WEBVTT 00:00:00.000 --> 00:00:03.140 Imagine an AI that doesn't just talk, but actually, 00:00:03.140 --> 00:00:05.200 you know, works while you sleep. Oh, yeah. That 00:00:05.200 --> 00:00:07.740 is the dream, right? Think about that for a second. 00:00:07.780 --> 00:00:10.140 It's the fundamental difference between having 00:00:10.140 --> 00:00:13.779 a simple chat with an AI and building a functioning 00:00:13.779 --> 00:00:16.920 independent system. If you want a digital employee, 00:00:17.260 --> 00:00:19.820 you basically have to build it a workspace. Precisely. 00:00:19.940 --> 00:00:22.820 And that's exactly what we are doing today with 00:00:22.820 --> 00:00:25.760 Claude Cowork. Right. Welcome back to the Deep 00:00:25.760 --> 00:00:29.100 Dive. Today, we are exploring building your first 00:00:29.239 --> 00:00:32.219 Claude Cowork Agent. We're going to cover the 00:00:32.219 --> 00:00:34.659 setup, the folder logic, and the automation steps. 00:00:34.859 --> 00:00:36.960 Yeah, you probably use Claude every day. But 00:00:36.960 --> 00:00:38.840 honestly, you're likely using it wrong, or at 00:00:38.840 --> 00:00:41.479 least you're only scratching the surface. Totally. 00:00:42.159 --> 00:00:45.479 Today is a literal roadmap for moving beyond 00:00:45.479 --> 00:00:48.479 that traditional chat box. We're going to explore 00:00:48.479 --> 00:00:51.759 how you can use Cowork to build a real, functional, 00:00:52.000 --> 00:00:55.340 daily AI news dashboard. OK, let's unpack this. 00:00:55.600 --> 00:00:57.700 Because when I first heard about Cowork, I honestly 00:00:57.700 --> 00:00:59.679 thought it was just I don't know, another UI 00:00:59.679 --> 00:01:01.640 update. Right, like just a visual thing. Yep, 00:01:01.759 --> 00:01:04.359 exactly. Another place to tap a question. But 00:01:04.359 --> 00:01:07.540 Cowork is totally different. It's built for repeatable 00:01:07.540 --> 00:01:11.200 file -based jobs. It lives inside your folders, 00:01:11.700 --> 00:01:15.280 it reads, it writes, and it saves documents rather 00:01:15.280 --> 00:01:17.599 than just existing in a fleeting conversation 00:01:17.599 --> 00:01:20.299 window where you have to constantly copy and 00:01:20.299 --> 00:01:22.079 paste everything. What's fascinating here is 00:01:22.079 --> 00:01:25.700 the shift in computing behavior. We are all used 00:01:25.700 --> 00:01:28.299 to standard cloud chat. Which is great, obviously. 00:01:28.379 --> 00:01:31.140 Oh, it's fantastic for quick brainstorming, drafting 00:01:31.140 --> 00:01:33.579 a single email, or just answering a one -off 00:01:33.579 --> 00:01:37.030 question. To transition from getting one -time 00:01:37.030 --> 00:01:39.310 help to building automated business workflows, 00:01:39.870 --> 00:01:41.989 you are no longer a conversational partner. Right. 00:01:42.150 --> 00:01:45.090 You are a systems architect. I love that. A systems 00:01:45.090 --> 00:01:47.269 architect. So if I'm playing that role, what 00:01:47.269 --> 00:01:49.650 are my actual building materials? Let's talk 00:01:49.650 --> 00:01:52.109 prerequisites for you guys listening. First, 00:01:52.370 --> 00:01:54.549 you need the cloud desktop app installed. Yes. 00:01:54.750 --> 00:01:56.769 This isn't running in a browser tab. Second, 00:01:56.810 --> 00:01:59.489 you need a paid plan. Pro is the bare minimum 00:01:59.489 --> 00:02:01.189 here. And we should probably explain exactly 00:02:01.189 --> 00:02:03.349 why that free tier simply won't cut it. Yeah, 00:02:03.430 --> 00:02:07.510 please do. Core workflows inherently require 00:02:07.510 --> 00:02:11.710 much higher usage limits and significantly more 00:02:11.710 --> 00:02:14.509 processing context. Right. When you run a dashboard 00:02:14.509 --> 00:02:18.870 workflow, the AI is reading multiple files. It's 00:02:18.870 --> 00:02:20.789 cross -referencing your instructions against 00:02:20.789 --> 00:02:23.409 several external articles, and it's generating 00:02:23.409 --> 00:02:26.389 heavily formatted outputs. It's doing a lot at 00:02:26.389 --> 00:02:29.430 once. Exactly. It needs a massive working memory 00:02:29.430 --> 00:02:31.930 to hold all that data simultaneously. That makes 00:02:31.930 --> 00:02:33.789 total sense. And that actually brings up what 00:02:33.789 --> 00:02:36.259 I like to call the work Golden Roll, you need 00:02:36.259 --> 00:02:39.479 a dedicated workspace folder on your actual computer. 00:02:39.740 --> 00:02:41.860 Yes, absolutely. You might name it something 00:02:41.860 --> 00:02:44.860 like Claude Cookework News Dashboard, but I want 00:02:44.860 --> 00:02:46.919 to make a hard stop right here because the temptation 00:02:46.919 --> 00:02:49.199 is just to give the AI access to everything. 00:02:49.379 --> 00:02:52.039 Oh, big mistake. Do not just point Claude at 00:02:52.039 --> 00:02:54.000 your whole computer or your entire documents 00:02:54.000 --> 00:02:56.360 folder. That is like letting a brand new intern 00:02:56.360 --> 00:02:58.379 rifle through all your personal filing cabinets, 00:02:58.520 --> 00:03:01.099 your tax returns, your family photos, just to 00:03:01.099 --> 00:03:03.879 find a single press release. Yeah, that is an 00:03:03.879 --> 00:03:06.919 incredibly dangerous trap for new users. It really 00:03:06.919 --> 00:03:08.699 is. If we connect this to the bigger picture 00:03:08.699 --> 00:03:11.139 of how language models operate, giving the AI 00:03:11.139 --> 00:03:14.000 too much access just creates massive confusion. 00:03:14.219 --> 00:03:16.219 Right, because it's reading everything. Exactly. 00:03:16.620 --> 00:03:19.340 Every time it reads an irrelevant file, it eats 00:03:19.340 --> 00:03:21.759 up its processing context, its working memory. 00:03:22.080 --> 00:03:24.599 It has to sift through your grocery lists and 00:03:24.599 --> 00:03:27.080 old spreadsheets just to figure out what a news 00:03:27.080 --> 00:03:30.780 update means. Wow. That dilutes its focus. You 00:03:30.780 --> 00:03:33.219 want to isolate the work space so the AI only 00:03:33.219 --> 00:03:35.680 sees exactly what it needs to execute the task. 00:03:36.259 --> 00:03:38.659 Nothing more, nothing less. And speaking of focus, 00:03:39.080 --> 00:03:41.259 you also have to dictate which underlying model 00:03:41.259 --> 00:03:43.840 is actually doing the processing. The sources 00:03:43.840 --> 00:03:46.500 are very clear about using the strongest model 00:03:46.500 --> 00:03:49.580 available for research tasks. So we are talking 00:03:49.580 --> 00:03:52.500 about using something like Claude 3 .5 Opus or 00:03:52.500 --> 00:03:55.000 Sonnet rather than the faster, lighter Haiku 00:03:55.000 --> 00:03:57.460 model, right? Yes. Precision matters immensely 00:03:57.460 --> 00:04:00.199 here. When you ask an AI to analyze multiple 00:04:00.199 --> 00:04:02.759 sources, rank them, and synthesize the information, 00:04:03.219 --> 00:04:05.319 you are fighting against a phenomenon called 00:04:05.319 --> 00:04:08.259 prompt drift. Wait, prompt drift. Explain prompt 00:04:08.259 --> 00:04:10.479 drift for me. Sure. Think of it like a game of 00:04:10.479 --> 00:04:12.659 telephone happening inside the AI's own head. 00:04:12.810 --> 00:04:15.889 Over the course of a complex multi -step task, 00:04:16.490 --> 00:04:19.730 a weaker model slowly forgets or misinterprets 00:04:19.730 --> 00:04:22.189 the original instructions. Oh, I've seen that 00:04:22.189 --> 00:04:24.350 happen. Right. It might start out strong on the 00:04:24.350 --> 00:04:26.790 first two articles, perfectly applying your formatting. 00:04:27.290 --> 00:04:30.209 But by the time it reads the 10th article, its 00:04:30.209 --> 00:04:33.149 working memory is stretched, and it starts hallucinating, 00:04:33.629 --> 00:04:36.509 summarizing things in the wrong tone, or just 00:04:36.509 --> 00:04:39.389 completely ignoring your formatting rules. That's 00:04:39.389 --> 00:04:41.970 so frustrating. It is. So using the strongest 00:04:41.970 --> 00:04:44.829 model, provides the immense cognitive stamina 00:04:44.829 --> 00:04:47.389 required to hold onto your rules from start to 00:04:47.389 --> 00:04:50.629 finish. Got it. So you've restricted the AI's 00:04:50.629 --> 00:04:52.970 access to your tax returns, and you've selected 00:04:52.970 --> 00:04:55.110 a model with enough stamina to avoid drifting 00:04:55.110 --> 00:04:57.689 off course. Right. Now you actually have to build 00:04:57.689 --> 00:05:00.430 the physical structure inside that isolated folder. 00:05:00.670 --> 00:05:02.870 Exactly. Think of the folder structure like an 00:05:02.870 --> 00:05:05.529 assembly line in a factory. To get this digital 00:05:05.529 --> 00:05:08.709 employee working smoothly, you need three main 00:05:08.709 --> 00:05:10.889 folders on that assembly line. Okay, what are 00:05:10.889 --> 00:05:14.410 they? about, outputs, and templates. I have to 00:05:14.410 --> 00:05:17.649 admit, creating three separate folders and a 00:05:17.649 --> 00:05:19.990 bunch of markdown files just to get a new summary 00:05:19.990 --> 00:05:22.930 feels incredibly tedious. It sounds like a lot 00:05:22.930 --> 00:05:25.930 of work. It really does. Why can't I just paste 00:05:25.930 --> 00:05:28.810 a massive three page prompt into the regular 00:05:28.810 --> 00:05:31.009 clawed chat window every morning and get the 00:05:31.009 --> 00:05:33.029 exact same result? Well, because copying and 00:05:33.029 --> 00:05:35.750 pasting a massive prompt every day isn't a system, 00:05:35.829 --> 00:05:38.709 it's a chore. Fair point. If you put all your 00:05:38.709 --> 00:05:40.949 rules into a chat prompt, that knowledge vanishes 00:05:40.949 --> 00:05:43.569 the moment you close the window. Yeah. By building 00:05:43.569 --> 00:05:45.790 the structure on your hard drive, you are creating 00:05:45.790 --> 00:05:49.569 permanence. The about folder serves as the permanent 00:05:49.569 --> 00:05:52.350 blueprint for the assembly line. Yeah. The templates 00:05:52.350 --> 00:05:55.720 folder is the mold. If you don't build the track 00:05:55.720 --> 00:05:58.879 correctly and permanently, the robotic arm claw 00:05:58.879 --> 00:06:00.959 just flails around in the dark every single time 00:06:00.959 --> 00:06:03.100 you start a new session. OK, that clarifies it. 00:06:03.319 --> 00:06:06.339 We are trading upfront effort for long term automation. 00:06:06.560 --> 00:06:08.620 I can get behind that. Exactly. So let's look 00:06:08.620 --> 00:06:11.360 at that blueprint, the ABOUT folder. Our sources 00:06:11.360 --> 00:06:13.519 say this is where you store the high signal rules 00:06:13.519 --> 00:06:16.120 that filter out the noise before the AI even 00:06:16.120 --> 00:06:19.970 begins to work. Yes. The internet is filled with 00:06:19.970 --> 00:06:22.670 garbage data. So much garbage. Left to its own 00:06:22.670 --> 00:06:24.709 devices, and AI will try to summarize all of 00:06:24.709 --> 00:06:28.529 it. The files in your about folder are your hard 00:06:28.529 --> 00:06:30.569 filters to prevent that. Let's walk through those 00:06:30.569 --> 00:06:33.129 filters. First up is your source list dot dash 00:06:33.129 --> 00:06:36.569 MD file. And in this file, you explicitly limit 00:06:36.569 --> 00:06:39.410 the AI to five to eight trusted sources. You 00:06:39.410 --> 00:06:42.579 organize them by priority. Priority 1 is official 00:06:42.579 --> 00:06:45.899 company blogs like Antropic or OpenAI. Priority 00:06:45.899 --> 00:06:48.240 2 is research repositories like Hugging Face. 00:06:48.720 --> 00:06:51.100 Priority 3 is general business and tech news. 00:06:51.139 --> 00:06:53.360 Right. But I'm looking at this limitation and 00:06:53.360 --> 00:06:55.800 wondering why handcuff the AI? Why not just say, 00:06:56.420 --> 00:06:58.240 search the whole internet for AI news today? 00:06:58.410 --> 00:07:01.430 Because a highly curated source list produces 00:07:01.430 --> 00:07:03.990 dramatically cleaner reports. If you tell an 00:07:03.990 --> 00:07:06.269 AI to search the entire web, it will pull from 00:07:06.269 --> 00:07:09.889 SEO, spam sites, unverified Reddit threads, and 00:07:09.889 --> 00:07:12.110 out -of -date opinion blogs. That's true. By 00:07:12.110 --> 00:07:14.790 restricting it to, say, five official URLs, you 00:07:14.790 --> 00:07:17.149 are guaranteeing the data quality. You are also 00:07:17.149 --> 00:07:19.129 explicitly instructing it, do not treat rumors 00:07:19.129 --> 00:07:21.629 as facts. OK, so we've restricted the sources. 00:07:22.250 --> 00:07:24.529 But even within five trusted official sites, 00:07:24.670 --> 00:07:27.129 there's a ton of garbage. Oh, absolutely. Companies 00:07:27.129 --> 00:07:30.139 published market. fluff all the time. How do 00:07:30.139 --> 00:07:33.579 we stop the AI from summarizing a minor interface 00:07:33.579 --> 00:07:36.339 tweak like it's a major technological breakthrough? 00:07:36.740 --> 00:07:39.160 That brings us to the second file in your About 00:07:39.160 --> 00:07:43.379 folder, the DashboardRules .md file. Okay. This 00:07:43.379 --> 00:07:45.560 is where you define consequence. You instruct 00:07:45.560 --> 00:07:48.600 the AI to rank every single news item on a strict 00:07:48.600 --> 00:07:51.180 scale of one to five based on practical impact 00:07:51.180 --> 00:07:54.120 and urgency. So how does that play out in a real 00:07:54.120 --> 00:07:57.579 scenario? Let's say OpenAI drops a massive new 00:07:57.579 --> 00:08:00.060 model upgrade, but on the exact same day, they 00:08:00.060 --> 00:08:02.439 also release a minor bug fix for their billing 00:08:02.439 --> 00:08:05.319 page. In a standard chat prompt, the AI might 00:08:05.319 --> 00:08:07.279 just give you two bullet points of equal length, 00:08:07.519 --> 00:08:09.980 treating both events as equal news. Right. But 00:08:09.980 --> 00:08:12.399 with your dashboard rules file, the AI is forced 00:08:12.399 --> 00:08:14.879 to pause and evaluate. It looks at the massive 00:08:14.879 --> 00:08:17.399 model upgrade, sees that it alters capabilities, 00:08:17.779 --> 00:08:20.160 and scores it a 5. OK. It looks at the billing 00:08:20.160 --> 00:08:22.720 bug fix, realizes it has zero impact on output 00:08:22.720 --> 00:08:25.980 quality, and scores it a 1. Your rules can then 00:08:25.980 --> 00:08:28.500 dictate only include items that score a three 00:08:28.500 --> 00:08:31.139 or higher in the final report. You've just built 00:08:31.139 --> 00:08:33.879 an automated editorial filter. You're forcing 00:08:33.879 --> 00:08:36.960 the AI to have editorial judgment based on your 00:08:36.960 --> 00:08:40.000 criteria. I see. But even if it picks the right 00:08:40.000 --> 00:08:42.399 stories, there's still the problem of how it 00:08:42.399 --> 00:08:46.360 actually talks. Yes, the tone. Which leads to 00:08:46.360 --> 00:08:48.779 the third file in the blueprint, the writing 00:08:48.779 --> 00:08:52.559 style .md file. Yes, and this might be the most 00:08:52.559 --> 00:08:55.500 crucial file for long -term readability. You 00:08:55.500 --> 00:08:58.299 use this strict writing style file to enforce 00:08:58.299 --> 00:09:01.360 clear, direct, human -sounding English. Right. 00:09:01.480 --> 00:09:03.899 You demand short sentences. You demand factual 00:09:03.899 --> 00:09:05.960 reporting. And the absolute most fascinating 00:09:05.960 --> 00:09:08.519 part of this file is the banned word list. Oh, 00:09:08.519 --> 00:09:11.379 it's so good. You explicitly ban those cliché, 00:09:11.399 --> 00:09:13.980 exhausting AI buzzwords. You tell it under no 00:09:13.980 --> 00:09:15.720 circumstances are you allowed to use the words 00:09:15.720 --> 00:09:18.340 revolutionary, game -changing, unlock, unleash, 00:09:18.600 --> 00:09:21.139 seamless, or dive into. This is a brilliant structural 00:09:21.139 --> 00:09:23.720 hack. AI models naturally default to marketing 00:09:23.720 --> 00:09:26.019 speak. They really do. Think about their training 00:09:26.019 --> 00:09:29.000 data. They have ingested millions of press releases 00:09:29.000 --> 00:09:32.159 and marketing websites. When an AI sees a product 00:09:32.159 --> 00:09:35.480 update, its highest probability predictive text 00:09:35.480 --> 00:09:38.980 pathway is to describe it as game changing. So 00:09:38.980 --> 00:09:40.779 what does this all mean when we ban those words? 00:09:41.019 --> 00:09:42.519 Were you just making it sound less annoying? 00:09:42.840 --> 00:09:45.820 It goes much deeper than just tone. Banning those 00:09:45.820 --> 00:09:48.519 words physically forces the model's neural network 00:09:48.519 --> 00:09:51.620 to access different, more analytical pathways 00:09:51.620 --> 00:09:55.100 to complete its sentence. If the AI is literally 00:09:55.100 --> 00:09:57.799 prohibited from saying, a new tool will unleash 00:09:57.799 --> 00:10:01.059 your potential, it hits a roadblock. It is forced 00:10:01.059 --> 00:10:03.419 to look closer at the data and explain the actual 00:10:03.419 --> 00:10:05.519 mechanism. That makes a lot of sense. It has 00:10:05.519 --> 00:10:08.340 to say, this tool reduces rendering time by 40 00:10:08.340 --> 00:10:10.899 percent. It completely shifts the output from 00:10:10.899 --> 00:10:13.860 passive marketing fluff to active practical analysis. 00:10:13.940 --> 00:10:15.759 So you're saying we have to artificially constrain 00:10:15.759 --> 00:10:18.320 the AI to make it smarter. Yes. We take away 00:10:18.320 --> 00:10:20.679 its favorite adjectives. to rely on strong verbs 00:10:20.679 --> 00:10:23.179 and hard facts. That is a massive paradigm shift. 00:10:23.659 --> 00:10:25.759 OK. So we have our assembly line blueprint built. 00:10:26.220 --> 00:10:28.480 The About folder is locked. But how does the 00:10:28.480 --> 00:10:31.379 AI actually know where to find the blueprint 00:10:31.379 --> 00:10:34.480 and what to do with it? That is the job of global 00:10:34.480 --> 00:10:36.940 instructions. You find these in the Settings 00:10:36.940 --> 00:10:39.960 menu of Cowork. Think of global instructions 00:10:39.960 --> 00:10:42.519 as the permanent laws of your workspace. They 00:10:42.519 --> 00:10:45.350 act as the traffic director. Right, so your global 00:10:45.350 --> 00:10:47.370 instructions would read something like read the 00:10:47.370 --> 00:10:49.830 source list, the dashboard rules, and the writing 00:10:49.830 --> 00:10:52.490 style in the about folder. Use the template in 00:10:52.490 --> 00:10:55.409 the templates folder and save the final output 00:10:55.409 --> 00:10:58.049 to the ITPU test folder. Exactly. But here's 00:10:58.049 --> 00:11:00.389 my question. If we went through all the trouble 00:11:00.389 --> 00:11:03.409 to write these complex rules, why do the sources 00:11:03.409 --> 00:11:06.450 explicitly warn us to keep the global instructions 00:11:06.450 --> 00:11:09.200 relatively short? Why not just put the band words 00:11:09.200 --> 00:11:11.080 right there in the main settings? This raises 00:11:11.080 --> 00:11:13.639 an important point about system architecture. 00:11:14.299 --> 00:11:16.740 Global instructions exist only to define the 00:11:16.740 --> 00:11:19.480 workspace logic. If you shove your one to five 00:11:19.480 --> 00:11:21.860 ranking system, your band words, and your source 00:11:21.860 --> 00:11:24.460 URLs into the global instructions, the cognitive 00:11:24.460 --> 00:11:26.940 load is too heavy at the top level. I get it. 00:11:27.120 --> 00:11:29.480 The traffic director gets overwhelmed trying 00:11:29.480 --> 00:11:31.840 to build the car while directing traffic. Great 00:11:31.840 --> 00:11:34.799 analogy. keep the global instructions short so 00:11:34.799 --> 00:11:37.019 they act as a stable foundation just pointing 00:11:37.019 --> 00:11:40.879 to the folders. The complex, specific rules belong 00:11:40.879 --> 00:11:43.259 safely isolated inside those separate markdown 00:11:43.259 --> 00:11:45.740 files. The traffic director just points. The 00:11:45.740 --> 00:11:47.899 blueprint does the heavy lifting. I like that. 00:11:47.919 --> 00:11:50.500 So the traffic director points the AI to the 00:11:50.500 --> 00:11:53.039 sources, filters the news through the rules, 00:11:53.279 --> 00:11:56.139 and then point it to the template. Right. Inside 00:11:56.139 --> 00:11:58.559 your templates folder, you have a file called 00:11:58.559 --> 00:12:03.059 inews -template .md. Exactly. And a weak dashboard 00:12:03.059 --> 00:12:06.379 template just asks for a list of headlines. A 00:12:06.379 --> 00:12:09.139 powerful, useful dashboard demands structured, 00:12:09.440 --> 00:12:12.139 actionable intelligence. The sources show this 00:12:12.139 --> 00:12:14.620 template dictating a very specific final output. 00:12:15.000 --> 00:12:17.580 It must include an executive summary. It must 00:12:17.580 --> 00:12:19.539 display the impact scores for each item that 00:12:19.539 --> 00:12:21.960 we talked about earlier. But the most valuable 00:12:21.960 --> 00:12:25.120 part is the requirement for content angles. It 00:12:25.120 --> 00:12:27.519 literally forces the AI to provide a newsletter 00:12:27.519 --> 00:12:30.200 idea or a LinkedIn post idea based specifically 00:12:30.200 --> 00:12:32.610 on that news item. What's happening here is the 00:12:32.610 --> 00:12:34.950 transformation of passive information into active 00:12:34.950 --> 00:12:37.600 utility. You aren't just reading that Anthropic 00:12:37.600 --> 00:12:40.259 released a new feature. Your system is handing 00:12:40.259 --> 00:12:43.519 you immediate, actionable ways to use that news. 00:12:43.639 --> 00:12:46.539 Which is huge. Right. For example, if you are 00:12:46.539 --> 00:12:49.120 a developer, your template might demand a code 00:12:49.120 --> 00:12:51.639 integration angle. If you are a marketer, it 00:12:51.639 --> 00:12:54.539 demands a campaign angle. You are tailoring the 00:12:54.539 --> 00:12:57.039 news directly to your daily deliverables. It 00:12:57.039 --> 00:12:59.659 is brilliant. We have our isolated folder. We 00:12:59.659 --> 00:13:02.460 have our assembly line built. We have our traffic 00:13:02.460 --> 00:13:05.399 directing global instructions. Now it is time 00:13:05.399 --> 00:13:08.360 to flip the switch and watch this digital employee 00:13:08.360 --> 00:13:12.120 go to work. Let's talk about execution. This 00:13:12.120 --> 00:13:14.320 is where you see the magic of the architecture 00:13:14.320 --> 00:13:16.440 you've built. You open a new task in Cowork, 00:13:16.779 --> 00:13:19.580 ensure it is pointed at your specific Cloud Cowork 00:13:19.580 --> 00:13:22.460 News dashboard workspace, and give it a simple 00:13:22.460 --> 00:13:25.350 one -line instruction. Like what? Just create 00:13:25.350 --> 00:13:28.330 today's daily AI news dashboard. That's it. One 00:13:28.330 --> 00:13:30.710 sense. That's it. Because of all your setup, 00:13:31.049 --> 00:13:33.509 Claude knows exactly what to do. It checks the 00:13:33.509 --> 00:13:35.710 blogs on your approved source list. It evaluates 00:13:35.710 --> 00:13:38.389 the rumors versus the facts. It ranks the updates 00:13:38.389 --> 00:13:41.870 one to five. It forces its neural pathways to 00:13:41.870 --> 00:13:44.809 avoid the banned words. And it formats everything 00:13:44.809 --> 00:13:47.610 into your template. And along the way, it might... 00:13:47.450 --> 00:13:50.049 pause and ask you questions, right? Yes, and 00:13:50.049 --> 00:13:52.490 this is a vital part of the co -work process. 00:13:52.909 --> 00:13:55.490 It might ask clarifying questions before it starts 00:13:55.490 --> 00:13:57.690 generating the final report. Give me an example. 00:13:57.929 --> 00:14:00.289 Well, it might say, I found three updates that 00:14:00.289 --> 00:14:02.549 score a four today. Do you want me to expand 00:14:02.549 --> 00:14:05.389 the content angles for all three or just focus 00:14:05.389 --> 00:14:07.970 on the top story? OK, that's helpful. You should 00:14:07.970 --> 00:14:10.190 always answer these questions. Providing that 00:14:10.190 --> 00:14:12.990 context prevents the AI from making weak assumptions 00:14:12.990 --> 00:14:15.289 and filling in the gaps itself. But I have to 00:14:15.289 --> 00:14:17.899 push back on this step. If the ultimate goal 00:14:17.899 --> 00:14:20.720 of this entire deep dive is to have this thing 00:14:20.720 --> 00:14:23.720 work while I sleep, why am I sitting here answering 00:14:23.720 --> 00:14:26.639 clarifying questions? I know, it feels counterintuitive. 00:14:26.799 --> 00:14:29.860 Why is this first manual run so critical? Why 00:14:29.860 --> 00:14:32.440 can't I just build my folders, set my rules, 00:14:32.899 --> 00:14:34.960 and immediately automate it to run every morning? 00:14:35.279 --> 00:14:37.500 It's a very common impulse, but if you automate 00:14:37.500 --> 00:14:39.600 a flawed system, you just get flawed outputs 00:14:39.600 --> 00:14:42.200 faster. Oh, that's a good point. The manual testing 00:14:42.200 --> 00:14:45.049 phase is your safety net. It allows for early 00:14:45.049 --> 00:14:47.529 corrections before you commit to a fully automated 00:14:47.529 --> 00:14:50.710 schedule. During this first run, you watch the 00:14:50.710 --> 00:14:54.070 tool activity window. If the AI grabs a source 00:14:54.070 --> 00:14:56.629 that isn't on your list, or if it slips up and 00:14:56.629 --> 00:14:58.710 calls something game -changing, you catch it 00:14:58.710 --> 00:15:00.809 immediately. And if it does make a mistake, if 00:15:00.809 --> 00:15:04.210 it hallucinates or breaks a rule, is that just 00:15:04.210 --> 00:15:07.909 the AI being stubborn? Rarely. And this is a 00:15:07.909 --> 00:15:10.730 hard truth for a lot of users to accept. If the 00:15:10.730 --> 00:15:13.409 results feel off, it is almost always a thaw 00:15:13.409 --> 00:15:16.250 in the system design, not the AI itself. Interesting. 00:15:16.309 --> 00:15:19.070 It's an error in your instructions. Maybe your 00:15:19.070 --> 00:15:21.649 definition of a level five impact was too vague. 00:15:22.149 --> 00:15:24.710 Maybe your source list had a broken URL. The 00:15:24.710 --> 00:15:27.009 manual run lets you debug your own rules. OK, 00:15:27.049 --> 00:15:28.850 so we've debugged it. We tweaked the blueprint. 00:15:29.129 --> 00:15:31.710 We ran it manually. And the output is beautiful. 00:15:31.879 --> 00:15:34.519 It is short, it is actionable, it is verified, 00:15:34.620 --> 00:15:37.240 and there's no marketing fluff. Perfect. Now 00:15:37.240 --> 00:15:40.019 we reach the holy grail, scaling and automation. 00:15:40.279 --> 00:15:42.399 Once your manual run is perfect, you can utilize 00:15:42.399 --> 00:15:44.419 the scheduling feature. You literally just type 00:15:44.419 --> 00:15:46.799 slash schedule in the task bar. Wow. You can 00:15:46.799 --> 00:15:49.820 tell Claude to run this exact dashboard workflow 00:15:49.820 --> 00:15:53.299 every single morning at 8 0 a .m. Just typing 00:15:53.299 --> 00:15:56.220 slash schedule. That feels like magic. Yeah. 00:15:56.559 --> 00:15:59.000 But. We need to be incredibly clear about the 00:15:59.000 --> 00:16:00.759 technical reality here because this is where 00:16:00.759 --> 00:16:02.679 people get frustrated. Yeah, there's a catch. 00:16:03.080 --> 00:16:05.500 What are the strict requirements for a scheduled 00:16:05.500 --> 00:16:08.759 task to actually fire? The requirements are physical. 00:16:09.120 --> 00:16:11.740 This workflow is not running on a server somewhere 00:16:11.740 --> 00:16:13.659 out in the cloud while you are disconnected. 00:16:14.279 --> 00:16:17.639 Claude Cowork operates locally. Explain the mechanics 00:16:17.639 --> 00:16:20.240 of that. Why does my computer need to be involved 00:16:20.240 --> 00:16:23.309 if the AI lives in the cloud? Because the AI's 00:16:23.309 --> 00:16:26.129 brain is in the cloud, but your workspace, your 00:16:26.129 --> 00:16:28.370 about folder, your templates, your outputs lives 00:16:28.370 --> 00:16:30.669 on your physical hard drive. For the scheduled 00:16:30.669 --> 00:16:32.850 task to execute, it physically needs to access 00:16:32.850 --> 00:16:34.970 the files on your hard drive, read them into 00:16:34.970 --> 00:16:37.789 its memory, process the data, and save a new 00:16:37.789 --> 00:16:39.889 Markdown file back to your local drive. That 00:16:39.889 --> 00:16:42.370 makes sense. If your machine is asleep or if 00:16:42.370 --> 00:16:45.250 the desktop app is closed, the bridge between 00:16:45.250 --> 00:16:47.690 Claude's cloud brain and your local files is 00:16:47.690 --> 00:16:50.289 completely subbered. The digital employee can't 00:16:50.289 --> 00:16:53.269 get into the office, so your claw desktop app 00:16:53.269 --> 00:16:55.649 must be open and your computer must be awake. 00:16:55.830 --> 00:16:57.990 So if your laptop is closed in your backpack 00:16:57.990 --> 00:17:01.570 at 8 a .m., the scheduled task will simply fail. 00:17:02.110 --> 00:17:03.870 You have to treat it like a real machine that 00:17:03.870 --> 00:17:07.059 needs power and access. Precisely. But if we 00:17:07.059 --> 00:17:08.900 pull back and look at what we've accomplished, 00:17:09.299 --> 00:17:12.240 the core lesson here is profound. Good outputs 00:17:12.240 --> 00:17:15.059 come from good systems. For the past couple of 00:17:15.059 --> 00:17:17.160 years, the entire tech world has been obsessed 00:17:17.160 --> 00:17:20.059 with writing clever, complex prompts. We are 00:17:20.059 --> 00:17:22.319 finally moving away from simple prompts and moving 00:17:22.319 --> 00:17:25.079 toward robust processes. You are building an 00:17:25.079 --> 00:17:27.400 architecture, not just asking a question. That 00:17:27.400 --> 00:17:30.180 is the perfect summary. Let's distill the actionable 00:17:30.180 --> 00:17:32.539 advice for you to take away today. You can build 00:17:32.539 --> 00:17:34.900 this. Start your first workspace folder today. 00:17:35.400 --> 00:17:37.460 Build your assembly line. Put in the upfront 00:17:37.460 --> 00:17:39.700 time to create your source list, your ranking 00:17:39.700 --> 00:17:42.859 rules, and your strict writing style. Keep it 00:17:42.859 --> 00:17:45.019 incredibly focused. Demand that your outputs 00:17:45.019 --> 00:17:48.519 are short and verified. And above all, do not 00:17:48.519 --> 00:17:51.839 grant the AI access to unnecessary files. Keep 00:17:51.839 --> 00:17:53.839 it strictly contained to its workspace so it 00:17:53.839 --> 00:17:56.640 doesn't get confused by your personal data. Let 00:17:56.640 --> 00:17:59.319 the machine work for you, but only on the exact 00:17:59.319 --> 00:18:01.660 terms you define. And I'll leave you with a final 00:18:01.660 --> 00:18:03.980 thought to mull over. We just spent this time 00:18:03.980 --> 00:18:06.640 walking through how to build an AI agent that 00:18:06.640 --> 00:18:10.019 can independently read industry blogs, rank them 00:18:10.019 --> 00:18:12.039 by importance based on your custom criteria, 00:18:12.480 --> 00:18:15.160 and write a human sounding newsletter draft while 00:18:15.160 --> 00:18:17.980 you are asleep. Yeah. So ask yourself what other 00:18:17.980 --> 00:18:20.519 repetitive daily digital chores in your life 00:18:20.519 --> 00:18:22.700 or your business are just waiting to be turned 00:18:22.700 --> 00:18:24.900 into a scheduled co -work folder. What a question 00:18:24.900 --> 00:18:27.839 to leave on. The potential is massive. Thank 00:18:27.839 --> 00:18:29.440 you so much for joining us for this deep dive 00:18:29.440 --> 00:18:31.779 into building your digital employee. We will 00:18:31.779 --> 00:18:32.359 see you next time.