YT YouTube2Text

YouTube2Text → Transcripts → Complete Guide to building an AI Second Brain with Claude

Complete Guide to building an AI Second Brain with Claude — Transcript

by Ray Fu · 2,899 words · 424 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hey guys, welcome back to the channel.
0:01If you are new here, my name is Ray and
0:03I quit my software engineering job at
0:04Meta last year to help you learn AI and
0:06make money and run your own business
0:07with AI. I recently built a fully
0:09powered AI second brain that runs across
0:11every single part of my daily life and
0:12my business. It indexes every meeting
0:14that I take, triages my inbox, and buys
0:16me back at least 15 hours every single
0:18week. Today, I'm going step-by-step
0:19through how you can build the same AI
0:20brain that I use and become part of the
0:221% of AI users in the world. When people
0:25ask me, "What should I build with AI?
0:26I'm a beginner." I always tell them to
0:27build an AI second brain because it has
0:29tremendously helped so many people and
0:30is one of the most important things to
0:32build out when you're using AI. Now, I
0:34know what you're thinking. Why would you
0:35spend weeks configuring your second
0:36brain when dozens of productivity apps
0:38already exist? But because till now,
0:40everybody thinks that a second brain
0:42means spending $40 a month on SaaS
0:44subscriptions, adding highlights to
0:45Notion, and bookmarking YouTube
0:47tutorials that you'll really never
0:48watch. You're basically just dumping
0:49data that you'll never read. When I
0:51started my solo business, I was wasting
0:5210 hours a week organizing notes
0:54manually. I tried Notion, Apple Notes,
0:56Google Drive. I tried a lot of them
0:58several times and abandoned all of them.
1:00The main reason that I was stuck was
1:01because I was using these tools as
1:02storage instead of a proper system. And
1:04also, before you get started, like with
1:06all of my videos, I gave away a free max
1:08subscription of Codex or Claude to my
1:10viewers for every single one of the
1:11videos that I make. Make sure to keep
1:13watching to learn how to get it. Okay,
1:14so now we're picking the winner for last
1:15week's video for that ChatGPT max or
1:17Claude max subscription. Here is the
1:19winner. Boom. If you want, make sure to
1:20email me or DM me on Instagram and I'll
1:22send you over that gift card. Once
1:23again, I give out one max plan in every
1:25single one of my YouTube videos. If you
1:26didn't win, just comment for this week's
1:27video. To understand why you need a
1:29second brain, let's go a bit back. A few
1:31months ago, Andrej Karpathy posted about
1:33building a personal knowledge base with
1:34AI. It went viral and many people tried
1:36to build it. Either they failed or they
1:38built it and had no idea how to get the
1:40best out of it. The reason why it became
1:41so popular is because it's a system that
1:43really remembers everything that you've
1:44added in the past and helps you remember
1:46stuff that you didn't even know existed.
1:48For example, one time in the past, you
1:49might have, you know, added a YouTube
1:50tutorial to do a specific thing that you
1:52totally forgot about, it remembers it
1:54for the future. And it also proactively
1:56reminds you and pushes your goals
1:57forward based on the things that you
1:58added in the past. Today, we're actually
2:00going to be building this AI second
2:02brain. I'm going to walk you through the
2:03exact full setup, show you the
2:05three-layer architecture, and give you
2:06the exact commands, and show you how to
2:08automate the entire thing so it gets
2:10smarter and smarter by itself. Okay, to
2:12understand why this system works, you
2:13have to understand why standard racking
2:15tools like Notebook LM hit a wall trying
2:17to build a second brain. If you don't
2:18know what they are and haven't used
2:19them, make sure to look them up because
2:20they are also important tools that you
2:22need in your everyday lives. Most
2:23people's experience with working with
2:25documents is like you upload a
2:27collection of PDFs or notes, then you
2:28ask a few questions, and then AI
2:30retrieves relevant chunks from those
2:31files. It synthesizes an answer and
2:33presents it. That is how Notebook LM,
2:35ChatGPT file uploads, and standard
2:36vector databases work. The issue with
2:38that model is that the AI rediscovers
2:40knowledge from scratch every single time
2:42and with every single question. There is
2:43zero accumulation. If you ask a complex
2:45question that requires synthesizing five
2:47different documents, let's say, you
2:48uploaded over the last 3 months, the
2:50model has to piece together fragments of
2:51data every single run. The moment you
2:53close that tab, that synthesis
2:55disappears. Part of this concept is that
2:57instead of retrieving from raw documents
2:58at query time, you have the AI
3:00incrementally build and maintain a
3:01persistent wiki about yourself and
3:03everything that you know. Okay, so you
3:05can think of it like software
3:05engineering. Your raw notes and articles
3:07are your raw inputs, Claude Code is your
3:09programmer, wiki is your persistent code
3:12base, and Obsidian is the IDE where you
3:13view the output itself, much like a
3:15front-end display. When you add a new
3:17source, the AI does not just store it
3:18for search. It reads the source,
3:20extracts the core entities, and actively
3:22integrates into the existing wiki. The
3:23knowledge is compiled once, and then it
3:25keeps updating. But, the tedious part of
3:27building a personal knowledge is
3:29updating cross-references, filing notes,
3:30and keeping summaries up-to-date. But,
3:32with Claude Codex reading these files,
3:34you don't have to do all of those manual
3:35tasks every time. Okay, let's start
3:37building this AI brain now. To run this,
3:38you need two tools. First is Obsidian.
3:41You can download it at obsidian.com.
3:43It's completely free. It's a local
3:44markdown editor that lets you view your
3:46wiki through clean text files, folders,
3:48and visual graph views. The thing I love
3:50about Obsidian is that it is available
3:51on desktop and mobile, so I can add
3:53notes from any device I want. It's
3:55pretty handy to just dump any of my
3:56stuff anytime. Then, you need a Claude
3:58subscription. First, download the
3:59desktop app at claude.com. This is your
4:01AI agent that reads your files, executes
4:03your scripts, and maintains the wiki.
4:05And you can also use Claude X here if
4:06you're a ChatGPT lover. Now, we're going
4:08to talk about how to get that free $100
4:10max subscription. Basically, all you
4:11have to do is subscribe, like the video,
4:13and comment below how you use AI in your
4:15business or in your daily lives. I'll
4:16pick one person from the comments, and
4:18then I'll send them the credit/gift card
4:20in order to get that max subscription.
4:22Okay, after you have these two items set
4:23up, let's set up the vault. Open folder
4:25as vault. Create a new folder on your
4:27computer and name it whatever you want.
4:28I call mine Ray Brain. you dump all your
4:30big files into one folder, your AI will
4:32hallucinate. When I first tested this by
4:34dumping all of my files, Claude Code had
4:36a 60% accuracy rate when answering the
4:38questions. The second I applied the
4:40strict structural boundaries, that
4:41accuracy jumped to 95%. We'll also add
4:44three layers in here so AI gives you
4:45even more accurate answers personalized
4:47just for you. First, let's create a
4:49folder called raw sources. This is the
4:50dump folder you want to add articles,
4:52PDFs, cost transcripts, screenshots, and
4:55research papers, and anything you want
4:57AI to check while answering. One
4:58critical rule that you should add here
4:59is that AI reads from them, but is never
5:01allowed to edit or delete them. This is
5:03important. Okay, then let's create a
5:05second folder called wiki. This is where
5:07AI writes the organized knowledge base.
5:08It contains concept pages, entity
5:10breakdowns, and comparisons. Here, the
5:12rule is opposite. AI owns this folder
5:13entirely. It creates the files, updates
5:15them, and when new data arrives, it
5:17maintains the cross-references. Inside
5:19the wiki, you have two critical system
5:20files. First is index.md. This is just a
5:23structured catalog of every single page
5:25in the wiki with the link and a one-line
5:27summary grouped by category. Claude
5:29reads this file on every query you give
5:31to locate relevant files. Then, you have
5:33a log.md. This is just a chronological
5:36record of what happened and when. Every
5:38query and health check gets timestamped
5:39here so the AI knows when it has been
5:41modified, how recently. It's like a
5:43change log of your wiki, basically. For
5:45now, I've named these files and created
5:47a proper structure. I'll use Claude to
5:48populate these files with data. Finally,
5:50you have the single most important file
5:52in the entire repository. It lives in
5:54the root directory and tells Claude Code
5:55how your brain is structured, what the
5:57rules are, and what commands to follow.
5:59All you got to do is create a file
6:00called claude.md right in the root
6:02folder. In my free guide below, I give
6:03you the exact copy of the schema prompt.
6:05It's in the description. Just copy that
6:07prompt, paste it inside, and boom, your
6:09AI knows now the exact rules of how to
6:11build your brain. Now that you set up
6:13your Obsidian files, it's only an
6:14organized folder of basically markdown
6:16files, your AI reads all of that context
6:18and will still reply like a generic
6:20chatbot until you make Claude think,
6:22analyze, and write just like you. For
6:24this, you have to give it an explicit
6:26identity layer. In our system that lives
6:27in three files at the root of the vault.
6:29Let's go over these. First is the user
6:31at md. It defines who you are. In my
6:33file, it says, "Ray Fu, ex-Meta software
6:35engineer based in New York City, running
6:37an AI automations community and creating
6:38content. Prefers direct communication,
6:40fast iteration over perfectionism,
6:41concrete numbers over vague adjectives,
6:43and simple composable systems over
6:45complex architectures." You can create
6:47anything that you want based on how you
6:48want it to respond and write to you.
6:50Then, I have my soul at md, which is
6:52essentially defines who the AI is to
6:54you. In my file, it says, "You are an
6:55elite technical chief of staff and
6:57co-founder. You are direct, analytical,
6:59and honest. Challenge my assumptions.
7:00[snorts]
7:01Never hedge my words like try, hope,
7:03maybe, or depends. If an idea is
7:05over-engineered or sloppy, point it out
7:07immediately before executing." And last,
7:09I've a file that is super important to
7:10me. It's called the anti-AI style at md.
7:13It's a strict blacklist that prevents AI
7:15from producing generic AI slop. Every
7:17time Claude writes a draft and uses a
7:19phrase that makes me cringe or like
7:20those triple three M dashes, any of
7:22those words, I just open this file and I
7:23just ban those words. And look, you do
7:25not need to type these files yourself.
7:26Just open Claude code, turn on voice
7:28mode, or use a tool like WhisperFlow,
7:29and paste in the interview prompt from
7:31my guide. The link is in the description
7:32once again. Yep, for just 15 minutes
7:34answering all the questions Claude asks
7:35you, and then Claude creates these files
7:37for you. Now that the vault structure
7:38and your identity are locked in, you
7:40have to feed the brain with information
7:41every single day. To be honest, doing
7:43this manually every day is a huge
7:44hassle. For this, I use the Obsidian Web
7:47Clipper extension to scrape articles,
7:49blogs, and relevant info I come across
7:50every day. You can use any sort of web
7:52scraping you want, but this is just
7:53already in Obsidian, so it's really easy
7:55to connect to and use. To install this,
7:57go to Chrome, install the official
7:59Obsidian Web Clipper extension, then
8:01open it in settings, set its destination
8:02folder to raw sources. Whenever you find
8:04a great article, you just hit the
8:05clipper, and boom, it drops right into
8:08the raw sources. Then in your terminal,
8:09all you do is type in this one command
8:11from the guide, step away, and let
8:12Claude cook. Claude code scans the raw
8:14files, parses thousands of words in
8:16seconds, extracts the core concepts, and
8:18starts generating individual markdown
8:20files inside the wiki. It creates the
8:22master index, logs operations in the
8:24log.md, and links every concept using
8:25Obsidian wiki links. Your raw sources
8:27stay completely untouched. There are
8:29thousands of other web scrapers on the
8:31market that you can use to feed data
8:32into your second brain, but I highly
8:34recommend this one. Whenever you add a
8:35new article or PDF to raw sources, you
8:37just tell Claude that you've added a new
8:38source and instruct it to update the
8:40connected files. Okay, now that we added
8:42a bunch of new raw files, let's see how
8:44it connects to our existing files. To
8:46check this, open the graph view in
8:47Obsidian, and then boom, this is
8:49beautiful. Look at what just happened.
8:51Every single concept is connected by
8:53lines. This is why an LLM wiki destroys
8:55standard search. It functions as an
8:57associated knowledge graph. When you ask
8:59Claude code a question, it is not brute
9:00force search through across all of your
9:02files. It opens index.md, locates the
9:05relevant cluster, checks the wiki links,
9:07and gives you a synthesized answer with
9:09exact citations. However, the most
9:11important point is to remember the rule
9:12that makes this system compound. Always
9:14save good answers back into the wiki as
9:16new pages. If Claude generates a
9:17brilliant comparison table or a
9:18technical breakdown based on your
9:20questions, do not let it disappear. Tell
9:22Claude to save this analysis as a new
9:24page in the wiki and update index.md.
9:26Through this, every question you ask
9:27makes the knowledge base smarter over
9:29time for the next question. Okay, now
9:31look, you just saw the core blueprint,
9:33but I know how this usually goes. You
9:34watch a video like this, get excited,
9:36close the tab, and never actually
9:37implement it. Or you try to set it up
9:39and then hit an error and never get the
9:41effort to try again. I saw this and
9:42started a community of founders and
9:43operators. I help them make money in the
9:45community. People exchange ideas. It's
9:47just a brilliant network of people. I
9:49walk every member through every single
9:50step inside our community. You can join
9:52the community below at
9:53school.com/racfoods. It's also in the
9:55description below. And one more thing,
9:56inside the community you get these
9:57weekly live build calls where I
9:59troubleshoot your system and build
10:00custom workflows in real time. Also, I
10:02have the exact playbooks you can use to
10:04sell these automation systems to
10:05businesses for 3 to 10K. Here are
10:06examples of wins and reviews from my
10:08community already. Okay, sorry about the
10:09community plug. Let's go back to our
10:11setup. Once your wiki is running, you
10:13can add three automate workflows that
10:14turn into an agentic operating system.
10:16The first one is the 7:30 morning
10:18briefing. First, set up a script on
10:20simple cron job that runs every morning.
10:22When you open your laptop with your
10:23morning coffee, you shouldn't open a
10:25long task list or a messy email inbox.
10:27Just open your terminal and Claude hands
10:29you a clean summary of what needs
10:30attention and what new orders sent to
10:32your ecosystem. Two, processing meeting
10:34transcripts. This is the second workflow
10:36that you can set up. After any long
10:37meeting drag your raw transcript into
10:39your raw sources folder and tell Claude
10:41to add key points, actions, and
10:42decisions to the wiki. I love this
10:44personally because I have like 10 to 15
10:46meetings every day and so hard to
10:47remember every action item from these
10:49meetings even though I'm keeping
10:50transcripts of them. What I usually do
10:51is what you should do as well as connect
10:53a meeting note taker MCP to Claude so
10:55that automatically syncs all of your
10:56meeting transcripts. I use Fireflies and
10:58Calendly for this. Three, weekly health
11:01checks. This is the last one. Once a
11:02week, tell Claude code to optimize your
11:04vault. This prevents context rot. If an
11:07agent writes something slightly
11:08incorrect, it catches it before it
11:09corrupts downstream answers. This should
11:11cover your health check for your second
11:12brain. You got to update your brain and
11:14have your backups ready. To do this,
11:16just install the free Obsidian Git
11:18community plugin. Set it to auto commit
11:20and push your vault to a private GitHub
11:22repository every 15 minutes. This helps
11:24you have automated cloud backups and
11:26complete version history across all of
11:27your devices. And that's pretty much it.
11:29Now you know exactly how to build your
11:30AI second brain. You don't need to build
11:32this entire system in one day. Just
11:34start by creating a raw folder
11:35structure, adding the Claude.md, and run
11:37your first data ingest. I have created a
11:39complete setup guide with all the folder
11:41structures, prompts, and automation
11:42scripts. Once again, just click the link
11:44in the description to grab it right now.
11:45And if you want to build alongside me,
11:47join our community. Thanks so much for
11:48watching. Make sure to like and comment
11:50below how I can improve on my videos and
11:52what video ideas you want to see me do,
11:53what tutorials you want to see me make,
11:55and subscribe and I'll see you in the
11:56next video. Bye-bye.

About this transcript

This page contains the full transcript of Complete Guide to building an AI Second Brain with Claude by Ray Fu, generated from the public captions YouTube serves with the video. The transcript has 2,899 words across 424 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.