WEBVTT 00:00:00.000 --> 00:00:04.200 You use AI every day. Beat. But that doesn't 00:00:04.200 --> 00:00:06.320 actually mean you're an advanced user. Right. 00:00:06.339 --> 00:00:08.740 I mean, it definitely feels advanced, but daily 00:00:08.740 --> 00:00:11.199 casual use is really just the beginning. I think 00:00:11.199 --> 00:00:14.039 about this a lot. We type a quick prompt into 00:00:14.039 --> 00:00:17.679 a chat box. We get a really solid answer. We 00:00:17.679 --> 00:00:20.899 feel incredibly productive. Oh, yeah. Totally. 00:00:21.140 --> 00:00:23.839 Beat. But we are barely scratching the surface 00:00:23.839 --> 00:00:27.219 of what is possible. Welcome to this deep dive. 00:00:28.379 --> 00:00:30.899 Today, we're charting a very specific path. We 00:00:30.899 --> 00:00:33.619 are exploring the five levels of cloud proficiency. 00:00:34.020 --> 00:00:36.200 We're going to move step by step through this 00:00:36.200 --> 00:00:38.420 entire framework. We'll start with those simple 00:00:38.420 --> 00:00:40.920 everyday chats that will climb all the way up. 00:00:41.060 --> 00:00:43.399 We'll explore fully autonomous background workflows. 00:00:44.000 --> 00:00:46.299 Our goal is to help you pinpoint exactly where 00:00:46.299 --> 00:00:48.320 you are right now. Right. And more importantly, 00:00:48.520 --> 00:00:50.500 we want to show you how to level up safely. So 00:00:50.500 --> 00:00:53.079 let's start at the absolute baseline. Level one. 00:00:53.200 --> 00:00:55.659 the enthusiast. Yeah, which is honestly where 00:00:55.659 --> 00:00:57.619 most people tend to live right now. You might 00:00:57.619 --> 00:00:59.619 assume you're a total power user just because 00:00:59.619 --> 00:01:02.119 you chat with the AI every single morning. But 00:01:02.119 --> 00:01:03.960 realistically, this is just the ground floor. 00:01:04.239 --> 00:01:07.260 Level one is really built for quick daily tasks. 00:01:07.900 --> 00:01:09.840 You know, you need a fast email draft. You need 00:01:09.840 --> 00:01:12.569 a short summary of a long article. Right. Or 00:01:12.569 --> 00:01:14.290 maybe you just need a simple explanation of a 00:01:14.290 --> 00:01:16.390 complex topic. It just gives you a much faster 00:01:16.390 --> 00:01:18.890 starting point. You paste in some rough meeting 00:01:18.890 --> 00:01:22.030 notes. You ask the AI for a clean version. Yeah. 00:01:22.250 --> 00:01:24.230 A few minutes later, you have something totally 00:01:24.230 --> 00:01:26.950 useful. It helps you think faster. You create 00:01:26.950 --> 00:01:30.329 a solid first draft. And then you just finish 00:01:30.329 --> 00:01:33.409 the actual work yourself. There is a very specific 00:01:33.409 --> 00:01:36.370 tactic here that I love. One that works incredibly 00:01:36.370 --> 00:01:39.750 well when you're at level one. What's that? Uploading 00:01:39.750 --> 00:01:41.969 screenshots directly into the chat. Oh, yeah. 00:01:42.090 --> 00:01:44.370 That is a fantastic habit to build. You really 00:01:44.370 --> 00:01:46.250 need to stop typing out every single detail. 00:01:46.370 --> 00:01:48.650 It just takes way too much time. Exactly. Because 00:01:48.650 --> 00:01:50.849 you might accidentally miss something important. 00:01:51.030 --> 00:01:52.969 Just share exactly what you see on your screen. 00:01:53.469 --> 00:01:56.109 Exactly. So imagine you get a really confusing 00:01:56.109 --> 00:01:59.370 dashboard error, or you see a weird error message 00:01:59.370 --> 00:02:01.189 during your work. Right. You panic a little. 00:02:01.469 --> 00:02:03.950 Yeah. But you don't try to transcribe the code. 00:02:04.409 --> 00:02:07.370 You just snap a quick screenshot. You upload 00:02:07.370 --> 00:02:10.509 the image directly. Then you ask it to explain 00:02:10.509 --> 00:02:13.270 the error in simple English, and you ask for 00:02:13.270 --> 00:02:15.669 the very next step to fix it. It also works incredibly 00:02:15.669 --> 00:02:18.370 well for visual content, too. Say you're designing 00:02:18.370 --> 00:02:21.250 a new newsletter layout, you take a quick screenshot 00:02:21.250 --> 00:02:24.310 of the draft, you ask the AI to check the title 00:02:24.310 --> 00:02:27.030 length, or you ask if the call -to -action placement 00:02:27.030 --> 00:02:29.389 needs clearer wording. It's working straight 00:02:29.389 --> 00:02:32.110 from the real visual data. But there's a really 00:02:32.110 --> 00:02:35.740 massive limitation at this stage. a severe lack 00:02:35.740 --> 00:02:38.259 of ongoing context. Oh, completely. One -off 00:02:38.259 --> 00:02:40.879 chats are perfectly fine for single isolated 00:02:40.879 --> 00:02:45.219 tasks, but repeated ongoing work is entirely 00:02:45.219 --> 00:02:47.400 different. Exactly. Repeated work requires a 00:02:47.400 --> 00:02:50.819 very specific, consistent tone. It requires deeply 00:02:50.819 --> 00:02:53.860 knowing the target audience. You have internal 00:02:53.860 --> 00:02:56.840 company rules. Yeah. You have past business decisions 00:02:56.840 --> 00:03:00.430 to consider. And in a quick chat, the AI... basically 00:03:00.430 --> 00:03:03.270 forgets all of that. It has total amnesia. You 00:03:03.270 --> 00:03:05.789 have to repeat your tone and rules every single 00:03:05.789 --> 00:03:08.310 time. Honestly, I have a vulnerable admission 00:03:08.310 --> 00:03:10.150 to make here. Okay, let's hear it. Beep beep. 00:03:10.569 --> 00:03:12.250 I still wrestle with starting every repeated 00:03:12.250 --> 00:03:14.930 task in a blank chat. Oh, we all do it sometimes. 00:03:15.030 --> 00:03:16.969 It just feels faster in the heat of the moment. 00:03:17.069 --> 00:03:19.650 You just want an answer right now. Yeah, I do. 00:03:19.810 --> 00:03:22.129 But over a month you waste countless hours re 00:03:22.129 --> 00:03:24.610 -explaining yourself. It really makes me wonder 00:03:24.610 --> 00:03:27.370 about the true value here. Is level one basically 00:03:27.370 --> 00:03:30.550 just using Claude as a glorified short -term 00:03:30.550 --> 00:03:33.270 search engine? Pretty much. It's a fast starter, 00:03:33.409 --> 00:03:35.530 but it forgets everything by the next day. Which 00:03:35.530 --> 00:03:37.789 perfectly highlights the core problem we need 00:03:37.789 --> 00:03:41.849 to solve. We have to fix that blank chat amnesia. 00:03:42.150 --> 00:03:45.030 Yes. So let's step up to level two, the beginner. 00:03:45.389 --> 00:03:48.469 We fix this by building a persistent workspace. 00:03:48.780 --> 00:03:51.840 Level two is entirely focused on creating continuity. 00:03:52.360 --> 00:03:54.639 Your important work continues across multiple 00:03:54.639 --> 00:03:57.120 conversations. It continues across your files 00:03:57.120 --> 00:03:59.360 and your connected data sources. This level relies 00:03:59.360 --> 00:04:02.580 heavily on three core features, projects, memory, 00:04:03.020 --> 00:04:05.740 and past chat search. Right. Memory naturally 00:04:05.740 --> 00:04:08.099 retains your specific working preferences over 00:04:08.099 --> 00:04:11.680 time. Past chat search lets you easily find older 00:04:11.680 --> 00:04:13.539 conversations. If you can just open a chat and 00:04:13.539 --> 00:04:15.620 say, find the campaign direction from last week. 00:04:16.259 --> 00:04:18.939 Exactly. Then you ask it to prepare the final 00:04:18.939 --> 00:04:21.660 outline based on that. It seamlessly maintains 00:04:21.660 --> 00:04:23.959 the historical thread. But it gets really interesting 00:04:23.959 --> 00:04:26.220 when we bring in connectors. You can actually 00:04:26.220 --> 00:04:28.439 connect the AI to the information you already 00:04:28.439 --> 00:04:31.480 use. Yeah, this is a huge step up. Google Drive, 00:04:32.079 --> 00:04:36.579 Gmail, Slack, Google Calendar, Microsoft 365. 00:04:36.939 --> 00:04:40.519 Wow, that's a lot. Your real... actionable context 00:04:40.519 --> 00:04:43.160 already sits inside those apps. Connectors pull 00:04:43.160 --> 00:04:45.379 that securely approved information right into 00:04:45.379 --> 00:04:47.279 your active conversation. So you don't have to 00:04:47.279 --> 00:04:49.680 manually collect every single detail anymore. 00:04:49.819 --> 00:04:52.779 You don't have to constantly switch tabs to copy 00:04:52.779 --> 00:04:55.839 and paste. You can just ask it to review a specific 00:04:55.839 --> 00:04:59.240 client email thread. Right. Or review the latest 00:04:59.240 --> 00:05:01.579 strategic brief directly from Google Drive. Let's 00:05:01.579 --> 00:05:04.000 look at a very specific, practical example here. 00:05:04.399 --> 00:05:06.660 Creating a highly structured meeting summary. 00:05:07.000 --> 00:05:10.779 You upload your messy, raw client notes. Yeah, 00:05:10.839 --> 00:05:12.839 we've all been there. And you want a professional 00:05:12.839 --> 00:05:14.660 Word document out of it. The prompt for this 00:05:14.660 --> 00:05:17.759 needs to be very deliberate. You ask it to clearly 00:05:17.759 --> 00:05:20.439 include the meeting background. You explicitly 00:05:20.439 --> 00:05:22.759 ask for the confirmed decisions. You ask for 00:05:22.759 --> 00:05:25.579 a list of questions that still need confirmation. 00:05:26.519 --> 00:05:29.300 And finally, you demand a next actions table. 00:05:30.000 --> 00:05:32.259 That table must include the responsible person. 00:05:32.540 --> 00:05:35.500 and the specific deadline. But here is absolutely 00:05:35.500 --> 00:05:37.899 the most important part of that workflow. You 00:05:37.899 --> 00:05:41.079 give the AI a very strict, unbreakable rule. 00:05:41.480 --> 00:05:44.139 Which is? If any information is missing, Claude 00:05:44.139 --> 00:05:47.120 must write, not confirmed. It is absolutely not 00:05:47.120 --> 00:05:49.819 allowed to guess or fill in the blanks. I like 00:05:49.819 --> 00:05:52.449 to think about it this way. Level one is a public 00:05:52.449 --> 00:05:54.810 whiteboard that gets wiped clean every night. 00:05:54.810 --> 00:05:57.949 I like that. Level two is a dedicated office 00:05:57.949 --> 00:06:00.170 desk where your files stay exactly where you 00:06:00.170 --> 00:06:02.290 left them. That's a perfect way to picture the 00:06:02.290 --> 00:06:05.509 difference. It reliably turns messy inputs into 00:06:05.509 --> 00:06:08.170 clearly structured files of files you can actually 00:06:08.170 --> 00:06:11.790 review, trust, and use in your job. But I do 00:06:11.790 --> 00:06:14.189 have a lingering concern here. With all those 00:06:14.189 --> 00:06:16.810 connected apps, can we trust it not to just invent 00:06:16.810 --> 00:06:19.680 information? It's safer, but you still must review 00:06:19.680 --> 00:06:21.819 the structured output before sending it anywhere. 00:06:22.139 --> 00:06:23.980 Okay, so we've got our files beautifully structured 00:06:23.980 --> 00:06:26.439 on our metaphorical desk, but I'm still the one 00:06:26.439 --> 00:06:28.459 opening the folders and doing the actual manual 00:06:28.459 --> 00:06:30.279 labor. Right. I'm still clicking and dragging. 00:06:30.519 --> 00:06:32.579 Exactly. And that bottleneck is why you have 00:06:32.579 --> 00:06:35.680 to level up again. Welcome to level three, the 00:06:35.680 --> 00:06:39.240 intermediate. We've mastered text outputs. Now 00:06:39.240 --> 00:06:42.420 we give the AI actual hands. We're talking about 00:06:42.420 --> 00:06:44.980 a feature called Claude Cowork. This is where 00:06:44.980 --> 00:06:47.339 it actively starts operating your computer for 00:06:47.339 --> 00:06:51.420 you. It handles complex multi -step tasks directly. 00:06:51.680 --> 00:06:54.459 It moves way beyond just reading your text files. 00:06:54.560 --> 00:06:56.920 It actually starts manipulating the files on 00:06:56.920 --> 00:06:59.720 your machine. Okay. So imagine you have a wildly 00:06:59.720 --> 00:07:02.779 messy downloads folder. It's just stuffed with 00:07:02.779 --> 00:07:06.120 random invoices, notes, and PDFs. You can give 00:07:06.120 --> 00:07:08.480 Cowork a very specific multi -step directive. 00:07:08.810 --> 00:07:11.350 You literally tell it to sort that folder by 00:07:11.350 --> 00:07:13.769 file type. You tell it to rename every single 00:07:13.769 --> 00:07:16.110 file to a standard format, like year, month, 00:07:16.129 --> 00:07:19.069 day. Then you ask it to summarize the contents 00:07:19.069 --> 00:07:21.089 of all those files. And if a file is unreadable 00:07:21.089 --> 00:07:24.050 or unclear, it doesn't just crash, it neatly 00:07:24.050 --> 00:07:26.149 moves it to a dedicated folder. Like a needs 00:07:26.149 --> 00:07:28.829 review folder. Exactly. It does all the tedious 00:07:28.829 --> 00:07:31.430 heavy lifting for you. But we really need to 00:07:31.430 --> 00:07:34.509 talk about safe file access here. This is absolutely 00:07:34.509 --> 00:07:38.170 crucial. When you give an AI direct file access, 00:07:38.629 --> 00:07:41.250 you must start small. You have to set up isolated, 00:07:41.470 --> 00:07:44.370 protected folders. Specifically, you should have 00:07:44.370 --> 00:07:46.949 three distinct folders for this. Right. Source 00:07:46.949 --> 00:07:50.930 files, working files, and outputs. Let's break 00:07:50.930 --> 00:07:53.589 that down. Source files are the read -only documents 00:07:53.589 --> 00:07:56.310 it can look at. Working files are the temporary 00:07:56.310 --> 00:07:59.889 sandbox where it actively processes data. Outputs 00:07:59.889 --> 00:08:02.509 are the final finished results for you to manually 00:08:02.509 --> 00:08:04.959 check. Exactly. You never let it loose on your 00:08:04.959 --> 00:08:07.420 main hard drive. That's just asking for a disaster. 00:08:07.899 --> 00:08:10.240 This isolated structure leads to something really 00:08:10.240 --> 00:08:13.480 powerful, though. Reusable skills. Oh, yeah. 00:08:13.680 --> 00:08:15.980 When a tedious task repeats every single week, 00:08:16.459 --> 00:08:19.379 you save it as a skill. Let's use a weekly client 00:08:19.379 --> 00:08:21.879 report generator as a solid example. You build 00:08:21.879 --> 00:08:24.839 a custom skill to process marketing data every 00:08:24.839 --> 00:08:27.480 Friday. OK. You establish strict rules for the 00:08:27.480 --> 00:08:29.850 output. You require five specific sections every 00:08:29.850 --> 00:08:31.769 single time it runs. So you'd tell it, I need 00:08:31.769 --> 00:08:33.850 an executive summary at the very top. It's a 00:08:33.850 --> 00:08:36.090 short, crisp summary of the week's performance. 00:08:36.590 --> 00:08:39.470 Right. And then you force it to extract the key 00:08:39.470 --> 00:08:43.149 results. It pulls the important, verified metrics 00:08:43.149 --> 00:08:45.730 directly from the uploaded data. Right. And you 00:08:45.730 --> 00:08:47.269 want to compare those numbers to the previous 00:08:47.269 --> 00:08:49.549 week. So you add a section for important changes. 00:08:49.710 --> 00:08:52.509 Yeah, exactly. You're building a reliable, repeatable 00:08:52.509 --> 00:08:54.850 template. And you'd probably want it to flag 00:08:54.850 --> 00:08:56.909 anything that's missing, too. Definitely. That's 00:08:56.909 --> 00:09:00.230 your open questions section. It lists any unclear 00:09:00.230 --> 00:09:02.730 data points. Then it always ends with a clear 00:09:02.730 --> 00:09:05.250 next actions table. Action, owner, and deadline. 00:09:05.529 --> 00:09:09.129 Once you save that specific skill, your weekly 00:09:09.129 --> 00:09:12.009 workflow changes entirely. How so? You just upload 00:09:12.009 --> 00:09:14.610 the fresh raw data each week. You hit run on 00:09:14.610 --> 00:09:16.830 the skill. It automatically follows the exact 00:09:16.830 --> 00:09:18.950 same five -step process without you managing 00:09:18.950 --> 00:09:21.009 it. You can even schedule these simple tasks 00:09:21.009 --> 00:09:24.009 to run automatically. Say you're using a cloud 00:09:24.009 --> 00:09:26.250 desktop application. Yeah. You could easily set 00:09:26.250 --> 00:09:29.289 up a customized briefing for 7 .30 p .m. every 00:09:29.289 --> 00:09:31.629 weekday. It quietly checks your connected sources 00:09:31.629 --> 00:09:33.769 in the background. It stands Gmail. It checks 00:09:33.769 --> 00:09:36.049 your calendar. It reads your Slack messages. 00:09:36.169 --> 00:09:39.230 It pulls all that context together. Then it prepares 00:09:39.230 --> 00:09:41.710 a short summary of tomorrow's meetings and any 00:09:41.710 --> 00:09:44.309 pending actions. It feels incredibly powerful. 00:09:44.690 --> 00:09:47.610 But I really have to ask this. Isn't giving AI 00:09:47.610 --> 00:09:50.590 direct file access a recipe for accidentally 00:09:50.590 --> 00:09:53.110 deleting your hard drive? Only if you're reckless. 00:09:53.649 --> 00:09:56.350 Always use copy files in isolated folders to 00:09:56.350 --> 00:10:00.720 test first, sponsor. We're back! So we just covered 00:10:00.720 --> 00:10:03.960 automating desktop chores in Level 3. Now we 00:10:03.960 --> 00:10:06.460 really scale things up. We do. Welcome to Level 00:10:06.460 --> 00:10:09.240 4, the Advanced Builder. This is a major jump. 00:10:09.379 --> 00:10:11.799 This is where we introduce Claude code. We're 00:10:11.799 --> 00:10:14.220 moving away from simple folder sorting tasks. 00:10:14.399 --> 00:10:16.860 We're getting into complex multi -file project 00:10:16.860 --> 00:10:19.580 engineering. This requires a real dedicated project 00:10:19.580 --> 00:10:22.340 structure. And crucially, it requires a Claude 00:10:22.340 --> 00:10:25.230 .md file. Let's quickly define that jargon, a 00:10:25.230 --> 00:10:27.590 simple text file that tells the AI your project 00:10:27.590 --> 00:10:29.789 rules. It basically acts like a strict working 00:10:29.789 --> 00:10:32.529 guide for the system. It clearly defines the 00:10:32.529 --> 00:10:35.309 overarching project goal. It explicitly explains 00:10:35.309 --> 00:10:37.149 what all the main folders are actually used for. 00:10:37.149 --> 00:10:39.970 Right. It lists the specific files the AI is 00:10:39.970 --> 00:10:42.549 allowed to edit. And crucially, it lists the 00:10:42.549 --> 00:10:45.149 files it must absolutely leave unchanged. It 00:10:45.149 --> 00:10:47.809 also tells the AI exactly how to run your internal 00:10:47.809 --> 00:10:50.909 tests. And it defines what the final report format 00:10:50.909 --> 00:10:53.230 should look like. But there is a very critical 00:10:53.230 --> 00:10:55.850 step here that you cannot skip. You must ask 00:10:55.850 --> 00:10:59.269 the AI to plan and explain its approach first. 00:10:59.309 --> 00:11:02.710 Yes. It has to do this before it makes any actual 00:11:02.710 --> 00:11:05.049 edits. It's like asking a contractor for blueprints 00:11:05.049 --> 00:11:07.370 before they knock down a wall. You prompt it 00:11:07.370 --> 00:11:10.309 to read the folder structure. You ask for a detailed 00:11:10.309 --> 00:11:12.970 step -by -step plan. It needs to explain which 00:11:12.970 --> 00:11:15.370 files will change and exactly why. You review 00:11:15.370 --> 00:11:18.750 that plan. You approve it. Only then does the 00:11:18.750 --> 00:11:21.639 actual coding work begin. As these projects get 00:11:21.639 --> 00:11:24.019 bigger, things naturally get much more complicated. 00:11:24.419 --> 00:11:27.080 One single session can quickly become overloaded 00:11:27.080 --> 00:11:29.580 and confused. Yeah, for sure. So we have to deliberately 00:11:29.580 --> 00:11:32.360 split the workout. We split big work into parallel 00:11:32.360 --> 00:11:35.240 sessions. We use specialized tools like sub -agents 00:11:35.240 --> 00:11:37.879 and work trees. One stream handles the front 00:11:37.879 --> 00:11:40.259 -end layout. Another stream prepares the backend 00:11:40.259 --> 00:11:43.659 tests. A third stream carefully reviews the overall 00:11:43.659 --> 00:11:45.840 logic. This keeps the different work streams 00:11:45.840 --> 00:11:48.779 entirely separate and clean. They happen at the 00:11:48.779 --> 00:11:51.419 exact same time. but they don't accidentally 00:11:51.419 --> 00:11:53.240 interfere with each other's code. And you always 00:11:53.240 --> 00:11:55.840 have a mandatory non -negotiable review step 00:11:55.840 --> 00:11:59.279 built in. The AI absolutely must check the result 00:11:59.279 --> 00:12:01.700 before you accept the files. It must run all 00:12:01.700 --> 00:12:04.500 the available system checks. If a check fails, 00:12:05.039 --> 00:12:08.159 it has to explicitly explain the issue. It fixes 00:12:08.159 --> 00:12:11.019 the error itself. It runs the check again. Finally, 00:12:11.139 --> 00:12:13.519 it prints out a list of every single changed 00:12:13.519 --> 00:12:16.580 file. You carefully look at that list. You verify 00:12:16.580 --> 00:12:19.480 the changes. Then you formally consider the task 00:12:19.480 --> 00:12:21.559 complete. But I'm genuinely curious about the 00:12:21.559 --> 00:12:23.799 mechanics of this process. Why bother splitting 00:12:23.799 --> 00:12:26.019 tasks into parallel sessions instead of just 00:12:26.019 --> 00:12:28.940 one giant prompt? To keep the AI focused, preventing 00:12:28.940 --> 00:12:30.980 it from getting confused, and mixing up changes. 00:12:31.100 --> 00:12:33.399 That makes perfect sense. Clear rules. Clear 00:12:33.399 --> 00:12:36.980 inputs. Clear safety checks. Which finally brings 00:12:36.980 --> 00:12:39.179 us to the absolute summit of this framework. 00:12:39.700 --> 00:12:43.169 Level five, the architect. This is fully autonomous 00:12:43.169 --> 00:12:45.909 infrastructure. We're completely removing the 00:12:45.909 --> 00:12:48.809 human from the active day -to -day loop. We're 00:12:48.809 --> 00:12:51.610 finally entering the realm of true background 00:12:51.610 --> 00:12:53.669 automation. You've already set the strict rules 00:12:53.669 --> 00:12:57.049 in Level 4. Now, the complex workflow runs without 00:12:57.049 --> 00:12:59.950 you even sitting at the screen. But the hardest 00:12:59.950 --> 00:13:02.730 part of Level 5 isn't actually technical. Beat. 00:13:02.879 --> 00:13:05.820 It's trust. Trust is incredibly difficult here. 00:13:06.360 --> 00:13:09.419 You're letting an AI touch real files and critical 00:13:09.419 --> 00:13:12.019 business data in the background. Giving up that 00:13:12.019 --> 00:13:14.559 control is scary. You have to build this autonomy 00:13:14.559 --> 00:13:18.139 very, very slowly. At this advanced stage, tasks 00:13:18.139 --> 00:13:21.220 start from automatic triggers. a time schedule, 00:13:21.460 --> 00:13:24.179 a brand new email arriving, a specific GitHub 00:13:24.179 --> 00:13:27.340 event, or an external API call. We use a system 00:13:27.340 --> 00:13:29.639 called Cloud Routines for this. These are secure 00:13:29.639 --> 00:13:32.120 cloud -based workflows. They run entirely on 00:13:32.120 --> 00:13:34.299 Anthropic's own back -end infrastructure. They 00:13:34.299 --> 00:13:36.259 don't rely on your local desktop being turned 00:13:36.259 --> 00:13:38.200 on. So you can set a routine to automatically 00:13:38.200 --> 00:13:40.620 review changed files when a pull request opens. 00:13:41.120 --> 00:13:43.360 Or you could have it create a massive weekly 00:13:43.360 --> 00:13:45.899 project summary every Friday afternoon. But we 00:13:45.899 --> 00:13:48.220 desperately need safety mechanisms here. This 00:13:48.220 --> 00:13:50.460 is exactly where hooks come in. Let's define 00:13:50.460 --> 00:13:53.299 that term clearly. Safety gates that check the 00:13:53.299 --> 00:13:56.220 AI's actions before and after. Right. There's 00:13:56.220 --> 00:13:58.940 a pre -tool use hook. It acts just like a strict 00:13:58.940 --> 00:14:01.929 security gate. Before the AI actually runs a 00:14:01.929 --> 00:14:04.610 command, the hook checks it. If the proposed 00:14:04.610 --> 00:14:07.529 action is risky, like deleting important files, 00:14:07.889 --> 00:14:11.250 it blocks it instantly. It asks the AI to rethink 00:14:11.250 --> 00:14:13.470 and correct the action. Then there's the post 00:14:13.470 --> 00:14:16.049 tool use hook. After the tool successfully finishes, 00:14:16.509 --> 00:14:19.389 this hook handles the necessary cleanup. It might 00:14:19.389 --> 00:14:22.350 automatically format a text file. Or it might 00:14:22.350 --> 00:14:24.350 send you a quick Slack notification that the 00:14:24.350 --> 00:14:26.590 background task is done. The entire workflow 00:14:26.590 --> 00:14:29.399 is fully automated. But those precise hooks keep 00:14:29.399 --> 00:14:32.259 it tightly and safely controlled. You still want 00:14:32.259 --> 00:14:34.100 to start this process very safely though. We 00:14:34.100 --> 00:14:36.240 highly recommend starting with a daily private 00:14:36.240 --> 00:14:38.960 morning briefing. Yes. Just let it run quietly 00:14:38.960 --> 00:14:41.159 every morning at 6 a .m. It checks your upcoming 00:14:41.159 --> 00:14:43.600 calendar. It safely scans your important emails. 00:14:43.899 --> 00:14:46.120 Right. It reviews your missed slack threads. 00:14:46.480 --> 00:14:49.440 It prepares a comprehensive summary. But it sends 00:14:49.440 --> 00:14:52.190 that summary only to you. It doesn't email your 00:14:52.190 --> 00:14:54.990 boss. You review it over coffee before taking 00:14:54.990 --> 00:14:57.250 any action. Yeah, you let it run like this for 00:14:57.250 --> 00:15:00.269 a full week. You check if it misses nuances or 00:15:00.269 --> 00:15:02.730 messes up your actual priorities. You carefully 00:15:02.730 --> 00:15:05.110 adjust the background instructions. Only after 00:15:05.110 --> 00:15:07.870 definitively proved reliable do you give it a 00:15:07.870 --> 00:15:11.570 more important workflow. Exactly. Beat. Whoa. 00:15:11.909 --> 00:15:15.070 I mean, imagine truly scaling this up. Imagine 00:15:15.070 --> 00:15:17.990 an AI autonomously running your complex workflows 00:15:17.990 --> 00:15:20.700 in the background while you sleep. It's wild. 00:15:21.120 --> 00:15:23.679 Perfect checks. Perfect balance. Waking up to 00:15:23.679 --> 00:15:26.159 finished work. It's an absolutely amazing thought, 00:15:26.620 --> 00:15:29.620 but it clearly demands immense caution in planning. 00:15:29.740 --> 00:15:31.799 Which naturally raises a really big question 00:15:31.799 --> 00:15:34.539 for me. How do we stop an autonomous system from 00:15:34.539 --> 00:15:36.940 going rogue and sending a bad email? You use 00:15:36.940 --> 00:15:39.720 those hooks to block risky actions and strictly 00:15:39.720 --> 00:15:42.080 require human review first. Okay, let's bring 00:15:42.080 --> 00:15:43.799 this all back down to earth for a minute. We 00:15:43.799 --> 00:15:46.620 need a big idea recap to tie this together. The 00:15:46.620 --> 00:15:48.799 core insight from all of this source material 00:15:48.799 --> 00:15:52.340 is surprisingly clear. Advanced use isn't really 00:15:52.340 --> 00:15:55.279 about writing better, more clever prompts. It's 00:15:55.279 --> 00:15:58.080 entirely about clearer workflow design. You don't 00:15:58.080 --> 00:16:00.120 just magically jump to level five. You don't 00:16:00.120 --> 00:16:02.620 blindly rush into full automation. You have to 00:16:02.620 --> 00:16:04.460 build the foundational structure first. Yeah, 00:16:04.460 --> 00:16:06.820 you add proper folder structure, you add firm 00:16:06.820 --> 00:16:09.340 safety limits, you add mandatory review steps. 00:16:09.620 --> 00:16:12.419 That's how this technology becomes truly, reliably 00:16:12.419 --> 00:16:15.240 useful. So here is your practical takeaway for 00:16:15.240 --> 00:16:18.279 today. Look closely at your own workflow. Pick 00:16:18.279 --> 00:16:21.240 just one repeated week task that annoys you. 00:16:21.259 --> 00:16:25.159 Just one task and upgrade it by exactly one single 00:16:25.159 --> 00:16:27.340 level. If you're stuck at level one, make a project. 00:16:27.580 --> 00:16:29.759 If you already have a project, connect one live 00:16:29.759 --> 00:16:32.419 data source. If you have connected sources, try 00:16:32.419 --> 00:16:35.240 a basic co -work task. Build a safer, more structured 00:16:35.240 --> 00:16:38.899 workflow. Add a mandatory review step. One single 00:16:38.899 --> 00:16:41.559 intentional upgrade will teach you way more than 00:16:41.559 --> 00:16:44.299 trying to build a massive autonomous system on 00:16:44.299 --> 00:16:46.659 day one. It forces you to truly understand the 00:16:46.659 --> 00:16:48.379 underlying mechanics. It forces you to actually 00:16:48.379 --> 00:16:51.200 build trust with the tool. Step by step, level 00:16:51.200 --> 00:16:53.019 by level. That's how you actually get advanced. 00:16:53.500 --> 00:16:55.360 It's been a truly fascinating journey through 00:16:55.360 --> 00:16:58.159 these five levels today, from the eager enthusiast 00:16:58.159 --> 00:17:00.259 all the way up to the architect. But I want to 00:17:00.259 --> 00:17:03.019 leave you with a final slightly provocative thought 00:17:03.019 --> 00:17:06.059 to mull over. Beat. Let's say you successfully 00:17:06.059 --> 00:17:08.970 level up. You reach a point where AI flawlessly 00:17:08.970 --> 00:17:11.589 remembers your context, where it perfectly plans 00:17:11.589 --> 00:17:14.069 your complex projects, where it autonomously 00:17:14.069 --> 00:17:16.210 executes your daily routines with perfect hooks 00:17:16.210 --> 00:17:17.109 and safety checks.