WEBVTT 00:00:00.000 --> 00:00:02.660 You build a specialized AI assistant. It answers 00:00:02.660 --> 00:00:04.960 with total confidence. It uses the exact right 00:00:04.960 --> 00:00:08.740 vocabulary. And yet, it completely misses the 00:00:08.740 --> 00:00:12.320 point. Yeah, it's incredibly frustrating. Welcome 00:00:12.320 --> 00:00:15.140 to the Deep Dive. Today, we're unpacking a March 00:00:15.140 --> 00:00:19.339 2026 guide by Max Anna. It's titled Master AI 00:00:19.339 --> 00:00:21.899 Skill Building. We're going to explore the definitive 00:00:21.899 --> 00:00:24.739 protocol. It stops you from constantly re -explaining 00:00:24.739 --> 00:00:27.559 yourself to AI. We're fixing the confident but 00:00:27.559 --> 00:00:30.300 clueless chatbot problem forever. I'm so ready 00:00:30.300 --> 00:00:32.340 for this. We're looking at a highly specific 00:00:32.340 --> 00:00:35.539 workflow today. It uses Notebook LM's deep research 00:00:35.539 --> 00:00:38.420 to build specialized, reusable expert skills. 00:00:38.659 --> 00:00:41.280 And you build them directly inside Claude. Okay, 00:00:41.340 --> 00:00:43.880 let's unpack this. We need to diagnose the exact 00:00:43.880 --> 00:00:46.520 failure point first. Right, which usually happens 00:00:46.520 --> 00:00:48.200 right at the beginning. Exactly. The problem 00:00:48.200 --> 00:00:50.240 is usually the setup, not the models themselves. 00:00:50.659 --> 00:00:53.219 Whether it's Claude or ChatGPT, the models are 00:00:53.219 --> 00:00:55.280 powerful. People just give them incredibly... 00:00:55.310 --> 00:00:57.429 vague instructions. They kind of just toss a 00:00:57.429 --> 00:00:59.390 prompt in, you know, and hope for the best. They 00:00:59.390 --> 00:01:01.429 really do. They hope the AI gods fill in the 00:01:01.429 --> 00:01:03.469 gaps. They want the model to magically understand. 00:01:03.750 --> 00:01:06.450 The exact context you're working in. Right. Two 00:01:06.450 --> 00:01:09.430 seconds. I have to admit, I still wrestle with 00:01:09.430 --> 00:01:12.090 vague prompts myself. I just hope the AI gods 00:01:12.090 --> 00:01:14.810 figure it out. Oh, we all do it. It's human nature 00:01:14.810 --> 00:01:16.709 to want a shortcut. But that's why we get those 00:01:16.709 --> 00:01:19.069 highly polished answers. Answers that completely 00:01:19.069 --> 00:01:22.430 fall apart under close inspection. Why does relying 00:01:22.430 --> 00:01:25.170 on the AI to fill the gaps create such a fragile 00:01:25.170 --> 00:01:28.189 result? Well, without an actual knowledge base, 00:01:28.269 --> 00:01:33.480 the AI is just guessing. It relies on broad statistical 00:01:33.480 --> 00:01:35.739 patterns. It wants to give you a complete plausible 00:01:35.739 --> 00:01:38.959 sentence immediately. So it prioritizes sounding 00:01:38.959 --> 00:01:41.280 right over being accurate. So without a knowledge 00:01:41.280 --> 00:01:43.560 base, it prioritizes sounding confident over 00:01:43.560 --> 00:01:46.260 actually being right. Exactly. And if we connect 00:01:46.260 --> 00:01:48.099 this to the bigger picture, things must change. 00:01:48.459 --> 00:01:50.859 We have to shift our mindset entirely. We have 00:01:50.859 --> 00:01:52.859 to move from writing tromps to building actual 00:01:52.859 --> 00:01:54.959 infrastructure. You're building a specialized, 00:01:55.200 --> 00:01:57.879 reusable expert skill, not just a one -off prompt. 00:01:58.120 --> 00:02:00.579 The guide concludes with a great rule. Build 00:02:00.579 --> 00:02:02.859 infrastructure, not just prompts. That's a really 00:02:02.859 --> 00:02:05.519 powerful way to frame it. Building this infrastructure 00:02:05.519 --> 00:02:09.340 is like stacking Lego blocks of data. You create 00:02:09.340 --> 00:02:12.180 a solid foundation. Rather than just asking the 00:02:12.180 --> 00:02:14.659 AI to magically build a house from scratch. Right, 00:02:14.699 --> 00:02:16.879 because a basic prompt is incredibly fragile. 00:02:17.039 --> 00:02:19.919 It lives and dies in a single browser window. 00:02:20.259 --> 00:02:22.360 What is the functional difference between a prompt 00:02:22.360 --> 00:02:25.240 and an infrastructure skill? A prompt is merely 00:02:25.240 --> 00:02:28.520 a temporary request for information. Infrastructure 00:02:28.520 --> 00:02:31.439 is a reusable system. It's grounded in specific 00:02:31.439 --> 00:02:34.860 curated context that stops the AI from guessing. 00:02:35.080 --> 00:02:37.319 Prompts are temporary wishes. Infrastructure 00:02:37.319 --> 00:02:39.819 is a permanent grounded system that prevents 00:02:39.819 --> 00:02:42.120 guessing. Spot on. You're removing the guesswork 00:02:42.120 --> 00:02:44.560 entirely. And to do that, you need the right 00:02:44.560 --> 00:02:46.180 tools. Here's where it gets really interesting. 00:02:46.669 --> 00:02:48.990 To build that infrastructure, you need massive 00:02:48.990 --> 00:02:51.389 research capabilities. That brings us to the 00:02:51.389 --> 00:02:53.969 first step of the blueprint. Step one is picking 00:02:53.969 --> 00:02:57.090 one niche and one job. You absolutely have to 00:02:57.090 --> 00:02:59.030 keep the agent focused. Yeah, you can't ask it 00:02:59.030 --> 00:03:01.129 to do everything at once. Yeah. Then you move 00:03:01.129 --> 00:03:04.169 into step two, running notebook LM deep research. 00:03:04.430 --> 00:03:07.830 Right. And deep research is having AI thoroughly 00:03:07.830 --> 00:03:10.449 analyze documents to build an actual knowledge 00:03:10.449 --> 00:03:14.639 base. Whoa. Imagine having an AI completely map 00:03:14.639 --> 00:03:16.699 out an entire knowledge base before it even starts 00:03:16.699 --> 00:03:19.080 guessing. It changes everything. It really does. 00:03:19.379 --> 00:03:22.159 But I want to pause on that first step. Why is 00:03:22.159 --> 00:03:25.460 it so critical to limit this to just one niche 00:03:25.460 --> 00:03:29.099 and one job? Because broadening the scope dilutes 00:03:29.099 --> 00:03:31.300 the agent's expertise. When it loses focus, it 00:03:31.300 --> 00:03:34.039 gets confused. That leads right back to the confident 00:03:34.039 --> 00:03:36.400 but shallow answers we're trying to avoid. Keep 00:03:36.400 --> 00:03:39.080 it narrow. Or the agent goes right back to producing 00:03:39.080 --> 00:03:42.379 shallow, confident guesses. Exactly. You have 00:03:42.379 --> 00:03:44.620 to aggressively protect the boundaries of the 00:03:44.620 --> 00:03:47.419 skill. Once Notebook LM gathers the data, we 00:03:47.419 --> 00:03:49.319 have to refine it. We have to transfer it to 00:03:49.319 --> 00:03:51.620 Claude to actually build the skill. What's fascinating 00:03:51.620 --> 00:03:54.159 here is step three. The text has an explicit 00:03:54.159 --> 00:03:56.819 warning. You have to synthesize the research. 00:03:56.879 --> 00:03:59.400 Do not just summarize it. What is the distinction 00:03:59.400 --> 00:04:02.139 Max -Anne makes between synthesizing and merely 00:04:02.139 --> 00:04:04.819 summarizing? Summarizing just shortens the text. 00:04:04.900 --> 00:04:07.009 It cuts things out. But synthesizing connects 00:04:07.009 --> 00:04:09.349 the core concepts into a cohesive framework, 00:04:09.550 --> 00:04:12.069 a framework that Claude can actually apply to 00:04:12.069 --> 00:04:14.789 new problems. Perfectly said. Then we get to 00:04:14.789 --> 00:04:17.629 steps four and five, moving the synthesized research 00:04:17.629 --> 00:04:20.670 into Claude to build the skill and the crucial 00:04:20.670 --> 00:04:23.329 step of letting Claude ask clarifying questions. 00:04:23.430 --> 00:04:25.589 Right. You literally tell Claude to interrogate 00:04:25.589 --> 00:04:28.310 you. You have it ask you five questions to calibrate 00:04:28.310 --> 00:04:30.649 its understanding. Yeah. It ensures Claude truly 00:04:30.649 --> 00:04:34.189 grasps the job because humans often don't know 00:04:34.189 --> 00:04:36.990 what the AI doesn't know. So what does this all 00:04:36.990 --> 00:04:39.209 mean? Even with the perfect handoff, the skill 00:04:39.209 --> 00:04:42.110 isn't finished. It has to survive contact with 00:04:42.110 --> 00:04:44.670 reality. Oh, absolutely. That is step six and 00:04:44.670 --> 00:04:47.430 seven, reviewing the skill and testing it on 00:04:47.430 --> 00:04:49.569 a real scenario. You have to look out for common 00:04:49.569 --> 00:04:52.129 mistakes, things that reduce the skill quality. 00:04:52.350 --> 00:04:54.569 This raises an important question, though. How 00:04:54.569 --> 00:04:57.170 do users actively identify when the skill quality 00:04:57.170 --> 00:04:59.769 is degrading? Yeah, because models can experience 00:04:59.769 --> 00:05:02.939 prompt drift over time. they slowly revert back 00:05:02.939 --> 00:05:04.980 to their generalized training. Right, they stop 00:05:04.980 --> 00:05:07.439 referencing your meticulously crafted documents. 00:05:07.819 --> 00:05:10.540 Why is testing on a real scenario the ultimate 00:05:10.540 --> 00:05:12.920 fail -safe for this workflow? Because artificial 00:05:12.920 --> 00:05:15.959 tests are, you know, usually too clean. They 00:05:15.959 --> 00:05:18.600 don't expose the subtle ways an AI might revert 00:05:18.600 --> 00:05:21.439 to its general training. Real scenarios reveal 00:05:21.439 --> 00:05:24.279 the actual gaps in the knowledge base. Real scenarios 00:05:24.279 --> 00:05:27.259 expose when the AI secretly abandons your research 00:05:27.259 --> 00:05:30.199 to use general training. Yes. Real data is messy. 00:05:30.339 --> 00:05:33.019 It forces the system to prove it's actually using 00:05:33.019 --> 00:05:35.259 the infrastructure you built. Before we wrap 00:05:35.259 --> 00:05:36.560 up, we're going to take a quick break. We'll 00:05:36.560 --> 00:05:39.019 be right back. This deep dive is brought to you 00:05:39.019 --> 00:05:42.660 by Aura VPN. In a world where data is constantly 00:05:42.660 --> 00:05:46.189 being scrubbed, you need solid protection. Aura 00:05:46.189 --> 00:05:48.750 VPN creates a secure, encrypted tunnel for your 00:05:48.750 --> 00:05:51.310 daily browsing. It stops trackers dead in their 00:05:51.310 --> 00:05:53.529 tracks. I use it every single day. It's incredibly 00:05:53.529 --> 00:05:55.730 fast and just runs quietly in the background. 00:05:55.949 --> 00:05:58.470 Protect your digital life today with Aura VPN. 00:05:58.850 --> 00:06:01.120 Okay, we're back. We've covered a massive amount 00:06:01.120 --> 00:06:03.079 of ground today. We really have. This workflow 00:06:03.079 --> 00:06:05.680 is a total game changer. Let's recap the big 00:06:05.680 --> 00:06:08.620 idea here. We have to stop endlessly re -explaining 00:06:08.620 --> 00:06:11.600 ourselves to AI. We must replace vague prompts 00:06:11.600 --> 00:06:14.500 with a highly structured workflow. Right, using 00:06:14.500 --> 00:06:17.519 Notebook LM for deep research synthesis. And 00:06:17.519 --> 00:06:20.579 using Claude for focused skill building. By doing 00:06:20.579 --> 00:06:23.339 that, you create a reliable agent grounded in 00:06:23.339 --> 00:06:27.259 reality. Beat. I want to leave you with one final 00:06:27.259 --> 00:06:29.639 provocative thought today. Way at odds. If our 00:06:29.639 --> 00:06:32.399 AI assistants start relying entirely on these 00:06:32.399 --> 00:06:35.540 deeply researched, custom curated knowledge bases 00:06:35.540 --> 00:06:38.439 instead of their default training, how will that 00:06:38.439 --> 00:06:40.860 change the way you value human curation versus 00:06:40.860 --> 00:06:43.720 raw AI generation? That is a fascinating question 00:06:43.720 --> 00:06:45.759 to sit with. Human curation might become the 00:06:45.759 --> 00:06:48.300 ultimate premium skill. I encourage you to try 00:06:48.300 --> 00:06:50.379 this blueprint today. Pick one niche and one 00:06:50.379 --> 00:06:52.980 job to test it out. Thank you for joining us 00:06:52.980 --> 00:06:54.579 on this deep dive. We'll catch you on the next 00:06:54.579 --> 00:06:54.699 one.