WEBVTT 00:00:00.000 --> 00:00:03.779 There is this strange modern trap that we all 00:00:03.779 --> 00:00:05.900 kind of fall into. Oh, the viral prompt trap. 00:00:06.099 --> 00:00:08.839 Yeah, exactly. You find this viral master prompt 00:00:08.839 --> 00:00:12.699 online. You copy it, paste it into an AI. And 00:00:12.699 --> 00:00:15.300 for like exactly five minutes, you feel like 00:00:15.300 --> 00:00:17.399 an absolute genius. Right, because it looks so 00:00:17.399 --> 00:00:19.699 complex. But then you read the output. I mean, 00:00:19.699 --> 00:00:21.879 it's grammatically perfect. It's highly structured. 00:00:22.519 --> 00:00:24.980 But you realize it sounds absolutely nothing 00:00:24.980 --> 00:00:27.379 like you. Yeah, it sounds like a machine doing 00:00:27.379 --> 00:00:29.719 a really cheap impression of a marketer. It's 00:00:29.719 --> 00:00:32.899 the ultimate trap, right? We're effectively outsourcing 00:00:32.899 --> 00:00:34.939 our thinking. We use a template built for someone 00:00:34.939 --> 00:00:37.179 else's brain, and then we wonder why the result 00:00:37.179 --> 00:00:40.380 feels so hollow. Well, welcome to today's Deep 00:00:40.380 --> 00:00:43.359 Dive. We are taking a completely different approach 00:00:43.359 --> 00:00:45.600 to AI today. A much better approach, honestly. 00:00:45.659 --> 00:00:48.259 I think so. The mission is simple, but pretty 00:00:48.259 --> 00:00:51.299 ambitious. We want to move past those generic 00:00:51.299 --> 00:00:54.700 plastic sounding outputs. So we're going to deconstruct 00:00:54.700 --> 00:00:57.359 how to build a personalized, highly specific 00:00:57.359 --> 00:01:00.350 master prompt. from the ground up. I am really 00:01:00.350 --> 00:01:02.310 excited to get into this. If you follow this 00:01:02.310 --> 00:01:04.969 process, it genuinely becomes a build once use 00:01:04.969 --> 00:01:07.760 forever system. It completely solves the biggest 00:01:07.760 --> 00:01:10.560 headache in modern content creation, which is 00:01:10.560 --> 00:01:13.560 just endlessly staring at a blinking cursor, 00:01:13.840 --> 00:01:16.780 trying to make the AI sound human. It really 00:01:16.780 --> 00:01:19.180 is a massive headache, and I'll be completely 00:01:19.180 --> 00:01:21.480 vulnerable here. I mean, I still wrestle with 00:01:21.480 --> 00:01:23.780 prompt drift myself. Oh, we all do. Even with 00:01:23.780 --> 00:01:25.439 custom instructions, I'll be working on a script, 00:01:25.459 --> 00:01:27.519 and suddenly the AI just kind of wanders off 00:01:27.519 --> 00:01:30.480 into this generic internet speak. Using words 00:01:30.480 --> 00:01:33.579 like delve or game changer. Right. So to fix 00:01:33.579 --> 00:01:35.680 this, we need to completely rethink think our 00:01:35.680 --> 00:01:38.159 tools. We all think we know what a master prompt 00:01:38.159 --> 00:01:40.439 is. Usually it's just a giant block of text we 00:01:40.439 --> 00:01:43.599 bought from some guru. But how would you actually 00:01:43.599 --> 00:01:45.920 define it, practically? I'll give you the absolute 00:01:45.920 --> 00:01:49.159 tightest definition. A master prompt is a reusable 00:01:49.159 --> 00:01:51.959 template that guides AI consistently. So it's 00:01:51.959 --> 00:01:54.719 not a magic spell. It's an architecture. Exactly. 00:01:55.099 --> 00:01:58.159 A strong master prompt has a really rigid six 00:01:58.159 --> 00:02:00.819 -part architecture. If you miss even one of these 00:02:00.819 --> 00:02:03.439 parts. The AI loses the plot. It just starts 00:02:03.439 --> 00:02:05.579 making things up. Yeah, it starts guessing. And 00:02:05.579 --> 00:02:08.599 when an AI guesses, it defaults to the mathematical 00:02:08.599 --> 00:02:11.120 average of the internet, which is boring. So 00:02:11.120 --> 00:02:13.539 let's build this architecture. What is the absolute 00:02:13.539 --> 00:02:16.240 foundation, the very first piece? Part one is 00:02:16.240 --> 00:02:19.280 the role. You have to define the specific specialist 00:02:19.280 --> 00:02:22.060 the AI is supposed to be. OK. Without a role, 00:02:22.280 --> 00:02:25.479 the model just defaults to a polite, hyper -generic 00:02:25.479 --> 00:02:28.659 assistant. But wait, I've tried giving it a persona 00:02:28.659 --> 00:02:31.490 before? I'll say, you know, act like a professional 00:02:31.490 --> 00:02:34.689 writer and the output still sounds like a caricature. 00:02:35.050 --> 00:02:37.889 It gets overly dramatic. Because professional 00:02:37.889 --> 00:02:40.330 writer isn't a specific role. I mean, it's way 00:02:40.330 --> 00:02:43.990 too broad. A weak role... creates a weak output. 00:02:44.629 --> 00:02:47.550 A strong role gives the AI a very specific worldview. 00:02:48.129 --> 00:02:50.889 So instead, you tell it, you are a YouTube script 00:02:50.889 --> 00:02:53.530 writer. You specialize in tutorial hooks for 00:02:53.530 --> 00:02:55.889 non -technical audiences. And these audiences 00:02:55.889 --> 00:02:58.629 are highly skeptical of big flashy promises. 00:02:58.930 --> 00:03:00.870 Oh, wow. That's fascinating. You're not just 00:03:00.870 --> 00:03:02.930 giving it a job title. You're giving it a philosophy. 00:03:03.189 --> 00:03:04.930 Right. You're giving it real boundaries. But 00:03:04.930 --> 00:03:06.990 even if it knows exactly who it is, it's still 00:03:06.990 --> 00:03:08.669 sort of shouting into a void if it doesn't know 00:03:08.669 --> 00:03:10.509 who's listening. Yeah. Which I assume brings 00:03:10.509 --> 00:03:13.960 us to part two. It does. Part two is context. 00:03:14.280 --> 00:03:17.340 This is where you describe the audience's mindset. 00:03:17.620 --> 00:03:19.800 And this is where most people fail, actually. 00:03:20.000 --> 00:03:22.060 Because they just list demographics. Exactly. 00:03:22.139 --> 00:03:24.800 They say, my audience is millennials who live 00:03:24.800 --> 00:03:28.219 in cities, which gives the AI almost nothing 00:03:28.219 --> 00:03:31.020 to work with emotionally. True context is about 00:03:31.020 --> 00:03:34.159 mindset. You need to list their distrusts. You 00:03:34.159 --> 00:03:36.979 need to list their actual desires. How does that 00:03:36.979 --> 00:03:38.900 look in practice, though? You'd say something 00:03:38.900 --> 00:03:41.849 like... You're writing for solo business owners 00:03:41.849 --> 00:03:45.810 age 30 to 50. They have tried AI tools before 00:03:45.810 --> 00:03:48.389 and were deeply disappointed. That's very specific. 00:03:48.530 --> 00:03:51.210 Yeah. And you add, they are extremely short on 00:03:51.210 --> 00:03:53.810 time. They want a specific actionable result, 00:03:54.370 --> 00:03:56.949 not a vague promise. That really grounds the 00:03:56.949 --> 00:03:59.289 whole interaction. No fluff allowed because the 00:03:59.289 --> 00:04:01.210 audience just won't tolerate it. So we have the 00:04:01.210 --> 00:04:03.449 role. We have the context. What's the third pillar? 00:04:03.629 --> 00:04:05.750 Part three is the objective. You have to state 00:04:05.750 --> 00:04:08.669 the exact format, the exact length, and the precise 00:04:08.669 --> 00:04:11.310 purpose. You never just say, like, help me write 00:04:11.310 --> 00:04:13.750 a video intro. No, never. Because it'll give 00:04:13.750 --> 00:04:16.870 you a three -page essay. Exactly. You say, write 00:04:16.870 --> 00:04:19.189 the opening hook for a 10 -minute YouTube tutorial. 00:04:19.610 --> 00:04:22.990 It must be spoken in under 30 seconds. You leave 00:04:22.990 --> 00:04:25.709 absolutely zero room for the AI to guess the 00:04:25.709 --> 00:04:27.670 container it's filling. Which leads perfectly 00:04:27.670 --> 00:04:30.769 into part four. And the source material highlights 00:04:30.769 --> 00:04:33.680 this as the most critical step. constraints. 00:04:34.240 --> 00:04:37.480 This tells the AI what to actively avoid. Yes. 00:04:37.800 --> 00:04:40.399 You have to block the bad patterns explicitly. 00:04:41.040 --> 00:04:42.759 Language models are basically prediction engines, 00:04:42.939 --> 00:04:45.600 right? They predict the next most likely word. 00:04:46.100 --> 00:04:47.720 If you don't build a fence, they will wander 00:04:47.720 --> 00:04:50.680 into the most highly probable and therefore overused 00:04:50.680 --> 00:04:53.319 vocabulary. I have a theory about this actually. 00:04:53.920 --> 00:04:56.540 Copying someone else's master prompt. without 00:04:56.540 --> 00:04:59.579 understanding these constraints is it's like 00:04:59.579 --> 00:05:01.319 wearing someone else's prescription glasses. 00:05:01.480 --> 00:05:02.939 Oh, that's a good way to put it. It's perfectly 00:05:02.939 --> 00:05:05.379 focused. It's just not focused for your eyes. 00:05:05.519 --> 00:05:07.560 You're adopting their boundaries, not your own. 00:05:07.980 --> 00:05:10.040 But let me ask you this. Why are constraints 00:05:10.040 --> 00:05:12.420 the part that most people completely skip, yet 00:05:12.420 --> 00:05:14.399 they're the most vital? Because without them, 00:05:14.819 --> 00:05:17.379 the model defaults to those hyped up words like 00:05:17.379 --> 00:05:21.639 game changer or revolutionary. Explicitly blocking 00:05:21.639 --> 00:05:24.720 bad patterns forces good ones to appear naturally. 00:05:24.990 --> 00:05:28.310 So telling the AI what not to do is the real 00:05:28.310 --> 00:05:30.490 secret. Absolutely. It clears out the weeds. 00:05:30.649 --> 00:05:34.250 When you forbid the AI from using generic buzzwords, 00:05:34.430 --> 00:05:37.589 it actually has to think. computationally, I 00:05:37.589 --> 00:05:40.970 mean, to find a more creative, precise way to 00:05:40.970 --> 00:05:43.410 express the idea. So we've blocked the bad habits. 00:05:43.470 --> 00:05:45.610 What is part five? Part five is examples. And 00:05:45.610 --> 00:05:47.569 this is your biggest lever for quality control. 00:05:48.230 --> 00:05:50.850 People think they need to feed an AI 50 examples 00:05:50.850 --> 00:05:53.069 for it to learn. Yeah, they just dump data into 00:05:53.069 --> 00:05:55.410 it. You really don't. Quality matters so much 00:05:55.410 --> 00:05:58.170 more than volume here. Two incredibly strong, 00:05:58.430 --> 00:06:00.589 perfectly calibrated examples are all you need. 00:06:00.949 --> 00:06:02.889 We'll dig into exactly how to find those examples 00:06:02.889 --> 00:06:04.589 in a minute. But let's finish the architecture. 00:06:04.829 --> 00:06:07.769 What's the sixth? and final piece. Part six is 00:06:07.769 --> 00:06:10.529 the output format. You tell the AI exactly how 00:06:10.529 --> 00:06:12.370 you want to read its response. If you don't do 00:06:12.370 --> 00:06:14.569 this, it will just dump a huge wall of text on 00:06:14.569 --> 00:06:16.370 you. Right. It gives you that standard bulleted 00:06:16.370 --> 00:06:19.410 list with emojis that screams, I use chat GPT. 00:06:19.560 --> 00:06:22.040 Exactly. So you control the format. You say, 00:06:22.439 --> 00:06:25.139 provide three distinct hook options. For each 00:06:25.139 --> 00:06:27.860 option, give me the full script. Below the script, 00:06:28.160 --> 00:06:30.680 identify the specific emotional trigger you used. 00:06:31.100 --> 00:06:33.560 And finally, state the specific promise made 00:06:33.560 --> 00:06:35.980 to the viewer. That is brilliant. It forces the 00:06:35.980 --> 00:06:39.100 AI to show its work. Like, it has to justify 00:06:39.100 --> 00:06:41.959 its own writing against the psychology of the 00:06:41.959 --> 00:06:43.899 audience. And it makes evaluating the output 00:06:43.899 --> 00:06:46.569 so much faster for you. Okay, so we have this 00:06:46.569 --> 00:06:50.509 beautiful pristine framework role context objective 00:06:50.509 --> 00:06:53.610 constraints examples output format But right 00:06:53.610 --> 00:06:56.350 now it's just an empty shell a framework needs 00:06:56.350 --> 00:06:59.189 data to actually run How do we fill this empty 00:06:59.189 --> 00:07:01.649 architecture with reality? This is where we introduce 00:07:01.649 --> 00:07:04.189 the tool stack and the best part is it's completely 00:07:04.189 --> 00:07:06.730 free The core engine we're gonna use is notebook 00:07:06.730 --> 00:07:09.589 LM We'll use Google Docs just to store our text 00:07:09.589 --> 00:07:12.209 and we'll use a platform like medium for psychological 00:07:12.209 --> 00:07:14.980 theory I want to stop right there Why notebook 00:07:14.980 --> 00:07:17.000 LM specifically? I think a lot of people listening 00:07:17.000 --> 00:07:18.740 are probably wondering, why not just build this 00:07:18.740 --> 00:07:21.199 in chat GPT or Claude? Why learn a whole new 00:07:21.199 --> 00:07:24.379 interface? Because of how notebook LM is engineered, 00:07:25.079 --> 00:07:28.519 it operates as a closed system. It uses a technique 00:07:28.519 --> 00:07:31.250 called source grounding. Meaning what? Exactly. 00:07:31.389 --> 00:07:33.370 When you ask it a question, it doesn't search 00:07:33.370 --> 00:07:37.529 its entire massive generic training data. It 00:07:37.529 --> 00:07:40.129 only reads the specific documents you upload 00:07:40.129 --> 00:07:42.810 to it. So it literally cannot wander off into 00:07:42.810 --> 00:07:45.370 the broader internet. Exactly. It completely 00:07:45.370 --> 00:07:48.589 eliminates that generic internet drift. It is 00:07:48.589 --> 00:07:52.129 legally bound, so to speak, to only use the reality 00:07:52.129 --> 00:07:55.009 that you provide. That is a really crucial distinction. 00:07:55.449 --> 00:07:57.589 OK, so we need to provide it with reality. How 00:07:57.589 --> 00:08:00.029 do we actually gather this data? Walk me through 00:08:00.029 --> 00:08:02.769 step one. Step one is manual data collection. 00:08:02.970 --> 00:08:04.370 You have to get your hands a little dirty here. 00:08:04.509 --> 00:08:06.870 You go to YouTube. You search for your specific 00:08:06.870 --> 00:08:09.009 niche. Let's say your niche is email productivity 00:08:09.009 --> 00:08:11.329 tips. OK, I search that. I get a million results. 00:08:11.490 --> 00:08:13.750 Right. So you filter those results by this year. 00:08:13.930 --> 00:08:16.529 So the data is fresh. Then you sort them by you 00:08:16.529 --> 00:08:19.829 count. But here is the critical part. You are 00:08:19.829 --> 00:08:21.689 not just looking for the biggest channels. What 00:08:21.689 --> 00:08:24.230 are we looking for? You are hunting for outliers. 00:08:24.730 --> 00:08:27.389 Define an outlier in this context. An outlier 00:08:27.389 --> 00:08:30.230 is a video that vastly outperformed its channel's 00:08:30.230 --> 00:08:33.769 baseline average. Imagine a channel with 2 ,000 00:08:33.769 --> 00:08:36.610 subscribers. Usually they get 500 views a video. 00:08:37.289 --> 00:08:40.399 But suddenly one video has 50 ,000 views. That 00:08:40.399 --> 00:08:42.580 means the algorithm didn't push it based on channel 00:08:42.580 --> 00:08:44.799 reputation. Yeah. It pushed it purely because 00:08:44.799 --> 00:08:47.379 the thumbnail and the hook were mathematically 00:08:47.379 --> 00:08:50.139 undeniable. Spot on. That specific hook broke 00:08:50.139 --> 00:08:53.519 through the noise. It is pure concentrated signal, 00:08:53.799 --> 00:08:56.940 and it's deeply worth studying. So you find 10 00:08:56.940 --> 00:08:59.679 to 15 of these true outliers. And what's the 00:08:59.679 --> 00:09:02.039 actual process of capturing them? It's very analog. 00:09:02.179 --> 00:09:04.860 You literally watch the first 30 seconds of each 00:09:04.860 --> 00:09:07.620 outlier video, you transcribe the spoken hook, 00:09:07.639 --> 00:09:10.220 and you paste that text into to a Google Doc 00:09:10.220 --> 00:09:12.500 right next to the video's title. So now I have 00:09:12.500 --> 00:09:15.440 a document filled with the raw proven text of 00:09:15.440 --> 00:09:18.100 what actually works in my niche. Yes. But here's 00:09:18.100 --> 00:09:20.460 the crucial twist that makes the system brilliant. 00:09:20.860 --> 00:09:23.480 You don't stop at the YouTube hooks. You open 00:09:23.480 --> 00:09:26.299 a new tab, go to Medium or a similar platform. 00:09:26.580 --> 00:09:28.919 You find two or three highly rated articles on 00:09:28.919 --> 00:09:31.360 human psychology or advanced copywriting principles. 00:09:31.679 --> 00:09:33.960 Like articles about the curiosity gap or how 00:09:33.960 --> 00:09:36.200 fear of missing out works in the brain. Exactly. 00:09:36.220 --> 00:09:39.200 You grab those URLs. So now you have your raw 00:09:39.200 --> 00:09:42.379 data. Step two is preparation. You open a new 00:09:42.379 --> 00:09:45.000 notebook in Notebook LM. You upload your Google 00:09:45.000 --> 00:09:47.500 Doc full of those 15 successful successful hooks, 00:09:48.019 --> 00:09:50.700 you paste in the URLs for the psychology articles, 00:09:50.960 --> 00:09:54.120 and finally you paste in the blank text of that 00:09:54.120 --> 00:09:57.019 six -part master prompt framework we just went 00:09:57.019 --> 00:09:59.419 over. I really love this approach. It anchors 00:09:59.419 --> 00:10:01.940 the language model entirely in empirical evidence. 00:10:02.100 --> 00:10:04.620 It's not sitting there guessing what might work 00:10:04.620 --> 00:10:07.899 based on old forum posts. It's being fed proven 00:10:07.899 --> 00:10:10.519 reality. Exactly. But I want to clarify something 00:10:10.519 --> 00:10:13.080 here. If we already have the successful YouTube 00:10:13.080 --> 00:10:15.320 hooks, which is the raw proof of what gets views, 00:10:15.909 --> 00:10:18.429 Why do we need to upload the psychology articles, 00:10:18.450 --> 00:10:21.210 too? Because the hooks show the AI what worked, 00:10:21.309 --> 00:10:23.730 but the psychology articles teach the AI why 00:10:23.730 --> 00:10:25.769 it works, giving it the theoretical layer. It 00:10:25.769 --> 00:10:27.929 gives the AI the underlying theory, not just 00:10:27.929 --> 00:10:31.529 the raw data. Exactly. It maps the fast -paced 00:10:31.529 --> 00:10:35.409 action of a YouTube hook to the deep, slow psychology 00:10:35.409 --> 00:10:38.450 of human behavior. It learns that a specific 00:10:38.450 --> 00:10:41.389 pause in a sentence isn't just a pause, it's 00:10:41.389 --> 00:10:47.759 a tension -building mechanism. Okay, so let's 00:10:47.759 --> 00:10:50.360 visualize this. We have our raw materials locked 00:10:50.360 --> 00:10:53.639 inside Notebook LM. We have 10 to 15 outlier 00:10:53.639 --> 00:10:56.059 hooks. We have deep psychology articles. We have 00:10:56.059 --> 00:10:58.620 this six -part framework. It's a massive wall 00:10:58.620 --> 00:11:01.259 of text. It is a lot of text. How do we synthesize 00:11:01.259 --> 00:11:03.740 all this without spending an entire week analyzing 00:11:03.740 --> 00:11:05.799 it ourselves? This is where the magic really 00:11:05.799 --> 00:11:09.120 happens. Step three is pattern analysis. We give 00:11:09.120 --> 00:11:12.000 Notebook LM a highly specific prompt. We command 00:11:12.000 --> 00:11:14.320 it to read all the uploaded sources and cross 00:11:14.320 --> 00:11:16.509 -reference them. What exactly are we instructing 00:11:16.509 --> 00:11:18.830 it to extract. We tell it to identify the five 00:11:18.830 --> 00:11:21.009 most common patterns across all those successful 00:11:21.009 --> 00:11:23.509 hooks. We want it to analyze the underlying sentence 00:11:23.509 --> 00:11:26.289 structure. We want it to identify the exact emotional 00:11:26.289 --> 00:11:28.789 triggers being used and link them to the psychology 00:11:28.789 --> 00:11:31.230 articles. Like recognizing when a creator is 00:11:31.230 --> 00:11:33.529 agitating a specific pain point before introducing 00:11:33.529 --> 00:11:36.600 the solution. Exactly. We also ask it to analyze 00:11:36.600 --> 00:11:38.980 the literal rhythm of the words. How long are 00:11:38.980 --> 00:11:41.480 the sentences? Where do they pause? We ask for 00:11:41.480 --> 00:11:43.639 a deeply analytical breakdown of the mechanics. 00:11:43.919 --> 00:11:46.039 That's incredible. It's essentially reverse engineering 00:11:46.039 --> 00:11:49.799 the charisma of a viral video. It is. And then... 00:11:49.850 --> 00:11:52.789 Based purely on those extracted patterns, we 00:11:52.789 --> 00:11:55.769 ask it to draft five strict constraints to help 00:11:55.769 --> 00:11:58.269 us avoid sounding hyped or generic. And finally, 00:11:58.370 --> 00:12:01.190 we ask it to pull the two absolute best examples 00:12:01.190 --> 00:12:03.690 from our raw data that perfectly embodied these 00:12:03.690 --> 00:12:06.570 successful patterns. Wow. Whoa. Just imagine 00:12:06.570 --> 00:12:09.610 scaling to a billion queries or just extracting 00:12:09.610 --> 00:12:12.330 the exact DNA of viral hooks in under a minute. 00:12:12.710 --> 00:12:14.909 Doing that by hand would take half a day. Right. 00:12:15.090 --> 00:12:16.970 I mean, you'd be staring at transcripts forever. 00:12:17.289 --> 00:12:19.370 You are distilling pure signal from the noise. 00:12:19.210 --> 00:12:22.250 here. Yeah. So notebook LM spits out this incredible 00:12:22.250 --> 00:12:24.610 pattern analysis. It gives us the patterns, the 00:12:24.610 --> 00:12:26.669 constraints, and the two perfect examples. What 00:12:26.669 --> 00:12:30.149 is step four? Step four is assembly, generating 00:12:30.149 --> 00:12:33.289 the actual master prompt. You immediately run 00:12:33.289 --> 00:12:36.169 a second prompt in notebook LM. You say, now 00:12:36.169 --> 00:12:38.470 use the patterns, the constraints, and the two 00:12:38.470 --> 00:12:41.250 examples you just identified. Fill out a complete 00:12:41.250 --> 00:12:43.590 master prompt using our six -part framework. 00:12:43.889 --> 00:12:47.029 So it acts as its own architect. It writes out 00:12:47.029 --> 00:12:49.559 the role. It defines the context based on what 00:12:49.559 --> 00:12:51.220 it learned about the audience from the hooks. 00:12:51.419 --> 00:12:54.240 Right. And it locks in the objective, lists the 00:12:54.240 --> 00:12:56.840 new constraints, inserts the examples, and sets 00:12:56.840 --> 00:12:59.120 up the output format. It writes the whole thing 00:12:59.120 --> 00:13:01.970 for you. That's amazing. It gives you a finalized, 00:13:02.049 --> 00:13:04.169 ready -to -use prompt that you can just copy 00:13:04.169 --> 00:13:06.289 and paste directly into your daily driver, whether 00:13:06.289 --> 00:13:08.990 that's Claude or ChatGPT. It's like stacking 00:13:08.990 --> 00:13:11.730 Lego blocks of data. You're letting the AI build 00:13:11.730 --> 00:13:13.710 a machine that will eventually build your content. 00:13:14.009 --> 00:13:16.470 And because it's built inside Notebook LM, the 00:13:16.470 --> 00:13:18.409 constraints are infinitely sharper because they 00:13:18.409 --> 00:13:21.269 come from hard data, not just our vague guesses 00:13:21.269 --> 00:13:23.460 about what sounds good. And the examples are 00:13:23.460 --> 00:13:25.419 bulletproof. They weren't just pulled from the 00:13:25.419 --> 00:13:27.899 ether. They were isolated as the mathematical 00:13:27.899 --> 00:13:30.360 best performers from your own curated research. 00:13:30.820 --> 00:13:32.740 I do have a question about that curation, though, 00:13:33.200 --> 00:13:35.840 about the examples. What happens if we get a 00:13:35.840 --> 00:13:37.879 bit lazy during the manual collection phase? 00:13:38.360 --> 00:13:40.799 Say we just feed it five average hooks instead 00:13:40.799 --> 00:13:43.399 of hunting down those two incredibly strong outliers. 00:13:43.639 --> 00:13:46.899 That will ruin the entire system. The AI learns 00:13:46.899 --> 00:13:49.340 from the patterns it's given. If you feed it 00:13:49.340 --> 00:13:52.820 mediocre examples, it will perfectly consistently 00:13:52.820 --> 00:13:56.000 replicate mediocrity. Quality is a far bigger 00:13:56.000 --> 00:13:58.659 liver than volume. Right. Garbage in means garbage 00:13:58.659 --> 00:14:02.019 out. So always choose absolute quality. Always. 00:14:02.179 --> 00:14:03.860 You want to feed it the gold standard because 00:14:03.860 --> 00:14:06.240 that becomes its new baseline. OK. So let's say 00:14:06.240 --> 00:14:08.840 we did the work. We have this beautifully crafted 00:14:08.840 --> 00:14:12.200 data -backed master prompt sitting in our clipboard. 00:14:12.830 --> 00:14:14.409 But I have to play the skeptic for a minute. 00:14:15.009 --> 00:14:17.690 Theory is great. It looks amazing on paper. But 00:14:17.690 --> 00:14:19.789 does it actually survive contact with the real 00:14:19.789 --> 00:14:22.830 world? That brings us to step five, the crucible, 00:14:23.330 --> 00:14:25.850 the test. You take your shiny new master prompt 00:14:25.850 --> 00:14:28.470 out of Notebook LM. You open your primary AI, 00:14:28.529 --> 00:14:31.169 let's say Chai GPT. You paste the master prompt 00:14:31.169 --> 00:14:33.730 in. And what do we ask it to generate? Do we 00:14:33.730 --> 00:14:35.750 ask it for an email productivity hook, since 00:14:35.750 --> 00:14:38.899 that was our research? No. That is the one crucial 00:14:38.899 --> 00:14:41.539 rule of the testing phase. You must test the 00:14:41.539 --> 00:14:43.779 prompt on a completely new topic, something that 00:14:43.779 --> 00:14:45.740 was absolutely not mentioned in your original 00:14:45.740 --> 00:14:48.539 source data. Wait, why? Wouldn't you want to 00:14:48.539 --> 00:14:51.279 test it on what it knows? Because if you test 00:14:51.279 --> 00:14:54.000 it on the same topic, you don't know if the AI 00:14:54.000 --> 00:14:56.919 is relying on the structural framework or if 00:14:56.919 --> 00:14:59.279 it's just regurgitating the specific facts it 00:14:59.279 --> 00:15:01.740 just read. You have to isolate the variable. 00:15:01.899 --> 00:15:03.740 OK, that makes perfect sense. Yeah. So give me 00:15:03.740 --> 00:15:05.740 an example of a good stress test. Let's stick 00:15:05.740 --> 00:15:08.789 with our scenario. All your source data was about 00:15:08.789 --> 00:15:11.529 email productivity tips. For the test, you give 00:15:11.529 --> 00:15:14.350 it the topic. Write a YouTube hook about using 00:15:14.350 --> 00:15:17.149 the Notion app to manage freelance client projects. 00:15:17.429 --> 00:15:19.370 A completely different software, different workload, 00:15:19.590 --> 00:15:22.110 a different target user. Exactly. You run the 00:15:22.110 --> 00:15:24.669 prompt, and then you aggressively audit the output. 00:15:25.070 --> 00:15:27.789 You look for three specific things. First, does 00:15:27.789 --> 00:15:30.730 the tone match those strict constraints we set? 00:15:31.019 --> 00:15:33.740 Second, does the structure actually follow the 00:15:33.740 --> 00:15:35.860 psychological patterns notebook LM identified? 00:15:36.440 --> 00:15:38.639 And third, and most importantly, does it sound 00:15:38.639 --> 00:15:41.500 like a real breathing human being? Rather than 00:15:41.500 --> 00:15:45.019 an AI trying to sound human. I mean, we all know 00:15:45.019 --> 00:15:48.559 that slightly too enthusiastic AI voice. We do. 00:15:49.320 --> 00:15:52.799 If the output sounds too formal, or if it slips 00:15:52.799 --> 00:15:56.559 in a dreaded word like powerful or crucial, you 00:15:56.559 --> 00:15:59.100 know your framework has a leak. I actually want 00:15:59.100 --> 00:16:01.100 to push back slightly here on how we fix those 00:16:01.100 --> 00:16:04.100 leaks. Measuring constraints is notoriously difficult. 00:16:04.580 --> 00:16:07.419 If the AI sounds too formal, a lot of people 00:16:07.419 --> 00:16:09.399 will just add a constraint that says don't be 00:16:09.399 --> 00:16:11.740 overly formal. But the source material has a 00:16:11.740 --> 00:16:14.519 great tip about this. Constraints must be strictly 00:16:14.519 --> 00:16:17.230 measurable. Walk me through what you mean by 00:16:17.230 --> 00:16:19.870 measurable. Well, you cannot give an AI a subjective 00:16:19.870 --> 00:16:22.070 command. Don't be wordy is a vague constraint. 00:16:22.570 --> 00:16:25.029 Vague constraints yield vague outputs because 00:16:25.029 --> 00:16:27.070 the AI doesn't know your personal definition 00:16:27.070 --> 00:16:28.750 of wordy. You have to give it a mathematical 00:16:28.750 --> 00:16:31.110 boundary. Ah, right. You have to say word cap 00:16:31.110 --> 00:16:35.409 of exactly 75 words or absolutely no sentences 00:16:35.409 --> 00:16:38.830 longer than 15 words. That is a brilliant clarification. 00:16:39.450 --> 00:16:42.250 The AI needs a hard boundary to bounce off of. 00:16:42.269 --> 00:16:44.759 So if it fails your test, You don't throw the 00:16:44.759 --> 00:16:47.259 prompt away. You go back to Notebook LM, you 00:16:47.259 --> 00:16:50.679 say, the output was too wordy. Update the constraint 00:16:50.679 --> 00:16:53.419 section to include a heart maximum of 12 words 00:16:53.419 --> 00:16:56.779 per sentence. Usually, one round of structural 00:16:56.779 --> 00:16:59.820 feedback completely fixes the leak. And once 00:16:59.820 --> 00:17:02.159 it finally passes the test, once you get that 00:17:02.159 --> 00:17:05.960 perfect, punchy, human -sounding hook. That is 00:17:05.960 --> 00:17:09.089 step six. You reuse it forever. You save that 00:17:09.089 --> 00:17:11.609 finalized text to a prompt library, keep it in 00:17:11.609 --> 00:17:13.390 Google Doc or a Notion dashboard, wherever you 00:17:13.390 --> 00:17:16.390 work, and you use this exact same workflow for 00:17:16.390 --> 00:17:18.869 everything you do. Gathering outliers, extracting 00:17:18.869 --> 00:17:21.289 patterns, building a framework. Exactly. You 00:17:21.289 --> 00:17:23.730 could build one for YouTube thumbnails, one for 00:17:23.730 --> 00:17:26.390 email subject lines, one for landing page headlines. 00:17:26.609 --> 00:17:29.710 One build infinite reuses. Next week, when you 00:17:29.710 --> 00:17:31.349 need a new hook for a video, you don't sit there 00:17:31.349 --> 00:17:33.230 staring at a blank screen. No, you don't start 00:17:33.230 --> 00:17:34.970 from scratch at all. You just grab your master 00:17:34.970 --> 00:17:37.269 prompt, paste it in, add your new topic, and 00:17:37.269 --> 00:17:39.470 you get high -quality, perfectly calibrated output 00:17:39.470 --> 00:17:42.049 in about five minutes. But how do we know for 00:17:42.049 --> 00:17:45.690 absolutely sure that this master prompt is genuinely 00:17:45.690 --> 00:17:48.849 a reusable system and not just a one -trick pony 00:17:48.849 --> 00:17:51.420 that got lucky once? By testing it on a topic 00:17:51.420 --> 00:17:54.619 that is wildly outside your usual range, if the 00:17:54.619 --> 00:17:57.740 tone and quality hold up without the AI hallucinating 00:17:57.740 --> 00:18:00.740 or breaking character, the framework is mathematically 00:18:00.740 --> 00:18:03.779 sound. If it works perfectly on a totally unrelated 00:18:03.779 --> 00:18:07.240 topic, your system is solid. It proves the architecture 00:18:07.240 --> 00:18:09.859 works independently of the subject matter. It 00:18:09.859 --> 00:18:12.160 proves you've captured the physics of good communication. 00:18:12.480 --> 00:18:14.420 Let's take a step back and look at the big picture 00:18:14.420 --> 00:18:17.619 here. The journey we just went on is deeply fascinating 00:18:17.619 --> 00:18:20.220 to me. We started with the frustration of copying 00:18:20.220 --> 00:18:23.079 other people's templates. And what we discovered 00:18:23.079 --> 00:18:25.819 is that the best prompt isn't found. It isn't 00:18:25.819 --> 00:18:28.519 a secret code you copy from a viral post. It 00:18:28.519 --> 00:18:31.180 is built. Built from the ground up, with intention. 00:18:31.299 --> 00:18:34.359 Yes. It is built from existing, proven evidence 00:18:34.359 --> 00:18:37.339 in your specific niche. It's constructed strictly 00:18:37.339 --> 00:18:39.900 around your specific audience's fears, their 00:18:39.900 --> 00:18:42.420 deep distrusts, and their actual desires. It 00:18:42.420 --> 00:18:44.420 fundamentally changes our relationship with the 00:18:44.420 --> 00:18:46.859 technology. It really does. It transforms AI 00:18:46.859 --> 00:18:49.500 from this generic, unpredictable slot machine 00:18:49.500 --> 00:18:51.680 where you pull the lever and just sort of pray 00:18:51.680 --> 00:18:55.740 for a good result into a reusable, highly calibrated 00:18:55.740 --> 00:18:58.000 engine. An engine that actually respects your 00:18:58.000 --> 00:18:59.880 time and your intelligence. It stops being a 00:18:59.880 --> 00:19:01.799 parlor trick and becomes a reliable piece of 00:19:01.799 --> 00:19:03.900 daily infrastructure. Which leaves me with a 00:19:03.900 --> 00:19:07.220 thought I cannot quite shake. It's a bit philosophical, 00:19:07.440 --> 00:19:10.200 but bear with me. If we can use this technology 00:19:10.200 --> 00:19:13.400 to so perfectly isolate, analyze, and automate 00:19:13.400 --> 00:19:17.420 the exact psychology of our most authentic human 00:19:17.420 --> 00:19:19.740 communication, what happens to our own voice 00:19:19.740 --> 00:19:22.920 in the long run? Does having a perfect digital 00:19:22.920 --> 00:19:25.799 clone, a framework that speaks exactly like us 00:19:25.799 --> 00:19:28.759 on our best day, force us to become even more 00:19:28.759 --> 00:19:31.000 deeply human, more unpredictable in our actual 00:19:31.000 --> 00:19:33.440 daily lives? It's something to think about as 00:19:33.440 --> 00:19:36.099 these tools get sharper. That is a profound question, 00:19:36.240 --> 00:19:37.940 and honestly, the best way to explore it is to 00:19:37.940 --> 00:19:39.680 get your hands dirty. While you ponder that, 00:19:39.799 --> 00:19:41.740 I highly recommend you take just 10 minutes today. 00:19:41.900 --> 00:19:44.099 Go to YouTube, search your niche, gather your 00:19:44.099 --> 00:19:46.380 first five outlier examples, run them through 00:19:46.380 --> 00:19:48.759 Notebook LM, test this out. See the mechanics 00:19:48.759 --> 00:19:51.660 for yourself. It is incredibly eye -opening once 00:19:51.660 --> 00:19:54.119 you see it work. Thank you for joining us on 00:19:54.119 --> 00:19:56.779 this deep dive. Take care, and we will catch 00:19:56.779 --> 00:19:57.259 you next time.