WEBVTT 00:00:00.000 --> 00:00:03.600 You spend 20 minutes perfectly explaining a task. 00:00:04.179 --> 00:00:06.259 You get the details just right. The formatting 00:00:06.259 --> 00:00:08.820 is flawless. Then you wake up the next morning, 00:00:09.560 --> 00:00:13.199 you open a new chat window, and it has completely 00:00:13.199 --> 00:00:15.259 forgotten who you are. Yeah, it's the classic 00:00:15.259 --> 00:00:17.440 amnesiac chat bot problem. I mean, we've all 00:00:17.440 --> 00:00:19.820 been there. It's incredibly frustrating, honestly. 00:00:20.300 --> 00:00:22.739 You're basically training a new employee from 00:00:22.739 --> 00:00:26.719 scratch every single day. Well, welcome to today's 00:00:26.719 --> 00:00:29.399 deep dive. We're shifting away from those temporary 00:00:29.399 --> 00:00:32.280 AI chats. The goal today is building a permanent 00:00:32.280 --> 00:00:35.259 living environment. We're calling it a co -work 00:00:35.259 --> 00:00:37.859 operating system, and we're doing it using Claude 00:00:37.859 --> 00:00:40.399 projects. So we have a really fascinating journey 00:00:40.399 --> 00:00:42.520 mapped out for you today. First, we'll clearly 00:00:42.520 --> 00:00:44.299 define the difference between standard Claude 00:00:44.299 --> 00:00:46.820 chat and this new co -work concept. Then we'll 00:00:46.820 --> 00:00:49.100 unpack the underlying architecture. It relies 00:00:49.100 --> 00:00:51.679 on three surprisingly simple text files that 00:00:51.679 --> 00:00:54.539 give the AI persistent memory. After that, we're 00:00:54.539 --> 00:00:57.020 mapping out structured workstations. And finally, 00:00:57.039 --> 00:00:58.659 we'll look at the advanced workflows to keep 00:00:58.659 --> 00:01:00.679 this whole system incredibly fast and cheap. 00:01:01.020 --> 00:01:03.039 Before we jump into the mechanics, I have to 00:01:03.039 --> 00:01:06.239 confess something. I still catch myself pasting 00:01:06.239 --> 00:01:09.120 the exact same five paragraph instruction prompt 00:01:09.120 --> 00:01:12.000 every morning. I'm just so conditioned to the 00:01:12.000 --> 00:01:14.480 old way of working. I still wrestle with prompt 00:01:14.480 --> 00:01:17.439 drift myself. Listen, it makes complete sense 00:01:17.439 --> 00:01:20.200 that you do. We've all been trained over the 00:01:20.200 --> 00:01:22.640 last few years to treat AI like a disposable 00:01:22.640 --> 00:01:26.060 tool. You use it, you close the tab, and it vanishes. 00:01:26.379 --> 00:01:29.700 But to build a true AI operating system, you 00:01:29.700 --> 00:01:32.359 have to completely change the environment. We 00:01:32.359 --> 00:01:34.260 need to start by looking at the three distinct 00:01:34.260 --> 00:01:36.659 flavors of Claude that currently exist. Right. 00:01:36.840 --> 00:01:38.920 Let's break down that baseline. We need to understand 00:01:38.920 --> 00:01:40.819 the environment before we can build the system. 00:01:41.120 --> 00:01:43.900 Exactly. So first, you have Cloud Chat. This 00:01:43.900 --> 00:01:45.620 is the web browser version. Everybody knows. 00:01:45.640 --> 00:01:47.420 It's great for doing some quick market research. 00:01:47.719 --> 00:01:49.579 It's helpful if you just need to draft a quick 00:01:49.579 --> 00:01:52.659 apology email to a client. But it starts fresh 00:01:52.659 --> 00:01:55.400 every single session. It has absolutely zero 00:01:55.400 --> 00:01:57.599 memory of what you did yesterday. It's essentially 00:01:57.599 --> 00:02:00.319 a blank slate every time you hit Enter. Then 00:02:00.319 --> 00:02:02.579 you have Cloud Code. Now, this is a very different 00:02:02.579 --> 00:02:04.799 beast. It's a terminal -based tool built specifically 00:02:04.799 --> 00:02:08.120 for developers. Meaning a text -only screen where 00:02:08.120 --> 00:02:10.719 programmers type direct system commands. Precisely. 00:02:10.830 --> 00:02:12.789 It connects directly to a software code base. 00:02:12.909 --> 00:02:15.310 It can actually autonomously edit files, run 00:02:15.310 --> 00:02:17.770 tests, and execute commands. But it's highly 00:02:17.770 --> 00:02:20.050 specialized for software engineering. It's definitely 00:02:20.050 --> 00:02:22.909 not built for your general day -to -day workflow. 00:02:23.289 --> 00:02:25.629 So if chat is entirely temporary and code is 00:02:25.629 --> 00:02:27.669 strictly for programmers, there has to be a middle 00:02:27.669 --> 00:02:30.250 ground, something built for persistent, everyday 00:02:30.250 --> 00:02:32.409 knowledge work. And that brings us to the third 00:02:32.409 --> 00:02:35.590 flavor, Claude Cowork. This is the actual desktop 00:02:35.590 --> 00:02:38.280 application. It's engineered specifically for 00:02:38.280 --> 00:02:41.960 long -term compounding workflows. The true magic 00:02:41.960 --> 00:02:44.860 here isn't just the AI itself. It's the ability 00:02:44.860 --> 00:02:47.340 to natively mount local folders from your computer. 00:02:47.719 --> 00:02:50.759 It maintains a persistent, evolving memory. And 00:02:50.759 --> 00:02:52.919 it uses external connectors to link directly 00:02:52.919 --> 00:02:55.199 to your gisting tools, things like Gmail, your 00:02:55.199 --> 00:02:58.139 Google Calendar, or your Notion databases. OK, 00:02:58.139 --> 00:03:01.080 let me try to contextualize this. Standard Cloud 00:03:01.080 --> 00:03:04.000 Chat is basically a temp worker. They show up, 00:03:04.020 --> 00:03:06.750 do a decent job. but they get a complete memory 00:03:06.750 --> 00:03:09.270 wipe every single night. It's literally like 00:03:09.270 --> 00:03:13.949 that movie, 50 First Dates. Beat. But co -work 00:03:13.949 --> 00:03:15.770 is fundamentally different. It's like hiring 00:03:15.770 --> 00:03:18.270 a full -time partner. You're giving them a dedicated 00:03:18.270 --> 00:03:20.490 desk. You're giving them a permanent filing cabinet. 00:03:21.009 --> 00:03:23.490 That desk analogy is perfect. You're giving the 00:03:23.490 --> 00:03:26.530 AI a physical home on your machine. It doesn't 00:03:26.530 --> 00:03:29.469 exist in a vacuum anymore. It has a spatial awareness 00:03:29.469 --> 00:03:32.060 of your digital life. But that raises a pretty 00:03:32.060 --> 00:03:35.039 massive privacy concern for me. Does mounting 00:03:35.039 --> 00:03:37.759 a local folder mean the AI is just freely reading 00:03:37.759 --> 00:03:39.919 everything on my computer? Is it scanning my 00:03:39.919 --> 00:03:42.080 personal photos while I sleep? No, absolutely 00:03:42.080 --> 00:03:44.879 not. The permissions are strictly ring -fenced. 00:03:45.360 --> 00:03:47.759 It only reads the specific folder you explicitly 00:03:47.759 --> 00:03:50.580 mount and allow inside the application. It has 00:03:50.580 --> 00:03:52.560 zero access to the rest of your hard drive. So 00:03:52.560 --> 00:03:54.719 it's a fenced -in playground, not a master key 00:03:54.719 --> 00:03:56.719 to my hard drive. Right. You dictate the boundaries. 00:03:56.840 --> 00:03:58.680 It only sees what you place inside the fence. 00:03:59.099 --> 00:04:01.580 Now that we understand Co -Work acts like a dedicated 00:04:01.580 --> 00:04:03.740 filing cabinet, we need to know what actually 00:04:03.740 --> 00:04:06.039 goes inside those folders. How does it physically 00:04:06.039 --> 00:04:08.379 remember us? This is where it gets genuinely 00:04:08.379 --> 00:04:11.240 fascinating. You would think an advanced AI operating 00:04:11.240 --> 00:04:15.360 system requires massive, complex databases. But 00:04:15.360 --> 00:04:17.540 the whole system runs on a surprisingly simple 00:04:17.540 --> 00:04:21.139 architecture. You literally just need three plain 00:04:21.139 --> 00:04:24.180 text files sitting inside a root folder. We'll 00:04:24.180 --> 00:04:26.790 call that master folder your Co -Work OS. Just 00:04:26.790 --> 00:04:29.089 three files. I mean, that feels almost too simple. 00:04:29.170 --> 00:04:31.189 Yeah, but they're just markdown files. Meaning 00:04:31.189 --> 00:04:34.189 basic text files with simple formatting headers 00:04:34.189 --> 00:04:36.550 and bullet points. Right. There's no complex 00:04:36.550 --> 00:04:39.509 programming logic. No Python scripts. The first 00:04:39.509 --> 00:04:43.089 file is called Claude .md. Think of this file 00:04:43.089 --> 00:04:46.069 as the master operating manual or the global 00:04:46.069 --> 00:04:48.910 routing map. When you open a new session, this 00:04:48.910 --> 00:04:51.589 file tells Claude what to load and where to look. 00:04:51.730 --> 00:04:53.769 It establishes the fundamental ground rules for 00:04:53.769 --> 00:04:56.480 how it should behave. Yes, and it has three core 00:04:56.480 --> 00:04:58.959 components. First is the memory system itself. 00:04:59.180 --> 00:05:01.160 It instructs Claude to always check your pass 00:05:01.160 --> 00:05:04.120 context before answering. Second is the routing 00:05:04.120 --> 00:05:07.100 map. This prevents the AI from guessing. If you 00:05:07.100 --> 00:05:09.579 ask a question about your taxes, the map routes 00:05:09.579 --> 00:05:11.959 it straight to your financial folder. Third, 00:05:12.339 --> 00:05:14.709 you have references. These are pointers that 00:05:14.709 --> 00:05:17.149 tell Claude where to find extra context, but 00:05:17.149 --> 00:05:19.509 only when it actually needs it. Two secs silence. 00:05:19.529 --> 00:05:21.509 I really have to push back here. It feels wild 00:05:21.509 --> 00:05:24.269 to me. The absolute cutting edge of artificial 00:05:24.269 --> 00:05:26.810 intelligence relies on basic text files from 00:05:26.810 --> 00:05:29.790 the 1990s to function. And the sources are very 00:05:29.790 --> 00:05:31.930 strict about this. They say we absolutely must 00:05:31.930 --> 00:05:35.949 cap this Claude .md file at 300 lines. Why cap 00:05:35.949 --> 00:05:37.970 it? If this is my master operating manual, why 00:05:37.970 --> 00:05:40.709 not dump every single rule, preference, and workflow 00:05:40.709 --> 00:05:42.829 I have in there? It's a really common trap people 00:05:42.829 --> 00:05:45.069 f— fall into to understand the cap, you have 00:05:45.069 --> 00:05:46.949 to understand how token limits actually work 00:05:46.949 --> 00:05:49.170 under the hood. Meaning the maximum amount of 00:05:49.170 --> 00:05:51.870 text the AI can process simultaneously. Think 00:05:51.870 --> 00:05:54.350 of it like the AI's cognitive battery power for 00:05:54.350 --> 00:05:57.410 a single conversation. Your large root files 00:05:57.410 --> 00:06:00.149 load every single time you start a new session. 00:06:01.009 --> 00:06:03.470 Every time you say hello, Claude has to silently 00:06:03.470 --> 00:06:07.120 read that entire Claude .md file first. If your 00:06:07.120 --> 00:06:10.420 manual is 20 pages long, it slowly drains your 00:06:10.420 --> 00:06:12.680 token limits before you even start working. It 00:06:12.680 --> 00:06:15.160 costs more money. It severely slows down the 00:06:15.160 --> 00:06:17.720 response time. Leaving barely any processing 00:06:17.720 --> 00:06:19.959 power to actually complete the task I asked for. 00:06:20.139 --> 00:06:22.079 Exactly. You're giving it cognitive overload. 00:06:22.420 --> 00:06:24.399 That's why you keep that main routing file under 00:06:24.399 --> 00:06:27.500 300 lines. It should be a traffic cop, not encyclopedia. 00:06:27.639 --> 00:06:29.399 Well, that makes a lot of sense. So if Claude 00:06:29.399 --> 00:06:31.980 .md is the traffic cop, what's the second file 00:06:31.980 --> 00:06:35.649 handling? The second file is memory .md. This 00:06:35.649 --> 00:06:37.930 functions as your running ledger. It holds your 00:06:37.930 --> 00:06:40.649 active, ongoing projects and your saved personal 00:06:40.649 --> 00:06:43.490 preferences. This specific file is the magic 00:06:43.490 --> 00:06:45.509 ingredient that stops your conversations from 00:06:45.509 --> 00:06:48.470 starting at zero every day. It's the actual continuity 00:06:48.470 --> 00:06:50.810 mechanism. Yeah, it tracks the reality of your 00:06:50.810 --> 00:06:53.209 current work. We usually split it into two sections. 00:06:53.649 --> 00:06:56.089 One section tracks active projects so it knows 00:06:56.089 --> 00:06:57.769 your upcoming newsletter is currently in the 00:06:57.769 --> 00:06:59.689 rough draft stage and you're struggling with 00:06:59.689 --> 00:07:03.050 the conclusion. The second section tracks granular 00:07:03.050 --> 00:07:07.160 preferences. Like knowing a specific client absolutely 00:07:07.160 --> 00:07:09.939 hastes bullet points and prefers numbered lists. 00:07:10.339 --> 00:07:12.360 Okay, and the third file in this architecture? 00:07:12.660 --> 00:07:15.759 That one is called VoicePrinciples .md. We usually 00:07:15.759 --> 00:07:17.779 drop this into a subfolder called Resources. 00:07:18.110 --> 00:07:21.910 This file's entire job is teaching the AI exactly 00:07:21.910 --> 00:07:24.250 how you write and communicate. We'll definitely 00:07:24.250 --> 00:07:26.050 dig into that voice file in a minute because 00:07:26.050 --> 00:07:29.410 sounding human is huge. But first, just managing 00:07:29.410 --> 00:07:32.250 these individual markdown files sounds potentially 00:07:32.250 --> 00:07:34.910 messy. If I'm manually editing text files all 00:07:34.910 --> 00:07:37.050 day, I feel like I'm stepping backward in productivity. 00:07:37.420 --> 00:07:39.699 That's a totally valid concern, and it's why 00:07:39.699 --> 00:07:42.220 a lot of power users rely on an app called Obsidian. 00:07:42.779 --> 00:07:45.699 It's a completely free markdown editor. You don't 00:07:45.699 --> 00:07:47.959 have to migrate anything. You just open your 00:07:47.959 --> 00:07:51.680 existing Cowork OS folder as a vault inside Obsidian. 00:07:52.300 --> 00:07:54.980 It instantly gives you a beautiful visual interface. 00:07:55.139 --> 00:07:58.160 You can easily view, link, and manage all these 00:07:58.160 --> 00:08:01.600 text files in one clean dashboard. But practically 00:08:01.600 --> 00:08:04.120 speaking, do I have to manually open that dashboard 00:08:04.120 --> 00:08:07.639 and type into memory .md every single time I 00:08:07.639 --> 00:08:10.339 finish a task? That feels like a lot of friction. 00:08:10.639 --> 00:08:12.540 No, not at all. That's the beauty of the system. 00:08:12.759 --> 00:08:14.480 You just tell Claude in your normal chat window 00:08:14.480 --> 00:08:17.319 to remember this. The AI takes your natural language 00:08:17.319 --> 00:08:20.259 instruction. reformats it, and silently writes 00:08:20.259 --> 00:08:22.360 it to the Markdown file automatically. You just 00:08:22.360 --> 00:08:24.420 talk to it and it updates its own memory notes. 00:08:24.600 --> 00:08:26.639 It completely removes the administrative burden. 00:08:26.839 --> 00:08:29.040 You just do your work and the system documents 00:08:29.040 --> 00:08:31.100 itself. Okay, so the AI can take its own notes. 00:08:31.420 --> 00:08:33.519 But let's circle back to that third file. How 00:08:33.519 --> 00:08:36.039 do we actually stop it from sounding like a generic, 00:08:36.320 --> 00:08:39.279 overly enthusiastic robot? And as we scale this 00:08:39.279 --> 00:08:41.440 up, how do we keep it from mixing up completely 00:08:41.440 --> 00:08:43.720 different parts of your life? Those are two separate 00:08:43.950 --> 00:08:46.490 but incredibly important challenges. That's where 00:08:46.490 --> 00:08:48.429 the voice system and the concept of workstations 00:08:48.429 --> 00:08:50.690 come into play. Let's tackle the voice system 00:08:50.690 --> 00:08:53.549 first. How do we actually build that voice principles 00:08:53.549 --> 00:08:56.110 file without it taking weeks of training? Well, 00:08:56.330 --> 00:08:59.269 you have to feed the AI real historical examples 00:08:59.269 --> 00:09:02.529 of your authentic writing. If your Gmail is already 00:09:02.529 --> 00:09:05.230 connected to your cowork app, you can literally 00:09:05.230 --> 00:09:07.870 just have it analyze your last 50 cent emails. 00:09:08.350 --> 00:09:09.870 If you don't want to connect your email, just 00:09:09.870 --> 00:09:12.970 copy and paste five to 10 past articles or reports 00:09:12.970 --> 00:09:15.009 you've written right into the chat. You ask it 00:09:15.009 --> 00:09:18.509 to reverse engineer your tone. Yes. But here's 00:09:18.509 --> 00:09:21.750 the critical failure point for most people. Vague 00:09:21.750 --> 00:09:25.350 rules fail miserably. If you tell Claude to be 00:09:25.350 --> 00:09:27.750 warm and professional, you're going to get terrible 00:09:27.750 --> 00:09:30.919 generic AI output. It'll sound like every corporate 00:09:30.919 --> 00:09:33.500 memo ever written. You have to extract highly 00:09:33.500 --> 00:09:36.139 specific mechanical rules. You tell it, keep 00:09:36.139 --> 00:09:39.120 paragraphs under three sentences. Or you instruct 00:09:39.120 --> 00:09:41.740 it, never use passive voice and avoid exclamation 00:09:41.740 --> 00:09:44.419 points. Because specificity is what creates repeatable 00:09:44.419 --> 00:09:46.779 behavior. A machine doesn't understand warmth, 00:09:46.940 --> 00:09:49.419 but it perfectly understands a maximum word count. 00:09:49.799 --> 00:09:52.840 And it's an iterative process. Over the first 00:09:52.840 --> 00:09:55.580 few weeks, you'll catch it making mistakes. You 00:09:55.580 --> 00:09:58.019 correct the output. You tell it to remember that 00:09:58.019 --> 00:10:01.779 specific correction. Over time, that voice principles 00:10:01.779 --> 00:10:05.480 file evolves into a living, incredibly accurate 00:10:05.480 --> 00:10:07.899 record of your actual voice. So that solves the 00:10:07.899 --> 00:10:11.440 robot tone. What about keeping my life organized? 00:10:11.820 --> 00:10:14.399 That requires building out workstations. Think 00:10:14.399 --> 00:10:17.120 of workstations as dedicated subfolders for specific 00:10:17.120 --> 00:10:19.799 domains of your life. I really like comparing 00:10:19.799 --> 00:10:22.759 workstations to the rooms in a physical house. 00:10:22.879 --> 00:10:24.559 You wouldn't keep your kitchen blender sitting 00:10:24.559 --> 00:10:27.460 on your bed. It makes no sense. Similarly, your 00:10:27.460 --> 00:10:29.899 highly confidential mortgage refinancing notes 00:10:29.899 --> 00:10:32.000 absolutely shouldn't bleed into your weekly creative 00:10:32.000 --> 00:10:34.519 writing drafts. That's a brilliant way to conceptualize 00:10:34.519 --> 00:10:37.840 it. You need structural boundaries. And we generally 00:10:37.840 --> 00:10:40.379 see two distinct types of workstations you can 00:10:40.379 --> 00:10:42.740 build within this system. So if one type handles 00:10:42.740 --> 00:10:45.159 broad categories, what does a universal workstation 00:10:45.159 --> 00:10:47.779 look like? Universal workstations cover broad 00:10:47.779 --> 00:10:50.240 functional areas that touch multiple parts of 00:10:50.240 --> 00:10:52.960 your life. An email HQ is the perfect example. 00:10:53.279 --> 00:10:55.240 You send emails for your day job, for your personal 00:10:55.240 --> 00:10:57.879 life, maybe for side hustle clients. The core 00:10:57.879 --> 00:10:59.840 rules of what makes a good email for you, like 00:10:59.840 --> 00:11:02.019 your sign off and formatting, apply everywhere. 00:11:02.440 --> 00:11:04.519 So that lives in a universal folder. And the 00:11:04.519 --> 00:11:06.559 second type, I imagine it's something highly 00:11:06.559 --> 00:11:09.659 siloed. Right. Dedicated workstations. These 00:11:09.659 --> 00:11:12.879 cover one highly specific contained domain. a 00:11:12.879 --> 00:11:15.500 folder strictly for personal finances, a folder 00:11:15.500 --> 00:11:17.980 dedicated to a complex upcoming international 00:11:17.980 --> 00:11:21.620 trip, or a folder locked down for one specific 00:11:21.620 --> 00:11:24.159 client launch. And each one of these individual 00:11:24.159 --> 00:11:26.720 workstations gets its own isolated setup. Yes. 00:11:27.039 --> 00:11:29.259 Each dedicated workstation gets its own local 00:11:29.259 --> 00:11:33.019 CLAW .md file and its own localized memory .md 00:11:33.019 --> 00:11:35.340 file. The structure perfectly mirrors your main 00:11:35.340 --> 00:11:37.960 root folder, but the rules inside only apply 00:11:37.960 --> 00:11:40.059 when you're actively working inside that specific 00:11:40.059 --> 00:11:42.659 space. Let me test the logic here. If I create 00:11:42.659 --> 00:11:45.500 a really fantastic rule for formatting emails 00:11:45.500 --> 00:11:48.299 inside my finance workstation, should I copy 00:11:48.299 --> 00:11:50.720 paste that rule up into my universal email HQ? 00:11:51.200 --> 00:11:54.240 No, never duplicate. You should never repeat 00:11:54.240 --> 00:11:56.820 global rules in multiple places across the system. 00:11:57.240 --> 00:11:59.909 It's a core design principle. Duplicate rules 00:11:59.909 --> 00:12:02.610 waste your token limits, and worse, they cause 00:12:02.610 --> 00:12:05.330 inconsistent AI behavior when the system tries 00:12:05.330 --> 00:12:07.889 to reconcile two slightly different versions 00:12:07.889 --> 00:12:10.450 of the same rule. Write the rule once at the 00:12:10.450 --> 00:12:12.850 highest level. Never duplicate it. Exactly. It 00:12:12.850 --> 00:12:15.549 keeps the entire architecture clean, efficient, 00:12:15.850 --> 00:12:19.460 and predictable. Sponsor's hearing. OK, so we've 00:12:19.460 --> 00:12:21.460 built the house. We furnished the individual 00:12:21.460 --> 00:12:24.860 rooms with workstations. What does actually living 00:12:24.860 --> 00:12:27.879 and working inside this system look like on a 00:12:27.879 --> 00:12:30.840 random Tuesday afternoon? And crucially, how 00:12:30.840 --> 00:12:33.120 do we pay the electric bill? Meaning, how do 00:12:33.120 --> 00:12:35.539 we manage the token costs so we don't go broke 00:12:35.539 --> 00:12:37.379 running this thing? Let's look at some advanced 00:12:37.379 --> 00:12:39.960 daily workflows. When you integrate those external 00:12:39.960 --> 00:12:43.240 connectors, the capability styrockets. The first 00:12:43.240 --> 00:12:45.720 amazing use case is automatic meeting follow 00:12:45.720 --> 00:12:47.740 -ups. Walk me through the actual mechanics of 00:12:47.740 --> 00:12:49.919 that. How does it happen? Imagine your Google 00:12:49.919 --> 00:12:52.669 Calendar is securely connected to Cowork. You 00:12:52.669 --> 00:12:55.429 also use a standard tool that records and transcribes 00:12:55.429 --> 00:12:57.830 your Zoom meetings. You simply open your chat 00:12:57.830 --> 00:13:00.129 and ask Claude to handle the follow -up for your 00:13:00.129 --> 00:13:03.090 2 p .m. meeting. Just one simple natural language 00:13:03.090 --> 00:13:05.169 request. That's it. Claude pings the calendar 00:13:05.169 --> 00:13:08.509 API to see who is in the 2 p .m. slot. It pulls 00:13:08.509 --> 00:13:11.110 the WAW transcript from your local folder. It 00:13:11.110 --> 00:13:13.799 reads the entire conversation. identifies the 00:13:13.799 --> 00:13:16.259 agreed -upon next steps, and figures out who 00:13:16.259 --> 00:13:18.899 is responsible for what. It then dynamically 00:13:18.899 --> 00:13:21.360 routes to your email HQ workstation to grab your 00:13:21.360 --> 00:13:24.240 communication rules. Finally, it drafts a perfectly 00:13:24.240 --> 00:13:26.600 formatted follow -up email using your exact tone, 00:13:27.100 --> 00:13:29.700 all from one single prompt. That is incredible 00:13:29.700 --> 00:13:31.940 leverage. What about applying this to project 00:13:31.940 --> 00:13:34.419 management? The sources mention Notion. Notion 00:13:34.419 --> 00:13:37.350 auto projects are a huge unlock. Let's say you 00:13:37.350 --> 00:13:39.169 tell Claude you're traveling to Boston for a 00:13:39.169 --> 00:13:41.909 tech conference in June. Claude automatically 00:13:41.909 --> 00:13:44.679 connects to your Notion workspace. Because it 00:13:44.679 --> 00:13:47.080 checks your memory file, it already knows exactly 00:13:47.080 --> 00:13:49.379 how you like your travel properties formatted. 00:13:50.080 --> 00:13:52.419 It knows you like a column for confirmation numbers 00:13:52.419 --> 00:13:55.360 and a separate view for daily itineraries. It 00:13:55.360 --> 00:13:59.240 builds the complete customized travel board autonomously. 00:13:59.399 --> 00:14:01.519 It's not just generating text anymore. It's generating 00:14:01.519 --> 00:14:04.000 customized software environments based on my 00:14:04.000 --> 00:14:06.259 historical preferences. Precisely. But if you 00:14:06.259 --> 00:14:08.279 want to see the true power of persistent memory, 00:14:08.600 --> 00:14:10.500 the finance tracking is where it gets really 00:14:10.500 --> 00:14:12.519 wild. Tell me about the finance workstation. 00:14:12.669 --> 00:14:14.470 How is that different from just using a spreadsheet 00:14:14.470 --> 00:14:18.070 macro? Whoa. Beat. Imagine it just silently watching 00:14:18.070 --> 00:14:21.090 a messy folder of bank statements. You literally 00:14:21.090 --> 00:14:23.730 just drop raw PDF statements into the mounted 00:14:23.730 --> 00:14:26.309 folder. It instantly reads them. It categorizes 00:14:26.309 --> 00:14:28.429 an entire year of transactions. It groups your 00:14:28.429 --> 00:14:30.250 software subscriptions, your travel expenses, 00:14:30.389 --> 00:14:32.570 your dining out. Then it generates a clean Excel 00:14:32.570 --> 00:14:35.370 summary. It's doing hours of tedious manual data 00:14:35.370 --> 00:14:39.129 entry in seconds. But AI hallucinates. What happens 00:14:39.129 --> 00:14:42.149 when it inevitably makes a mistake? What if it 00:14:42.149 --> 00:14:45.009 categorizes a business lunch as a personal grocery 00:14:45.009 --> 00:14:46.990 expense? You just correct it once in the chat. 00:14:47.110 --> 00:14:48.870 You say, hey, anything from this specific restaurant 00:14:48.870 --> 00:14:51.490 is a business expense. Remember this category 00:14:51.490 --> 00:14:54.409 rule. It instantly writes that logic to the local 00:14:54.409 --> 00:14:57.389 memory .md file inside the finance workstation. 00:14:57.690 --> 00:14:59.769 It will never make that specific mistake again. 00:15:00.169 --> 00:15:02.870 The system actually learns. Two -sex silence. 00:15:03.830 --> 00:15:07.740 That is genuine compounding leverage. But processing 00:15:07.740 --> 00:15:11.179 dozens of PDFs, reading calendar APIs, managing 00:15:11.179 --> 00:15:13.919 Notion databases, doing all of this continuously 00:15:13.919 --> 00:15:16.600 must cost an absolute fortune in API credits. 00:15:16.980 --> 00:15:18.799 It definitely can if you aren't careful with 00:15:18.799 --> 00:15:20.700 your architecture. You need strict cost control 00:15:20.700 --> 00:15:22.559 measures in place. And the biggest secret to 00:15:22.559 --> 00:15:24.899 keeping the bill low is intelligent model selection. 00:15:25.100 --> 00:15:26.620 Explain how we manage that. Are we restricting 00:15:26.620 --> 00:15:28.919 what it can do? Not restricting, just optimizing. 00:15:29.279 --> 00:15:32.000 You should use Claude 3 .5 Sonnet for 80 % of 00:15:32.000 --> 00:15:34.820 your routine tasks. Sonnet is incredibly capable. 00:15:34.860 --> 00:15:37.580 It's very fast. and it is significantly cheaper 00:15:37.580 --> 00:15:40.340 to run. You save the heavier model, Claude Erpus, 00:15:40.820 --> 00:15:43.539 exclusively for highly complex multi -step reasoning 00:15:43.539 --> 00:15:47.580 tasks. So, Sonnet is for the daily grind. Sorting 00:15:47.580 --> 00:15:50.379 the emails, summarizing the transcripts, Obis 00:15:50.379 --> 00:15:53.350 is for the heavy lifting. like analyzing a complex 00:15:53.350 --> 00:15:56.769 legal contract. Exactly. You don't need a supercomputer 00:15:56.769 --> 00:15:59.529 to draft a calendar invite. And the other cost 00:15:59.529 --> 00:16:01.509 -saving measure brings us back to those text 00:16:01.509 --> 00:16:04.210 files. Keep your route instructions ruthlessly 00:16:04.210 --> 00:16:07.450 short. Let the routing map pull detailed contests 00:16:07.450 --> 00:16:10.090 from deeper subfolders only when it is actually 00:16:10.090 --> 00:16:12.590 needed for the task at hand. Before we wrap up, 00:16:12.870 --> 00:16:14.769 the sources mentioned a specific habit to build. 00:16:15.590 --> 00:16:18.409 What exactly is a session audit before closing 00:16:18.409 --> 00:16:21.299 a chat window? It's a really powerful, simple 00:16:21.299 --> 00:16:23.860 prompt to use at the end of your day. Before 00:16:23.860 --> 00:16:26.059 you hit the close button and wipe that temporary 00:16:26.059 --> 00:16:28.500 context window, you just ask Claude to scan our 00:16:28.500 --> 00:16:30.919 current conversation. You ask it to find anything 00:16:30.919 --> 00:16:33.379 worth saving to the memory files and update them. 00:16:33.580 --> 00:16:35.700 It's a quick debrief to ensure nothing valuable 00:16:35.700 --> 00:16:38.500 gets forgotten. Exactly. It ensures the system 00:16:38.500 --> 00:16:41.000 keeps learning and compounding without you having 00:16:41.000 --> 00:16:43.659 to sit there doing manual data entry or trying 00:16:43.659 --> 00:16:45.659 to remember what you discussed three hours ago. 00:16:45.879 --> 00:16:48.559 This whole conversation really highlights a massive 00:16:48.490 --> 00:16:51.610 shift in how we interact with machines. We're 00:16:51.610 --> 00:16:54.230 looking at a profound paradigm shift. For the 00:16:54.230 --> 00:16:56.110 last two years, everyone has been obsessed with 00:16:56.110 --> 00:16:58.669 finding the perfect combination of words. The 00:16:58.669 --> 00:17:01.250 age of obsessing over individual prompt engineering 00:17:01.250 --> 00:17:04.190 is ending. It really is. The focus on the magic 00:17:04.190 --> 00:17:07.569 prompt was a symptom of temporary AI. We're entering 00:17:07.569 --> 00:17:10.970 a completely new era. The new era is context 00:17:10.970 --> 00:17:13.849 engineering. The ultimate utility of artificial 00:17:13.849 --> 00:17:17.559 intelligence no longer depends on perfectly wording 00:17:17.559 --> 00:17:20.839 a single prompt in an empty chat box. It depends 00:17:20.839 --> 00:17:23.700 entirely on building an organized, logical folder 00:17:23.700 --> 00:17:26.019 infrastructure around the AI. It's about building 00:17:26.019 --> 00:17:28.859 a robust foundation. If the foundation is solid, 00:17:29.140 --> 00:17:31.619 the AI does the heavy lifting effortlessly. Your 00:17:31.619 --> 00:17:34.160 Co -Work OS is essentially a living database 00:17:34.160 --> 00:17:36.660 of your professional life, and it compounds in 00:17:36.660 --> 00:17:39.390 value over time. every single session, every 00:17:39.390 --> 00:17:41.549 single correction, teaches the system a little 00:17:41.549 --> 00:17:43.269 more about how you operate and how you think. 00:17:43.450 --> 00:17:45.230 And that's the most beautiful part of this entire 00:17:45.230 --> 00:17:47.529 framework. You don't need a perfect, massive 00:17:47.529 --> 00:17:49.750 system on day one. In fact, you shouldn't try 00:17:49.750 --> 00:17:51.549 to build one. You just need a working system. 00:17:51.789 --> 00:17:54.630 So here is your call to action for today. Start 00:17:54.630 --> 00:17:57.470 incredibly small. Build just your master root 00:17:57.470 --> 00:18:00.869 folder. Set up your claw .md file. Maybe add 00:18:00.869 --> 00:18:03.609 two simple workstations, perhaps an email HQ 00:18:03.609 --> 00:18:06.230 and a project folder. Get comfortable with how 00:18:06.230 --> 00:18:08.250 the active memory talks to the routing file. 00:18:08.450 --> 00:18:11.349 Yeah, let your real actual daily needs dictate 00:18:11.349 --> 00:18:13.869 the structure. Don't build elaborate workstations 00:18:13.869 --> 00:18:15.410 you don't actually use just because they look 00:18:15.410 --> 00:18:17.750 nice in a folder tree. It all comes back to that 00:18:17.750 --> 00:18:19.910 initial frustration we talked about at the start, 00:18:20.309 --> 00:18:23.170 the amnesiac chatbot. You don't have to live 00:18:23.170 --> 00:18:25.109 with a tool that forgets you every single morning 00:18:25.109 --> 00:18:27.369 anymore. You can build a system that actually 00:18:27.369 --> 00:18:29.539 remembers. Right. You're building a permanent 00:18:29.539 --> 00:18:31.680 partner. But stepping back, it leaves us with 00:18:31.680 --> 00:18:34.980 something much deeper to consider. If this persistent 00:18:34.980 --> 00:18:37.960 compounding system perfectly mimics your writing 00:18:37.960 --> 00:18:41.000 style over time, if it learns your organizational 00:18:41.000 --> 00:18:43.680 habits and begins anticipating your decision 00:18:43.680 --> 00:18:46.480 -making processes, will there come a day where 00:18:46.480 --> 00:18:48.680 you actually forget where your natural workflow 00:18:48.680 --> 00:18:51.660 ends and the AI's structural influence begins? 00:18:52.160 --> 00:18:54.700 Beat. What happens when your artificial coworker 00:18:54.700 --> 00:18:56.940 knows your own habits and preferences better 00:18:56.940 --> 00:18:57.500 than you do?