Skip to content

Podcast

Claude Code Can Be Your Second Brain

AI & I

Source ↗ ← All highlights
  • Obsidian On A Home Server
    • Noah moved his Obsidian vault to a home server and runs Claude Code on top to interact with thousands of notes.
    • He uses that setup primarily to read, research, and think rather than just to generate final artifacts. Transcript: Noah Brier Us through that? Yeah, I wouldn’t. I’m not sure about actually the vibe coding part of it. I have a sort of fairly heavy duty cloud code set but actually mostly not for code. So since those days of super organizers, like many people, I’ve abandoned Evernote and switched over to Obsidian. And one of the big advantages with obsidian as a note-taking platform is that it’s a bunch of markdown files and a bunch of folders um and uh they can then be synced with git and you can do Lots of other fun kinds of things and so actually probably my number one cloud code use is using it as a tool to interact with my notes. And so that I’ve got a fairly serious cloud code setup that I use with Obsidian. (Time 0:02:40)
  • Research By Voice In The Car
    • Noah uses Grok voice mode (including in his Tesla) for long-form research and finds it far better than other voice modes.
    • He conducted multi-hour research sessions by voice while driving and gained deep explanations like for self-attention. Transcript: Noah Brier And, you know, I find that that as a sort of like alternative way of working through problems, I have a Tesla, so now it’s baked into the Tesla. And, you know, it’s just, I, and, you know, obviously all the sort of other chat GPT and Claude and all these things of just being able to sort of like go and do research and really think And, and explore things in this device that’s always been useful, but like not useful for deep work, I think is probably something most people would agree with is that the phone is not Been the best place to kind of do deep coding and research work. And I feel like it, it, it’s really changed my ability to do that. Dan Shipper Wait, I got to stop you. So you’re using Grok voice mode and, and is that specifically because it’s built into your Tesla or using it in situations where you could also use, for example, ChadGBT voice mode? Noah Brier No, I’m using it because it way better than any of the other voice modes, and I will fight anybody who says anything different. Okay. Dan Shipper No, tell me, what do you like about it? Why is it better? Noah Brier To be fair, OpenAI launched their real-time API, which may or may not be baked into ChatGPT voice now. It’s not totally clear. But the old voice mode was based on 4.0 and i just found it to be completely unusable um and uh uh gemini’s voice mode i just didn’t find to be smart enough um and i just found grok’s voice Mode to be significantly smarter than anybody else is um you know i’m using grok two three four i don’t even remember whatever whatever the latest latest one yeah yeah but not the super I don’t have the most expensive super heavy or whatever i don’t have super heavy um but i just find it to be much better it does tool calling way better than any of the other ones that’s that’s I found to be a major weakness of the voice models is that they don’t do great tool calling and research and umk’s team steps all that. So no, even before it was loaded in my Tesla, I dropped my daughter off at summer camp this summer up in New Hampshire. So I had a five hour drive on my own and I spent like two hours researching and essentially like working through a piece. And I did it by just like connecting it to Bluetooth and just sort of sitting there in the car. And, um, I found it to be by far the, uh, the best of the voice modes. I, I hope these other models catch up there. Cause I would, I would love more really good voice modes. I mean, I had a mind blowing session this weekend and, uh, I’m giving a talk and I’m sort of have some ideas. I think it’s generally going to be about transformers eating the world. And so I was sort of catching myself up on self-attention and exactly how it works. And I did like an hour session and it really, I like was by far the sort of best explanation I’ve ever read for it or ever heard, I guess. (Time 0:05:25)
  • Force ‘Thinking Mode’ Before Writing
    • Tell the model explicitly when you are in thinking mode to prevent it from drafting the artifact prematurely.
    • Instruct subagents to ask questions and log thinking rather than produce final copy. Transcript: Noah Brier And so there’s some stuff in here where I’ve specifically told, I think it’s in the front matter actually, where I’ve told Claude Code, like, don’t help me write anything right now. And I generally find this to be a big thing with all these models is like, they immediately jump to wanting to help you with the artifact. And, you know, when you’re just in thinking mode, you have to be very explicit in like, hey, I just want you to help me think and ask me questions. And so, yeah, what you can see here is like there’s a bunch of files in (Time 0:16:28)
  • Run Claude At Your Vault Root
    • Start Claude Code at your Obsidian root so the agent can access your entire vault and pull relevant notes.
    • Add package.json commands to Obsidian to enable custom code operations and streamline workflows. Transcript: Noah Brier No, I’m, so I’m starting it in the full Obsidian vault. So like if we, like this is coming, if I step out of here, right, we’re in the root directory. All this stuff is in the root directory, my Obsidian. Dan Shipper I see it. I get it. Noah Brier And my Obsidian setup is also a little more intense for what it’s worth because I’ve also realized you can add a package.json to add a bunch of custom code commands to your Obsidian that You can then run. And then you could use those code commands and slash commands and all of these other things. So there are a bunch of other moving pieces in here. But generally, it’s a fairly straightforward… (Time 0:19:18)
  • Archive LLM Chats As Research
    • Noah saves full transcripts of chats with different LLMs into his Obsidian project folder for later review.
    • He instructs Claude to review those transcripts as part of project research rather than discarding them. Transcript: Noah Brier Step is i actually have an agent in here so if we go to i’ll do continue for now um and so for people who are listening so you’re just starting up claude you’re using the continue flag so uh Dan Shipper You’re you’re starting Claude by continuing the last session that you were in and now you’ve got, and you Claude code gives us ability to do sub agents. So those are like little mini Claude’s that you can spawn and you have a, you have a thinking partner, thinking partner, sub agent. Okay. How does that work? Noah Brier Yeah. And so this is the whole thing where I’m like, Hey, you’re a collaborative thinking partner, specializing in helping people explore complex problems. Your role is to facilitate thinking and basically don’t try to write the thing. And so after I had that initial set of things, I flipped to this and it’s like, okay, let’s get into a flow, ask me the kinds of questions, help me think through it. You know, this is also where I’ve got a chats folder in here. So it’s like, it’s not just happening here. I was also having like a, I’ve got a whole, sorry, I’m just backing out. So like if we go into, if we go into chats, like these are a whole bunch of the chats that I had, like with the interviewer. No, these are chats I was having with chat GPT and Claude and Grok and all of these different things that I went and just grabbed the full transcript of. So, you know, I was also having all these other conversations and then, you know, I’m specifically telling the interviewer, like review all these other things. (Time 0:23:58)
  • AI’s Reading Is Undervalued
    • Noah argues models’ reading ability is underappreciated compared to their writing ability.
    • He finds reading and synthesizing existing notes far more useful for daily thinking than producing artifacts. Transcript: Noah Brier One of the things I say to a lot of people is just like, I think partially because we call it generative, there’s entirely too much focus on its ability to write and not enough focus on its Ability to read. Um, you know, and it’s like its ability to read is incredible. Right. And, and I think, you know, arguably sort of like much more useful on a day to day basis. Like we produce artifacts far less frequently than we just like think about things. And so, yeah, I do this a lot. This is definitely a complaint I have about all the models is like, you know, even when you very specifically tell it not to try to do your work, it still often still tries to do your work. And so you have to like, really, really be like, no, I said no. (Time 0:27:57)
  • Use A Thinking Subagent
    • Build a subagent that behaves as a collaborative thinking partner and instruct it to log questions and insights.
    • Have the subagent keep an ongoing running log to catch you up when returning to a project. Transcript: Dan Shipper Or yeah. Noah Brier So that thinking agent is sort of told to, as it asks me questions, kind of make notes about the questions that it’s asking me and keep a kind of running log of what I’m uncovering and how I’m thinking about it and all those sorts of things. Dan Shipper Got it. And then, you know, you come back the next day and you’re like, oh, I just want to go down this rabbit hole on XYZ thing about the, you know, this wild bill guy. Um, and that you started a new chat, maybe with it, maybe with the sub vision, maybe, maybe not. And that becomes its own new file on that topic. Yeah, exactly. Noah Brier So like I, I, there’s the, um, the wild bill stuff started as like deep research in chat GPT. Um, and I, you know, had it go out and I’m reading the wild bill book right now. There’s a, like one sort of particularly famous biography of him. And, uh, um, you know, I’m kind of thinking about the bits and pieces and trying to make, and I think I made a kind of interesting connection in there, um, uh, where, you know, a big part Of what sort of he seems to have been after with the OSS and, and, you know, the sort of inspiration for the special forces was like empowering individuals. Um, (Time 0:29:59)
  • AI Bridges Organizational Tools
    • Noah frames AI as a glue that lets teams keep their preferred tools while interoperating through models.
    • He suggests AI reduces organizational friction by translating between disparate systems instead of forcing one platform. Transcript: Noah Brier I think what I was playing with here is this idea that bureaucracy was actually like an innovation, right? That like we look at bureaucracy as a negative and generally we talk about it as a negative. And I think often it is a negative. But, you know, bureaucracy was a sort of like huge innovation for how companies operate. And ultimately, it sort of represents kind of hierarchy and structure and a whole bunch of things that are like actually like pretty positive for operating at a large scale. And so, you know, again, my kind of whole thesis on AI around all this bureaucracy stuff is that what’s interesting about it is that as opposed to kind of past technologies, which kind Of forced you to make a decision about whether you wanted to kind of like use your existing sort of structure and build that technology into your existing structure or adopt the new structure. And most of the time, the new software required you adopted the new structure. And that’s why so many sort of software projects failed for so long. At least that’s sort of part of what I think. And, you know, I think part of what’s interesting about AI, what I find so interesting is like that you can kind of keep letting everybody work in whatever way they want. You know, it’s like a classic problem inside large companies is like one team wants to use Asana and one wants to use Jira and one wants to use Linear, right? And so then at some point there’s like a, there’s a, there’s a huge project and they bring in some big consulting firm and they decide they’re going to all centralize on this one thing. And now two thirds of the company is unhappy and like they’ve all made sacrifices. And, you know, you’re sort of in this like very non-ideal state. And I think what’s really interesting about AI and I this is a little more sort of theoretical because I think we’re so early in this, is that I think it’s very possible you could just say, Well, everybody just keep doing what you’re doing. We’re going to stick sort of some models in the middle. They don’t care what you use because it’s all just data structures to them. And so we can then have this sort of central thing. And when we talked about Percolate at the beginning, Percolate was a content marketing platform, worked with very large companies. So it was an enterprise software product. And it’s like, you know, at the end of the day, this is sort of the fundamental challenge of enterprise software is about like, you know, adoption and change management. And I just think, I think, and I hope, and again, this is sort of the optimist in me that like AI kind of lets us just not worry so much about these things. And rather than trying to make everybody change the ways that they work, kind of let them work in these ways and let AI sort of it’s my I call it my Thomas’s English muffin theory of AI, which Is that it like gets into the nooks and crannies. And so, yeah, that’s so anyway, but I have no idea what bureaucracy is positional encoding means yet. (Time 0:34:28)
  • Build Intuition By Using Models
    • Noah emphasizes we must build new intuitions for probabilistic AI tools through hands-on use.
    • He compares learning models to gaining ‘fingertip feeling’ that only practice provides. Transcript: Noah Brier And I think actually like, I think one of the really damaging sort of things out there is that, um, I think there are a lot of people who think we’re way further along in this than we are. And so, you know, I think particularly the people who are sort of scared, you know, we run, we work with like fortune 50 companies. And so, you know, when we’re sort of like out there and we’re talking to people inside the organization, a lot of people feel like they’ve already been left behind. And it’s like, no, you can like literally go sign into chat GPTT and like do something like nobody’s thought about doing with this thing yet, because there’s just so much white space To explore. And you might discover some totally new way of using it and or like totally new trick. And I don’t know, that’s just that’s and, you know, I think to be fair to some people, that’s sort of very intimidating. And I don’t think by and large, the models do any favors to themselves in helping those people get their feet wet in that, like, you know, I think people go on there and they, you know, it’s Like you ask it to write you a poem and then it writes you a poem and you’re like, okay, it wrote me a poem. But I don’t know that feeling of like, that feeling of, yeah, it’s like being on the frontier, right? Totally. Dan Shipper And yeah, I think that your point about intuitions and getting intuitions is the big thing. And I think people, what we don’t realize is when you’re dealing with something fundamentally new, you can’t trust like how you reason about it without experiencing it. (Time 0:44:21)
  • Ten-Year-Old Building With Vibe Code
    • Noah gave his 10-year-old his phone to build an app with vibe coding and she iterated heavily and learned data modeling concepts.
    • He sees this as evidence that low-code tools let kids meaningfully create and learn programming ideas. Transcript: Noah Brier Um, this weekend, actually, for the first time, my 10 year old, um, uh, she was really eager to be every year. Um, uh, my wife and her sister and brother and mom and, and all the cousins that we all get together and we do Christmas together. And so it’s too many presents to give to everybody. So we do a kind of like not secret secret Santa where everybody chooses one person. And my 10 year old really wanted to be the one who got to be the chooser. Um, and, uh, I encouraged her to, uh, vibe code in app to do it. Um, and so I just gave her my phone and V zero. And, um, honestly, that was like so amazing to watch, like not just cause it was so cool to see her do that and build it. And she went through, she was having so much fun. She did 75 revs on V0. So she like really got it going. Polished the app. She also like started to get into like really interesting kind of like computer science ideas without knowing it. So in one of the things like the adults give presence to adults and the kids give presence to kids, but she wanted this to be a more generalized app. So she realized that like, rather than having adults and kids, you need to call them groups. Right. And like, you know, so she’s like getting into data modeling and like all of this, I’m like watching this conversation happen. And, um, I just thought that was so awesome. Um, and you know, also just like a real pet beam of mine right now. And I’ve sort of gotten this argument with a bunch of people is like, there seems to be a big conversation that like, there’s a, a bubble in vibe coding and like, because one company or Another might have too high of a valuation. And my take on that is like, I just, I could not care less what the valuations of these companies are. I think like fundamentally, if there’s a tool that can allow a 10 year old to like build an app, there can’t, that can’t be a bubble. (Time 0:51:02)
  • Prioritize Media Literacy Over Cheating Fears
    • Teach media literacy and skepticism to kids so they can detect hallucinations and misinformation from AI.
    • Focus education on meta-skills like critical evaluation rather than rote memorization. Transcript: Noah Brier So, you know, both at the sort of schools they go to, and then also I went to NYU and I’ve sort of been having more and more conversations with the dean of the school I went to there. And there’s a lot of fear inside schools right now about AI and about cheating. And there’s a big thing, you know, so some parents in my town, they, you know, they wanted to have more of a conversation about it. And, you know, as someone who I’ve thought about this a lot, but I’ve also just like been, I’ve spent my entire life thinking about sort of technology and its effects on culture. And I think I’m like relatively grounded in these things of like, I’ve put in good hours of thinking. I mean, I know that for sure. And so, you know, my take on it is like one that sort of you can’t hide technology that won’t be hidden. Right. So it’s like, you know, putting our head in the sand is not the best solution. And then, you know, my bigger one though, is like, I was out, um, a friend of mine asked me to come, um, talk to a school two years ago about AI, um, uh, out in LA. And, um, afterwards I was talking to a English teacher there and she was like, what do I do? Like, what do I do about all these kids, you know, using AI? And I was like, look, I don’t really know what your job is because like, I mean, being an English teacher for 11th graders sounds really much harder than my job. Um, but on a really fundamental level, like I don’t actually think your job is to teach these kids to write. Cause that’s like a lifelong pursuit. I think your job is to convince them that it’s worth learning to write. And so in that way, like, I, I’m not sure that anything fundamentally changes because of AI. Like, I think that, you know, and again, this is my very optimistic take, but like, I am, I think that there are so many parts of the education system that AI really just exposes the sort Of flaws in the way that we teach. Like, why are there so many tests on these kinds of things instead of encouraging thinking and learning and coming to love to write and research and whatever, you know, it’s like, we’re So focused on teaching kids the, you know, five paragraph essay, you know, while every adult who is a writer has long abandoned that. And because like, it’s all about sort of discovering that you like to write and you, you know, what your own style is and how to do it. And, you know, it’s like, I’m sure it’s like, you know, a big part of working with AI to write is like telling it is ignoring it. Cause you’re like, no, that’s not me. I’m not into that. Like I, like I, I, I’m totally comfortable saying really here. I know you don’t think it’s a good idea, but like, I’m cool with it. And so, you know, I don’t know. I I’ve been sort of having a lot of these different thoughts. Actually pitching a class for the fall of 26 at NYU to kind of the idea for it is code is essay. And my sort of point of view is like this sort of opens up this new way to express yourself. Um, and that, you know, like we have all these other ways that we celebrate to express ourselves, but code has been like long shut off from people because it’s, but actually it’s kind Of amazing. And it lets you express yourself in all these different kinds of ways. And like, you know, this is what my 10 year old was doing this weekend. So anyway, those are all the bits and pieces as a parent. I’ve been thinking about the one other thing I will plug is media literacy. I think it’s a big piece of it. A lot of people are afraid of these models and hallucinations. And there’s a book by a guy named Tim Harford who writes for the FT and he’s an economist. And he has a book called The Truth Detective, which is an adaptation of The Data Detective, but it’s written for kids. And it’s the best media literacy book I’ve ever read for adults or for children. And I think, you know, a lot of what this AI conversation exposes is like how bad a job we do with helping and arming our kids and our adults to be sort of like truly media and technology literate. And, you know, like being really good at knowing what’s real on social media turns out to be also really useful for like differentiated between hallucinations and non-hallucinations In chat GBD, right? Like this is a sort of, to me, a kind of very central skill that like we need to arm everybody with. (Time 0:53:18)
  • Phone-First Deep Work Setup
    • Noah SSHes into a basement mini-PC via Tailscale VPN and runs Claude Code against his synced Obsidian repo from his phone.
    • He routinely edits notes, pulls research, and even pushes code changes from his phone while away from his desk. Transcript: Noah Brier Okay. So here we go. So I am going into an app called Termius. Termius is just a terminal. Um, and, uh, what is allowing all this to happen behind the scenes is in my basement, I have a mini PC. And on that mini PC, I have a thing called Tailscale running. And Tailscale lets you set up these very simple VPNs. So I’m currently inside, like if I scroll down here, you see I’m on a VPN. That’s my personal VPN. I’m not like on an outside VPN. So the only way to access this machine is through my VPN. Okay. And so then when I go in there, because I sync my Obsidian with Git, so I put it, it’s on GitHub, in a private GitHub. I can then sync it back down to here. And so then I can just call up Claude. And now I’m just in Claude code, talking and thinking. And I can just be like, what’s new in the last two days? I can access any of my agents. I can do anything. And again, this is in my Obsidian, but I can use this anywhere, right? So I’ll be like on the fly. I’ve got other repos in here. I like a link was broken on my conference site. And so I, uh, um, I just, uh, opened the repo. I pulled it down. I asked CloudCo to make the changes and I was able to do it right here. Um, so this has been like completely wild to me because again, this is very much in that, like, like on Tuesday of this week, um, we had Monday off Tuesday. I dropped the, I dropped the kids off of the bus. And then I, uh, I went and I sat and had breakfast and I literally sat on my phone and worked on this talk for like two hours. Um, and I did it through here, right. Like on my phone where I was like doing real thinking and research and pulling things in and pasting things in and doing all this kinds of stuff. And, you know, I’m able to do it all. And it just, it doesn’t seem like I could do that kind of thing without that. (Time 1:02:00)
  • Automate Box Setup With Claude Helpers
    • Build Claude Code helpers to automate OS and editor setup so new machines match your preferences quickly.
    • Teach the agent your dotfiles, plugins, and package choices so it can bootstrap environments for you. Transcript: Noah Brier I’m playing with this Omarchi, which is DHH’s Linux distribution. And I’m not like super comfortable in here. And so I got this whole, this is a Cloud Code project specifically to help me configure this box. And it’s like so nice because I’m like, oh, how do you do this in Linux again? Or like, what’s the Neo Vim command? Or like, can you change this? Can you help me install this plugin for Neo Vim or whatever it is? And so now I have one on my Mac too, where it’s like, can you clean up all the homebrew things? Or like, I switched from like, which Python package manager I was using. And it was like, that would have been super overwhelming for me. And I was like, I want to switch from using PIP to using UV. Can you like, just make that happen? And like, just did all this stuff for me. And it knows my preferred settings. And so I actually have a version of this where now I’ve got it so tuned that if I want to launch a new box for doing something, it’ll just have all my settings ready to go. And then I can log into Cloud Code and the Cloud Code can then set up anything else that didn’t get set up in the initial process. (Time 1:07:21)
  • Use LLMs To Clean Local Data
    • Noah built an attachments cleaner using Gemini Flash to rename and index Obsidian file attachments.
    • He leverages LLMs to normalize messy local data and rewrite links automatically. Transcript: Noah Brier It puts them in a table in the attachments folder and then it renames all the attachment links back. So just like cleans everything up. It just does it through Gemini Flash. And so it’s like, I don’t know, stuff like that’s kind of amazing. (Time 1:09:42)