The argument in one line.
You can build a self-updating personal knowledge system that grounds your journal responses and CRM in your own saved articles, videos, and transcripts by combining Obsidian, Codex, and hourly AI automation.
Read if. Skip if.
READ IF YOU ARE…
- You save articles, videos, and transcripts regularly but rarely revisit them, and want a searchable wiki that surfaces relevant notes during conversations.
- A founder or consultant who journals daily about work decisions and wants AI-grounded reflection that pulls from your accumulated knowledge base.
- You meet people frequently at events or calls and struggle to retain conversation details and context for future interactions without a formal system.
SKIP IF…
- You're already using a mature PKM system like Roam Research or Logseq and have built custom workflows — this covers foundational setup, not advanced customization.
- You don't regularly capture external content (articles, videos, transcripts) from the web into a central store — the system's value depends on consistent input.
- You need this to work with your existing tech stack and can't adopt Obsidian + Codex as your primary tools.
The full version, fast.
A passive note vault is a graveyard; the fix is to turn captured material into a queryable wiki you can journal against and that responds grounded in your own saved knowledge. The build stacks Obsidian as the markdown front end, Codex as the AI back end, and the Obsidian web clipper to dump articles and YouTube transcripts into a raw folder, then follows Andrej Karpathy's LLM-Wiki architecture so an hourly automation summarizes sources, extracts people, tools, and themes, cross-links related notes, and moves processed files aside. Extending the same agents.md prompt adds a journal that pulls advice from your wiki and a lightweight CRM. Commit the vault to a private GitHub repo for backup, and the system compounds with every clip.
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 – 00:33
01 · Cold open — what the system does
Results-first preview: a wiki he can chat with, a journal that responds grounded in his vault, a CRM that remembers people.
00:33 – 01:35
02 · Why most second brains fail
The dumping-ground problem — info goes in, never gets reviewed. Sets up the three pillars: Wiki, CRM, Journal.
01:35 – 03:07
03 · The three-pillar diagram
Wiki at the center, CRM and Journal as connected modules. Inputs: articles, YouTube transcripts, meeting notes, tweets, podcasts.
03:07 – 04:47
04 · The full system spec
Save → summarize → extract entities (people, companies, tools, ideas, themes) → auto-link → journal-grounded responses → pattern detection.
04:47 – 06:23
05 · Wiki concept walkthrough
Entity pages (tools, people, companies) get auto-generated from raw saves; clicking a tool surfaces every video that mentioned it. Auto-linking creates a Zettelkasten-style graph.
06:23 – 08:23
06 · Sponsor — Hostinger + OpenClaw
Sponsored segment: one-click deployment of OpenClaw AI agents on Hostinger, code 'MattWolf' for 10% off.
08:23 – 09:40
07 · Credit to Karpathy + tool stack
The whole LLM-Wiki idea is Andrej Karpathy's. Required tools: Codex (IDE), Obsidian (markdown vault), Obsidian Web Clipper (Chrome extension).
09:40 – 10:31
08 · Obsidian Web Clipper demo
Pulls full YouTube transcripts into Obsidian with one click. Creates a fresh 'second brain' vault, deletes the welcome note, opens it as a Codex project.
10:31 – 12:00
09 · Build the wiki bones in Codex
Prompts Codex with the Karpathy LLM-Wiki GitHub URL. First pass over-builds 51 files; reprompted with 'remove all the extra crap.' Resulting structure: raw/, wiki/, agents.md, index.md, log.md.
12:00 – 14:30
10 · Configure the web clipper + first ingest
Dials in the clipper settings (vault name, default template, raw/ destination, front-matter fields). Ingests the LLM-Wiki page itself as the first source. Adds rule to capture YouTube channel name.
14:30 – 18:00
11 · First processing run
Processes raw/ — generates wiki pages (compounding knowledge base, environment design, identity-led goals, temporal discounting, temptation bundling). Index and log update automatically. The graph view starts forming connections.
18:00 – 19:30
12 · Batch ingest more videos
Pulls in 6 more videos from his watch history through the clipper. Six-minute processing run. Wiki + index expand; concept pages start linking to multiple sources.
19:30 – 21:20
13 · Chat with the wiki
Asks the vault for motivation tips for hard tasks. Codex queries the index, answers grounded in saved sources, then writes the answer back into the wiki as a reusable page.
21:20 – 23:00
14 · Two refinements — processed folder + back-linking
Adds a raw/processed/ archive so the inbox stays clean. Fixes the channel-name placement (front matter of the source, not the wiki page). Adds cross-linking so wiki pages back-reference their source notes.
23:00 – 25:00
15 · Wire up Journal + CRM in agents.md
Prompts Codex to extend the agent: 'journal' prefix opens a journal entry mode; CRM instructions add or update person records. Both get their own index.md and folder. agents.md grows three operating modes.
25:00 – 27:00
16 · CRM live test — Matthew Berman
Adds Matthew Berman to the CRM with three meeting touchpoints. Codex creates the record, updates the CRM index, logs the change. Demonstrates recall by asking 'where did I meet Matthew Berman?'
27:00 – 28:50
17 · Journal entry — clickbait dilemma
Brain-dumps the title-vs-clickbait struggle into a journal session. Response is grounded in saved creator-strategy notes plus LLM knowledge — names two prior vault pages (YouTube value of death, creator persistence) and structures advice around the 'two fears braided together' frame.
28:50 – 30:50
18 · Reprocess + Codex automations
Reprocesses raw/ to apply the new rules. Sets up a Codex hourly automation: 'if anything is in raw/, process it now.' Pipeline becomes hands-off — clip from the browser, the rest happens.
30:50 – 33:00
19 · GitHub backup layer
Creates a private GitHub repo, prompts Codex to commit + push. Extends the hourly automation to commit after each processing run — vault becomes versioned and backed up automatically.
33:00 – 33:56
20 · Recap + sign-off
Reviews what was built, teases that the graph view gets denser over weeks. Stack summary: Obsidian + Codex (or Claude Code / Cowork). Standard subscribe CTA.
Lines worth screenshotting.
- Most second brain systems are storage, not thinking — information goes in and dies there because retrieval requires the user to already know what they're looking for.
- A self-updating wiki built on top of saved knowledge is the architectural step that converts a dump folder into something an AI can reason against.
- Piping articles, YouTube transcripts, and podcasts through a Chrome web clipper into a raw folder, then letting hourly automations build the wiki, makes capture frictionless and processing automatic.
- A journal that queries your own knowledge base when you write in it turns journaling from reflection into retrieval — the system responds with what you've already learned that's relevant to what you're thinking through.
- A lightweight CRM built inside a second brain means context about people you meet lives next to the ideas they sparked, not in a separate tool.
- Grounding a journal AI in your vault instead of the model's training data means advice comes from your sources, your thinking, and your saved knowledge — not from generic internet consensus.
- Obsidian as the front end and Codex as the AI back end with hourly automation running between them is a minimal stack with no subscription cost beyond the AI API.
- Andrej Karpathy's LLM-Wiki architecture is the open-source foundation that lets anyone build a personal wiki that an AI can query — the innovation is the extensions, not the core.
- YouTube transcripts clipped into a second brain are more valuable than bookmarked videos because text is searchable, indexable, and usable as context for AI queries.
- A second brain that responds to chat queries grounded in your saved content answers the question differently than the same model with no context — the grounding is the entire product.
- Building a CRM inside your knowledge system rather than using a standalone CRM tool means people data and idea data share the same retrieval layer.
- The hourly automation that processes raw clips into wiki entries is what makes the system feel alive rather than static — the vault grows while you work on other things.
- Personal knowledge management compounds only when the retrieval layer matches the capture layer — a vault with no way to surface the right thing at the right time is just a better-organized file system.
- Showing the entire build process in 34 minutes proves the stack is buildable by a non-engineer — the existence proof is what makes the tutorial valuable.
- A second brain that answers from your vault rather than hallucinating from training data is a trust calibration — you can verify any answer by looking at the source it came from.
Wire Obsidian and Codex Into a Self-Updating Knowledge Base
Second brain build
Matt Wolfe builds a personal knowledge system on Karpathy's LLM-Wiki architecture — a vault that ingests, summarizes, cross-links, and chats back using only your own saved content.
01Cold open — what the system does
- Results first: a wiki you can chat with, a journal that answers from your vault, a CRM that remembers people — all from saved markdown
02Why most second brains fail — The dumping-ground problem
- Information goes in and never gets reviewed — the system must process and connect, not just store
- Three pillars fix the architecture: Wiki for knowledge, CRM for people, Journal for reflection
03The three-pillar diagram
- Wiki at the center, CRM and Journal as connected modules — inputs include articles, transcripts, meeting notes, tweets, podcasts
04The full system spec
- Save, summarize, extract entities, auto-link, answer grounded in sources, detect patterns — six stages in sequence
- Each stage adds signal that the next stage uses — the pipeline compounds on itself
05Wiki concept walkthrough
- Entity pages for tools, people, companies, and ideas get auto-generated from raw saves
- Auto-linking creates a Zettelkasten-style graph — clicking a tool surfaces every video that mentioned it
07Credit to Karpathy + tool stack
- Required stack: Codex as the IDE, Obsidian as the markdown vault, Obsidian Web Clipper for ingestion — all three are free
08Obsidian Web Clipper demo
- One-click pulls full YouTube transcripts into Obsidian — then open the vault as a Codex project
- Start clean: create a fresh vault, delete the welcome note, and let the system build its own structure
09Build the wiki bones in Codex
- Prompt Codex with the Karpathy GitHub URL — it builds the skeleton, then reprompt to remove unnecessary files
- Core structure: raw/, wiki/, agents.md, index.md, log.md — five components, each with a clear job
11First processing run
- One processing command turns raw saves into linked wiki pages — the graph view starts forming connections immediately
- Index and log update automatically — you see the system growing without any manual curation
13Chat with the wiki
- Ask the vault any question — Codex queries the index and answers grounded in your actual saved sources
- Answers get written back into the wiki as reusable pages — the vault gets smarter with each query
14Two refinements — processed folder + back-linking
- Move processed files to raw/processed/ so the inbox stays clean and the history is preserved
- Back-linking from wiki pages to source notes makes the graph bidirectional and more useful for research
15Wire up Journal + CRM in agents.md
- Journal prefix opens a date-stamped entry mode that responds grounded in your vault
- CRM instructions in agents.md add person records and update them — the same processing pipeline handles people and knowledge
Terms worth knowing.
- Second brain
- A personal knowledge management system — typically software — used to capture, organize, and retrieve notes, articles, and ideas outside of working memory.
- Obsidian
- A local-first note-taking application that stores files as plain Markdown on your computer, with a graph view that links related notes together.
- Codex (AI backend)
- In this context, an AI-powered reasoning layer that sits on top of a knowledge base and answers questions by retrieving and synthesizing stored information.
- LLM-Wiki
- A knowledge base architecture (associated with Andrej Karpathy) that uses large language models to automatically build and maintain a wiki from raw ingested content.
- RAG (Retrieval-Augmented Generation)
- A technique where an AI model searches a private document store for relevant context before generating an answer, grounding responses in specific saved material rather than general training data.
- Knowledge vault
- The full collection of saved notes, articles, transcripts, and documents that make up a personal knowledge management system, used as the source of truth for AI queries.
- CRM
- Customer Relationship Management — software that stores contact details, conversation history, and relationship notes about people, typically used in sales but applicable to any networking context.
- Chrome web clipper
- A browser extension that captures web pages, articles, or videos directly to a note-taking app or folder with one click, without manual copy-paste.
- Knowledge base
- A centralized collection of organized information — documents, notes, articles — that can be searched or queried to answer questions or surface relevant context.
- Journaling (AI-grounded)
- A journaling practice where an AI reads both the journal entry and the user's knowledge base simultaneously, offering responses or reflections informed by the user's own saved writing and research.
Things they pointed at.
09:02toolCodex (IDE / AI coding environment)
09:28toolObsidian Web Clipper (Chrome extension)
06:20productHostinger OpenClaw managed plan
10:32toolGranola (meeting note recorder)
25:00channelMatthew Berman (referenced as CRM test contact)
07:40toolDGX Spark (where Matt runs OpenClaw locally)
30:15toolGitHub (private repo for vault backup)
33:20toolAnthropic Cowork / Claude Code (alternatives to Codex)
Lines you could clip.
00:41
“Most second brain systems are just like storage. You dump your YouTube transcripts and your articles and your podcasts into one place. Problem is that's kind of where the information just goes to die.”
04:10
“The knowledge base sits at the center, and then everything else sort of connects to it.”
08:23
“This whole LLM knowledge base idea came straight from Andrej Karpathy.”
27:40
“I see you're struggling with ideas for videos. Well, you saved this video three days ago that says you should do this.”
30:15
“Whenever I come across stuff I wanna save, I just use the Obsidian web clipper and clip it into my raw folder. And every hour, it's gonna ingest that and turn it into one of the wiki pages.”
33:10
“All you really need is Obsidian and Codex. Anthropic's Cowork, or Claude Code, also works.”
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.
metaphoranalogystory
00:00So I just built a second brain knowledge management system that has an entire wiki built in that I can chat with. It will pull any information from that second brain when I chat. It's got a built in CRM.
00:11I can journal, and it will actually look at my Wiki knowledge base and try to help me with whatever issues I'm going through from my journal by looking inside of the Wiki. And, of course, it's got all of the content that I've saved from around the web, including YouTube videos and articles and tweets and podcasts and just tons of stuff that I've injected into this.
00:31It is all accessible directly from a chat or from journaling. It is really, really sweet.
00:36And I'm gonna break down how the whole thing works and how you could build one for yourself right now. Most second brain systems are just like storage. Right?
00:43You dump your YouTube transcripts and your articles and your blog posts and your podcasts and just everything that you're interested in, you just dump it all into one place. Problem is that's kind of . Unless you're, like, actively going back through and reviewing the notes all the time and searching through your second brain, it's just a dumping ground for information that I never go back and look at later.
01:08So for the knowledge management system that I'm gonna build, there's three core pillars that I wanna build for mine. Number one is the wiki slash knowledge base. This is where I'm gonna store, like, everything from around the web that I find.
01:19YouTube transcripts, articles, podcast transcripts, tweets, you name it.
01:24It all goes into this wiki knowledge base section. Number two is my CRM. So whenever I go to events and I meet people or I jump on Zoom calls with people, I wanna remember those conversations I wanna be able to recall them in the future.
01:37I also wanna store details about those people. How I met them, where I met them, some of the discussions we had, any sort of contact details I got from them, email, phone number, address, whatever.
01:48They can all live in this sort of CRM element of this bigger second brain that I'm building. And the third element is where this all gets pulled together, and that's the journal. Now, I'm a big journaler.
01:57I journal pretty much every single day. When I have good days, I journal about what went right, the things I'm excited about, gratitude, that of stuff. When I have rough days, I journal about the things that are bothering me.
02:08My videos not performing as well as I want them to. Having a creative block and not knowing what to make videos about. I do a lot of travel, and I debate a lot about whether the travel is gonna be worth it or not.
02:18I journal about pretty much everything in my business. So these are my three ideal inputs. For you, it might be clients or workouts, research papers, recipes, sales calls, classroom notes.
02:30The point is the , and then everything else sort of connects to it. The two elements that I think are probably the most useful to the most amount of people are gonna be the wiki and the journal. Maybe the CRM CRM isn't what you need.
02:43Again, maybe it's your classroom notes, your workouts, your recipes, etcetera. So here's a rough drawing of what I have in mind. So you've got your knowledge base that lives at the center of all of this.
02:53All of this knowledge is going to live in Obsidian. I'll get into the whole building process in a second. You're gonna save articles from around the web, YouTube videos, you know, podcast notes.
03:03Whatever you find around the web that's relevant to you, you're going to save it with a simple Chrome web clipper, and it's gonna save into your knowledge base. The CRM that I just mentioned, notes about people you met and where you met them and all that kind of stuff gets saved to the knowledge base. Meeting notes, I personally use granola to record my meetings and take notes for me.
03:21Those meeting notes can automatically be injected into the knowledge base, and then you have journal entries. This is the layer where you actually interact with your knowledge base. You journal on what you're dealing with right now, and ideally, it's gonna pull from the knowledge base that has all of this other information in it to ground the responses to your journal entries.
03:40This will make more sense as I go. Please excuse the PowerPoint style slide here, but I really wanna explain what I'm trying to build. So here's the system I imagine.
03:48You save your articles, your transcripts, etcetera, into this system. We'll do that using a web clipper. The AI layer in the background that we're gonna build then summarizes this stuff for us.
03:58So it's not just a giant transcript. It's actually sort of the bullets and just the information we need to know. The AI is also going to extract people, companies, tools, ideas, and themes, and sort of break those off.
04:09That's where it becomes kinda like a wiki. You could click into the tools page, and it will list off all the tools that have been mentioned across everything that we've saved. You can click into one of the tools, and it will mention where and what video that came from.
04:22And that goes for all of these little categories here. I also wanted to auto link related notes. So if I have multiple videos about how to build something with Open Claw, they all get cross referenced to each other, and I can click around and sort of jump into others.
04:36If you're familiar with other second brain systems or like the Zettelkasten system, it's essentially that same concept of interlinking. I'm then going to let the journal directly into the system.
04:46So when I do write journals, it responds like chat GPT, but it's actually grounded in my own saved knowledge. So it's not going to just respond with what chat GPT would have responded with. It's going to respond with something like, I see you're struggling with ideas for videos.
05:01Well, you saved this video three days ago that says you should do this, this, and this. You also saved this video two weeks ago that gave this advice when you're struggling with video ideas. And it will actually pull from the knowledge that I've saved, making it more tailored to exactly what interested me and what I found valuable over time as I save stuff into the system.
05:21It's also gonna use AI to find patterns from past journal entries. So if I'm constantly journaling on the same thing over and over again, the same struggle, it's gonna see that as a pattern and take that into account when it's responding to my next journal entries. It should resurface relevant ideas when I need them, and it should also let me save notes about people and connect those people to ideas, companies, events, and conversation.
05:41So the CRM and the meeting notes that I'm constantly saving, that should also be connected to ideas and journal entries and all of it sort of pulled together in this big mashup of, like, here's the conversations I've had, here's what I'm journaling on, here's what I've saved that I found interesting, and it's all in this big old soup of content that I'm saving.
05:59But that that I'm saving is the grounded information that is getting pulled when I write my journal entries and then get a chat response based on those journal entries. Alright.
06:09Now that I've completely overexplained this, let's just jump into the process I'm going through to build this. There's a cool new feature from Hostinger that's made for deploying AI agents at home. So if you've ever wanted to run something like OpenClaw but don't know where to start or find it really complicated, this is going to make the process a lot easier.
06:26I actually use OpenClaw myself so I can attest that the setup can get pretty technical. Right now, it's deployed on my DGX Spark, and my favorite thing it built, this simple CRM that when I meet people at events, I can save quick notes in Slack and then later ask it to remind me what we talked about.
06:43So it's great for automating everyday things like that to full on businesses and these huge agentic systems, which I've also seen people do. And this new feature from Hostinger makes it a lot easier to set up. You can either pick a managed OpenCLOB plan or a VPS option.
07:00The managed version is honestly the easiest. It's a one click deployment that includes built in AI credits, web scraping already connected, and they handle the updates, backups, and security for you.
07:11You just go to the Openclaw landing page on Hostinger, click get Openclaw, choose your plan.
07:16And by the way, if you do the twelve or twenty four month option, it's even cheaper per month. Then during setup, you choose your AI provider, pick your communication channel like Telegram or WhatsApp, and launch.
07:28And then once it's live, your agent runs twenty four seven. AI agents like this used to feel very gate kept. Like, only the really technical people or big enterprises were able to use them.
07:38But now it's pretty cool how easy it's become to set it up and use it for yourself at home. So if you've wanted your own AI assistant without the headache of managing servers manually, this is probably the easiest way I've seen to do it.
07:51Check it out at the link in the description box. And to save even more money, use my code Matt Wolf for an additional 10% off. And thank you so much to Hostinger for supporting my channel and sponsoring this portion of today's video.
08:03Now before I go any further on this, do wanna give credit where credit is due. This whole LLM knowledge base idea came straight from Andre Carpathi. I specifically took the idea of using Obsidian as the front end.
08:14Obsidian sort of helps organize and easily read markdown files. I'm just sort of extrapolating off of this idea and adding my journaling element and my CRM element to the wiki concept that Andre laid out here.
08:27Now in order to build this, you're gonna need a couple tools. I'm gonna build this in codex here. This has been sort of my IDE of choice lately to do coding and projects like this.
08:36This is free to download, and you get a certain amount of usage on the free chat GPT plan, but you're gonna get the most out of it if you do end up upgrading. I'm also using Obsidian here. Again, this is just a giant markdown organizer and reader.
08:48This is totally free to get. You can find it over at obsidian.md. You're also going to want the Obsidian web clipper.
08:55So if you're on the obsidian.md website, I can scroll all the way to the bottom here and you can see there's a link for web clipper. Click in there, click add to Chrome, and you'll get this little Obsidian web clipper that you can see here in my Chrome.
09:07And if I click this, this is what automatically creates a new markdown file, a new note for whatever page you're on. And the cool thing that I really like about this Obsidian web clipper is that it automatically pulls the transcripts from any YouTube video.
09:20So if I come over here to YouTube and click on one of my recent YouTube videos, and then I come up to the Obsidian web clipper, it'll take a second to load, but it will eventually load the entire transcript for this whole video straight into Obsidian, and that makes it really easy to inject any YouTube video or any article you find around the web directly into your Obsidian vault.
09:41So once you have Obsidian installed on your computer, create a brand new vault and I'm gonna call it second brain, and I'm gonna save it in a folder on my computer called second brain. Now I'll open that, and we'll create this new vault, and you can see I have a fresh blank vault with nothing in it yet except for a little welcome message.
09:59Now it's important to remember where on your computer you just save this vault because that's gonna be necessary in the next step here. I'm gonna delete this welcome message. It's not gonna be necessary.
10:09And now we have a purely empty clean vault. So for the next step, I'm gonna jump into codex and we're gonna actually build the dang thing. So over on the left here inside of codex, I'm gonna click on add new project and then I'm gonna select use an existing folder.
10:24It's gonna open up my browser here to pick the folder and I'm gonna go to the exact folder that we just set our Obsidian vault up with. So for me, it's this second brain folder that I created here, and we'll go ahead and open that. And then you can see I now have a project over here called second brain.
10:40So start this off, we're gonna build the basic bones of our wiki. And luckily, Andre Carpathi generously gave us this GitHub page that explains exactly how the Wiki architecture works.
10:51So the initial sort of hard part of building the Wiki is already figured out for us. We can just take this URL to this GitHub post here, open up Codex, make sure we're in our second brain project folder here.
11:03I'm getting it the prompt build out the wiki architecture based on Carpathi's LLM Wiki here. I'm linking to that page on GitHub that we were just looking at, and then I said the current second brain folder is the folder that Obsidian is connected to. It is currently empty, so we're building from scratch.
11:17And let's go ahead and let it build out the sort of architecture bones for us based on what Carpathi's already figured out. Alright.
11:25So it worked for about five minutes. It actually built out a whole bunch of extra files that it didn't need to build. I don't know why it created 51 files.
11:34The architecture is actually supposed to be pretty small for this. So I literally prompted it, please remove all the extra crap and just build what's explicitly called for in Carpathi's game plan. And it says done.
11:44I pruned it back to the minimal Carpathi game plan, and now we just have these files built in. If we pop open Obsidian here, you could see we've got just the folders we need. We have the raw folder.
11:55We have the wiki folder. We have our agents dot m d file, our index dot m d file, and our log dot m d file. We can see here exactly what each of these is for.
12:05The raw folder is for the immutable source material. This is where the original stuff goes. Raw slash assets.
12:11This is for optional local obsidian attachments. You got the wiki. This is the AI generated markdown files that is pulling from the raw content that we're inputting.
12:21You have the agents dot m d file, which basically explains how this whole thing works. So we can see it's got the operations when the user adds a source and asks LLM to process it. It does all these things.
12:31When the user asks a question, it queries it this way. So it basically tells it how this agent should operate. You have the index dot m d file.
12:39This is basically the catalog of everything that's in the wiki, and then you have the log file where whenever you make updates or changes or add things, it updates the log file. Super super simple. We're starting bare bones here.
12:50If I look directly in the folder, we just have what you see inside of Obsidian. So now, I'm gonna make sure that my Obsidian web clipper is dialed in, so I'll go ahead and click on this. We'll click on settings.
13:01Make sure you add the name of your vault right here under the vault list. If you're in Obsidian down in the very bottom left corner down here, you can see this is the name of the vault, so make sure it's the same name exactly. And then over under default, you've got the templates over here.
13:15Click on the default template and make sure that you select that second brain vault or whatever you titled it. And then I'm having mine pull in these properties, the source title, the source URL, the date that it was created. That's the date that I'm saving it in the web clip or not the date that the article was actually written, and it's adding an automatic web clip tag to it.
13:33And then for the note content, it's just pulling in the content. A lot of this might actually be set for you by default, but if it's not, this is what it should look like. Under note location, we're gonna change this to just say raw because that is the folder inside of our Obsidian vault that we want it to dump it inside of.
13:50Alright. So I can close out of this and for the very first thing I'm going to ingest, might as well ingest the instructions for how to build one of these wikis. I know it's very meta, but I want it inside of my wiki.
14:01I'll click on my little Obsidian clipper button and you can see the source title, LLM wiki. We've got our source URL, the date I'm pulling this in, and the tags for web clip. And then here is all of the content of this page here.
14:12We'll go ahead and click add to Obsidian and we can see it added it directly inside of the raw folder here inside of Obsidian. Now, nothing's going to happen automatically. We actually need it to tell it to process the files inside of raw for anything that actually happened.
14:26But let's add a few more things. I'm gonna look through my YouTube history and ingest some of the recent videos that I've watched, like this video called how to trick your brain into becoming so disciplined your friends will be shocked by your success. I love a lot of psychology and mindset type videos.
14:41If I want to ingest this, come up to my web clipper and you can see it's going to pull in the entire transcript here. So I'm gonna go ahead and click add to Obsidian. And once again, we've got another file here under raw.
14:52Now there's one issue that's going to pop up when I pull in YouTube video is it's not going to properly know the channel name because it's not automatically pulling it in into any of these properties. But I can go to codex here and give it some additional instructions. When I save a video from YouTube using the Obsidian web clipper, and then you go and process the files, make sure it also pulls the channel name from YouTube and adds it as one of the front matter fields.
15:20Alright. So let's go ahead and do that. So now let's go ahead and do a quick test.
15:23We've got two source files in here. So I'm gonna jump into Codex and go ahead and tell it to process the files inside the raw folder. Let's see how well it does right now.
15:33Alright. So it took about three minutes to process and it created a few new sections. So let's just go ahead and pull open Obsidian here and we can see it left the original source material here, but then it started to build out the wiki of everything else.
15:47So we've got our compounding knowledge base, which was clearly pulled from the explanation from Andre, discipline without willpower. This was pulled from this channel, Aaron Miller study. Let me just double check that it got the channel name correct.
16:01Yep. Aaron Merrill study. Environment design.
16:04We could see this was from the source discipline without willpower, which was one of the concepts that it saved, which came from this original video that we saved. Identity led goals, LLM Wiki, temporal discounting, and temptation bundling. So we can see our Wiki is starting to get built out, and then we have our index here, our various sources, the LLM Wiki and the discipline without willpower.
16:25It actually renamed it because it was originally called how to trick your brain into becoming so disciplined your friends will be shocked by your success, but it decided discipline without willpower was a better name for it. We could see the concepts here and it's starting to build out.
16:38And if we look at our log, we can see what it registered in our log so far since we started building this. One thing that I actually like to do as this gets bigger and bigger is you've got this graph view here that starts really small when you first build it. And over time, you'll see this build out and build out and things get more interconnected with each other and it just gets really cool over time.
16:58Now, I'm gonna go see this with some more content. I'm gonna go through my watch history and pull in some of the other videos that I watched recently using your money to be happier, the art of tripod filmmaking, how to become addicted to doing hard things.
17:10If you think you're too busy, watch this. How to become a lucky person, and then build your own self improving AI Wiki in eleven minutes.
17:17I know that's very meta, but let's go ahead and import that. So I'm just gonna go through and inject every single one of these like we just saw. I'm gonna let it process them all, and then they'll all be in the wiki.
17:27And then we'll move on to the next steps, which are building out the journal and the CRM elements that I mentioned earlier. Okay. So it's done ingesting all of those videos that I just saved.
17:36It took about six minutes here, and this is what my Obsidian looks like now. You can see all of the assets of stuff that I ingested into it, and the Wiki is getting built out quite a bit more. We've got our index here, and as you can see, the index is also getting built out more as well.
17:52If I click into, like, Hermes agent here, we can see we've got key ideas from this original video plus related content inside of our wiki. So codex capabilities, I click on this one and it jumps to the video from Riley Brown and the details around that one. Again, this is the very very simple basic setup of Carpathi's LLM Wiki.
18:13Now, I come into Codex again, we can essentially chat with the Wiki. So I come to my second brain folder, click on new chat, and I can ask questions like what are some tips for motivation when I don't feel like doing the hard task today?
18:25I know I saved a couple videos about this exact topic and we can see it's already saying I'll treat this as a wiki query. First, I'm checking the vault index, then I'll answer from anything already captured and add the reusable bit back into the wiki if it isn't there yet. Here's our final response.
18:40When you don't feel like doing the hard task, don't wait for motivation to arrive first. Treat it as a task design problem. Make the first few minutes smaller, easier, and more rewarding.
18:48Try this. Gives me a handful of tips, and this was all pulled and grounded from updated the Wiki based on the question that I asked.
18:57You could see that it changed the index dot m d file, the log dot m d file, and the Wiki motivation for hard tasks. So opening up my Obsidian vault here again, looking in my log, we can see that it actually logged this query motivation when avoiding a hard task, answered a query about motivation, and it even updated the index with it.
19:15And it created motivation for hard tasks and linked back to the original sources that it found this information from. So as questions, the wiki further and further and further builds out based on the questions you were asking.
19:27Now, there's a few things that I wanna do to clean this up a little bit and make it slightly more useful for me because right now, once it processes something, it just leaves it in this raw folder and this is just gonna build up and build up. And so what I wanna do is under this raw folder here, I'm gonna go ahead and create a new folder and I'm gonna call it process.
19:46Whenever it processes one of these files and adds it to the wiki, I wanted to move it to the process folder so I know that that has already been ingested. So now that I've got this process folder, I can simply come down to my agents file here and then tweak what happens when the file is processed. So if I come down here, we've got operations ingest when the user adds a source and asks the LLM to process it, read the source from raw, create or update wiki pages, update relevant entity concept topic overview synthesis or comparison pages, update index dot m d, append an entry to log dot m d.
20:18Well, now I can just add a number six and say move the source file from the root raw directory to raw slash processed. By adding that extra bit to the little prompt here, now it's gonna go through all these steps, but then move it into the processed folder.
20:34It also misunderstood me when I said to add the channel name. It thought I wanted it to add the channel name to the actual wiki generated page, but I wanted to add the channel name to the original source. That's what makes the most sense to me.
20:46So I'm gonna come to my agents section here and just tweak that as well. So right now it says for YouTube videos clipped with Obsidian Web Clipper, also open or inspect the YouTube source URL and add the channel name to the generated wiki page front matter. But instead of that, I'm gonna say add the channel name to the original source page front matter.
21:04I also want it to link back to the original source. So I'm gonna add a step right after step three here and say, cross link any Wiki pages generated or updated to the original source page. Basically, don't want these pages orphaned.
21:17If there's a new Wiki page here, I want it to link to the original page here. So that's the manual way to update the agents dot m d file, but you can also do it by prompting it inside of Codex. So if I come to my second brain project, create a new chat here, I can give it instructions on additional things that I want to happen.
21:35So I mentioned my journal and I mentioned my CRM. So let me go ahead and build the bones for that here. I will close these folders to clean everything up.
21:43I will create a new folder called journal and a new folder called CRM. Now, can come to Codex and say, update the agents MD file to handle these items. Number one, if I start a chat with journal, add the text of that chat and subsequent conversation as a new MD file within the journal folder.
22:03The entire conversation should be added to the markdown file. Create an index file in the journal folder that's similar to the wiki index file. Each new journal entry gets added to the index file.
22:12Decide on a short title for the journal entry based on the contents of the journal and use the date and the title as the journal entry file name. Add the date and title to the index and link to the entry. Also, log the journal entry title and short summary in the log dot m d file.
22:27Your response to my journal entry should be grounded in content from the wiki in the same way you view the index and respond to my chat questions based on what's in the wiki, provide advice and insights to my journal entries based on what's available in the wiki, as well as your own LLM knowledge. Provide helpful advice, insights, guidance, tactics, and ideas using what you know along with what's available from the wiki, past journal entries, and the CRM.
22:49So when I journal, I want it to look in the wiki, find information that's helpful to what I just journaled on. I want it to look in past journal entries to see if there anything relevant I have journaled on in the past, and I wanted to look at my CRM and see if there's conversations I've had with people about what I'm journaling about.
23:04Two, for the CRM. If I tell you I'm giving you information for the CRM, either update the person in the CRM or add the person to the CRM. CRM files should always be a person's name.
23:14I will share details about a person, their name, contact details I have for them, details about where or how we met, things that I know about them, etcetera. Create or update the contact record in the CRM with whatever details I give you. In the CRM folder, create an index file similar to our other index files with the name of the people in the CRM listed in alphabetical order and a short bio of what information I have about that person.
23:34This will allow me to ask questions about contacts that are inside the CRM. I wanted to update these two things in our agents dot m d file which will make it so that whenever I chat with my second brain project here in Codex, it's either one going to answer the question that I asked it using the sort of query task that's already built into the agent, it's going to handle it as a journal if I preempt it with journal here or if I tell the chatbot that it's for the CRM, it will update the CRM section.
24:03So I'm gonna let it go ahead and update our agents dot m d file, create the various index files, and that should build out the system for these elements. Okay. So we can see that it updated the agents m d file with our journal rules, our CRM rules, and if we open up Obsidian once again, I can open my journal folder and you can see we've got an index with date entry and summary.
24:24And if I open up my CRM, we've got our index with the person and the summary, and where most of the updates happened was this agents dot m d file. So now below the wiki section, we have a journal section. It built all the rules that I just asked it to build for me, and then we also have our CRM section.
24:40So it created all the rules for the CRM section now. Let's jump into Codex one more time, and let's test this out. I'm just gonna say add to CRM, and then I'll put Matthew Berman, and then I'll just put met at Qualcomm event in 2024, hung out at CES in 2025, got lunch during TechCrunch disrupt in 2025.
25:07And then I would obviously add whatever sort of contact details I have about him, but no point in doing that on camera. I'll just tell it to add that to the CRM and we could see updating the CRM record for Matthew Berman, creating a new record, and it says, great. I created that.
25:20If I look in my Obsidian here, we could see under CRM, we now have a Matthew Berman CRM here with our summary. No contact details provided, how we met, relationship context, and, you know, other stuff that's still empty right now.
25:33And then if I look at my index, can see it added the person Matthew Berman and a quick summary there. It also added it to the log down here that the CRM was updated. But the idea being is, let's say I'm going to a conference and I bump into somebody and I'm like, oh, shoot.
25:47Where do we meet? I could jump into my second brain and say, where did I meet Matthew Berman again? I'll check the vault and CRM records.
25:54Met You Matthew Berman at Qualcomm event in 2024 according to your CRM note. So a really good way to reference information later. I'm gonna go ahead and make a pretend journal.
26:03So what I'll do is I'll click on a new chat inside of my second brain project here and I'll start it with journal and then I'll go to another line and I'll just brain dump some thoughts. So this is just a simple demo journal entry, but you know, it's got some reality to it. I have a ton of ideas for videos but decided not to make them because I'm worried they won't get a good view count.
26:23I also talk about how I know how to write titles that will get a lot of clicks, but I also hate the whole clickbait thing. So I end up making titles that I know will get less clicks because they're more literal to what the video is about, and I constantly grapple with making videos that will get a lot of clicks, but knowing I'll get hate in the comments because they're too clickbaity versus making videos that the title is very literal, so I won't get the hate comments, but I'll also get a lot of Let's views.
26:51And that's something that I grapple with, and I know pretty much every YouTuber grapples with. And so that's my short journal entry. So let's go ahead and submit that and see what happens.
27:00I'll treat this as a journal entry and grab the response in the vault first. Going to read the indexes, look for relevant prior notes around creating video attention, then I'll write the journal file before answering. The vault has several creator strategy pages.
27:12I'm pulling those in rather than answering from a blank slate. There aren't prior journal entries yet, and the CRM doesn't appear relevant to this one. So it's just pulling from the wiki.
27:20Okay. And check out this journal response here. I think there are here, creative integrity and channel safety.
27:26One part of you does not wanna betray the video with a title that feels gross. Another part is afraid that a few underperformers will teach YouTube not trust you. The creator notes in the vault push back on the second fear.
27:35YouTube value of death and creator persistence both frame visible metrics as lagging indicators. A few lower view videos are not a downward spiral, blah blah blah. So it is actually grounding the response to my journal and what is existing in the wiki, and then it actually goes on to just give a response that chat GPT would normally give, but it's pulling from both.
27:55The knowledge built into the LLM as well as the knowledge in the wiki. So that's all working exactly as intended. And then if I wanna pop open Obsidian here, I can open my journal folder and we can see here is my journal entry.
28:08Here is what the assistant replied with. Here's some synthesis and here's some related content about what I'm journaling about. If I look at my index here, it automatically logged the date, the entry title, and a quick like one paragraph summary of what my journal entry is about.
28:23Now, I want some of this to happen on autopilot. I don't wanna have to jump into codex constantly and tell it to process everything that I saved in here.
28:32And there's an easy solution for that. Before I show you, I'm gonna do one thing. I'm gonna go ahead and clean up my existing Obsidian Wiki because remember, I want these to all get moved into processed.
28:42And I also want it to save the YouTube channel name up here in the front matter. So I'm gonna tell it, please reprocess all of the files in the raw directory following the recent updated instructions on how to process them.
28:59So I'm gonna let that process. It's gonna clean up my wiki real quick. Alright.
29:02Now that that's finished and my Obsidian vault is cleaned up and we can see all of my videos have been moved to process and it added the channel name to the original videos here. Let's automate some of this. So in codecs here, you've got a feature called automations.
29:16This is where you can set it up to do recurring tasks. So if I click into automations here, select new automation. For automation title, we'll call it process second brain raw files.
29:26For work tree, I'm gonna set it on local so it runs directly in the selected project. For our project, we're gonna select second brain. Here, we'll select when we want it to run.
29:34I'm gonna go ahead and set mine to hourly, but you can do it at whatever cadence you want. And I'm just going to say if there are any unprocessed files inside the raw directory, please process them now.
29:46For the model, I'm gonna set it to GPT 5.5. I recommend just using the strongest model you have available. I'm gonna set it on high reasoning and I will create this automation.
29:56Now, it's gonna run every hour. See if there's anything in my raw folder that's unprocessed and then it will process it and that's it. That's the whole process now.
30:05Whenever I come across stuff I wanna save, I just use the Obsidian web clipper and clip it into my raw folder automatically. And every hour, it's gonna ingest that and turn it into one of the wiki pages. If I wanna add somebody to my CRM, I just open up Codex, create a new chat inside of this project, and add the CRM details.
30:21If I feel like journaling right now, I can journal straight into my second brain, and it will ground the response in what's available in my wiki, past journals, and within my CRM. If you want an extra layer of backup, you can also go to GitHub, create a new repository on GitHub. I'm gonna go ahead and call this one second brain.
30:39I'm gonna set this to private so it's only available to me and I'll create the repository. Now, if I copy this URL, jump back over to Codex here, I could say commit this current version to my private GitHub repo here, paste in the URL of the private GitHub repo I just created and then go ahead and submit it.
30:58Now, I've previously attached the GitHub plugin, so it should just work out of the box. You can see I'm already synced with GitHub, but if you haven't done that already, just add the GitHub plugin and go through the motions to get that set up and you should be good to go. And it went ahead and pushed it to GitHub.
31:12So if I open my browser here and refresh, we could see everything I've created is all saved on GitHub now. If I jump back into Codex, I can go to my automation and edit this automation and say if there are any unprocessed files in the raw directory, please process them now.
31:28Once everything is processed, please commit and push the current version of the directory to the main branch on GitHub. So now it's gonna process everything in the raw directory and then once it's done processing, it's going to update GitHub so that backup is constantly happening every hour. And there you have it.
31:42It. There's the whole second brain process. Not only do you have a wiki of all of the information you're finding and saving from around the internet, but now you have a journal and a CRM that's built on top of it as well.
31:52And if you ever wanna tweak how it operates, you just open up Obsidian, which is your visibility layer to see how everything is built, and you go into agents.md and you just tweak the instructions.
32:03This is all just prompts at the end of the day. You just change how it gets prompted. And in the short amount of time that we've been working on it, this is what our graph view looks like now and we can start to see all of these things interconnect with each other a little bit more.
32:15Do this for a few days and a few weeks and the next thing you know, you have a vault that looks like this. Yeah. Pretty insane crazy vault that just has a ton of information saved inside of it.
32:26So I know that video was long. There was a lot of details. I wanna make sure it was very clear and you got the whole process and I didn't skip any steps.
32:33But I wanna show you how I've been building this sort of second brain concept that I can journal on top of, that I have a CRM on top of, where all of the wiki elements are interconnected. You could even dial it in more by building in separate folders and telling it to break out people from different pieces of content and break out companies from different pieces of content and really really dial in that wiki more and more and more.
32:56But really really cool concept. All you really need is Obsidian and Codex. Anthropics, Cowork, or Claude Code also works whatever your sort of front end chat platform of choice is.
33:07I've been really liking Codex lately, so that's what I've been using. But, yeah, go build it. It's really, really cool and a lot simpler than you think.
33:15And over time, it just gets smarter and smarter and smarter and more and more powerful. So that's what I got for you. This isn't like my normal videos.
33:22I normally make end of week news breakdowns where every single Friday, I break down all of the news that happened in the AI world for the week. I drink from the fire hose all week, keep up with all of the news, keep myself completely looped in so that other people don't have to feel overwhelmed. I'll take on the overwhelm and just report what I think is the most interesting.
33:41I put those videos out every single Friday. If stuff like that as well as tutorials like this are something that interest you, maybe consider liking this video and subscribing to this channel. It really, really helps me out a lot.
33:51But, again, that's what I got. Thanks for hanging out with me, nerding out with me. Hopefully, I'll see you in next one.
33:55Bye bye.
The Hook
The bait, then the rug-pull.
Matt Wolfe opens with a 30-second results-first montage: a vault that he chats with, journals into, and stores a CRM inside — all grounded in his own saved knowledge. Then he flips into 'and here's exactly how I built it,' positioning the next 33 minutes as a follow-along rather than a flex.
Named ideas worth stealing.
01:35list
The Three Pillars of a Useful Second Brain
- Wiki / Knowledgebase
- CRM
- Journal
Wiki holds saved knowledge; CRM holds people; Journal is where you interact with the system and let it ground responses in everything else.
Steal forAny personal-knowledge product positioning — most PKM tools sell pillar 1 and stop. The unlock is pillars 2 and 3 turning storage into a loop.
08:23model
Karpathy's LLM-Wiki Architecture
- raw/ (immutable sources)
- wiki/ (AI-generated entity pages)
- agents.md (the operating instructions)
- index.md (catalog)
- log.md (audit trail)
Five-element folder/file structure that turns a markdown vault into a self-extending wiki. raw/ holds the originals; wiki/ holds derived entity pages auto-generated by an LLM that follows agents.md.
Steal forDirect template for ModBrain — Joe could clone this structure exactly and bolt his own modules on (clips, scripts, characters) the same way Matt bolted on CRM + Journal.
17:00list
The Six Ingest Operations (agents.md spec)
- Read source from raw/
- Create or update wiki entity pages
- Cross-link wiki pages to original source
- Update index.md
- Append entry to log.md
- Move source from raw/ to raw/processed/
Numbered checklist the agent follows on every ingest. Acts as the contract between the human and the LLM — change the checklist, change the behavior.
Steal forAny agent prompt that needs predictable side-effects. Number the steps; the model will follow the list.
27:50concept
Two Fears Braided Together (journal response)
When the AI saw Matt's clickbait journal entry, it reframed the problem as two distinct fears stacked together: creative integrity vs. channel safety. Naming the two fears separately is the unlock.
Steal forUse this 'name the two forces' move as a journaling prompt OR a content-strategy lens. Works anywhere you're stuck between two values.
04:37concept
Zettelkasten-style Auto-Linking
Every entity page links to every source that mentioned it. Click a tool, see every video it appeared in. Same actor as Zettelkasten / Andy Matuschak's note-graph, but built by the LLM at ingest time instead of by hand.
Steal forIf you're building a content library, every clip/quote should auto-link to every entity it mentions — characters, frameworks, products. Cross-reference is the moat.
How they asked for the click.
VERBAL ASK
33:36subscribe
“If stuff like that as well as tutorials like this are something that interest you, maybe consider liking this video and subscribing to this channel.”
Soft and standard — no urgency or special promise. Earns trust by being low-pressure, but leaves audience-development upside on the table.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
