WEBVTT 00:00:00.000 --> 00:00:01.840 Imagine buying a state -of -the -art professional 00:00:01.840 --> 00:00:04.400 kitchen today. Oh, yeah. You spend an absolute 00:00:04.400 --> 00:00:07.059 fortune on the equipment, and then you only use 00:00:07.059 --> 00:00:09.400 it to toast bread. Right, which sounds completely 00:00:09.400 --> 00:00:11.960 ridiculous out loud. It really does. But honestly, 00:00:12.160 --> 00:00:15.380 that is exactly how we use AI today. Welcome 00:00:15.380 --> 00:00:18.600 to our deep dive. Glad to be here. We are unpacking 00:00:18.600 --> 00:00:21.260 a really fascinating article by Max Anne today. 00:00:21.359 --> 00:00:24.140 It covers a major shift happening in the industry 00:00:24.140 --> 00:00:26.839 right now. A massive shift, yeah. We are moving 00:00:26.839 --> 00:00:30.760 away from simple AI chat interfaces completely. 00:00:31.199 --> 00:00:34.259 We are heading toward a fully autonomous AI workforce 00:00:34.259 --> 00:00:36.619 instead. It is a totally different way of working. 00:00:36.759 --> 00:00:39.539 And we will explore four specific open source 00:00:39.539 --> 00:00:42.479 blueprints to get there. We are looking at superpowers, 00:00:42.719 --> 00:00:45.799 G -Stack, Hermes Agent. And paperclip. Okay, 00:00:45.859 --> 00:00:47.679 let's untack this. We really need to understand 00:00:47.679 --> 00:00:50.659 the baseline problem first. A regular AI assistant 00:00:50.659 --> 00:00:54.539 is purely reactive by its very nature. You ask 00:00:54.539 --> 00:00:57.039 a direct question and you get an immediate answer. 00:00:57.200 --> 00:00:59.740 Which works perfectly for simple, highly isolated 00:00:59.740 --> 00:01:02.299 daily tasks. You just want a quick recipe or 00:01:02.299 --> 00:01:05.379 maybe a code snippet. It is fast. It is undeniably 00:01:05.379 --> 00:01:09.000 very useful for those tiny micro tasks. Yeah, 00:01:09.000 --> 00:01:12.180 but real meaningful work is rarely that incredibly 00:01:12.180 --> 00:01:14.840 simple. Building a software feature requires 00:01:14.840 --> 00:01:18.459 intense planning and rigorous testing. Exactly. 00:01:18.700 --> 00:01:21.620 It requires maintaining memory across many different 00:01:21.620 --> 00:01:23.719 sequential work sessions. And this is exactly 00:01:23.719 --> 00:01:26.780 where single chat AI falls apart completely. 00:01:27.439 --> 00:01:30.599 The entire illusion of intelligence just breaks 00:01:30.599 --> 00:01:32.900 down under structural pressure. I always compare 00:01:32.900 --> 00:01:35.439 it to a talented but forgetful intern. Oh, that 00:01:35.439 --> 00:01:37.599 is a great way to frame it. They might be brilliant 00:01:37.599 --> 00:01:40.439 at writing a single block of code. But they lack 00:01:40.439 --> 00:01:42.879 the architectural scaffolding to manage their 00:01:42.879 --> 00:01:46.040 own memory. Right. They need exact, precise instructions 00:01:46.040 --> 00:01:49.060 for every single subsequent step. If we connect 00:01:49.060 --> 00:01:51.560 this to the bigger picture, the next massive 00:01:51.560 --> 00:01:54.340 productivity jump is a completely different paradigm 00:01:54.340 --> 00:01:57.349 entirely. It is not about building one marginally 00:01:57.349 --> 00:02:00.209 smarter foundational model. It is about coordinating 00:02:00.209 --> 00:02:02.909 multiple capabilities with a very rigid structure. 00:02:03.170 --> 00:02:05.450 I mean, I get that it forgets things over time, 00:02:05.569 --> 00:02:08.430 but why does the single chat model break down 00:02:08.430 --> 00:02:11.270 so quickly on big projects? Well, it comes down 00:02:11.270 --> 00:02:14.169 to how context windows actually operate mechanically. 00:02:14.229 --> 00:02:17.479 Okay. An isolated chat window. just cannot hold 00:02:17.479 --> 00:02:20.240 overarching project goals. It predicts the next 00:02:20.240 --> 00:02:22.719 word based on the immediate previous text. So 00:02:22.719 --> 00:02:25.680 it just reacts blindly to the very last prompt 00:02:25.680 --> 00:02:28.979 you typed. Exactly. It has no true long -term 00:02:28.979 --> 00:02:31.740 plan. Right. Complex work needs memory and multi 00:02:31.740 --> 00:02:34.479 -step processes, not just instant answers. Beat. 00:02:34.800 --> 00:02:37.219 We need a fundamental shift in our daily approach. 00:02:37.560 --> 00:02:40.060 We really need to build guardrails around how 00:02:40.060 --> 00:02:42.479 the model operates. Which brings us to fixing 00:02:42.479 --> 00:02:45.550 the sloppy code generation first. But how do 00:02:45.550 --> 00:02:48.110 we actually force that discipline onto an LLM? 00:02:48.229 --> 00:02:51.090 This brings us directly to a project called Superpowers. 00:02:51.289 --> 00:02:53.949 Right. It was created by a very clever developer 00:02:53.949 --> 00:02:57.669 who goes by Obra. Superpowers adds real software 00:02:57.669 --> 00:03:00.349 engineering discipline directly to Claude Code. 00:03:00.569 --> 00:03:02.870 I still wrestle with prompt drift myself when 00:03:02.870 --> 00:03:05.750 asking AI for code to Sex Island. Yeah, it happens 00:03:05.750 --> 00:03:07.590 constantly. It just kind of wanders off the main 00:03:07.590 --> 00:03:09.969 architectural path entirely. We all do. That 00:03:09.969 --> 00:03:11.810 is the fundamental nature of the probabilistic 00:03:11.810 --> 00:03:14.879 piece. AI coding agents are incredibly fast right 00:03:14.879 --> 00:03:17.939 now, but that raw speed often brings hidden, 00:03:18.060 --> 00:03:21.000 deeply frustrating structural errors. Because 00:03:21.000 --> 00:03:23.680 they skip crucial structural tests just to save 00:03:23.680 --> 00:03:26.159 a little time. Exactly. They introduce weird, 00:03:26.319 --> 00:03:29.360 convoluted shortcuts that inevitably break things 00:03:29.360 --> 00:03:31.580 later on. They produce code that works perfectly 00:03:31.580 --> 00:03:34.219 fine for today, but it becomes totally impossible 00:03:34.219 --> 00:03:37.460 for a human to maintain by tomorrow. But superpowers 00:03:37.460 --> 00:03:40.259 actually forces the AI to slow down completely. 00:03:40.539 --> 00:03:43.879 It really does. It demands a totally clean, isolated 00:03:43.879 --> 00:03:47.219 workspace before writing anything. It uses Git 00:03:47.219 --> 00:03:49.319 work trees to keep the main environment safe. 00:03:49.500 --> 00:03:51.680 Which are isolated folders for safely testing 00:03:51.680 --> 00:03:54.000 different code changes. Right. It forces the 00:03:54.000 --> 00:03:56.599 AI to brainstorm before writing actual logic. 00:03:56.840 --> 00:03:59.539 Then it writes a highly detailed step -by -step 00:03:59.539 --> 00:04:02.340 implementation plan. The documentation explains 00:04:02.340 --> 00:04:05.800 this core goal in a very direct way. The plan 00:04:05.930 --> 00:04:08.289 must be incredibly detailed and logically robust. 00:04:08.590 --> 00:04:11.449 Even an inexperienced human engineer could follow 00:04:11.449 --> 00:04:13.449 it correctly. It essentially removes all the 00:04:13.449 --> 00:04:16.269 dangerous, hallucinatory guesswork. It enforces 00:04:16.269 --> 00:04:19.149 a strict test -driven development cycle on the 00:04:19.149 --> 00:04:21.819 model. Right. Writing automated tests before 00:04:21.819 --> 00:04:24.560 writing the actual software code. Exactly. Humans 00:04:24.560 --> 00:04:26.480 honestly hate doing test -driven development 00:04:26.480 --> 00:04:28.939 because it feels slow. Oh, absolutely. But an 00:04:28.939 --> 00:04:31.439 AI doesn't have the capacity to feel bored. It 00:04:31.439 --> 00:04:33.939 will happily write a failing test and then fix 00:04:33.939 --> 00:04:36.399 it. And then it explicitly requests a thorough 00:04:36.399 --> 00:04:39.379 code review. It forces the branch to finish properly 00:04:39.379 --> 00:04:41.779 and cleanly. Right. It prevents the model from 00:04:41.779 --> 00:04:44.519 just abandoning a half -finished thought. I get 00:04:44.519 --> 00:04:47.319 that slowing it down sounds good in theory, but... 00:04:47.680 --> 00:04:49.579 aren't we just stripping away its main speed 00:04:49.579 --> 00:04:53.060 advantage? Why is slowing the AI down actually 00:04:53.060 --> 00:04:56.500 an upgrade here? Because raw speed often introduces 00:04:56.500 --> 00:04:59.379 hidden vulnerabilities and sloppy logical shortcuts. 00:04:59.579 --> 00:05:02.459 Oh, I see. Slowing down forces the AI to formally 00:05:02.459 --> 00:05:05.540 verify its own logic first. It makes the final 00:05:05.540 --> 00:05:07.920 output actually trustworthy and structurally 00:05:07.920 --> 00:05:10.079 maintainable. Got it. Forcing a pause prevents 00:05:10.079 --> 00:05:13.060 sloppy, hard -to -maintain code later on. Beat. 00:05:13.600 --> 00:05:16.579 So, we have highly disciplined coding mechanics 00:05:16.579 --> 00:05:19.560 established now. The model is writing clean, 00:05:19.819 --> 00:05:22.889 tested, and verifiable software code. But perfectly 00:05:22.889 --> 00:05:25.870 written code for a terrible idea is still terrible. 00:05:26.149 --> 00:05:29.089 How do we evaluate the actual ideas being coded? 00:05:29.370 --> 00:05:31.670 Well, we have to give the AI different hats to 00:05:31.670 --> 00:05:33.730 wear. And this is where GStack enters the conversation 00:05:33.730 --> 00:05:36.410 perfectly. It addresses the exact blind spot 00:05:36.410 --> 00:05:38.350 you just mentioned. Yeah, it was created by Y 00:05:38.350 --> 00:05:41.180 Combinator president Gary Tan. VStack gives the 00:05:41.180 --> 00:05:44.399 AI a roster of highly specific roles. You are 00:05:44.399 --> 00:05:47.420 building a one -person startup team, essentially. 00:05:47.680 --> 00:05:50.480 Your AI gets a bunch of different highly specialized 00:05:50.480 --> 00:05:53.839 job titles. We're talking CEO, engineering manager, 00:05:54.100 --> 00:05:57.120 and lead designer. It also acts as a QA lead 00:05:57.120 --> 00:05:59.660 and security officer. It follows a very strict 00:05:59.660 --> 00:06:03.120 and methodical sprint flow. Think, plan, build, 00:06:03.399 --> 00:06:07.240 review, test, ship, and finally reflect. What's 00:06:07.240 --> 00:06:09.180 fascinating here is the underlying mathematics 00:06:09.180 --> 00:06:12.250 of the prompt. When you assign a specific persona, 00:06:12.490 --> 00:06:15.389 you shift token probabilities. A general agent 00:06:15.389 --> 00:06:18.209 gives loose, generalized, and very broad responses. 00:06:18.430 --> 00:06:21.209 A role -based setup forces highly focused specific 00:06:21.209 --> 00:06:24.029 domain decisions. Because each role looks at 00:06:24.029 --> 00:06:25.889 the exact same problem from a different angle. 00:06:26.149 --> 00:06:28.430 The security officer agent strictly looks for 00:06:28.430 --> 00:06:31.430 potential injection flaws. The designer agent 00:06:31.430 --> 00:06:33.790 only cares about user experience and interaction 00:06:33.790 --> 00:06:37.310 flow. It stops the AI from rushing into an immediate 00:06:37.310 --> 00:06:40.600 generic implementation. It forces it to pause 00:06:40.600 --> 00:06:43.240 and evaluate from multiple perspectives. Inside 00:06:43.240 --> 00:06:46.100 Cloud Code, you access this through simple slash 00:06:46.100 --> 00:06:49.120 commands. You just type slash C stack or slash 00:06:49.120 --> 00:06:52.279 office hours to trigger the roles. It guides 00:06:52.279 --> 00:06:55.120 the AI through deeply structured analytical steps. 00:06:55.300 --> 00:06:57.879 It prevents the model from trying to be everything 00:06:57.879 --> 00:07:00.529 at once. I have to offer a massive warning right 00:07:00.529 --> 00:07:02.790 here, though. Do not ask it to build an entire 00:07:02.790 --> 00:07:05.430 app on day one. Well, definitely not. It is exactly 00:07:05.430 --> 00:07:07.910 like hiring a brand new project manager. You 00:07:07.910 --> 00:07:09.949 cannot demand a shipped product on their first 00:07:09.949 --> 00:07:13.290 afternoon. That usually leads to incredibly messy 00:07:13.290 --> 00:07:16.310 and logically broken results. Yeah. The context 00:07:16.310 --> 00:07:18.509 window gets overwhelmed by too many competing 00:07:18.509 --> 00:07:22.029 priorities. So start small. Evaluate a tiny market 00:07:22.029 --> 00:07:24.470 opportunity or reframe a basic product idea. 00:07:24.709 --> 00:07:27.410 Get a design review or run a simple debugging 00:07:27.410 --> 00:07:30.759 workflow. Does assigning a fake CEO title actually 00:07:30.759 --> 00:07:33.680 change the underlying code the AI writes? Yes, 00:07:33.800 --> 00:07:36.160 because it radically changes the mathematical 00:07:36.160 --> 00:07:39.759 context of the prompt. Okay. The AI evaluates 00:07:39.759 --> 00:07:42.180 the coding problem through a completely different 00:07:42.180 --> 00:07:45.459 semantic lens. It prioritizes different logical 00:07:45.459 --> 00:07:48.939 metrics based on that specific assigned persona. 00:07:49.240 --> 00:07:52.199 Makes sense. Different roles force the AI to 00:07:52.199 --> 00:07:55.100 catch its own blind spots. Beat. We have discipline 00:07:55.100 --> 00:07:57.709 coding. And we have structured, specialized roles. 00:07:57.949 --> 00:08:00.790 But there is still a massive missing piece to 00:08:00.790 --> 00:08:03.329 this puzzle. Right. Mid -role sponsor read. We 00:08:03.329 --> 00:08:04.850 need to talk about what happens when you close 00:08:04.850 --> 00:08:07.430 your laptop. Usually the AI just forgets everything 00:08:07.430 --> 00:08:09.730 you just did together. The context window wipes 00:08:09.730 --> 00:08:12.189 clean and you start from absolute zero. Let's 00:08:12.189 --> 00:08:14.529 fix that. This brings us to a project called 00:08:14.529 --> 00:08:17.170 Hermes Agent. It was built by the brilliant team 00:08:17.170 --> 00:08:19.730 over at Noose Research. Hermes is a completely 00:08:19.730 --> 00:08:23.170 self -improving AI agent framework. Most AI tools 00:08:23.170 --> 00:08:25.500 start... totally fresh every single time you 00:08:25.500 --> 00:08:28.040 open them. Hermes is trying to be the exact opposite 00:08:28.040 --> 00:08:30.220 of that temporary paradigm. It creates specific 00:08:30.220 --> 00:08:32.860 skills from its past interactions with you. Yeah, 00:08:32.919 --> 00:08:34.519 it searches through all your past conversations 00:08:34.519 --> 00:08:37.480 seamlessly using vector databases. Which builds 00:08:37.480 --> 00:08:40.299 a much better long -term picture of how you actually 00:08:40.299 --> 00:08:43.679 work. It features a truly unified messaging gateway 00:08:43.679 --> 00:08:45.940 as well. You can connect it directly to Telegram 00:08:45.940 --> 00:08:50.019 or your company Slack. You can use Discord, WhatsApp, 00:08:50.240 --> 00:08:52.929 Signal. or the standard command line so you do 00:08:52.929 --> 00:08:55.389 not have to constantly switch between five different 00:08:55.389 --> 00:08:58.330 browser windows you interact with the exact same 00:08:58.330 --> 00:09:01.759 intelligent agent everywhere you go oh imagine 00:09:01.759 --> 00:09:04.279 it's scaling to remember your specific workflow 00:09:04.279 --> 00:09:07.500 across every single app you use it changes the 00:09:07.500 --> 00:09:09.240 entire fundamental relationship you have with 00:09:09.240 --> 00:09:12.039 the machine yeah you are no longer asking highly 00:09:12.039 --> 00:09:15.019 isolated randomly generated questions you are 00:09:15.019 --> 00:09:17.480 developing a persistent system that organically 00:09:17.480 --> 00:09:19.779 evolves alongside you but we need to talk about 00:09:19.779 --> 00:09:22.299 the actual onboarding reality here we have to 00:09:22.299 --> 00:09:24.759 issue a quick very serious security warning first 00:09:24.759 --> 00:09:27.240 definitely it is always worth reviewing open 00:09:27.240 --> 00:09:29.860 source installation scripts very You must do 00:09:29.860 --> 00:09:31.940 this before piping them directly into your local 00:09:31.940 --> 00:09:34.200 shell. Taking a moment to inspect the script 00:09:34.200 --> 00:09:37.340 reduces unnecessary local risk. Because you are 00:09:37.340 --> 00:09:39.860 giving an autonomous agent access to your file 00:09:39.860 --> 00:09:42.480 system? And you really need to be incredibly 00:09:42.480 --> 00:09:46.000 patient with this specific tool. Do not expect 00:09:46.000 --> 00:09:49.299 magically strong personalized results on the 00:09:49.299 --> 00:09:52.399 very first day. It behaves way more like onboarding 00:09:52.399 --> 00:09:54.720 a real human teammate. You have to correct it 00:09:54.720 --> 00:09:56.860 when it makes a logical mistake. It is absolutely 00:09:56.860 --> 00:09:58.840 not an instant feature toggle you just casually 00:09:58.840 --> 00:10:01.039 flip. People always expect these memory tools 00:10:01.039 --> 00:10:04.200 to be instantly telepathic. How long does it 00:10:04.200 --> 00:10:06.919 realistically take for this self -improving loop 00:10:06.919 --> 00:10:10.059 to actually feel useful? It usually takes several 00:10:10.059 --> 00:10:13.299 weeks of consistent daily interaction to calibrate. 00:10:13.320 --> 00:10:16.179 Wow, weeks. Yeah. The system needs enough varied 00:10:16.179 --> 00:10:18.700 data to understand your specific workflow patterns. 00:10:18.879 --> 00:10:21.899 It cannot learn your deep, nuanced habits from 00:10:21.899 --> 00:10:24.759 just two short conversations. So it's true onboarding. 00:10:24.759 --> 00:10:27.259 It builds rich context through real use over 00:10:27.259 --> 00:10:30.200 time. Beat. So we have disciplined role -playing 00:10:30.200 --> 00:10:32.740 agents equipped with persistent cross -platform 00:10:32.740 --> 00:10:35.179 memory. Let's put them all in one digital building 00:10:35.179 --> 00:10:37.100 and track the budget. We're talking about Paperclip 00:10:37.100 --> 00:10:39.059 now. Right. It is easily the most experimental 00:10:39.059 --> 00:10:41.559 and fragile project on this list. It orchestrates 00:10:41.559 --> 00:10:44.080 multiple specialized agents as if they were a 00:10:44.080 --> 00:10:51.179 re - Yeah. It gives you an actual visual dashboard 00:10:51.179 --> 00:10:56.399 for everything. Exactly. Inside this interface, 00:10:56.659 --> 00:10:59.899 agents take on major corporate roles. You have 00:10:59.899 --> 00:11:04.179 a CEO. a cmo and the cto actively working together 00:11:04.179 --> 00:11:06.879 it includes a full organizational chart and a 00:11:06.879 --> 00:11:09.360 functional ticketing system it has strict governance 00:11:09.360 --> 00:11:12.399 controls and granular financial budget tracking 00:11:12.399 --> 00:11:16.139 you can see exactly how much each specific generative 00:11:16.139 --> 00:11:19.259 task actually costs which is crucial here's where 00:11:19.259 --> 00:11:21.700 it gets really interesting i look at all this 00:11:21.700 --> 00:11:24.539 corporate scaffolding and i have to ask Is this 00:11:24.539 --> 00:11:27.559 just corporate process theater applied to AI? 00:11:27.899 --> 00:11:30.980 That is a very fair and deeply necessary critical 00:11:30.980 --> 00:11:33.240 question. I have to give the honest warning straight 00:11:33.240 --> 00:11:35.539 from the source text. Okay. This is absolutely 00:11:35.539 --> 00:11:37.960 not a magic money machine at all. You will not 00:11:37.960 --> 00:11:40.179 wake up to sudden unexpected surprise profits. 00:11:40.639 --> 00:11:43.960 The project is still incredibly rough and highly 00:11:43.960 --> 00:11:46.799 experimental software. It is meant for those 00:11:46.799 --> 00:11:50.000 genuinely curious about multi -agent orchestration 00:11:50.000 --> 00:11:53.330 at scale. Its true purpose is high -level coordination, 00:11:53.669 --> 00:11:55.929 not just blind, reckless automation. Because 00:11:55.929 --> 00:11:59.509 as soon as you use multiple agents, massive structural 00:11:59.509 --> 00:12:02.549 problems appear, responsibilities blur quickly, 00:12:02.870 --> 00:12:05.809 and agents get stuck in infinite conversational 00:12:05.809 --> 00:12:08.269 loops. And when they get stuck in loops, EPI 00:12:08.269 --> 00:12:11.210 costs quietly skyrocket. Oh, they will happily... 00:12:11.690 --> 00:12:13.950 Burn through your open AI credits arguing with 00:12:13.950 --> 00:12:16.169 each other. Paperclip attempts to bring visible 00:12:16.169 --> 00:12:18.909 structure to that massive underlying complexity. 00:12:19.409 --> 00:12:22.009 It places everything in one single, unified, 00:12:22.110 --> 00:12:25.110 and highly trackable interface. Why do we even 00:12:25.110 --> 00:12:27.809 need a whole dashboard just to manage AI agents? 00:12:27.850 --> 00:12:30.110 Because multiple agents working simultaneously 00:12:30.110 --> 00:12:33.509 create massive, unreadable chaos very quickly. 00:12:33.669 --> 00:12:36.320 Right. Goals drift rapidly without central oversight 00:12:36.320 --> 00:12:39.500 and very clear financial boundaries. A dashboard 00:12:39.500 --> 00:12:42.320 makes the invisible chaotic work of agents visible 00:12:42.320 --> 00:12:45.240 and trackable. Exactly. It brings visible coordination 00:12:45.240 --> 00:12:49.399 and budget tracking to multi -agent chaos. Beat. 00:12:49.559 --> 00:12:51.620 We should probably synthesize the ultimate pattern 00:12:51.620 --> 00:12:53.879 behind all these different tools. Let us step 00:12:53.879 --> 00:12:57.179 back and look at the actual big picture. We just 00:12:57.179 --> 00:12:59.879 covered four radically different open source 00:12:59.879 --> 00:13:02.500 software projects today. But they all share the 00:13:02.500 --> 00:13:05.240 exact same underlying philosophical approach. 00:13:05.419 --> 00:13:08.580 Every single one makes the exact same major technological 00:13:08.580 --> 00:13:11.519 bet. The future of AI is not a race for the smartest 00:13:11.519 --> 00:13:14.200 individual model. Right. It is about building 00:13:14.200 --> 00:13:16.820 a much better structure around how models work. 00:13:17.179 --> 00:13:19.879 G -Stack gives your general models incredibly 00:13:19.879 --> 00:13:23.039 clear, structurally defined roles. Hermes gives 00:13:23.039 --> 00:13:25.539 them deep. persistent, and highly accessible 00:13:25.539 --> 00:13:28.460 cross -platform memory. Superpowers gives them 00:13:28.460 --> 00:13:30.899 reliable, highly disciplined software engineering 00:13:30.899 --> 00:13:33.379 process. And Paperclip gives them a massive, 00:13:33.460 --> 00:13:35.740 visible organizational management structure. 00:13:36.059 --> 00:13:38.419 The industry focus is shifting away from raw, 00:13:38.600 --> 00:13:41.179 generalized benchmark intelligence. It is moving 00:13:41.179 --> 00:13:43.700 toward consistent, highly predictable behavior 00:13:43.700 --> 00:13:46.600 in real -world use. That shift might feel slightly 00:13:46.600 --> 00:13:48.879 less dramatic than a viral benchmark screenshot, 00:13:49.240 --> 00:13:52.159 but it is way more important for anyone actually 00:13:52.159 --> 00:13:54.080 building real things. We should probably help 00:13:54.080 --> 00:13:55.779 people figure out where to actually start. I 00:13:55.779 --> 00:13:57.700 want to explicitly lay out the recommended order 00:13:57.700 --> 00:13:59.940 of operations here. Definitely. I do not want 00:13:59.940 --> 00:14:02.000 you to lose your entire weekend to confusing 00:14:02.000 --> 00:14:05.700 documentation. Do not try to install all four 00:14:05.700 --> 00:14:08.580 on a Saturday afternoon. That is a truly great 00:14:08.580 --> 00:14:11.480 way to end up with absolutely nothing useful. 00:14:11.799 --> 00:14:13.740 First, you need to start with the superpowers 00:14:13.740 --> 00:14:16.419 framework. Do this if you already use Claude 00:14:16.419 --> 00:14:19.419 code on a daily basis. It is the absolute fastest 00:14:19.419 --> 00:14:22.720 way to see immediate practical coding value. 00:14:23.039 --> 00:14:25.120 The improvement to your daily development process 00:14:25.120 --> 00:14:28.379 is stark and visible right away. Second, you 00:14:28.379 --> 00:14:30.720 should try implementing the GStack role system. 00:14:31.100 --> 00:14:33.779 Add those structured, specialized roles to your 00:14:33.779 --> 00:14:36.820 newly disciplined coding process. It pairs naturally 00:14:36.820 --> 00:14:39.220 with superpowers once you understand the basic 00:14:39.220 --> 00:14:42.139 operational layers. Third, move on to the Hermes 00:14:42.139 --> 00:14:44.419 agent persistent framework. That gives you the 00:14:44.419 --> 00:14:46.639 powerful cross -platform memory you really need. 00:14:46.799 --> 00:14:48.899 It is absolutely perfect if you want an agent 00:14:48.899 --> 00:14:51.080 that runs scheduled background tasks. Finally, 00:14:51.159 --> 00:14:53.019 you can carefully explore the paperclip dashboard. 00:14:53.419 --> 00:14:56.120 Only do this if you are genuinely ready for experimental 00:14:56.120 --> 00:14:59.700 multi -agent orchestration. It is the most powerful 00:14:59.700 --> 00:15:02.220 conceptual framework, but definitely the roughest 00:15:02.220 --> 00:15:04.519 tool. Yeah, save it for last. Following this 00:15:04.519 --> 00:15:06.980 specific order helps you see practical value 00:15:06.980 --> 00:15:10.139 very quickly. You gradually understand exactly 00:15:10.139 --> 00:15:13.100 how each structural layer contributes to the 00:15:13.100 --> 00:15:15.419 whole. The future probably won't look like one 00:15:15.419 --> 00:15:18.179 singular genius chat assistant. No, it will look 00:15:18.179 --> 00:15:21.120 exactly like a small orchestra of highly specialized 00:15:21.120 --> 00:15:24.320 systems. Each has a very clear role and a rigidly 00:15:24.320 --> 00:15:26.659 defined process. They will smoothly hand off 00:15:26.659 --> 00:15:29.460 complex work. to the next digital player. The 00:15:29.460 --> 00:15:31.940 massive operational advantage goes to those who 00:15:31.940 --> 00:15:34.919 design these workflows early. If AI is rapidly 00:15:34.919 --> 00:15:37.080 becoming the orchestra of the modern workforce, 00:15:37.480 --> 00:15:39.740 what skills do you need to start developing today 00:15:39.740 --> 00:15:42.399 to become the conductor? That is a perfect, profound 00:15:42.399 --> 00:15:44.539 question to leave them with today. Thanks for 00:15:44.539 --> 00:15:47.200 joining us on this deep dive. Stay curious. Out 00:15:47.200 --> 00:15:47.820 to your own music.