WEBVTT 00:00:00.000 --> 00:00:02.899 You know, usually when you need quick directions, 00:00:03.120 --> 00:00:05.700 you just roll down your window and ask a stranger 00:00:05.700 --> 00:00:08.859 on the street. It's fast. It's highly transactional. 00:00:08.939 --> 00:00:10.800 They point you to the coffee shop, and you just 00:00:10.800 --> 00:00:12.359 never see them again. Yeah, you get the immediate 00:00:12.359 --> 00:00:14.679 answer. But tomorrow, if you need coffee again, 00:00:14.720 --> 00:00:17.320 you have to find a completely new stranger and 00:00:17.320 --> 00:00:19.679 start from zero. I mean, there is absolutely 00:00:19.679 --> 00:00:24.519 no retained memory of your preferences or the 00:00:24.519 --> 00:00:27.219 interaction itself. OK, let's unpack this. What 00:00:27.219 --> 00:00:30.289 if? Instead of relying on a random stranger, 00:00:30.730 --> 00:00:33.850 you had a dedicated personal assistant. Oh, totally 00:00:33.850 --> 00:00:35.750 different game. Right, like someone who already 00:00:35.750 --> 00:00:38.829 knows exactly where you like to go, what routes 00:00:38.829 --> 00:00:41.270 to avoid because of morning traffic, and just 00:00:41.270 --> 00:00:44.109 remembers to bring you your specific coffee order 00:00:44.109 --> 00:00:46.710 every single day without you even having to ask. 00:00:46.929 --> 00:00:49.469 Yeah, that's the dream. That leap from a one 00:00:49.469 --> 00:00:52.429 -off transaction to a dedicated automated system, 00:00:52.750 --> 00:00:54.909 that is exactly what we are dissecting today. 00:00:55.659 --> 00:00:59.439 Welcome to our deep dive. We are looking at a 00:00:59.439 --> 00:01:01.939 really fascinating instructional playbook today 00:01:01.939 --> 00:01:05.760 titled Lesson 4, Build Your First Claude Cowork 00:01:05.760 --> 00:01:08.040 AI Agent. It's a great piece of source material. 00:01:08.079 --> 00:01:10.140 It really is. And the mission here is to guide 00:01:10.140 --> 00:01:12.819 you through a complete paradigm shift, transforming 00:01:12.819 --> 00:01:16.459 Claude from just a... chat window you casually 00:01:16.459 --> 00:01:20.040 talk to, into a structured workspace that executes 00:01:20.040 --> 00:01:22.739 a daily automated workflow right on your heart 00:01:22.739 --> 00:01:25.200 rate. Yep. Specifically, we're building a fully 00:01:25.200 --> 00:01:28.840 automated daily AI news dashboard. Right. And 00:01:28.840 --> 00:01:31.439 to make that leap, we have to fundamentally change 00:01:31.439 --> 00:01:34.700 how we view the technology. I mean, most users 00:01:34.700 --> 00:01:39.060 are currently trapped in this chat mindset. Yeah, 00:01:39.060 --> 00:01:41.709 exactly. Clawed chat or, you know, really any 00:01:41.709 --> 00:01:44.209 standard LLM chat interface. It's brilliant for 00:01:44.209 --> 00:01:47.090 brainstorming or getting a complex concept explained 00:01:47.090 --> 00:01:50.010 to you or drafting a one -off email. The quick 00:01:50.010 --> 00:01:52.969 transactional stuff. Exactly. But structurally, 00:01:53.170 --> 00:01:55.670 it all happens inside a single ephemeral conversation 00:01:55.670 --> 00:01:58.189 window. Once it generates an answer, the burden 00:01:58.189 --> 00:02:01.349 is entirely on you to copy and paste that output 00:02:01.349 --> 00:02:04.310 into a Word document or your CMS or an email 00:02:04.310 --> 00:02:06.450 client to actually put it to use. When you start 00:02:06.450 --> 00:02:09.009 a new chat the next morning, the AI just has 00:02:09.009 --> 00:02:11.259 amnesia. You're back to square one. Which is 00:02:11.259 --> 00:02:13.479 why Claude Cowork operates on a totally different 00:02:13.479 --> 00:02:15.960 mechanical level. What's fascinating here is 00:02:15.960 --> 00:02:19.340 that it is an execution workspace. Instead of 00:02:19.340 --> 00:02:21.939 living in a browser tab, it actually integrates 00:02:21.939 --> 00:02:25.139 with your local environment. It works inside 00:02:25.139 --> 00:02:28.159 a designated folder on your computer. Wait, on 00:02:28.159 --> 00:02:31.180 your actual computer? Yes. It autonomously reads 00:02:31.180 --> 00:02:34.719 files you've prepared, processes that data, Generics 00:02:34.719 --> 00:02:36.879 deliverables, and this is the kicker, actually 00:02:36.879 --> 00:02:39.180 saves those finished documents directly to your 00:02:39.180 --> 00:02:41.979 local drive. Oh, wow. The underlying philosophy 00:02:41.979 --> 00:02:44.740 here is cognitive offloading for both you and 00:02:44.740 --> 00:02:47.759 the AI. A co -work agent only becomes powerful 00:02:47.759 --> 00:02:50.400 because the workspace eliminates the AI's need 00:02:50.400 --> 00:02:54.680 to guess. your context before the task even starts. 00:02:54.939 --> 00:02:57.039 But wait, isn't regular chat good enough if I 00:02:57.039 --> 00:02:59.699 just write a really long, detailed prompt every 00:02:59.699 --> 00:03:01.360 morning? Like I know a lot of developers and 00:03:01.360 --> 00:03:03.639 marketers who just use a three -page mega prompt 00:03:03.639 --> 00:03:05.500 saved in Apple Notes. Oh yeah, the mega prompt. 00:03:05.620 --> 00:03:08.159 Right. They just copy it, paste it into a fresh 00:03:08.159 --> 00:03:10.400 chat window every morning, drop in some links 00:03:10.400 --> 00:03:13.580 and let it run. Why go through the hassle of 00:03:13.580 --> 00:03:15.740 building a localized folder architecture if I 00:03:15.740 --> 00:03:18.900 can just paste a massive detailed prompt to solve 00:03:18.900 --> 00:03:21.080 the context problem? Well, it seems like a logical 00:03:21.080 --> 00:03:23.379 workaround, right? Yeah. But the mega prompt 00:03:23.379 --> 00:03:26.680 approach fundamentally misunderstands how large 00:03:26.680 --> 00:03:29.120 language models actually process information. 00:03:29.199 --> 00:03:32.080 How so? When you drop a giant wall of text into 00:03:32.080 --> 00:03:34.900 a fresh chat window, you are demanding that the 00:03:34.900 --> 00:03:39.340 AI parse, interpret, weigh, and prioritize all 00:03:39.340 --> 00:03:42.340 of those instructions simultaneously, every single 00:03:42.340 --> 00:03:45.000 time. Sounds exhausting. It is. You are flooding 00:03:45.000 --> 00:03:47.639 its attention mechanism. In an LLM, the more 00:03:47.639 --> 00:03:50.099 context you stuff into a single prompt, the more 00:03:50.099 --> 00:03:52.240 the model struggle to retrieve specific nuances. 00:03:52.699 --> 00:03:54.759 It's a well -documented phenomenon known as the 00:03:54.759 --> 00:03:55.960 lost -in -the -middle problem. Probably just 00:03:55.960 --> 00:03:57.759 forgets the stuff in the middle of your huge 00:03:57.759 --> 00:04:00.219 prompt. Exactly. It gets overwhelmed, which leads 00:04:00.219 --> 00:04:02.259 to inconsistent formatting and forgotten rules. 00:04:02.439 --> 00:04:04.819 So the AI essentially gets distracted by its 00:04:04.819 --> 00:04:08.580 own instructions. Precisely. A workspace physically 00:04:08.580 --> 00:04:11.800 partitions the instructions. By isolating the 00:04:11.800 --> 00:04:14.780 rules, the sources, and the templates into distinct 00:04:14.780 --> 00:04:17.639 purpose -built files, you're effectively creating 00:04:17.639 --> 00:04:21.759 a localized retrieval augmented generation, or 00:04:21.759 --> 00:04:25.920 RG system. Right. The AI knows exactly which 00:04:25.920 --> 00:04:29.040 file to query for the rules and which file to 00:04:29.040 --> 00:04:31.399 query for the output format. It doesn't have 00:04:31.399 --> 00:04:33.860 to hold the entire universe of instructions in 00:04:33.860 --> 00:04:35.959 its active memory all at once. It just pulls 00:04:35.959 --> 00:04:38.259 what it needs when it needs it. That physical 00:04:38.259 --> 00:04:40.300 setup requires a specific environment though. 00:04:40.519 --> 00:04:42.540 The playbook notes you need the Claw desktop 00:04:42.540 --> 00:04:44.439 app installed on your machine and it requires 00:04:44.439 --> 00:04:46.980 a paid tier pro is the baseline here. Yeah, pro 00:04:46.980 --> 00:04:48.740 is fine for this. They mentioned a max plan but 00:04:48.740 --> 00:04:50.779 clarify that's only necessary if you are running 00:04:50.779 --> 00:04:53.819 like massive token heavy workflows down the line. 00:04:54.279 --> 00:04:56.399 But as we start setting up this architecture, 00:04:56.839 --> 00:04:59.120 the source material throws up a massive red flag 00:04:59.120 --> 00:05:01.959 regarding permission. Huge reflect. Because Cowork 00:05:01.959 --> 00:05:04.579 runs locally on your desktop, it prompts you 00:05:04.579 --> 00:05:07.860 to select a workspace directory. And the absolute 00:05:07.860 --> 00:05:10.959 worst mistake you can make is giving the AI access 00:05:10.959 --> 00:05:13.480 to your entire computer. Here's where it gets 00:05:13.480 --> 00:05:16.519 really interesting. Giving the AI your entire 00:05:16.519 --> 00:05:18.980 hard drive is like calling a plumber to fix your 00:05:18.980 --> 00:05:21.300 sink and handing them the keys to your filing 00:05:21.300 --> 00:05:24.439 cabinet and your car. That is a perfect analogy 00:05:24.439 --> 00:05:26.959 because if you select your root directory or 00:05:26.959 --> 00:05:30.199 say your entire documents folder, you are creating 00:05:30.199 --> 00:05:33.519 a disaster on two fronts. First, there's the 00:05:33.519 --> 00:05:35.319 obvious security implication. You really don't 00:05:35.319 --> 00:05:37.420 want an automated agent having free reign over 00:05:37.420 --> 00:05:40.079 your tax returns or your personal photos. Definitely 00:05:40.079 --> 00:05:43.420 not. But the secondary. An arguably bigger issue 00:05:43.420 --> 00:05:46.519 for this workflow is context pollution. Context 00:05:46.519 --> 00:05:48.519 pollution. Yeah, they're going to get lost. If 00:05:48.519 --> 00:05:51.300 the AI has access to your entire system and you 00:05:51.300 --> 00:05:54.220 ask it to summarize today's AI news, it might 00:05:54.220 --> 00:05:57.920 randomly pull context from an old college essay 00:05:57.920 --> 00:06:00.720 you wrote about robotics in 2014 simply because 00:06:00.720 --> 00:06:02.620 it found the key word. Oh, because it's just 00:06:02.620 --> 00:06:04.620 searching everything it has access to. Right. 00:06:05.399 --> 00:06:07.800 The LLM doesn't inherently know what files are 00:06:07.800 --> 00:06:10.540 relevant until you define the boundaries. The 00:06:10.540 --> 00:06:13.620 playbook is incredibly rigid about this. You 00:06:13.620 --> 00:06:17.899 must create one clean, isolated folder. Start 00:06:17.899 --> 00:06:20.980 small. Name it something hyper -specific, like 00:06:20.980 --> 00:06:23.279 Claude Cowork News Dashboard. Keep it contained? 00:06:23.660 --> 00:06:26.959 Exactly. Inside that isolated room, every single 00:06:26.959 --> 00:06:30.420 file has a clear, defined purpose. The AI can't 00:06:30.420 --> 00:06:33.120 get distracted by irrelevant data because the 00:06:33.120 --> 00:06:35.420 irrelevant data physically does not exist in 00:06:35.420 --> 00:06:38.019 its environment. Okay, so once we have that isolated 00:06:38.019 --> 00:06:40.860 room, we have to furnish it. We have to build 00:06:40.860 --> 00:06:43.120 the brains of the operation. The fun part. Yeah. 00:06:43.480 --> 00:06:45.579 The source lays out a very deliberate architecture 00:06:45.579 --> 00:06:48.339 to put inside our main workspace. You need three 00:06:48.339 --> 00:06:51.680 specific folders named About, Outputs, and Templates. 00:06:51.759 --> 00:06:53.980 It's such an important question though. Why these 00:06:53.980 --> 00:06:55.980 specific folders? Well, this is where we shift 00:06:55.980 --> 00:06:58.000 from just using an AI to actually programming 00:06:58.000 --> 00:07:00.379 an AI without writing a single line of traditional 00:07:00.379 --> 00:07:02.779 code. The About folder holds your behavioral 00:07:02.779 --> 00:07:07.610 logic. The AudiCoputes folder is obviously the 00:07:07.610 --> 00:07:10.470 destination for the finished daily reports. And 00:07:10.470 --> 00:07:12.490 the templates folder provides the structural 00:07:12.490 --> 00:07:15.019 anchor. because language models are inherently 00:07:15.019 --> 00:07:17.220 stochastic. Meaning they introduce randomness. 00:07:17.379 --> 00:07:19.300 Exactly, they introduce variation to sound natural. 00:07:19.879 --> 00:07:22.600 So a template forces the model to adhere to a 00:07:22.600 --> 00:07:25.360 rigid structure every single time it runs. Let's 00:07:25.360 --> 00:07:28.319 dig into that About folder because the specific 00:07:28.319 --> 00:07:30.839 files you put in there totally dictate the success 00:07:30.839 --> 00:07:33.579 of the entire dashboard. The instructions require 00:07:33.579 --> 00:07:37.120 creating three Markdown files. And if you're 00:07:37.120 --> 00:07:39.120 listening and thinking, why Markdown and not 00:07:39.120 --> 00:07:41.980 just a Word document? It really comes down to 00:07:41.980 --> 00:07:45.180 efficiency. Markdown, which uses the MD extension, 00:07:45.519 --> 00:07:48.139 is a lightweight plain text format. Super clean. 00:07:48.459 --> 00:07:51.279 It strips away all the invisible clunky formatting 00:07:51.279 --> 00:07:54.319 code that Microsoft Word or like a PDF uses. 00:07:54.540 --> 00:07:57.040 It is incredibly cheap for a language model to 00:07:57.040 --> 00:07:59.199 read, which saves on token usage and processing 00:07:59.199 --> 00:08:01.519 time. Yeah, it is basically the native language 00:08:01.519 --> 00:08:03.779 of these models. So inside that About folder, 00:08:04.139 --> 00:08:06.199 the first Markdown file you create is sourcelist 00:08:06.199 --> 00:08:09.279 .md. This is your boundary for reality. The text 00:08:09.279 --> 00:08:11.759 suggests limiting this to five to eight highly 00:08:11.759 --> 00:08:14.339 trusted sources. Places like the Anthropic Blog, 00:08:14.980 --> 00:08:17.639 Google DeepMind, TechCrunch, The Verge. You are 00:08:17.639 --> 00:08:20.240 explicitly telling the AI, do not scrape the 00:08:20.240 --> 00:08:22.879 open internet for rumors. Only pull data from 00:08:22.879 --> 00:08:25.579 these specific URLs. And the rules written inside 00:08:25.579 --> 00:08:27.500 that file are aggressively strict about truth. 00:08:27.759 --> 00:08:30.500 You write explicit commands, no duplicate stories, 00:08:30.660 --> 00:08:33.500 do not treat rumors as facts, and crucially, 00:08:33.879 --> 00:08:36.580 if a claim is uncertain or the AI cannot verify 00:08:36.580 --> 00:08:39.340 a data point across those specific sources, it 00:08:39.340 --> 00:08:41.740 must explicitly tag the item with the phrase, 00:08:41.899 --> 00:08:44.460 needs verification. You are engineering a safeguard 00:08:44.460 --> 00:08:46.879 against hallucination. Look, LLMs are designed 00:08:46.879 --> 00:08:49.779 to predict the next most likely word. They naturally 00:08:49.779 --> 00:08:51.799 want to give you a confident, complete answer, 00:08:52.039 --> 00:08:53.960 even if they have to invent the details to do 00:08:53.960 --> 00:08:55.879 so. Right, they want to please you. Exactly. 00:08:56.080 --> 00:08:59.179 By writing need certification into the system 00:08:59.179 --> 00:09:02.340 prompt, you are giving the AI a sanctioned off 00:09:02.340 --> 00:09:04.580 -ramp. You are giving it permission to admit 00:09:04.580 --> 00:09:07.100 it doesn't know something. Which is huge. So 00:09:07.100 --> 00:09:09.259 we've established where the AI gets its reality, 00:09:09.500 --> 00:09:11.759 but a list of sources isn't enough, because to 00:09:11.759 --> 00:09:15.279 an AI, a minor bug fix in a software update looks 00:09:15.279 --> 00:09:18.019 just as valid as, like, a sweeping international 00:09:18.019 --> 00:09:20.019 AI regulation. Right, you can't tell the difference. 00:09:20.039 --> 00:09:22.080 It needs a rubric. And that brings us to the 00:09:22.080 --> 00:09:26.840 second file. Dashboard -rules .md. This file 00:09:26.840 --> 00:09:30.399 acts as the AI's editorial judgment. You instruct 00:09:30.399 --> 00:09:32.799 the AI to score every piece of news on a scale 00:09:32.799 --> 00:09:34.980 of one to five based on practical impact and 00:09:34.980 --> 00:09:37.940 urgency. You are defining what important actually 00:09:37.940 --> 00:09:40.379 means in the context of your specific business. 00:09:40.860 --> 00:09:43.259 You might write rules prioritizing major model 00:09:43.259 --> 00:09:47.000 releases or new enterprise agents while explicitly 00:09:47.000 --> 00:09:49.279 instructing it to filter out pure opinion pieces, 00:09:49.580 --> 00:09:52.440 funding rumors, or tiny incremental updates that 00:09:52.440 --> 00:09:54.789 don't change user behavior. Yeah. Without this 00:09:54.789 --> 00:09:56.929 scoring rubric, the AI lacks the capacity to 00:09:56.929 --> 00:09:59.230 weigh impact against hype. It simply processes 00:09:59.230 --> 00:10:02.309 text. The dashboard -rules .md file gives the 00:10:02.309 --> 00:10:04.950 AI an analytical lens. But, you know, identifying 00:10:04.950 --> 00:10:07.330 a crucial, highly -scored software update doesn't 00:10:07.330 --> 00:10:09.529 help much if the AI summarizes it, sounding like 00:10:09.529 --> 00:10:12.070 a used car salesman. Oh, the AI voice. We have 00:10:12.070 --> 00:10:14.570 to actively suppress its natural tone, which 00:10:14.570 --> 00:10:17.009 is where the third file comes in. Writing -style 00:10:17.009 --> 00:10:20.500 .md. The playbook's instructions here are honestly 00:10:20.500 --> 00:10:23.720 ruthless. It demands clear, direct English with 00:10:23.720 --> 00:10:26.720 short sentences. And it includes a hilariously 00:10:26.720 --> 00:10:29.600 specific banned word list. I love this part. 00:10:29.789 --> 00:10:32.809 You explicitly forbid the AI from using generic 00:10:32.809 --> 00:10:35.769 hype words like revolutionary, game -changing, 00:10:36.190 --> 00:10:39.509 unlock, unleash, seamless, and dive into. Banning 00:10:39.509 --> 00:10:42.129 those words is arguably the most critical usability 00:10:42.129 --> 00:10:44.850 step in the entire playbook. We all instantly 00:10:44.850 --> 00:10:47.950 recognize that overly enthusiastic, artificially 00:10:47.950 --> 00:10:50.809 polished AI tone. Oh yeah? It feels like a corporate 00:10:50.809 --> 00:10:53.129 press release, and human brains are now trained 00:10:53.129 --> 00:10:55.429 to just immediately tune it out. Let me challenge 00:10:55.429 --> 00:10:58.500 that, though. If we ban exciting words like revolutionary 00:10:58.500 --> 00:11:01.240 or game -changing and force it to only use short 00:11:01.240 --> 00:11:03.980 factual sentences, won't the final report be 00:11:03.980 --> 00:11:06.299 incredibly boring to read? Doesn't it strip the 00:11:06.299 --> 00:11:08.240 energy out of reading the news? Well, boring 00:11:08.240 --> 00:11:11.220 in the context of a daily business workflow is 00:11:11.220 --> 00:11:13.860 actually highly efficient. Consider the end user 00:11:13.860 --> 00:11:16.980 of this dashboard, a busy founder, a creator 00:11:16.980 --> 00:11:20.000 or a marketing director. They are reading this 00:11:20.000 --> 00:11:23.200 at 8 a .m. to make decisions. Right. Artificial 00:11:23.200 --> 00:11:25.779 enthusiasm is friction. It forces the reader 00:11:25.779 --> 00:11:27.840 to parse through adjectives just to find the 00:11:27.840 --> 00:11:30.659 verb. When you ban fluff words like unleashed 00:11:30.659 --> 00:11:33.080 seamless workflows, you force the language model 00:11:33.080 --> 00:11:35.120 to explain the actual mechanics of the news. 00:11:35.679 --> 00:11:38.440 It has to say, this tool reduces rendering time 00:11:38.440 --> 00:11:42.210 by 20%. Directness makes the information actionable. 00:11:42.549 --> 00:11:44.830 OK, I see. It forces the AI to rely on facts 00:11:44.830 --> 00:11:47.889 instead of vibes. I love that. Exactly. So we've 00:11:47.889 --> 00:11:50.169 forced it to be factual, analytical, and direct. 00:11:50.450 --> 00:11:52.490 We take all of that, and we tie it together with 00:11:52.490 --> 00:11:54.629 the templates folder. Inside there, you build 00:11:54.629 --> 00:11:57.190 unused -dashboard -template .md. This is the 00:11:57.190 --> 00:11:59.759 literal mold for the output. You'd find the headers, 00:12:00.419 --> 00:12:02.419 executive summary, the 1 to 5 impact score, why 00:12:02.419 --> 00:12:04.419 it matters. You could even add custom angles 00:12:04.419 --> 00:12:07.220 to the template, like forcing the AI to draft 00:12:07.220 --> 00:12:09.519 a specific LinkedIn post based on the top -scored 00:12:09.519 --> 00:12:12.320 story of the day. It transforms a raw list of 00:12:12.320 --> 00:12:14.720 links into a highly -structured utilitarian tool. 00:12:14.960 --> 00:12:17.320 So we have the rulebooks written. The folder 00:12:17.320 --> 00:12:20.240 architecture is locked. But an architecture is 00:12:20.240 --> 00:12:22.779 completely useless if the AI just ignores it. 00:12:22.879 --> 00:12:25.480 Right. We have to guarantee the AI actually reads 00:12:25.480 --> 00:12:27.919 these markdown files before jumping into its 00:12:27.919 --> 00:12:31.000 task. And this is where we set up global instructions. 00:12:31.620 --> 00:12:33.460 Global instructions function as the operating 00:12:33.460 --> 00:12:36.500 system for this specific workspace. You find 00:12:36.500 --> 00:12:38.639 them in the Settings menu under the Cowork tab. 00:12:39.299 --> 00:12:41.639 This is where you establish immutable ground 00:12:41.639 --> 00:12:44.120 rules for the folder. So it happens in the background. 00:12:44.559 --> 00:12:46.740 Exactly. You write a command that effectively 00:12:46.740 --> 00:12:49.860 says, whenever you execute a task in this workspace, 00:12:50.460 --> 00:12:52.960 you must seamlessly ingest the files in the About 00:12:52.960 --> 00:12:55.960 folder, you must format your output using the 00:12:55.960 --> 00:12:58.659 Templates folder, and you must save the final 00:12:58.659 --> 00:13:01.299 markdown file directly into the IDPutJust folder. 00:13:01.659 --> 00:13:03.299 It eliminates the need for you to repeat the 00:13:03.299 --> 00:13:05.620 ground rules every morning. The playbook highlights 00:13:05.620 --> 00:13:08.240 the stark difference between a weak task prompt 00:13:08.240 --> 00:13:11.259 and a strong one. A weak task is opening the 00:13:11.259 --> 00:13:14.279 workspace and just typing, make an AI news report. 00:13:14.440 --> 00:13:17.779 A strong task triggers the architecture. Create 00:13:17.779 --> 00:13:21.340 today's daily AI news dashboard using the approved 00:13:21.340 --> 00:13:24.679 source list, editorial rules, and template. Ask 00:13:24.679 --> 00:13:27.220 me questions first if anything is unclear. We 00:13:27.220 --> 00:13:29.700 connect this to the bigger picture. That directive, 00:13:30.059 --> 00:13:33.059 ask me questions first, changes the entire dynamic 00:13:33.059 --> 00:13:35.629 of the interaction. When you initiate the task, 00:13:35.909 --> 00:13:38.110 you will see the cowork interface light up. It 00:13:38.110 --> 00:13:40.110 shows you its internal monologue. Which is really 00:13:40.110 --> 00:13:42.009 cool to watch. It is. You'll see it actively 00:13:42.009 --> 00:13:44.929 reading the source list .md, scanning the dashboard 00:13:44.929 --> 00:13:48.269 dash rules .md, and pulling the data. But a properly 00:13:48.269 --> 00:13:50.570 designed agent will often pause its execution 00:13:50.570 --> 00:13:53.409 to ask you clarifying questions. Right. It might 00:13:53.409 --> 00:13:55.809 stop and say, I found a major story about a new 00:13:55.809 --> 00:13:58.710 developer tool and a story about marketing regulations. 00:13:59.149 --> 00:14:01.309 Which audience matters most to your strategy 00:14:01.309 --> 00:14:03.500 today? A lot of users get frustrated by that, 00:14:03.600 --> 00:14:05.379 though. They feel like, hey, I built this system 00:14:05.379 --> 00:14:07.299 so I wouldn't have to think, just do the work. 00:14:07.740 --> 00:14:10.700 So what does this all mean? Well, global instructions 00:14:10.700 --> 00:14:13.299 are essentially an employee handbook, and the 00:14:13.299 --> 00:14:15.980 daily task prompt is your morning stand -up meeting. 00:14:16.179 --> 00:14:18.779 That's a great way to put it. If the AI asks 00:14:18.779 --> 00:14:21.620 clarifying questions during that stand -up, it 00:14:21.620 --> 00:14:24.600 means it's actually contextualizing the job rather 00:14:24.600 --> 00:14:27.559 than blindly agreeing to execute a task it doesn't 00:14:27.559 --> 00:14:29.480 fully grasp. Yeah, because the alternative to 00:14:29.480 --> 00:14:32.080 a clarifying question is a hallucinated assumption. 00:14:32.720 --> 00:14:35.840 If an LLM encounters a gap in its context and 00:14:35.840 --> 00:14:38.720 isn't permitted to ask for clarity, its underlying 00:14:38.720 --> 00:14:41.500 programming compels it to just guess. And it's 00:14:41.500 --> 00:14:43.519 usually a bad guess. Right. It will fill the 00:14:43.519 --> 00:14:45.799 void with generalized lowest common denominator 00:14:45.799 --> 00:14:49.000 data. By forcing the agent to pause and ask, 00:14:49.440 --> 00:14:51.639 you are ensuring the final output is tightly 00:14:51.639 --> 00:14:54.620 aligned with your specific immediate needs. It's 00:14:54.620 --> 00:14:56.840 a feature of a robust system. It is not a flaw. 00:14:57.049 --> 00:15:00.110 So we run the task, we answer the quick clarifying 00:15:00.110 --> 00:15:02.389 question, we watch the agent read the sources, 00:15:02.690 --> 00:15:05.330 filter the noise using our 1 -5 rubric, strip 00:15:05.330 --> 00:15:07.570 out the hype words, format it perfectly into 00:15:07.570 --> 00:15:10.289 our template, and then silently drop a finished 00:15:10.289 --> 00:15:14.490 markdown file into our O2P UTS folder. It's incredibly 00:15:14.490 --> 00:15:17.409 satisfying. It really is. But doing this manually 00:15:17.409 --> 00:15:20.100 every morning is still a loop. The ultimate goal 00:15:20.100 --> 00:15:22.940 of the playbook is absolute automation, taking 00:15:22.940 --> 00:15:25.100 yourself out of the repetitive loop entirely. 00:15:25.240 --> 00:15:27.860 Yes, taking the human out of the execution phase. 00:15:28.740 --> 00:15:30.820 Once you have run the task manually a few times 00:15:30.820 --> 00:15:33.240 and verify that the AI perfectly adheres to the 00:15:33.240 --> 00:15:35.519 boundaries of the workspace, you crystallize 00:15:35.519 --> 00:15:38.039 the workflow. Okay. You save your strongest task 00:15:38.039 --> 00:15:41.789 construction into a dedicated file. like dailynewsdashboard 00:15:41.789 --> 00:15:45.090 -task .md. Instead of typing a prompt, your only 00:15:45.090 --> 00:15:47.389 input becomes telling the co -work agent, run 00:15:47.389 --> 00:15:49.649 the daily news task. But the real apex of this 00:15:49.649 --> 00:15:51.730 playbook is the schedule feature. The source 00:15:51.730 --> 00:15:53.830 explains that on supported accounts, you simply 00:15:53.830 --> 00:15:56.470 type schedule into the command bar and instruct 00:15:56.470 --> 00:15:59.529 the system to run the entire workflow at, say, 00:15:59.750 --> 00:16:02.950 8 0 a .m. every single day. It's magic. You just 00:16:02.950 --> 00:16:06.269 wake up, open your laptop, and a highly synthesized, 00:16:06.570 --> 00:16:09.389 custom curated industry report is just sitting 00:16:09.389 --> 00:16:13.029 locally on your hard drive. But there are physical 00:16:13.029 --> 00:16:15.269 realities to how this operates, right? Yeah, 00:16:15.350 --> 00:16:18.309 there's a catch, because cowork is executing 00:16:18.309 --> 00:16:21.080 locally. meaning it is interacting with your 00:16:21.080 --> 00:16:23.299 desktop environment to read and write files. 00:16:24.100 --> 00:16:26.240 It is bound by the physical state of your machine. 00:16:26.320 --> 00:16:28.980 Right. It operates similarly to a local Cron 00:16:28.980 --> 00:16:32.019 job or a background script. For a scheduled task 00:16:32.019 --> 00:16:34.899 to trigger the API, access the local folders, 00:16:35.019 --> 00:16:37.220 and generate the document, your computer must 00:16:37.220 --> 00:16:40.019 be awake and the Claw desktop app must be running. 00:16:40.500 --> 00:16:43.100 If your laptop is powered down in your bag, the 00:16:43.100 --> 00:16:45.240 automated trigger will just fail. The playbook 00:16:45.240 --> 00:16:48.000 also highlights several pretty catastrophic mistakes 00:16:48.000 --> 00:16:50.240 users make when trying to automate too quickly, 00:16:50.639 --> 00:16:52.919 like writing bloated context files that confuse 00:16:52.919 --> 00:16:55.820 the model or failing to explicitly define the 00:16:55.820 --> 00:16:58.039 output directory, which results in the agent 00:16:58.039 --> 00:16:59.960 generating the report and then essentially throwing 00:16:59.960 --> 00:17:01.759 it into a digital void where you can't even find 00:17:01.759 --> 00:17:04.119 it. Yeah, it just vanishes. But the most dangerous 00:17:04.119 --> 00:17:06.640 mistake is scheduling the agent before manually 00:17:06.640 --> 00:17:09.339 stress testing it. I mean, I'd be terrified to 00:17:09.339 --> 00:17:11.559 just let this run on autopilot while I'm asleep. 00:17:12.099 --> 00:17:14.680 What if it hallucinates a major news story? If 00:17:14.680 --> 00:17:18.660 a massive, complex tech story breaks at 3 a .m., 00:17:18.660 --> 00:17:21.539 I am relying entirely on an AI to synthesize 00:17:21.539 --> 00:17:24.140 reality without my oversight. Which loops us 00:17:24.140 --> 00:17:26.940 entirely back to the folder architecture. The 00:17:26.940 --> 00:17:29.019 fear of an autonomous agent hallucinating facts 00:17:29.019 --> 00:17:30.680 while you're asleep is exactly why you spent 00:17:30.680 --> 00:17:33.700 the time engineering those about files. Ah, right. 00:17:34.160 --> 00:17:37.259 The playbook specifically flags skipping verification 00:17:37.259 --> 00:17:40.660 as a fatal error in system design. You commanded 00:17:40.660 --> 00:17:43.460 the AI, separate facts from analysis, do not 00:17:43.460 --> 00:17:45.420 invent missing numbers, and most importantly, 00:17:45.579 --> 00:17:48.559 if a claim is unclear, flag it as needs verification. 00:17:48.900 --> 00:17:50.880 So the skepticism is built directly into the 00:17:50.880 --> 00:17:53.759 prompt's DNA? It has to be. You're not blindly 00:17:53.759 --> 00:17:55.980 trusting a language model to be truthful. You 00:17:55.980 --> 00:17:58.220 have engineered a local ecosystem that forces 00:17:58.220 --> 00:18:01.019 the model to expose its own uncertainties. If 00:18:01.019 --> 00:18:02.960 the agent cannot corroborate a breaking story 00:18:02.960 --> 00:18:05.920 across your five specific sources, it is structurally 00:18:05.920 --> 00:18:08.339 obligated to visually warn you in the final output. 00:18:08.460 --> 00:18:11.299 And that perfectly illustrates the core philosophy 00:18:11.299 --> 00:18:14.539 of this entire playbook. It outlines a very clear 00:18:14.539 --> 00:18:17.059 evolutionary path for how we interact with this 00:18:17.059 --> 00:18:20.640 technology. Prompt, then process, then system, 00:18:20.839 --> 00:18:24.440 then automation. Flawless outputs don't materialize 00:18:24.440 --> 00:18:27.180 from a magic three -page megaprompt. No, they 00:18:27.180 --> 00:18:30.019 don't. They are the result of meticulously designed 00:18:30.019 --> 00:18:32.599 systems. Exactly. The architecture does the heavy 00:18:32.599 --> 00:18:34.940 lifting, not the chat window. Prompt, process, 00:18:35.240 --> 00:18:38.240 system, automation. We have walked through how 00:18:38.240 --> 00:18:40.680 organizing local files and forcing rigid templates 00:18:40.680 --> 00:18:43.579 and defining strict behavioral rules transforms 00:18:43.579 --> 00:18:46.200 a simple chat interface into a highly reliable, 00:18:46.579 --> 00:18:48.900 autonomous digital employee, one that delivers 00:18:48.900 --> 00:18:51.059 a customized briefing straight to your hard drive. 00:18:51.200 --> 00:18:53.559 And while the playbook focuses entirely on an 00:18:53.559 --> 00:18:56.960 AI news dashboard, I mean, the mechanical framework 00:18:56.960 --> 00:18:59.079 is universally applicable. Oh, absolutely. Once 00:18:59.079 --> 00:19:01.000 you understand how to build a localized workspace, 00:19:01.599 --> 00:19:03.799 you can construct agents for daily lead generation, 00:19:04.160 --> 00:19:06.480 competitor price analysis, or parsing hundreds 00:19:06.480 --> 00:19:08.960 of weekly customer support tickets into actionable 00:19:08.960 --> 00:19:11.740 product updates. It fundamentally rewires how 00:19:11.740 --> 00:19:14.400 you perceive the tool. You transition from asking, 00:19:14.579 --> 00:19:18.539 what question can I ask this AI, to what automated 00:19:18.539 --> 00:19:21.019 system can I build with this engine, which leaves 00:19:21.019 --> 00:19:24.119 us with a really lingering thought. If AI can 00:19:24.119 --> 00:19:27.079 now perfectly execute these complex, repetitive 00:19:27.079 --> 00:19:29.880 synthesis tasks on autopilot tasks that used 00:19:29.880 --> 00:19:32.500 to take human analysts hours of reading and formatting, 00:19:33.279 --> 00:19:35.559 how does that alter our core role in the workplace? 00:19:35.880 --> 00:19:37.859 That's the big question. Right. If we are moving 00:19:37.859 --> 00:19:40.500 away from being the information gatherers, it 00:19:40.500 --> 00:19:42.480 seems inevitable that we need to become system 00:19:42.480 --> 00:19:45.059 designers instead. The stranger on the street 00:19:45.059 --> 00:19:47.059 just became your dedicated personal assistant. 00:19:47.539 --> 00:19:49.319 The only question left to explore on your own 00:19:49.319 --> 00:19:51.539 is, what are you going to build with all that 00:19:51.539 --> 00:19:54.019 reclaimed time? You transition from performing 00:19:54.019 --> 00:19:57.039 the labor to managing the intelligence that performs 00:19:57.039 --> 00:19:59.700 the labor. Something to carefully ponder as you 00:19:59.700 --> 00:20:01.819 begin mapping out your own workspaces. We'll 00:20:01.819 --> 00:20:02.359 leave you to it.