a subfolder on how to build slides we call that a skill maybe you built it yourself then again my my is a little different right if i said Build me a slide deck for my investor into an AI that had none of this folder structure, none of this knowledge.
Guess what? Doesn't know who you are. Might know how to build a slide deck okay.
Definitely doesn't know who your investor is or your company. Again, you can swap these words out. But if I have a nice little folder structure, then guess what?
I can simply describe what that is. So let's go ahead and expand me out. Again, folder inside a subfolder.
Much simpler here. So under me, I might have something about the things that I work with, the people that I work with. And in this case, I have a folder about my investors for my company.
Or you could swap it out for clients. But either way, I have maybe a couple markdown files, text files that describe who they are. Maybe there's links to their websites just pasted in there, all of that sort of stuff.
And so that folder... or subfolder is part of this kind of process, my company, the way I like to work things. And every time I ask my AI, whether it's Claude or Anthropic or anything, put it straight into the chat, as long as it has access to this folder, it will go through and go build me.
Okay, what do you like to build and who are you? A slide deck. Okay, how do I build slide decks?
For my investors. Okay, my... You have investors.
Let's read who those are. Traditionally, over the last two or three years, this was a much more complicated process. Now, traditionally, if I wanted my AI model to be able to do a task like this, we would have to create some sort of agent, or at least that's what a lot of people have done.
So they will take the AI. And the model on itself knows nothing about you, maybe sucks at building slide decks, definitely doesn't know anything about your investors unless it has to go do a bunch of research. So over the last two or three years, what people have done is they would create a database that maybe has a whole bunch of information on you.
And this could be an SQL database, which for my data nerds out there, you know, it's just a really efficient way to use code to capture data and bring it towards you. They might use a big giant Google Drive that has a whole bunch of stuff, maybe a SharePoint or any sort of other kind of data place. Other times we put it into these text files called JSON and YAML.
We don't need to go into that, but just know it's a little bit more complex and structured. than those markdown files or text files that we were just talking about. Problem is, these are all behind an API or some sort of basically gateway that says, hey, in order to access me, in order to get to me, you have to say these words in this order to make sure you get the right stuff.
That goes for all of these databases. So the AI doesn't know how to do that. So we got to create a little package for it that lets it do that.
And usually we would write that in code. So this could be Python, C sharp, a computer language. And I've written quite a lot of these, right?
I used to do this all the time. I have a master's thesis, which I have a couple of videos on. You can go check them out, where I took a whole bunch of different psychometric scales, personality tests, and essentially created a bunch of pipelines to access all those tests and send them to the AI model.
Now, again, The reason we had to do this is because the AI doesn't have this information trained into it. You can train a good bit of it, but even then, it's still not as efficient.
Sometimes it messes things up if you purely just train it on data. So sometimes it's more efficient and better to give it to it. If we don't give it that data, it doesn't know anything about me.
The next problem that we have to deal with is building a slide deck. So it has the information on me. So now...
We're done here. We built this really cool Python connector. We've connected it with code It's able if I ask hey, who am I it's able to pull that information.
We solved that problem now I need it to be able to build a slide deck So then I create another bit of code that says this is how to build a slide deck Here's how it connects to Microsoft power came. Let's use a red one there. So we'll do I don't know if PP is the right word here, but we're going to go for it anyway.
So we've got a PowerPoint, and you could do Google Slides. There's a lot of steps. There's assets.
There's clickable buttons. There's so many different things. This is it using another software.
And so at first, we tried to get the AI to control making PowerPoints. I worked with a large company, Fortune 500. We have an NDA, so we can't talk about it too much.
But I trained over 1 ,500 of their employees over the course of a year, and we were using Microsoft Copilot. which had a pipeline, data connection, and a very specific way to manage PowerPoints. The worst PowerPoints I have seen an AI make ever.
They were awful because it couldn't fully control the whole system. It's gotten significantly better now. But the problem is it can't interact with PowerPoint software well.
So to make that better, we have to create a separate tool that allows it to either... write code that gets converted into a PowerPoint, or take pictures that are pasted into a PowerPoint. And so common things were to write HTML code, the stuff that websites are written with, or to just build it out in some sort of HTML or image model, take a screenshot, and then send that and paste it into a PowerPoint.
Because the AI just could not really work with PowerPoint. Well, it just, it didn't work, we didn't have the systems. This...
With enough iterations, now some of them do work pretty well. They can decently mess with PowerPoint in that way, and they have a decent amount of tools. But really, all you needed was a couple of these tools.
And I'll say why that's important in a second. But again, we're still not done. It had all the information on me, but what about these investors?
Well, there's websites, there's maybe different LinkedIn profiles they might have, all of this stuff. So we need to give it a tool that lets it scrape the web. And so that might be more code, more Python, more C Sharp, whatever it is.
And it rolls out and then it finally finds that investor. Hopefully. Sometimes it didn't.
So then we had to have error catching built into it to make sure that it would find the right person based on my question. Also, my investors needs to go back over here to pool and double check that it's actually calling them. And so that's what it would do.
And it'd go out and it would find it. Finally, once it's taken all of that information in, it consumes it all. And then it says, hey, here's who your investors are.
Here's the slide deck. Here's what you've made. Now, I don't know if I need to tell you this.
That's one agent. If I change this word to a PDF document. Oh, I don't have a tool for PDFs.
So I have to come back over and build a whole new one for how that works, the instructions for how I like to do it, connect it to these ones, make sure it's reconnected. So now I have agent one, which builds slide decks and knows about my investors and handles all these things, and agent two that can build PDFs in the same way.
Now, I will say, once you built the architecture for the first one, it's easier to build subsequent agents. But why would we do that when instead we can flip it all on its head?
So let's go back to the folder again and why I think it's such a powerful tool in the AI world. Between 1969 and 1970, you have two men named Ken Thompson and Dennis Ritchie who discovered a really good philosophy called Unix. And basically they said, if you can separate data, put it into a nice file tree structure, and put it onto a page that you can click through, you have a great way to interact with a computer.
And they created all the structure, all the ideas, thought about how you would interact it with your mouse, how it would work on operating systems and Windows, because again, you might take for granted, folders on your laptop didn't always exist. Folders weren't always an option for people, right? They haven't always existed, and we take for granted some of the most simple software that we use.
but it's often the most powerful because we take it for granted. I mean, think about it. If you didn't have folders, wouldn't your entire workflow collapse?
I know most of these multi -million dollar companies, it goes straight down the drain real quickly. If that's the case, if something invented in the 70s is still fundamental today, just because an AI is reading the text, navigating it, using the software, doesn't mean that fundamental disappears. It just means something else is using it.
So. To be able to do that, I did some testing. And I said, well, instead of making all this crazy agentic stuff, what are the minimum tools that an AI needs to be able to read a folder?
And I'm not the only one who's doing this. Google is doing this. Anthropic is doing this.
OpenAI is diving deeper into it a lot now because they kind of have to because of how many people and researchers are exploring it. But the idea was we can build a harness. So a harness is simply a set of general tools.
that seem to be used almost over every workflow. So a tool that lets you read folder files, a tool that lets you write into folder files, maybe one tool that lets you search the web or search online. And those are essentially a package.
And no matter what AI you use, where you use it, all of the best ones at least have this little harness attached. to the model. Most AI that we talk about today is just a big model with a whole bunch of numbers that when you put something in, it decides the best next step to give back to you, right?
And it does that thousands of times. So when I say build me a PDF for my investors, as long as it has a couple of these tools, I don't need to build a whole bunch of agents. It'll go through and read, write, and search for what it needs.
And if you can give it what it needs before it has to go look elsewhere, Guess what? You don't need to build all this crazy stuff.
So let me show you how that would work. And so in order to show you the massive, interesting side of this, let's try to condense what the AI model is so that you can appreciate why giving it a harness does so much. So you have, in 2017, a paper called Attention is All You Need.
This came out from Google Labs, Google DeepMind, I believe. And essentially, without diving into the huge amount of math sitting inside this paper, there has been decades of AI research.
I mean, all the way back in the 1950s, when we first coined artificial intelligence, all the way to the 1980s and 1990s, where we started talking about neural networks. We copied what we thought neural pathways in the brain would look like as code to modern day transformers. which is what was talked about in that 2017 paper, they basically said, hey, taking this decades and decades of research on how AI should work, what predicting the next process should be, how can I use language to control a computer?
They said, well, we can just take all of that data. We can take every time you make a decision, every time you decide something, and we can give it a number. And that number...
can be a percentage of its possibility happening again. So if I was to say, the sky is blank, I can have a list of words based off thousands of sentences that have ever been said and have a percentage. Obviously, the highest percentage is probably going to be blue.
Let's say that's 60%. The next highest percentage might be beautiful.
or gray or any of those. And so each of these is going to have a different percentage of possibility when I ask a question or say a statement. And so what we did is we took all of those up and then we said, that's one set of numbers.
But now, what if I added a sentence in front of this? The UK sky or the night sky. Context.
changes the percentage. So let's go put a whole bunch of numbers around all of that. And so they did that billions and billions of times to get every possible combination and context layer and said, okay, well, if we can do that enough and we can get enough numbers, every time you put in a question, the model should output the most likely answer, which more often than not, is the correct answer.
The thing is, the more we put layers on it, the more we did that, the more it gave not the most likely answer, but the most seemingly fitting answer, which there's a difference there. When I believe something is correct, I'm thinking based on what I have learned. So if you add one more layer to this, which is the person using it, you know a lot more context information.
which means the right answer, predictively, the UK sky is gray, most often. But if they know it's me and I'm studying in spring in Edinburgh and it was a beautiful day and I talk about how much it is, then my most likely answer is going to be beautiful. The right most likely answer is not too much different from how we think, how we look at that.
The hard part is creating the massive, massive databases. and the algorithms that look through all of that. That is years and years of research, years of effort, years of failure.
And yes, sometimes they suck. Sometimes they are wrong. But to sit there and diminish it as simply a word predictor is to diminish the years of effort and research.
To diminish it down to that is to diminish the years of research that people and effort have put into it. The time and the love that we have looked at because we're all looking at engineering, meaning, structure. Which brings us back to the original thing.
Little side note there. When I make a folder, To hold all of my context who I am what I'm doing how I like to do things I'm providing another layer of data to help it come to my right answer what I think is best.
And this is the opposite of what a lot of people are worried about with cognitive thought and the models controlling you. We can flip it on its head and we can decide what its context is. We can be the one in control and it becomes much more of a tool rather than a predictor.
So why would I want to do that? Let's go back to the original folder structure. If I sit there and I have my instructions for building.
I have my instructions for who I am and what I like. I have my instructions for how PDFs and things should be built. And then I have my harness that has the ability to read, right?
That might actually be code, right? This is the one place where code is needed, but this is the same as downloading a software or tool. It's really not too hard nowadays.
Most of those models in AI already have them built in to write and to search. then technically with just these few things, as long as my instructions are there, it can do almost anything I need. So if we're looking at opinions, instructions, and data, and we're just describing them in text files inside of a folder, when I say build me a PDF for my investors, it goes into the folder and only reads the opinions associated with what I asked.
what my instructions are and what my data is. So let's say I have hundreds of different questions and things that I like to do on a daily basis. Instead of having to build a whole bunch of crazy connectors and integrations, I can just add a folder and describe what I want to do.
And that's it. I just added a new feature, a new agent, essentially. As the model, as the AI goes through the folder, goes through the instructions, It's reading what it needs to and becomes your PDF agent, your slide deck agent, your marketing agent.
Instead of having to build agents for every process, you're building instructions for each process and letting one agent, one AI, become what it needs when you need it, when you want it to. Because in a world full of cheap answers, Questions become valuable.
Instructions become valuable. This places the person at the center of decision -making. This is an augmentation rather than just an automation.
I think we're pretty clear. It's much simpler to build out a folder to handle a lot of processes. Obviously, there are specific scenarios, but let's start following it down into the rest of the computer.
We talked about the instructions. We talked about the folders and where it is, but let's talk about... the other side of this.
How is this being processed? What does that look like? What did we have to do beforehand to actually have this code?
Now, we're not going to dive in too deep. I'm not going to get absurdly technical, but I think it's important to understand the thousands of lines of code that actually happen when you put this in. And so instead, we're going to replace this with traditional code, one of the most basic ones that everyone learns about.
And I actually have another video on this where I go in more detail. Let's just talk about it real quick. Print, hello, world.
This is one of the most well -known computer statements that you learn in almost every programming class. For this situation, this is written in Python. Now, Python was created in the 1990s.
It was actually named after Monty Python, not the snake. The creator of it really much enjoyed the movies and figured that would be a funny way to talk. And essentially, if we talk and I say print, Hello world.
This is a command this command says hey computer You need to display in an image below this the words hello and world There's a lot going on here now when you learn it in computer programming all you see Come out of the computer This hello world like ah it worked This is actually thousands of lines of code that took 40 years to develop, not in Python, but in everything else behind it.
In the same way, if I asked an AI to write this code statement, in the same way people are vibe coding now or asking it to write code, when that code is ran, even after the AI built it, there's hundreds of things happening. So let's go ahead and look at what that looks like. So Python is a computer language.
Now, before we go into that print hello world, it's important to understand why someone would make Python and what a computer language is. So let's go even farther back to 1800s. I know, so far away.
There was a man called Jacquard. And he created something called the Jacquard loom. I believe it was somewhere in France.
I can't remember the name off the top of my head right now. And essentially he said, hey, right now we've got all these people and they're using these crazy looms and they're going back and forth and they have to keep moving these levers back and forth to change the design. So what if I could make a beautiful little card that had some holes in it that you could slot in and essentially...
all the springs on those little levers would go up or down whether or not there was a hole. So anywhere there wasn't a hole was flat, the lever would be up, and anywhere there was a hole, it would slot into the hole and go down, which on the other side meant it would lift up, changing the thread pattern. And if you did that enough times and you made certain areas flat, you would eventually have a card that would have a pattern on it.
Up. and down, created an image, created a series of interactions. So as I sat there and folded back and forth, the loom would roll over itself in a certain pattern.
This is one of the first ever computer programs. Granted, no electricity, no programs technically, but it took logic. He had to understand where each hole would end up on the other side.
He had to understand how you could translate something physical, an image or a pattern, into a series of up and down. Fast forward a little bit.
Only a couple years, about 30 years, we had someone called Ada Lovelace. Absolutely brilliant mathematician. And she worked on something called the analytical engine.
Basically, it was this giant physical gear thing with a whole bunch of gears in it. And each gear had a little sprocket on it. that would represent a number 0 through 9.
And so if you had one gear here, it would be 0, 1, 2, 4, depending on where that gear is, and as you turned it, it would change them in an order. They would use some sort of card that had holes in it to decide the starting pattern of the numbers, and it could then do calculations. Of course...
The amount of cards that you would have to make to change all of the numbers based on gearing was massive. And so she was tasked with writing an amazing book around it. Actually, specifically the notes, because, well, they didn't really like women being authors of books at the time, unfortunately, despite she was absolutely brilliant.
So instead, she wrote the notes of the book. The book was about 120 pages long. Her notes were 470 pages long.
So the real book was the actual notes. What's really cool is I did a research project on her. I read through a whole bunch of her notes.
We saw patterns with her mathematics still being used in a lot of algorithms we look at today. But the coolest thing is, every so often she'd be writing notes, all these mathematic notations, and the next thing you know, hmm. You know, if you assigned the right variables in the right place, I bet you this could produce music.
It could pluck a string in the right order. It could do an orchestra. And then she'd go right back into the math and she'd be writing and writing and writing and algorithms.
And then she'd stop again and she would return back to that thought and just write another note and say, well, does that mean if it can make music that they're thinking? Well, what would it need to be to be a thinking machine? And that statement and what she argues about there actually goes back.
before the 1800s. This is important because if you can learn the history of the why, the how becomes much easier to understand. I could sit here and try to teach you how to program in Python, how to build an AI, how to run it.
But if you don't understand why, how doesn't help you at all. If you don't understand the why, You don't understand how something breaks.
You don't understand what it took to get there and why we do certain things certain ways. It also means you don't know what you can skip, what you don't need to learn, because someone has already figured it out and it's just a process in the giant cog. So that idea of thinking machines that Ada Lovelace talked about actually was brought up for the first time in the 1640s, I believe 1642.
by a man named René Descartes. You know, I think, therefore I am. You might have heard that before.
Well, he didn't just write on philosophy, religion, and mathematics. He also wrote on thinking machines. In a book called Discourse on the Method, he wrote about two possible ways to know if a machine was thinking.
Mind you, he didn't test it. He was just watching a whole bunch of cool puppets and weird steam machines and these da Vinci Renaissance engines going around. And he said, if it had a human limb, if it had a human face, what if we were to build all the organs that a human would have?
How do I know it's not human? How do I know it's not thinking? And even if it's not human, what's the difference?
And so he said, albeit in 17 pages, much longer than I'm able to describe it here, well, first of all, it would have to, and I quote, not communicate like a brute. And what he meant by that is have a conversation, understand what I mean, not just respond to commands, but to add to them, to be able to be a partner in thought.
But he said, we still might be able to create that in a machine. He said, if we could organize every single word and made sure it organized the right way, then maybe we could do that. So it needs to do a second thing.
It needs to be able to do something outside of what we designed it to do or what it was designed to do. So a clock is not thinking because it can tell the time. But being able to tell the time is a very powerful thing for thinking beings.
We built the clock because we couldn't keep track of it. So if the machine can do both of those things, build something outside of itself and what it was built to do, and could talk to us, well then, you know what? Maybe it is thinking.
Ada Lovelace, going back to her and her lovely Note G is one of the most famous ones where she talks about this. If you look up analytical engine and Google Note G, you'll find a lot of information on it and you can go read it yourself or you can go check out, I'll have some links in the bio I made some papers on so you can just jump straight to that.
But she said, you know, it might be able to make music. It might be able to predict what it could be doing. It might be able to do all sorts of fantastical things, but it can't originate.
anything. It always needs the intention of me, of a human. And even if we built all of these machines, they were built with the intention of a human.
So the spark of consciousness, the spark of light, no matter how advanced they get, was with us. So there is no origination. Funny enough, if you fast forward all the way to the 1940s, someone picked up on this long conversation.
1940s. Maybe you've heard of him before. There's quite a few movies about him, but his name is Alan Turing.
There is a specific test called the Turing test. Now, he did not call it that. He called it the imitation game.
But he said, if you, as a computer, could talk to a person, and on the other side of it, you don't know if it's a human or a person, It's past the consciousness test. It is the imitation of thought that is more important than the actual thought being done.
Because he believed we still don't know what thought is. You could sit here and describe it and what it is. And he actually specifically references Ada Lovelace in one of his papers, I believe it was his 1952 paper, where he says it's called the Ada Lovelace objection.
He goes, you know, she didn't believe that they could originate anything. But there is nothing original. under the sun.
And that's a great Greek philosopher. All of us are listening to programs of school, of where we grew up, of what is being given to us by the environment. So what is original to begin with?
He didn't attack saying that the AI couldn't originate. He said that there's no such thing as original, that all of this is imitations and building together and movements. It wasn't necessarily an argument that they could be consciousness.
Rather, he said it is a common fallacy. that both mathematicians and philosophers fall to. That once all the facts of a system are known, all of the consequences are known with it.
But that's not true. We often do research for decades. We find every little thing.
And we come to find out that, in fact, we are surprised. We missed something. Oh, we didn't get all the data.
And his argument is that machines could surprise us too, despite of what we know. But around this time... There's another important person that came about, and this is where we come into programming and processes.
At this point, machines had come together well. We had decided that we could take the Jaccard loom and we could add zeros and ones together. We could, instead of up and down, let's assign it to a number value.
And then if we assign patterns of those zeros and ones to another number, we can start doing math with it. So we assigned something called binary. And so from zero to 15, We could use four indications off and on to create 15 numbers.
That's a really cool logic. And if we use the algorithms and the processes of Ada Lovelace and Jacquard, and then we put it into electronics, switches and cables, we could start making the computers do math. It could do algorithms.
It could do all of these things. And Alan Turing used this to break the Nazis' enigma. He created a massive machine that could calculate its cryptography, all the secret message.
And it was impossible to do that hand by hand because there were so many possible calculations. But if a machine did it, it could do it faster. But they still augmented Turing.
They didn't automate it. And so what ended up happening is saying, how do we make that smaller? How do we make it more condensed?
Now, many of you might be using an AI called Claude. Well... That is actually named after a wonderful individual called Claude Shannon.
Now, when we talk about information theory, when we talk about logic and ideas and switches, Claude is one of the pioneers in all of the major ways we still even use today. Claude talked about the ways in which we interact with semantics and meaning are completely separate from the engineering problem. If it can be described in logic, it can be made into hardware.
Claude decided to go ahead and write all the papers we would possibly need on that. Not all of them, but a really huge amount of them. Now, I don't know what was in the water in the 1940s or 50s, but I can tell you we had some heavy hitters.
There's another one called Grace Hopper, an absolutely brilliant individual. But Hopper was dealing with a similar concept that I'm dealing with today. So Grace Hopper came up with something called the compiler.
And essentially what a compiler does is it says, hey, you have all of these words and commands. What if we routed that to all of that crazy machine logic you had before? One of her first compilers was something called the Flowmatic.
Now, essentially, you had all of these crazy instructions, right? These zeros, these ones, these ifs, these ends, these RBNs, these moves, blah, blah, blah. We had a whole list of things that we were doing in machine language, assembly language, things like that.
And she said, well, there's a lot of patterns. In the same way we saw patterns in zeros and ones and wrote machine code, she said, well, it doesn't just have to be arithmetic. we can assign it very specific sentences.
We can say move or guide or place or insert small little sentences that would stand for a whole bunch of specific actions, specific code being ran. But in order to make sure that that all mapped well, in the same way we had to map those zeros, those ones, those ideas to the logic and write that all out, she had to write out all of the ways in which this would map to hundreds of code in different orders.
And what you're starting to notice here, hopefully, is there's a similarity. Language to action. Language to action.
Right? We're building on it here. And so...
She created one of the first real compilers, as in, I will take all of your words and compile it down, take it all together, and decide what actions the machine could take. And there's thousands of lines of code that would happen with that, right? Funny enough, there's a quote from her.
And she says, I had a working compiler and no one would touch it. They said... computers could only ever do arithmetic.
Grace Hopper had built the system, had seen how it worked, had gotten it working, and decided, you know, this should be something shared with other people. And the traditional engineers of the 30s, 20s, and 40s didn't believe her. They said, it's too unknown.
You can't control the code. It's going to make too many errors. It's not in control of it anymore.
You're not thinking about the code anymore. Does that sound familiar? I don't know.
Maybe. We'll find out. Well, long story short, it ended up being one of the most important inventions and led to all sorts of amazing languages.
We got Fortran, which is still being used by NASA today to be able to send ships up into space. We got all sorts of amazing computer languages that allowed us to write instructions in code. that would handle all of that crazy binary and absurd things we were talking about earlier.
Now, once we understood that we could compile something, as in we could take pieces of code, pieces of patterns, and it didn't all need to be written down all the time, we just needed to catch the patterns, someone else could write it down, and we could compile that into machine -readable language, as in those silicon transistor, those files, those processors go, oh, zeros and ones, I get that.
it doesn't understand English. We started building a lot more. And people realized, well, why are we, you know, this is still very mathematic languages.
Even Fortran and all of these lower level languages were still really hard for people to use and learn. It was more math than it was language. So that's when languages like C and Python started coming about.
Now, C came out, and it was adding a couple more English words, words that we could actually use to do things, like the word print being something that is a command that controls a whole bunch of assembly and machine code to make the computer do something, make calculations, put an image on a screen, and handle text. You know, it was still not there.
It was still a little challenging. So our lovely friend I talked about earlier said, well, let's see if we can make it even more readable. Let's focus around objects.
And we call it object -oriented language, as in we have a thing that does certain stuff, i .e. separations of concerns. Let's just put that behind a word.
So let's take all of this C stuff and actually just... make that into a word as well. So people were building a whole bunch of different programs that did really good stuff in C, and they were a whole bunch of lines of code.
Well, we keep reusing that program again. What if we made that a statement made that a word, right? And so Python kind of does that it's not exactly like that.
So you do want to take this with a grain of salt. But essentially, when you write something in Python, it's actually taking that and it's deciding what else did you write?
What was here? And it's making a tree and it's deciding, okay, well, based on this is a certain type of command or output or process, and it's assigning that a value. And it's just doing that for every single sentence you write.
And then it's compiling that. down into bytecode, which is then getting compiled technically, down into machine -readable code, which is then getting compiled down into the zeros and ones, which is then just electrical signals moving back and forth, which is really just off and on. So every single time you type a key, the letter K, the letter J, into a keyboard, that is hundreds of lines of code being executed.
You are living on the top of giants right now. Every process has been logically built out. We have been using that process all the way up in today, constantly abstracting one idea and building on top of it and building on top of it and building on top of it on the ever journey to get closer so that one thing, what is the goal of computer language?
To turn our language into something a machine can do. That has always been its goal.
That has always been its direction. We've tried to do that hundreds of times before, and we keep doing it. But it also automates the processes of intent.
Each stage was complicated. So we added that into a sentence, into a word, and brought that in to a smaller version of the next stage, which became more complicated. And then we condensed all of that down until finally we find ourselves here today.
We are at one of the epitome of computer languages that people had wanted so long ago, which is, I can say, build me a PDF for my investors and it will run all of this code. That is a line of code now. Our English language has become code.
But there's one problem. AI doesn't know what you think is best. what you want.
Your intention, just like with all the other software, is still the main goal. So how can we standardize giving my intention to this AI? My folder method that I explained all the way at the beginning of the video is one method in the history of time to standardize the way in which we structure our intent, our logic, our zeros.
and ones. All the symbols around it. Because when I say build, it's compiling down into a whole bunch of stuff.
I'm simply describing that in folders. When I say PDF, I'm simply describing how to build a PDF. And it is connecting to tools that already exist.
The best engineering in the world happening with AI is happening in English language. The reason computer language has been so powerful for us, the more people that can use it, the more things we can build to automate to give to people to augment. Our goal for today, our goal for the next decade is to learn how do my intentions and my instructions and AI fit into the next step.
Because what happens when we turn this into a word called print. What does that look like?
What is the next abstraction? And the only people who are going to figure that out is us, because we're the ones standing on the back of giants. And in the future, people will be standing on our backs.
And until next time, happy learning.
The Hook
The bait, then the rug-pull.
The whole video hangs on one sentence boxed at the top of a whiteboard: build me a slide deck for my investors. Everything that follows is an inventory of what a machine has to know before that sentence means anything, and an argument that the inventory is shorter than the industry has been selling.
Frameworks
Named ideas worth stealing.
01:00list
The three unknowns behind any prompt
How you like the artifact built
Who you are
Who the artifact is for
Any real request to an AI hides three pieces of missing information. Every agent architecture is an attempt to deliver these three, and all three are expressible as plain text.
Steal forauditing whether a proposed AI integration is solving a real retrieval problem or a writing problem
02:30list
Data, context and tools
Data
Context
Tools
His standing shorthand for what makes a model stop producing generic output. Referenced as the frame his earlier videos established and the one folder structure satisfies.
Steal fora diagnostic checklist when a model's output is technically correct but useless
01:30model
The folder stack
Build instructions folder
Per-artifact skill subfolder
Me folder: brand voice, history, origin
People subfolder: investors, clients, collaborators
A nested directory of markdown and text files where each level answers one of the three unknowns. The model reads only the branches a given request touches.
Steal forthe first context repo for any repeated AI workflow
12:00concept
The harness
Read a file
Write a file
Search
The minimum general tool set that covers nearly every workflow. It is the only place code is genuinely required, and it already ships with every major model.
Steal fordeciding whether a proposed custom tool is actually new capability or redundant plumbing
18:40list
Opinions, instructions, data
Opinions
Instructions
Data
The three kinds of content a folder holds. A request pulls only the opinions, instructions and data relevant to what was asked, rather than loading everything.
Steal fornaming and splitting files inside a context repo so retrieval stays scoped
27:40list
Descartes' two tests for a thinking machine
It must not communicate like a brute, meaning it adds to a conversation rather than answering commands
It must do something outside what it was designed to do
Written in Discourse on the Method in the 1640s, long before any machine could be tested against it. Both tests are still the ones people reach for when arguing about model capability.
Steal forframing a debate about what current models can and cannot do
38:20model
The abstraction ladder
Punched card
Binary
Machine and assembly language
Compiler
Fortran
C
Python
English
Every rung took a repeating pattern from the rung below and collapsed it into a shorter instruction. The claim is that natural language prompting is simply the newest rung, not a break in the pattern.
Steal forexplaining to a skeptic why prompting is continuous with programming rather than opposed to it
CTA Breakdown
How they asked for the click.
VERBAL ASK
28:20link
“I'll have some links in the bio. I made some papers on it so you can just jump straight to that.”
The only ask in 43 minutes, dropped mid-sentence inside the Lovelace section and never repeated. The community link and seven source links live in the description, not the audio. The sign-off is a bare 'until next time, happy learning' with no subscribe prompt, which fits the lecture framing but leaves the entire conversion load on the description.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
An 11-lesson course that takes a complete beginner from typing into a chat box to running an AI agent inside their own folders, built around one idea: chat, then skills, then folders and one agent.
A mastermind member's question about landing clients turns into Jake Van Clief's longest, most honest answer on how to go from zero to paid, and why buyers really open their wallets.
A 90-minute live breakdown of Interpretable Context Methodology, why multi-agent swarms fail at enterprise scale, and how plain filesystem folders replace orchestration frameworks.
A 91-minute VIP community session: London Tech Week debrief, live ICM routing demo, Chicago executive workshop findings, and a first look at Microsoft SkillOpt.
00:44
01:20
01:42
02:30
02:56
04:02
04:17
05:05
05:23
06:09
06:27
07:13
07:38
08:17
08:39
09:28
10:26
11:30
11:51
12:51
13:30
14:12
14:55
15:47
16:20
16:43
17:11
17:50
18:59
19:22
19:59
20:37
21:09
22:26
22:35
23:26
24:00
24:05
25:10
26:00
26:33
26:47
27:28
28:05
28:49
29:01
29:32
30:14
31:08
31:44
32:07
32:52
33:27
33:46
34:57
35:21
35:58
36:44
37:11
37:38
38:33
39:25
40:05
40:30
40:46
41:35
42:00
54:04
16:25
1:31:31
23:18
1:31:43
22:03