Full transcript
0:01Hello everyone. My name is Jake Van
0:02Cleef and I'm going to do my best to
0:04explain ICM in under two or three
0:06minutes. Essentially, a lot of people
0:09are trying to build out contexts and
0:11agents and all of these things on top of
0:14the same model and the same harness and
0:16I argue those are just system prompts
0:18trapped inside code. Those are things
0:20that you don't need to do.
0:22Right now, a lot of the models
0:24themselves can already read and write
0:26files and are really capable on their
0:29own and building out these extra routing
0:31things for context is unneeded. Instead
0:34of building a planner agent or analysis
0:36agent or a writer agent that has to have
0:39all of that context stuck inside of it
0:41and it's not really mutable. Instead,
0:44you just take one good model with one
0:46good harness and point it at a folder
0:49structure that has all of those
0:50judgments, those prompts, those contexts
0:53inside of it. What do I mean by that?
0:56Well, essentially, I have an ICM that
0:58I'm going to run you through right now.
1:00Every weekend, I like to do a lecture
1:02where I talk about, you know, work or
1:04software, things that we're doing and I
1:06like to take that transcript and do a
1:08lot of different things with it. And
1:10traditionally, people would say, oh,
1:12well, let's make an agent that takes the
1:14transcript and converts it into all the
1:16different files so that every single
1:17time it does it the same way. And that's
1:20great, but I think that's the wrong
1:21mentality. Instead, I simply captured
1:24the way in which I like to edit the
1:26files, what I like to do with that and
1:28put those descriptions, those prompts
1:30into stages inside of a folder. So, if I
1:33was to do this manually, I would be
1:35creating sessions, I would be dropping
1:37my recording inside of those sessions
1:40and then I would be doing my work with
1:42it, right? So, in this case, I have my
1:44recording in here and then I would
1:46transcribe it manually, right? And I
1:48would get those transcriptions and then
1:50maybe I would do something with it. I'd
1:52look at the transcriptions and I'd add
1:54up all the questions that are in there
1:56or maybe I would create some sort of
1:58term sheets, or some sort of
2:01maybe some other kit or or digital
2:03artifact that people would want. Uh
2:06maybe I'd make some sort of slide deck,
2:07right? This is one of the slide decks
2:09that I actually made from one of my
2:11lectures.
2:13Um and I would need to be able to pull
2:14from various assets and things like
2:16that. Uh right? I would have uh I would
2:19have a certain way in which I like to
2:20design. I'd have fonts and UI kits. But
2:23if that's already all inside of folders,
2:25well then all I need to do is make one
2:27really good file that describes
2:30what all of those folders are and what I
2:33do inside of them, right? It's almost a
2:35simple file map uh as a markdown file
2:38that describes, "Hey, here's what's all
2:40in this folder. Here's what's in this
2:41folder. This is what I do in this
2:43workspace." And every time I go deeper
2:46into one of the folders, I might have
2:48another set of context or descriptions
2:51on what I do inside of that folder. And
2:54what that allows me to do uh is
2:56essentially automate that whole process
2:58by pointing one good model and one good
3:00AI at that folder, and it will do all of
3:04those things as if it was an agent. Yet
3:06I'm not stuck inside of an agent. I can
3:08use any AI. So when I point Codex or
3:11ChatGPT at this folder and say, "Can you
3:14create a uh package for the HighTide
3:1615?" It knows nothing about any of this.
3:18This is a fresh AI. But the first thing
3:21it reads is that agent.md or that
3:23claw.md.
3:25It immediately understands what the
3:27workspace looks like, what it needs to
3:29do, and it just starts working away,
3:31right? And then next thing you know, I
3:33end up with my finished product, my
3:35finished work or sets of works that I
3:38actually wanted. And the best part is,
3:40if I wanted to do this with any other
3:41AI, I don't have to change any code, I
3:44don't have to do anything advanced. I
3:45just point that AI, as long as it can
3:47read files and navigate them, at that
3:50folder, ask my question, and it becomes
3:54the agent I need based on just simply
3:56reading and navigating that folder and
3:59that structure. This allows me to also
4:01be able to edit and mold and do things
4:04at as I want with this structure. I can
4:07either automate the whole process, or I
4:10can just have the AI only do one part of
4:12the process, and I can go through and
4:14edit part of it. It allows me to go back
4:16and forth. It allows me to capture it
4:17all. It allows other people to access
4:20that folder simply uh and be able to
4:22navigate and use it if I use OneDrive or
4:24some sort of deployable folder software.
4:27Uh and it just allows me to avoid a lot
4:28of the agentic hubbub and work and
4:31simply just get straight to the outcome,
4:33straight to the work that I actually
4:35need. Just drop in the recording, and
4:37the AI will transcribe it, package it,
4:39and publish it as simply as I need, all
4:42inside of a folder structure, which is
4:44more editable by less technical people.
4:46It is more scalable, and it is simple to
4:48audit. Uh it's a little bit over uh 2
4:51minutes, but I think that gets the idea
4:52done. Uh if you want to learn more, uh
4:55just search in Jake Van Cleef
4:56Interpretable Context Methodology. I
4:59have research papers and everything else
5:00about it. Or you can listen in on any of
5:03the amazing lectures uh that are
5:05lecturing on all this as well. Thank you
5:07so much, everyone, and happy learning.