WEBVTT 00:00:00.000 --> 00:00:02.600 You know, there is a universal friction we all 00:00:02.600 --> 00:00:05.160 experience when we sit down to build. Oh, absolutely. 00:00:05.259 --> 00:00:08.080 You get completely in the zone. The ideas are 00:00:08.080 --> 00:00:10.759 flowing perfectly. But instead of actually building, 00:00:11.339 --> 00:00:13.859 you're forced to context switch. Yeah, the worst. 00:00:14.000 --> 00:00:16.559 You jump over to a terminal window. Then you 00:00:16.559 --> 00:00:18.839 flip to a browser tab to read some documentation. 00:00:19.019 --> 00:00:21.780 Right. Then you tab over to a separate chat box. 00:00:22.160 --> 00:00:25.460 Two secs silence. It completely fractures your 00:00:25.460 --> 00:00:28.260 flow state. We've spent decades adapting our 00:00:28.260 --> 00:00:30.420 brains to how computers organize information. 00:00:30.679 --> 00:00:33.100 Which is pretty unnatural, honestly. It is. So 00:00:33.100 --> 00:00:36.600 what if your entire digital workspace lived inside 00:00:36.600 --> 00:00:39.960 one intelligent window? That is the holy grail, 00:00:40.000 --> 00:00:42.799 isn't it? Yeah. We leak so much cognitive energy 00:00:42.799 --> 00:00:45.619 just shuttling context between different apps. 00:00:46.179 --> 00:00:47.960 Our brains really aren't wired for that kind 00:00:47.960 --> 00:00:50.640 of constant switching. Welcome to the Deep Dive. 00:00:50.759 --> 00:00:53.240 Today we're... We're exploring the new Claude 00:00:53.240 --> 00:00:55.979 Code desktop app. Yeah, and we are basing this 00:00:55.979 --> 00:00:58.240 on the architect's manual. But we aren't just 00:00:58.240 --> 00:01:00.259 going to read you a feature list today. We're 00:01:00.259 --> 00:01:02.420 exploring a fundamental shift in how builders 00:01:02.420 --> 00:01:04.930 operate. We're looking at how this app moves 00:01:04.930 --> 00:01:07.230 away from being a simple chat box. It acts more 00:01:07.230 --> 00:01:09.370 like a unified workshop. Yes. We're going to 00:01:09.370 --> 00:01:12.430 break down the philosophy behind plan mode. We'll 00:01:12.430 --> 00:01:14.349 look at the future of work through automated 00:01:14.349 --> 00:01:18.030 routines. And we'll honestly examine where the 00:01:18.030 --> 00:01:20.290 illusion breaks down. It's a massive behavioral 00:01:20.290 --> 00:01:23.250 shift. Yeah. I'm really eager to unpack the psychology 00:01:23.250 --> 00:01:26.069 here, how this tool actually changes our daily 00:01:26.069 --> 00:01:28.959 workflows. Let's start with the physical architecture 00:01:28.959 --> 00:01:32.859 of this new space. To understand how the AI operates 00:01:32.859 --> 00:01:34.859 differently here, we have to look at the layout. 00:01:35.239 --> 00:01:37.640 The manual highlights three specific tabs at 00:01:37.640 --> 00:01:42.400 the top, chat .co -work and code. Clicking that 00:01:42.400 --> 00:01:46.239 code tab opens a dark focused workspace. It immediately 00:01:46.239 --> 00:01:48.379 sets a different tone. It feels like a serious 00:01:48.379 --> 00:01:50.560 developer environment. It certainly doesn't feel 00:01:50.560 --> 00:01:53.359 like a conversational chatbot anymore. You're 00:01:53.359 --> 00:01:55.239 presented with this brilliant three -pane layout. 00:01:55.480 --> 00:01:57.540 So on the far left, you have mission control. 00:01:57.799 --> 00:01:59.840 That's where your project lists and active sessions 00:01:59.840 --> 00:02:02.400 live. Then the middle column is your active session 00:02:02.400 --> 00:02:05.019 that holds your chat thread and your input box. 00:02:05.620 --> 00:02:09.219 And the right side is this flexible zone for 00:02:09.219 --> 00:02:11.199 extra panes. That right pane is where the magic 00:02:11.199 --> 00:02:13.360 happens. You can open a preview pane to see your 00:02:13.360 --> 00:02:16.789 app running or... a plan view to see architecture 00:02:16.789 --> 00:02:20.090 or diff view to see exact code changes. The layout 00:02:20.090 --> 00:02:22.729 relies heavily on drag and drop. It's about reducing 00:02:22.729 --> 00:02:25.270 cognitive load. If you're reviewing a massive 00:02:25.270 --> 00:02:27.469 code refactor, you don't want to squint at a 00:02:27.469 --> 00:02:30.430 tiny box. Exactly. You just grab that diff view 00:02:30.430 --> 00:02:32.330 and drag it to take up the whole screen. Yeah. 00:02:32.810 --> 00:02:35.069 Or maybe you're tweaking user interface colors. 00:02:35.770 --> 00:02:38.129 You can drag the live preview right next to the 00:02:38.129 --> 00:02:41.270 chat. It completely eliminates that endless alt 00:02:41.270 --> 00:02:43.550 -tabbing we're all so used to doing. The analogy 00:02:43.550 --> 00:02:45.810 in the original manual didn't quite capture the 00:02:45.810 --> 00:02:48.509 mechanism for me. Oh, really? Yeah. Working with 00:02:48.509 --> 00:02:51.150 old AI web chats felt like working at a tiny 00:02:51.150 --> 00:02:54.569 cluttered desk. Right. But this new layout, it 00:02:54.569 --> 00:02:57.430 feels more like stacking Leo blocks of data on 00:02:57.430 --> 00:03:00.349 a massive multi -room factory floor. Oh, I like 00:03:00.349 --> 00:03:03.509 that. Because with the old web chats, the AI 00:03:03.509 --> 00:03:07.080 overheard every single conversation. It would 00:03:07.080 --> 00:03:09.460 hallucinate code from a marketing site you asked 00:03:09.460 --> 00:03:12.139 about three days ago. Yeah, that cross -contamination 00:03:12.139 --> 00:03:14.699 is incredibly frustrating. So this new layout 00:03:14.699 --> 00:03:17.860 uses a project -first model. The AI is placed 00:03:17.860 --> 00:03:20.759 in a soundproof room. It only has access to the 00:03:20.759 --> 00:03:22.539 specific blueprints you slide under the door. 00:03:22.939 --> 00:03:25.560 Work is strictly separated by folders. And that 00:03:25.560 --> 00:03:28.159 folder boundary is enforced mechanically. It 00:03:28.159 --> 00:03:30.219 uses something called Git work trees underneath 00:03:30.219 --> 00:03:32.400 the hood. Let's define that jargon really quickly. 00:03:32.460 --> 00:03:34.939 Sure. A Git work tree is basically a way to work 00:03:34.939 --> 00:03:37.060 on different parts of code at once. That's spot 00:03:37.060 --> 00:03:40.199 on. Agent A can be completely rewriting your 00:03:40.199 --> 00:03:43.439 payment module. At the exact same moment, Agent 00:03:43.439 --> 00:03:45.879 B is writing unit tests for your user login. 00:03:46.360 --> 00:03:48.259 Wow. Yeah, they don't step on each other's toes 00:03:48.259 --> 00:03:50.819 at all. So the workspace is organized, but an 00:03:50.819 --> 00:03:53.580 empty room doesn't write software. Why is this 00:03:53.580 --> 00:03:56.159 isolated folder -based context such a massive 00:03:56.159 --> 00:03:58.180 game changer compared to the old chat threads? 00:03:58.460 --> 00:04:01.099 Because language models are highly susceptible 00:04:01.099 --> 00:04:04.860 to noise, when you isolate the context to a single 00:04:04.860 --> 00:04:08.759 folder, the AI's attention mechanism becomes 00:04:08.759 --> 00:04:11.099 hyper -focused. It doesn't have to guess which 00:04:11.099 --> 00:04:13.479 version of a file you care about. So clean folders 00:04:13.479 --> 00:04:15.580 mean the AI never pulls context from the wrong 00:04:15.580 --> 00:04:18.139 project. It's a hard, impenetrable boundary. 00:04:18.279 --> 00:04:21.319 So the physical space is set, but how do we actually 00:04:21.319 --> 00:04:23.819 breathe life into it? The first decision a user 00:04:23.819 --> 00:04:26.620 makes actually dictates the AI's entire operating 00:04:26.620 --> 00:04:29.220 boundary. Right. And you have to set the foundation 00:04:29.220 --> 00:04:31.740 properly. This requires a fresh installation 00:04:31.740 --> 00:04:35.160 of the desktop app. And, crucially, it requires 00:04:35.160 --> 00:04:37.870 a Claude Pro subscription. The manual mentions 00:04:37.870 --> 00:04:40.089 the free tier is mostly just for testing the 00:04:40.089 --> 00:04:42.470 visual layout. Yeah, because the AI is reading 00:04:42.470 --> 00:04:44.949 your entire project folder, the context window 00:04:44.949 --> 00:04:46.949 fills up immediately. You'll hit the free limits 00:04:46.949 --> 00:04:49.470 in minutes. You really need the pro plan for 00:04:49.470 --> 00:04:51.209 this workflow. Let's pause and talk about the 00:04:51.209 --> 00:04:54.110 economics of that. The pro plan is roughly $20 00:04:54.110 --> 00:04:56.259 a month. According to the source, that gives 00:04:56.259 --> 00:05:00.339 you about 44 ,000 tokens per 5 -hour Watto. That 00:05:00.339 --> 00:05:02.720 translates to anywhere from 10 to 40 prompts. 00:05:02.959 --> 00:05:05.860 But I have to push back a bit here. That scarcity 00:05:05.860 --> 00:05:07.980 fundamentally changes the relationship, doesn't 00:05:07.980 --> 00:05:10.959 it? 44 ,000 tokens can run out fast on heavy 00:05:10.959 --> 00:05:14.120 code bases. Users really need to be strategic. 00:05:14.339 --> 00:05:17.459 Oh, it absolutely does. You can't just spam the 00:05:17.459 --> 00:05:20.699 AI with lazy half -baked thoughts anymore. You 00:05:20.699 --> 00:05:23.819 have to treat every single prompt like an expensive 00:05:23.819 --> 00:05:26.459 managerial directive. Which means choosing your 00:05:26.459 --> 00:05:29.040 operating environment carefully. When you boot 00:05:29.040 --> 00:05:32.139 up a session, you get three choices. Local, Remote, 00:05:32.379 --> 00:05:35.360 or SSH. Right. Let's explore the implications 00:05:35.360 --> 00:05:37.920 of those choices. Local is your speed layer. 00:05:38.180 --> 00:05:40.500 It runs directly on your local machine and reads 00:05:40.500 --> 00:05:43.240 your file system natively. It skips the delays 00:05:43.240 --> 00:05:45.180 of pushing and pulling from a cloud repository. 00:05:45.279 --> 00:05:48.120 It's incredibly fast for rapid iteration. The 00:05:48.120 --> 00:05:50.339 manual notes Windows users need Git installed 00:05:50.339 --> 00:05:52.899 for that while Macs have it natively. What about 00:05:52.899 --> 00:05:55.560 the remote environment? Remote operates completely 00:05:55.560 --> 00:05:58.319 differently. It runs safely on Anthropic's cloud 00:05:58.319 --> 00:06:00.420 infrastructure. The brilliant part is that it 00:06:00.420 --> 00:06:02.699 survives even if you close your laptop. Oh, wow. 00:06:02.839 --> 00:06:05.680 Yeah. It's much safer for delicate client projects 00:06:05.680 --> 00:06:08.040 because changes are stored in a dedicated branch 00:06:08.040 --> 00:06:10.959 with a full history. And SSH is the third option. 00:06:11.339 --> 00:06:13.699 SSH just allows the session to tunnel into a 00:06:13.699 --> 00:06:16.199 remote server that you control. It's great for 00:06:16.199 --> 00:06:18.120 folks managing their own bare metal hardware. 00:06:18.459 --> 00:06:20.990 So we have our environment. Next is the model 00:06:20.990 --> 00:06:24.889 selection. We have Claude Sonnet 4 .6 and Claude 00:06:24.889 --> 00:06:28.529 Opus 4 .7. Think of Sonnet as your highly caffeinated 00:06:28.529 --> 00:06:31.910 junior developer. It's fast, it's great for daily 00:06:31.910 --> 00:06:35.230 tasks, simple UI changes, and quick bug fixes. 00:06:35.490 --> 00:06:38.930 And Opus. Opus is your lead architect. You bring 00:06:38.930 --> 00:06:41.569 in Opus for deep reasoning, complex database 00:06:41.569 --> 00:06:44.089 migrations, or massive architectural shifts. 00:06:44.310 --> 00:06:46.170 There's also an effort selector. You can choose 00:06:46.170 --> 00:06:49.800 low, medium, high, or auto. They simplified that 00:06:49.800 --> 00:06:52.500 nicely. Low effort means the agent moves quickly 00:06:52.500 --> 00:06:55.079 and doesn't overthink small tweaks. High effort 00:06:55.079 --> 00:06:57.240 forces the model into a deeper chain of thought 00:06:57.240 --> 00:06:59.759 process before it touches a single line of code. 00:07:00.120 --> 00:07:02.300 But the real leverage seems to be hidden. There's 00:07:02.300 --> 00:07:04.540 a subtle menu inside the input box that a lot 00:07:04.540 --> 00:07:06.839 of people overlook. Oh, the context menu. It 00:07:06.839 --> 00:07:08.339 is the nerve center of the whole application. 00:07:08.540 --> 00:07:10.980 You click that little icon and a massive set 00:07:10.980 --> 00:07:13.790 of tools opens up. You can inject specific reference 00:07:13.790 --> 00:07:16.810 files. You can pull live GitHub issues directly 00:07:16.810 --> 00:07:19.610 into the AI's working memory. And you have connectors. 00:07:20.029 --> 00:07:22.930 You can seamlessly link the session to your Gmail, 00:07:23.129 --> 00:07:25.389 your Slack channels, or your Notion workspaces. 00:07:25.470 --> 00:07:27.550 Right. You also access your skills from that 00:07:27.550 --> 00:07:30.389 menu and the transcript mode where you can toggle 00:07:30.389 --> 00:07:33.879 between verbose, normal, or summary. It's an 00:07:33.879 --> 00:07:35.660 incredible amount of power right at the cursor. 00:07:36.060 --> 00:07:38.259 So for someone starting a brand new project this 00:07:38.259 --> 00:07:41.779 afternoon, how should they choose between local 00:07:41.779 --> 00:07:44.259 and remote environments? I always recommend starting 00:07:44.259 --> 00:07:46.800 local. You want that tight, immediate feedback 00:07:46.800 --> 00:07:49.199 loop when you're initially prototyping. Once 00:07:49.199 --> 00:07:51.319 the foundational architecture is stable, you 00:07:51.319 --> 00:07:53.360 shift the session to remote. That protects the 00:07:53.360 --> 00:07:55.680 code base with a continuous cloud history. Local 00:07:55.680 --> 00:07:58.579 for speed and testing, remote for safe, continuous 00:07:58.579 --> 00:08:02.279 cloud history. Exactly right. So we've established 00:08:02.279 --> 00:08:05.379 that tokens are precious. We can't afford sloppy 00:08:05.379 --> 00:08:08.939 meandering mistakes to sex islands. Jumping straight 00:08:08.939 --> 00:08:11.240 into coding is usually what causes those expensive 00:08:11.240 --> 00:08:14.180 errors. There is a specific feature designed 00:08:14.180 --> 00:08:19.199 to purposely slow the AI down. Plan mode. I genuinely 00:08:19.199 --> 00:08:21.439 believe it's the most profoundly undervalued 00:08:21.439 --> 00:08:24.079 feature in the entire ecosystem. To understand 00:08:24.079 --> 00:08:26.199 plan mode, we first have to look at the permission 00:08:26.199 --> 00:08:29.319 states. The default state is ask permissions. 00:08:29.420 --> 00:08:32.710 Right. It's incredibly safe. but painfully slow. 00:08:33.409 --> 00:08:36.870 The AI pauses to ask your permission before every 00:08:36.870 --> 00:08:39.250 single file edit. Then you have the opposite 00:08:39.250 --> 00:08:42.470 end of the spectrum. Auto accept edits. The agent 00:08:42.470 --> 00:08:45.769 runs wild. It writes files, edits code, and executes 00:08:45.769 --> 00:08:48.370 terminal commands freely. It only pauses for 00:08:48.370 --> 00:08:51.429 highly destructive actions, like deleting a directory. 00:08:51.870 --> 00:08:53.929 And sitting outside of both of those is plan 00:08:53.929 --> 00:08:56.250 mode. Plan mode forces the AI to put its hands 00:08:56.250 --> 00:08:59.059 behind its back. It drafts a highly detailed 00:08:59.059 --> 00:09:01.620 comprehensive roadmap and pure markdown. You 00:09:01.620 --> 00:09:03.519 get structured headings, you get code blocks, 00:09:04.200 --> 00:09:06.799 but it absolutely refuses to write project code 00:09:06.799 --> 00:09:08.860 immediately. The manual provides a brilliant 00:09:08.860 --> 00:09:11.080 example of this constraint in action. You tell 00:09:11.080 --> 00:09:12.820 the agent you want to build an on -chain token 00:09:12.820 --> 00:09:15.200 watch list application, but before it writes 00:09:15.200 --> 00:09:17.559 anything, you command it to ask you five highly 00:09:17.559 --> 00:09:20.139 targeted questions about scope, auth, and the 00:09:20.139 --> 00:09:22.740 tech stack. I love that workflow. You force the 00:09:22.740 --> 00:09:25.350 AI to interrogate you. It asks about your target 00:09:25.350 --> 00:09:27.690 audience. It asks about your preferred database 00:09:27.690 --> 00:09:30.309 structure. And the incredible part is, the AI 00:09:30.309 --> 00:09:32.970 actually stops and waits for your answers. It 00:09:32.970 --> 00:09:35.570 does. It sits there patiently. Once you answer, 00:09:35.590 --> 00:09:38.350 it generates a meticulous step -by -step build 00:09:38.350 --> 00:09:41.110 plan. It outlines the precise file structure, 00:09:41.529 --> 00:09:44.629 the API endpoints, and the UI components. It 00:09:44.629 --> 00:09:47.549 forces alignment before a single token is wasted 00:09:47.549 --> 00:09:50.149 on bad code. I'll make a vulnerable admission 00:09:50.149 --> 00:09:52.929 right here. beat. I still wrestle with prompt 00:09:52.929 --> 00:09:55.690 drift myself. Oh yeah. Yeah. I get so impatient 00:09:55.690 --> 00:09:57.990 when I try to skip the planning phase. I see 00:09:57.990 --> 00:10:00.610 that blinking cursor and I just want the dopamine 00:10:00.610 --> 00:10:02.549 hit of watching code generate on the screen. 00:10:02.690 --> 00:10:04.769 We are all guilty of that. We've been conditioned 00:10:04.769 --> 00:10:07.409 by standard chatbots to expect instant gratification. 00:10:07.950 --> 00:10:10.309 But skipping plan mode on a fresh build is a 00:10:10.309 --> 00:10:12.629 devastating trap. You inevitably spend three 00:10:12.629 --> 00:10:15.129 times as many tokens untangling a spaghetti code 00:10:15.129 --> 00:10:17.549 mess three hours later. You review the draft 00:10:17.549 --> 00:10:19.990 in the dedicated plan pane. You spot the missing 00:10:19.990 --> 00:10:22.549 logic. You correct it. Only then do you switch 00:10:22.549 --> 00:10:25.409 the permissions to accept edits. It's the ultimate 00:10:25.409 --> 00:10:28.730 preventative measure. But to be fair, you shouldn't 00:10:28.730 --> 00:10:31.129 use it for everything. Right. The manual notes 00:10:31.129 --> 00:10:33.629 that plan mode is gold for new builds, but it's 00:10:33.629 --> 00:10:36.370 massive overkill for fixing a simple CSS alignment 00:10:36.370 --> 00:10:39.009 issue or a tiny bug. If you just need a button 00:10:39.009 --> 00:10:41.870 move five pixels to the left, use auto accept 00:10:41.870 --> 00:10:44.970 edits. Feed the agent the exact file name and 00:10:44.970 --> 00:10:48.049 the exact line number. Don't overcomplicate small 00:10:48.049 --> 00:10:51.090 chores. That does make me wonder about the psychological 00:10:51.090 --> 00:10:54.110 friction. Does forcing yourself into plan mode 00:10:54.110 --> 00:10:56.970 kill the momentum of rapid prototyping? It definitely 00:10:56.970 --> 00:10:59.669 feels slower in the first 15 minutes, but it 00:10:59.669 --> 00:11:03.230 is exponentially faster by hour three. The AI 00:11:03.230 --> 00:11:05.789 catches fatal architectural flaws before they're 00:11:05.789 --> 00:11:08.269 ever coded into existence. Plan mode actually 00:11:08.269 --> 00:11:10.870 saves hours by preventing massive code rewrites 00:11:10.870 --> 00:11:13.500 later. Measure twice, cut once. It's timeless 00:11:13.500 --> 00:11:16.080 advice. Sponsor. So the workspace is set. The 00:11:16.080 --> 00:11:18.940 architectural plan is rock solid. But the execution 00:11:18.940 --> 00:11:20.679 phase is where this tool breaks out of being 00:11:20.679 --> 00:11:23.259 just a fancy text editor. This brings us to the 00:11:23.259 --> 00:11:26.899 new automation capabilities, routines, and skills. 00:11:27.419 --> 00:11:29.600 Routines are where we start seeing the future 00:11:29.600 --> 00:11:32.480 of software development. Anthropic rolled these 00:11:32.480 --> 00:11:35.960 out on April 14, 2026 as a research preview. 00:11:36.240 --> 00:11:39.279 Let's unpack what a routine actually is. At its 00:11:39.279 --> 00:11:42.279 core, a routine is a cloud -based configuration 00:11:42.279 --> 00:11:44.580 that runs automatically. It bundles together 00:11:44.580 --> 00:11:47.960 a specific prompt, a target repository, and the 00:11:47.960 --> 00:11:51.019 necessary API connectors. And crucially, it runs 00:11:51.019 --> 00:11:53.700 completely autonomously on Anthropic's cloud. 00:11:53.899 --> 00:11:55.840 Which means your laptop can be closed and stored 00:11:55.840 --> 00:11:57.980 in your backpack. Exactly. It operates entirely 00:11:57.980 --> 00:12:00.299 independently of your local hardware. The manual 00:12:00.299 --> 00:12:03.139 details three specific ways to trigger these 00:12:03.139 --> 00:12:05.159 routines. Let's walk through them. The first 00:12:05.159 --> 00:12:07.980 is schedule triggers. These use standard Krone 00:12:07.980 --> 00:12:10.679 expressions. You can configure a routine to run 00:12:10.679 --> 00:12:13.019 every night at midnight or every Monday morning. 00:12:13.600 --> 00:12:15.480 Anthropic currently caps the frequency at once 00:12:15.480 --> 00:12:17.960 per hour to manage compute loads. The second 00:12:17.960 --> 00:12:20.549 trigger is what really caught my eye. API triggers. 00:12:21.149 --> 00:12:23.750 Every single routine is assigned its own unique 00:12:23.750 --> 00:12:27.149 HTTP endpoint. Right. You just had a simple POST 00:12:27.149 --> 00:12:29.750 request with an authentication token and the 00:12:29.750 --> 00:12:33.509 AI session boots up in the cloud. Wow. Imagine 00:12:33.509 --> 00:12:36.950 wiring that API trigger directly up to your deploy 00:12:36.950 --> 00:12:40.629 hooks. You could literally have the AI review 00:12:40.629 --> 00:12:43.009 your code while you sleep. It's mind blowing. 00:12:43.429 --> 00:12:46.250 You literally wake up to a fully vetted perfectly 00:12:46.250 --> 00:12:50.169 documented pull request. It's essentially a tireless 00:12:50.169 --> 00:12:52.370 junior engineer working the graveyard shift. 00:12:52.370 --> 00:12:55.309 Yeah. And that perfectly segues into the third 00:12:55.309 --> 00:12:58.730 trigger type, GitHub triggers. The AI listens 00:12:58.730 --> 00:13:01.929 to the repository's heartbeat. Yes. It natively 00:13:01.929 --> 00:13:04.629 subscribes to specific repository events. If 00:13:04.629 --> 00:13:06.830 someone opens a new pull request or cuts a new 00:13:06.830 --> 00:13:09.470 release, the routine automatically springs into 00:13:09.470 --> 00:13:11.629 action. Wow. You just have to authorize the Cloud 00:13:11.629 --> 00:13:14.049 GitHub app first. The source material gives a 00:13:14.049 --> 00:13:16.289 brilliantly practical example of a schedule trigger, 00:13:16.750 --> 00:13:19.250 the 7 a .m. daily morning brief. It's such a 00:13:19.250 --> 00:13:21.809 clever quality of life hack. Every weekday morning 00:13:21.809 --> 00:13:24.269 at exactly 7 a .m., the routine quietly spins 00:13:24.269 --> 00:13:26.730 up. It uses the Gmail connector to read your 00:13:26.730 --> 00:13:29.610 inbox. It scans for urgent deadlines or missed 00:13:29.610 --> 00:13:31.570 payments. Then it reads your Google Calendar 00:13:31.570 --> 00:13:33.509 for the day. Without you touching a single key. 00:13:33.750 --> 00:13:36.629 Right. It synthesizes all that chaotic information 00:13:36.629 --> 00:13:40.450 into a clean, structured markdown document. Finally, 00:13:40.649 --> 00:13:43.090 it uses the Flac connector to drop that brief 00:13:43.090 --> 00:13:45.230 directly into your private channel before you 00:13:45.230 --> 00:13:47.590 even wake up. That's a beautiful example of cross 00:13:47.590 --> 00:13:50.090 -platform orchestration. But we also need to 00:13:50.090 --> 00:13:52.789 examine skills, which operate a bit differently. 00:13:52.990 --> 00:13:55.769 Skills are what transform clogged code from a 00:13:55.769 --> 00:13:58.529 coding assistant into a customizable work platform. 00:13:59.009 --> 00:14:02.009 A skill is essentially a deeply ingrained reusable 00:14:02.009 --> 00:14:05.009 set of instructions loaded on demand. The manual 00:14:05.009 --> 00:14:07.830 suggests creating a product -led Prad skill. 00:14:08.110 --> 00:14:10.590 When you invoke it, the AI instantly assumes 00:14:10.590 --> 00:14:13.529 the persona of a rigorous product manager. It 00:14:13.529 --> 00:14:16.149 automatically asks seven specific questions about 00:14:16.149 --> 00:14:18.610 user pain points, demographics, and technical 00:14:18.610 --> 00:14:20.690 constraints. So instead of copying and pasting 00:14:20.690 --> 00:14:23.169 that massive prompt from a notes app every single 00:14:23.169 --> 00:14:25.269 time you have an idea, you just type a slash 00:14:25.269 --> 00:14:28.190 command. Every new project gets the exact same 00:14:28.190 --> 00:14:30.309 rigorous treatment. You can load these skills 00:14:30.309 --> 00:14:33.129 via a zip file or by pasting a terminal command. 00:14:33.399 --> 00:14:36.159 or even using skills .sh. Yeah, it gives your 00:14:36.159 --> 00:14:38.679 tool stack immense flexibility. You just have 00:14:38.679 --> 00:14:41.059 to be careful not to install dozens of them at 00:14:41.059 --> 00:14:43.279 once, or they start competing for the model's 00:14:43.279 --> 00:14:45.779 attention. This raises an interesting operational 00:14:45.779 --> 00:14:47.820 question. How do these native skills actually 00:14:47.820 --> 00:14:49.879 differ from just keeping a Notion document full 00:14:49.879 --> 00:14:52.360 of your favorite copy -paste prompts? A Notion 00:14:52.360 --> 00:14:56.360 document is just dead text. A skill is a living, 00:14:56.379 --> 00:15:00.519 active tool. When you invoke a skill, it natively 00:15:00.519 --> 00:15:03.200 configures the agent's internal parameters and 00:15:03.200 --> 00:15:06.019 runs its built -in slash commands automatically 00:15:06.019 --> 00:15:08.299 within the session's memory. Skills are native 00:15:08.299 --> 00:15:10.899 reusable AI tools that automatically run their 00:15:10.899 --> 00:15:13.179 built -in slash commands. They literally become 00:15:13.179 --> 00:15:15.929 extensions of the agent's brain. It sounds like 00:15:15.929 --> 00:15:19.350 a deeply seductive, perfect system, but to actually 00:15:19.350 --> 00:15:21.350 master it, we have to look past the marketing. 00:15:21.690 --> 00:15:23.870 We have to understand the underlying mental model, 00:15:23.870 --> 00:15:25.950 and we have to be brutally honest about where 00:15:25.950 --> 00:15:28.789 the illusion currently breaks down. The underlying 00:15:28.789 --> 00:15:31.309 mental model is the skeleton key to this whole 00:15:31.309 --> 00:15:34.269 app. It breaks down into three very simple components, 00:15:34.909 --> 00:15:38.309 context, agents, outputs. Let's explore those. 00:15:38.669 --> 00:15:41.190 Context is simply the folder. Right, but it's 00:15:41.190 --> 00:15:44.370 not just storage. The folder is the AI's short 00:15:44.370 --> 00:15:46.379 -term memory. It holds your code base, sure. 00:15:46.759 --> 00:15:49.320 But it should also hold reference PDFs, brand 00:15:49.320 --> 00:15:52.720 guidelines, and old architecture docs. The richer 00:15:52.720 --> 00:15:55.500 the soil in that folder, the better the AI's 00:15:55.500 --> 00:15:58.399 output. The manual calls out a very specific 00:15:58.399 --> 00:16:00.980 widespread mistake people make regarding context. 00:16:01.779 --> 00:16:04.000 Staring at a completely empty folder and typing 00:16:04.000 --> 00:16:07.639 a massive 500 -word prompt, the agent has absolutely 00:16:07.639 --> 00:16:09.820 zero grounding. You have to invert that behavior. 00:16:10.159 --> 00:16:12.620 Fill the folder with dense context first. Right. 00:16:12.750 --> 00:16:15.549 Then write a very short, highly focused prompt. 00:16:15.850 --> 00:16:18.070 Let the agent connect the dots. The second pillar 00:16:18.070 --> 00:16:21.129 is agents. And an agent is not just the underlying 00:16:21.129 --> 00:16:23.549 language model. That's a crucial distinction. 00:16:24.090 --> 00:16:26.629 An agent is Claude Opus, combined with your specific 00:16:26.629 --> 00:16:28.610 permission settings, your active skills, and 00:16:28.610 --> 00:16:30.809 your enabled connectors. You can run the exact 00:16:30.809 --> 00:16:32.649 same underlying model with different rule sets, 00:16:32.690 --> 00:16:34.809 and you will get wildly different behaviors. 00:16:35.110 --> 00:16:37.330 And the final pillar is outputs. Code is the 00:16:37.330 --> 00:16:39.929 obvious answer here. But the output can be anything 00:16:39.929 --> 00:16:42.759 now. It can be a fully formatted markdown brief. 00:16:43.080 --> 00:16:45.539 It can be a visual Canva design. It can be a 00:16:45.539 --> 00:16:50.019 complex calendar invite. The AI seamlessly moves 00:16:50.019 --> 00:16:52.460 across mediums without ever leaving that single 00:16:52.460 --> 00:16:54.940 window. You really start treating the interface 00:16:54.940 --> 00:16:58.039 like a small, highly capable team. But let's 00:16:58.039 --> 00:17:00.799 look at the rough edges. The manual is surprisingly 00:17:00.799 --> 00:17:03.460 candid about what still feels clunky or broken. 00:17:03.600 --> 00:17:06.640 The most glaring omission is the lack of a proper 00:17:06.640 --> 00:17:09.779 file viewer. There is no native file tree panel 00:17:09.779 --> 00:17:12.180 inside the application yet. Which is maddening 00:17:12.180 --> 00:17:13.859 if you're a visual thinker. You can't just click 00:17:13.859 --> 00:17:15.559 folders to see how the project is structured. 00:17:15.599 --> 00:17:18.299 You have to ask the AI to print out the directory 00:17:18.299 --> 00:17:21.279 structure. So for now, you end up keeping standard 00:17:21.279 --> 00:17:24.299 VS Code open on a second monitor just to maintain 00:17:24.299 --> 00:17:26.400 your spatial awareness of the project. I have 00:17:26.400 --> 00:17:28.299 to thoughtfully challenge this all -in -one workspace 00:17:28.299 --> 00:17:31.099 claim then. Oh. If a builder still needs VS Code 00:17:31.099 --> 00:17:33.180 open to see their files, and they still need 00:17:33.180 --> 00:17:35.779 Chrome DevTools open to debug their CSS, then 00:17:35.779 --> 00:17:38.579 this isn't truly a single window workspace. It's 00:17:38.579 --> 00:17:40.980 just a prettier terminal. I see why it feels 00:17:40.980 --> 00:17:43.539 that way. But I think you're looking at the wrong 00:17:43.539 --> 00:17:46.680 layer of abstraction. Clod code isn't trying 00:17:46.680 --> 00:17:49.539 to replace your IDE's file explorer. It's trying 00:17:49.539 --> 00:17:52.240 to replace your hands on the keyboard. Think 00:17:52.240 --> 00:17:54.960 of it like a manager's control room. You don't 00:17:54.960 --> 00:17:56.619 need to hold the wrench if you're directing the 00:17:56.619 --> 00:17:59.200 mechanic. That's a fascinating reframe. The preview 00:17:59.200 --> 00:18:01.099 console is another area that feels a bit hollow. 00:18:01.200 --> 00:18:03.460 It is very lightweight. You only get basic session 00:18:03.460 --> 00:18:08.029 logs. If you hit a complex front -end bug, the 00:18:08.029 --> 00:18:10.369 internal preview won't give you the deep inspection 00:18:10.369 --> 00:18:12.349 tools you need. You still have to open the local 00:18:12.349 --> 00:18:14.910 server in a real browser. And the cloud sandboxes 00:18:14.910 --> 00:18:16.849 have their own friction. They feel sluggish on 00:18:16.849 --> 00:18:19.190 that initial boot. It can take a solid minute 00:18:19.190 --> 00:18:21.990 or two to spin up a fresh GitHub branch sandbox 00:18:21.990 --> 00:18:24.990 in the cloud. That's why local is so crucial 00:18:24.990 --> 00:18:28.009 for the fast, messy iteration phase. So, knowing 00:18:28.009 --> 00:18:30.109 all these intricate mechanics, the folder rules, 00:18:30.329 --> 00:18:33.180 the token limits, the missing file trees... How 00:18:33.180 --> 00:18:35.960 should users handle the situation when the AI 00:18:35.960 --> 00:18:38.700 inevitably generates a terrible output, rather 00:18:38.700 --> 00:18:41.319 than just blaming the model's intelligence? You 00:18:41.319 --> 00:18:43.599 have to audit the environment you built for it. 00:18:43.880 --> 00:18:46.400 Did you fail to provide a key reference document 00:18:46.400 --> 00:18:49.079 in the folder? Are your permission rules too 00:18:49.079 --> 00:18:53.099 restrictive? The model itself is incredibly capable. 00:18:53.900 --> 00:18:56.759 The failure is almost always in the managerial 00:18:56.759 --> 00:18:59.240 setup. Don't blame the model. Fix your folder 00:18:59.240 --> 00:19:01.980 context or adjust the permission rules. You are 00:19:01.980 --> 00:19:04.279 the orchestrator now. You have to manage the 00:19:04.279 --> 00:19:06.599 workspace. Let's pull all these threads together. 00:19:06.680 --> 00:19:08.619 We've covered a tremendous amount of ground. 00:19:09.059 --> 00:19:10.880 For someone listening right now, what is the 00:19:10.880 --> 00:19:13.059 ultimate takeaway? The builders who are moving 00:19:13.059 --> 00:19:15.339 incredibly fast right now are combining three 00:19:15.339 --> 00:19:19.589 specific things. Plan mode. Custom skills. and 00:19:19.589 --> 00:19:22.650 cloud routines. That specific formula is what 00:19:22.650 --> 00:19:25.150 separates the fast builders from the slow ones. 00:19:25.470 --> 00:19:27.750 Moving from a chat box mentality to treating 00:19:27.750 --> 00:19:29.930 the app like a small managed team that ships 00:19:29.930 --> 00:19:32.009 work for you. It's a complete paradigm shift. 00:19:32.269 --> 00:19:34.750 It requires patience to learn, but the compounding 00:19:34.750 --> 00:19:37.170 returns on your time are massive. So here's our 00:19:37.170 --> 00:19:39.930 call to action for you. Pick one tiny low stakes 00:19:39.930 --> 00:19:42.930 project today. Open the app, create a strictly 00:19:42.930 --> 00:19:45.849 isolated folder, boot up a session in plan mode, 00:19:46.230 --> 00:19:48.509 and write your first architectural prompt. Just 00:19:48.509 --> 00:19:52.130 get one small win. The initial friction disappears 00:19:52.130 --> 00:19:55.450 very quickly. Two secs silence. But before we 00:19:55.450 --> 00:19:56.910 wrap up, I want to leave you with a lingering 00:19:56.910 --> 00:19:59.880 thought. We started this deep dive talking about 00:19:59.880 --> 00:20:02.380 the universal frustration of context switching, 00:20:03.039 --> 00:20:05.819 the deep desire for one intelligent window that 00:20:05.819 --> 00:20:08.880 understands our intent. We've just explored how 00:20:08.880 --> 00:20:11.220 Claude code now handles the initial project planning. 00:20:11.339 --> 00:20:14.519 It writes the PRDs. It executes the raw code. 00:20:14.660 --> 00:20:17.279 It even summarizes its own results on Slack for 00:20:17.279 --> 00:20:20.279 you while you sleep. beat. If the AI is handling 00:20:20.279 --> 00:20:22.299 the planning, the execution, and the reporting, 00:20:22.839 --> 00:20:24.839 well what exactly is the human's core role in 00:20:24.839 --> 00:20:27.599 the software loop of 2027? That is a real question 00:20:27.599 --> 00:20:29.160 we're all going to have to answer. Out to your 00:20:29.160 --> 00:20:29.579 own music.