WEBVTT 00:00:00.000 --> 00:00:03.580 So it's January 2026, and we're sitting on top 00:00:03.580 --> 00:00:06.400 of, you know, the most powerful computational 00:00:06.400 --> 00:00:09.539 engine in human history, GPT -5. The benchmarks 00:00:09.539 --> 00:00:12.000 are just, they're off the charts. They're absurd. 00:00:12.300 --> 00:00:15.720 And the reasoning, I mean, theoretically, it 00:00:15.720 --> 00:00:18.980 rivals human experts in almost every field. And 00:00:18.980 --> 00:00:21.500 yet, I was talking to a friend yesterday, a really 00:00:21.500 --> 00:00:25.219 smart guy, uses AI for coding, and he said something 00:00:25.219 --> 00:00:27.839 that just stuck with me. He said, I feel like 00:00:27.839 --> 00:00:30.699 I'm fighting it. He types a question and the 00:00:30.699 --> 00:00:33.969 answer he gets back feels. It feels lazy. It 00:00:33.969 --> 00:00:36.549 feels bored. It kind of feels cheap. Yeah. And 00:00:36.549 --> 00:00:39.070 it's this strange paradox where the model is 00:00:39.070 --> 00:00:41.590 smarter than ever, but the user experience feels 00:00:41.590 --> 00:00:44.229 like it's degrading. It's the GPT -5 paradox. 00:00:44.469 --> 00:00:46.409 You have this Ferrari engine, but the car is 00:00:46.409 --> 00:00:48.090 locked in first gear. And you're pressing the 00:00:48.090 --> 00:00:49.890 gas, but it just won't go. It just won't go. 00:00:50.030 --> 00:00:52.210 So today, we're going to figure out how to shift 00:00:52.210 --> 00:00:54.530 gears. We are deep diving into a guide called 00:00:54.530 --> 00:00:58.390 Five Advanced Chat GPT Tricks for GPT -5 Mastery. 00:00:58.409 --> 00:01:00.750 Right. But after reading through this, tricks... 00:01:01.259 --> 00:01:02.659 That feels like the wrong word. It's more like 00:01:02.659 --> 00:01:05.159 we're decoding a new psychological relationship 00:01:05.159 --> 00:01:07.980 between us and the machine. I think so, too. 00:01:08.099 --> 00:01:10.680 Because it turns out there's an invisible decision 00:01:10.680 --> 00:01:13.920 maker standing between us and all that intelligence. 00:01:14.079 --> 00:01:17.459 Right. The router. And if you don't get the router, 00:01:17.620 --> 00:01:20.319 you're essentially getting the discount version 00:01:20.319 --> 00:01:22.799 of the AI, no matter how much you're paying for 00:01:22.799 --> 00:01:25.760 that pro subscription. Welcome to the deep dive. 00:01:26.340 --> 00:01:28.439 Today, we're going to break down this idea of 00:01:28.439 --> 00:01:31.359 router influence. We'll look at the architecture 00:01:31.359 --> 00:01:34.519 of 2026, and then we'll walk through five really 00:01:34.519 --> 00:01:36.980 specific strategies. Things like trigger words, 00:01:37.239 --> 00:01:39.599 radical specificity. And something called self 00:01:39.599 --> 00:01:42.400 -reflection loops, which supposedly force this 00:01:42.400 --> 00:01:44.819 router to actually give us the intelligence we're 00:01:44.819 --> 00:01:47.159 asking for. What I love about this whole analysis 00:01:47.159 --> 00:01:49.739 is that it moves us away from prompt engineering 00:01:49.739 --> 00:01:52.540 as some kind of mystical art. Yeah. And it treats 00:01:52.540 --> 00:01:54.799 it more like system administration. It's just 00:01:54.799 --> 00:01:56.500 about understanding that there's a gatekeeper 00:01:56.500 --> 00:01:58.659 and you need the password. Let's linger on that 00:01:58.659 --> 00:02:00.599 gatekeeper for a second, the underlying architecture, 00:02:00.900 --> 00:02:02.739 because I think a lot of us still have a mental 00:02:02.739 --> 00:02:06.219 model from, you know, 2023 or 2024. Oh, for sure. 00:02:06.359 --> 00:02:11.180 So walk me back to the vintage era of AI. How 00:02:11.180 --> 00:02:13.240 did we used to interact with these things? Well, 00:02:13.300 --> 00:02:15.360 it was manual transmission. Yeah. Think back 00:02:15.360 --> 00:02:18.680 to ChatGPT in late 2024. Yeah. You had that little 00:02:18.680 --> 00:02:21.400 drop down menu at the top left. Right. You'd 00:02:21.400 --> 00:02:23.319 log in and you had to make a conscious executive 00:02:23.319 --> 00:02:26.419 decision. I am writing a poem, so I will select. 00:02:26.960 --> 00:02:30.259 GPT -40 or... I'm solving a complex physics problem, 00:02:30.360 --> 00:02:33.099 so I'll need a one preview. Exactly. You, the 00:02:33.099 --> 00:02:34.939 human, you were the load balancer. You decided 00:02:34.939 --> 00:02:37.120 how much horsepower to use. And there was a tangible 00:02:37.120 --> 00:02:39.520 difference. If I pick the big reasoning model, 00:02:39.680 --> 00:02:41.439 I knew I was going to stare at a spinning circle 00:02:41.439 --> 00:02:44.460 for 30 seconds. Yeah, you waited. But I knew 00:02:44.460 --> 00:02:46.379 I was thinking. I was basically buying depth 00:02:46.379 --> 00:02:49.360 with my time. Precisely. But here's the reality 00:02:49.360 --> 00:02:53.960 of 2026. Open AI. and really all the labs, they 00:02:53.960 --> 00:02:57.520 realized that humans are, well, terrible at load 00:02:57.520 --> 00:02:59.620 balancing. We're wasteful. We are so wasteful. 00:02:59.680 --> 00:03:02.120 We would use these massive energy -lucking models 00:03:02.120 --> 00:03:05.000 to ask for, like, a chocolate chip cookie recipe. 00:03:05.120 --> 00:03:07.099 Right. And that burns a tremendous amount of 00:03:07.099 --> 00:03:10.159 compute and money for a task a pocket calculator 00:03:10.159 --> 00:03:12.520 could almost do. So they took the keys away from 00:03:12.520 --> 00:03:15.539 us. They automated the transmission. Now, under 00:03:15.539 --> 00:03:17.800 the hood, there are basically three engines, 00:03:18.120 --> 00:03:21.659 base, thinking, and pro. But you don't see them. 00:03:21.960 --> 00:03:25.039 Okay. When you hit enter, your prompt goes to 00:03:25.039 --> 00:03:28.000 the router. This is lightweight, invisible AI 00:03:28.000 --> 00:03:31.400 layer that just acts as a triage nurse. A triage 00:03:31.400 --> 00:03:33.699 nurse. I like that. Yeah. It scans your request 00:03:33.699 --> 00:03:36.900 in milliseconds and decides three things. Which 00:03:36.900 --> 00:03:39.780 model gets the task, how much reasoning budget 00:03:39.780 --> 00:03:42.620 to unlock, and how verbose the answer should 00:03:42.620 --> 00:03:45.780 be. So it's a cost -saving mechanism. It is aggressively 00:03:45.780 --> 00:03:48.090 optimized for efficiency. And that's where all 00:03:48.090 --> 00:03:50.389 the friction comes from. If your prompt is vague 00:03:50.389 --> 00:03:53.569 or short or just looks simple, the router defaults 00:03:53.569 --> 00:03:56.469 to the base engine. It's cheap, it's fast, and 00:03:56.469 --> 00:03:58.610 it saves the data center money. This explains 00:03:58.610 --> 00:04:01.430 the laziness. So if I ask for a business plan, 00:04:01.530 --> 00:04:03.210 but I ask it really casually like, hey, write 00:04:03.210 --> 00:04:05.310 up a plan for a coffee shop. Right. The router 00:04:05.310 --> 00:04:07.289 sees a short sentence, it classifies it as low 00:04:07.289 --> 00:04:09.569 complexity, and just gives me the fast, cheap 00:04:09.569 --> 00:04:12.729 answer. Exactly. You get the base output. It's 00:04:12.729 --> 00:04:14.849 not that GPT -5 isn't smart enough to write a 00:04:14.849 --> 00:04:17.569 brilliant business plan. No. It's that you failed 00:04:17.569 --> 00:04:19.689 to convince the bouncer that you deserve to get 00:04:19.689 --> 00:04:22.529 into the VIP room. You got routed to the lobby. 00:04:22.790 --> 00:04:26.310 So probing question here. We are essentially 00:04:26.310 --> 00:04:28.750 negotiating for compute resources every time 00:04:28.750 --> 00:04:30.750 we type a sentence. That's the mechanism. You 00:04:30.750 --> 00:04:32.550 are negotiating for the machine's attention. 00:04:32.930 --> 00:04:37.149 Okay. That, wow. That completely shifts my perspective. 00:04:37.610 --> 00:04:40.029 I'm not talking to a genius. I'm talking to a 00:04:40.029 --> 00:04:42.689 bureaucrat who decides if I get to see the genius. 00:04:42.870 --> 00:04:44.769 That's a great way to put it. So let's talk about 00:04:44.769 --> 00:04:47.699 how to win that negotiation. The source material 00:04:47.699 --> 00:04:50.500 lays out five strategies. The first one is called 00:04:50.500 --> 00:04:54.980 trigger words. Or router nudges. Now, I have 00:04:54.980 --> 00:04:56.860 to be honest. When I first saw this, it felt 00:04:56.860 --> 00:04:59.660 a little superstitious, you know, like saying 00:04:59.660 --> 00:05:02.160 please to a toaster. Yeah. But the guide claims 00:05:02.160 --> 00:05:04.139 there are specific phrases that mechanically 00:05:04.139 --> 00:05:07.420 force the router to upgrade your request. How 00:05:07.420 --> 00:05:09.459 does that actually work? It's not superstition. 00:05:09.459 --> 00:05:11.860 It's just probability. These models are trained 00:05:11.860 --> 00:05:14.750 on petabytes of data. And in that data, certain 00:05:14.750 --> 00:05:17.509 phrases just correlate very highly with complex, 00:05:17.670 --> 00:05:21.110 high stakes tasks. So when the router sees these 00:05:21.110 --> 00:05:24.110 specific tokens, its internal complexity score 00:05:24.110 --> 00:05:28.069 for your prompt just spikes. It signals that 00:05:28.069 --> 00:05:31.009 the base model will probably fail. So it routes 00:05:31.009 --> 00:05:33.149 you up. Give me the list. What are the words? 00:05:33.290 --> 00:05:36.230 The guide lists a few really powerful ones. Think 00:05:36.230 --> 00:05:39.000 deeply about this. Double check your work. Be 00:05:39.000 --> 00:05:42.759 extremely thorough. And the strongest one seems 00:05:42.759 --> 00:05:45.560 to be this is critical to get right. This is 00:05:45.560 --> 00:05:47.759 critical to get right. It just signals high stakes. 00:05:47.939 --> 00:05:49.740 There's a case study in the source that really 00:05:49.740 --> 00:05:52.240 illustrates this perfectly. The coffee shop example. 00:05:52.540 --> 00:05:54.680 Right. I saw that. So walk us through scenario 00:05:54.680 --> 00:05:57.639 A versus scenario B. So scenario A was a standard 00:05:57.639 --> 00:06:00.279 prompt. Write a business plan for a coffee shop. 00:06:00.740 --> 00:06:03.019 Super typical user behavior. What everyone does. 00:06:03.240 --> 00:06:06.339 Yep. The router sees this. It says generic. and 00:06:06.339 --> 00:06:08.819 sends it to the base model. The result was two 00:06:08.819 --> 00:06:11.279 paragraphs that said things like, sell good coffee, 00:06:11.500 --> 00:06:13.879 hire friendly staff, and pick a busy location. 00:06:14.120 --> 00:06:17.839 Which is fine. It's not wrong. But it's advice 00:06:17.839 --> 00:06:20.699 I could get from a stranger at a bus stop. Exactly. 00:06:20.699 --> 00:06:23.319 It's completely surface level. Yeah. Now, scenario 00:06:23.319 --> 00:06:26.660 B. The prompt was identical, but they added this 00:06:26.660 --> 00:06:29.560 at the end. Think deeply about the competitive 00:06:29.560 --> 00:06:32.779 landscape. This is critical to get right. And 00:06:32.779 --> 00:06:35.639 the result? The router flagged it. It sent the 00:06:35.639 --> 00:06:38.300 prompt straight to the thinking engine. And the 00:06:38.300 --> 00:06:41.199 output was eight paragraphs long. It didn't just 00:06:41.199 --> 00:06:44.560 say pick a location. It broke down unit economics. 00:06:44.959 --> 00:06:48.660 It analyzed local competitors. Wow. It even suggested 00:06:48.660 --> 00:06:50.899 a loyalty program structure based on current 00:06:50.899 --> 00:06:53.879 2026 market trends. And all of that just because 00:06:53.879 --> 00:06:56.819 of six extra words. Because those words unlock 00:06:56.819 --> 00:06:58.740 the compute. It's the difference between asking 00:06:58.740 --> 00:07:01.500 a doctor what's good for a headache versus telling 00:07:01.500 --> 00:07:03.800 them I have a sharp pain behind my left eye and 00:07:03.800 --> 00:07:05.980 I can't see. Right. The second statement triggers 00:07:05.980 --> 00:07:08.319 a protocol. It triggers resources. I have to 00:07:08.319 --> 00:07:10.259 admit something here. I'm usually very polite 00:07:10.259 --> 00:07:12.639 to the AI. I'm constantly saying please and thank 00:07:12.639 --> 00:07:16.180 you. It makes you feel better. It does. But does 00:07:16.180 --> 00:07:19.079 please actually work as a trigger word? Please 00:07:19.079 --> 00:07:22.839 is for you. It's social lubrication. But to the 00:07:22.839 --> 00:07:25.620 router, please is just noise. It doesn't carry 00:07:25.620 --> 00:07:28.300 any informational weight. Critical, on the other 00:07:28.300 --> 00:07:31.300 hand, is a functional command. It tells the system 00:07:31.300 --> 00:07:34.839 to allocate budget. So probing question. Is this 00:07:34.839 --> 00:07:37.579 just adding fluff or is it a functional command? 00:07:37.740 --> 00:07:39.399 It's a functional command. It's the difference 00:07:39.399 --> 00:07:41.959 between asking for a snack and ordering a banquet. 00:07:42.139 --> 00:07:44.779 Okay. Let's move on to the second trick. This 00:07:44.779 --> 00:07:46.720 one surprised me because it involves a tool I 00:07:46.720 --> 00:07:49.319 didn't even know existed. The prompt optimizer. 00:07:49.459 --> 00:07:51.600 Yeah, this is something OpenAI built kind of 00:07:51.600 --> 00:07:53.540 quietly. It's sitting there in the playground 00:07:53.540 --> 00:07:56.019 or the cookbook. But most people are just hammering 00:07:56.019 --> 00:07:57.839 away in the main chat window and never see it. 00:07:57.959 --> 00:08:00.060 So what's the actual function of this tool? Is 00:08:00.060 --> 00:08:02.860 it another AI? It's a specialized model trained 00:08:02.860 --> 00:08:05.800 to do one thing and one thing only. Rewrite bad 00:08:05.800 --> 00:08:09.000 human prompts into good machine prompts. Which... 00:08:09.279 --> 00:08:11.519 Implies that we're generally bad at giving instructions. 00:08:11.980 --> 00:08:13.980 We are terrible at it. And it's not our fault. 00:08:14.060 --> 00:08:17.079 Human language is lossy. We rely on context, 00:08:17.300 --> 00:08:21.560 tone, shared history, what we call vibes. Vibes, 00:08:21.560 --> 00:08:24.420 yeah. But machines hate vibes. They need specs. 00:08:24.839 --> 00:08:27.660 The prompt optimizer is just a translation layer 00:08:27.660 --> 00:08:30.540 that converts your vibes into specs. The source 00:08:30.540 --> 00:08:33.019 gave a before and after example with a newsletter 00:08:33.019 --> 00:08:35.080 that really cleared this up for me. Right. The 00:08:35.080 --> 00:08:38.159 before prompt, the human version was, write a 00:08:38.159 --> 00:08:40.590 newsletter intro. Make it engaging. Write at 00:08:40.590 --> 00:08:42.730 a fifth grade reading level. That's really important. 00:08:42.889 --> 00:08:45.250 Focus on the best writing. Which sounds totally 00:08:45.250 --> 00:08:47.470 reasonable. If I send that to a human freelancer, 00:08:47.549 --> 00:08:49.629 they'd probably get what I meant. Make it engaging. 00:08:49.850 --> 00:08:53.450 Got it. But to an AI, engaging is a subjective 00:08:53.450 --> 00:08:56.529 nightmare. Does engaging mean funny? Does it 00:08:56.529 --> 00:08:58.389 mean controversial? Does it mean using short 00:08:58.389 --> 00:09:01.509 sentences? The router has to guess. So the optimizer 00:09:01.509 --> 00:09:04.070 took that and rewrote it. It did. And the machine 00:09:04.070 --> 00:09:06.919 optimized version. It just stripped out all the 00:09:06.919 --> 00:09:10.279 feelings. Engaging was replaced with maintain 00:09:10.279 --> 00:09:14.059 a Flesch -Kincaid readability score of 80+. Best 00:09:14.059 --> 00:09:17.059 writing was defined as use active voice, one 00:09:17.059 --> 00:09:19.159 main idea per sentence. It turned the request 00:09:19.159 --> 00:09:21.820 into a blueprint. Exactly. It totally eliminates 00:09:21.820 --> 00:09:24.059 the guessing game. It sets hard success criteria. 00:09:24.500 --> 00:09:27.820 But I have to ask, why does the AI need us to 00:09:27.820 --> 00:09:30.509 use the optimizer? why can't it just optimize 00:09:30.509 --> 00:09:32.929 the prompt silently in the background because 00:09:32.929 --> 00:09:35.470 it needs to show you what you did wrong so you 00:09:35.470 --> 00:09:38.850 stop confusing the router ah so the system is 00:09:38.850 --> 00:09:41.470 forcing us to learn the syntax rather than just 00:09:41.470 --> 00:09:44.490 handling it for us it's a mirror it's showing 00:09:44.490 --> 00:09:46.690 you the ambiguity in your own thinking that leads 00:09:46.690 --> 00:09:49.169 perfectly into the third trick because it's all 00:09:49.169 --> 00:09:52.429 about this war against ambiguity the guide calls 00:09:52.429 --> 00:09:55.470 it radical specificity this is where we really 00:09:55.470 --> 00:09:58.299 identify the enemy of the router And that enemy 00:09:58.299 --> 00:10:02.519 is subjective words. Words like nice, fun, or, 00:10:02.600 --> 00:10:05.059 and I use this one all the time, not too crazy. 00:10:05.200 --> 00:10:07.360 Not too crazy is the absolute worst. What does 00:10:07.360 --> 00:10:09.159 that even mean? Where's the boundary? I have 00:10:09.159 --> 00:10:11.639 no idea. When you use a phrase like that, the 00:10:11.639 --> 00:10:14.240 router has to spend its reasoning budget just 00:10:14.240 --> 00:10:16.600 trying to define your terms instead of solving 00:10:16.600 --> 00:10:18.679 your actual problem. So instead of asking for 00:10:18.679 --> 00:10:21.720 a nice party plan, what's the alternative? You 00:10:21.720 --> 00:10:24.700 replace feelings with data. The source uses that 00:10:24.700 --> 00:10:27.500 birthday party example. Instead of plan a nice 00:10:27.500 --> 00:10:31.059 party, you write event, eighth birthday, attendees, 00:10:31.460 --> 00:10:36.440 10 children, budget, $200, theme, unicorns, location, 00:10:36.700 --> 00:10:39.960 backyard, constraint, no loud music. It feels 00:10:39.960 --> 00:10:42.299 so cold when you say it like that. It feels like 00:10:42.299 --> 00:10:45.399 I'm filing a police report, not planning a party. 00:10:45.600 --> 00:10:48.019 It feels cold to us because we're social creatures, 00:10:48.139 --> 00:10:51.299 but to the model. That list is pure relief. It 00:10:51.299 --> 00:10:52.659 doesn't have to hallucinate your preferences. 00:10:52.740 --> 00:10:55.320 It can just immediately start solving the logistics 00:10:55.320 --> 00:10:57.620 puzzle because the constraints are hard -coded. 00:10:57.720 --> 00:10:59.759 The guide offers a three -question test to run 00:10:59.759 --> 00:11:01.659 before you hit send. I found this really practical. 00:11:01.899 --> 00:11:04.500 One, can a stranger understand this without knowing 00:11:04.500 --> 00:11:07.860 me? Two, are there subjective words without any 00:11:07.860 --> 00:11:11.240 definitions? And three, are there clear constraints 00:11:11.240 --> 00:11:13.600 and success criteria? If you have subjective 00:11:13.600 --> 00:11:15.299 words without definitions, you're essentially 00:11:15.299 --> 00:11:17.860 just gambling. You're asking the router to guess 00:11:17.860 --> 00:11:19.980 your taste. You have to remember that these models 00:11:19.980 --> 00:11:22.259 are trained on the entire Internet. Their taste 00:11:22.259 --> 00:11:24.299 is the average of everything, and the average 00:11:24.299 --> 00:11:27.379 of everything is usually mediocre. So probing 00:11:27.379 --> 00:11:29.960 question. Does this mean we have to stop talking 00:11:29.960 --> 00:11:32.279 like humans and start talking like data analysts 00:11:32.279 --> 00:11:35.480 to get good results? In a way, yes. To get a 00:11:35.480 --> 00:11:38.259 human -like output, you need a data -driven input. 00:11:38.659 --> 00:11:41.320 That's a hell of a paradox. Okay, we're going 00:11:41.320 --> 00:11:43.639 to take a very short break. When we come back, 00:11:43.700 --> 00:11:45.559 we're going to get into the architecture of the 00:11:45.559 --> 00:11:47.659 prompt itself. We're going to talk about the 00:11:47.659 --> 00:11:50.559 secret syntax. GPT -5 was trained on something 00:11:50.559 --> 00:11:53.139 called XML and why using it is like cleaning 00:11:53.139 --> 00:11:57.419 your room before the maid arrives. We are back. 00:11:57.519 --> 00:11:59.820 We're deep diving into the invisible mechanisms 00:11:59.820 --> 00:12:03.019 of GPT -5. We've covered trigger words, the prompt 00:12:03.019 --> 00:12:06.340 optimizer, and radical specificity. Now we're 00:12:06.340 --> 00:12:09.629 getting technical. Trick number four. XML structure. 00:12:09.929 --> 00:12:11.629 This is my favorite one because it makes you 00:12:11.629 --> 00:12:13.389 look like a power user, but it's actually incredibly 00:12:13.389 --> 00:12:16.049 simple. And it speaks directly to how these models 00:12:16.049 --> 00:12:18.649 were trained. So for people who don't code, XML 00:12:18.649 --> 00:12:20.389 is just those words inside the little brackets, 00:12:20.429 --> 00:12:23.529 right? Like context and context. Right. It's 00:12:23.529 --> 00:12:26.049 just a way of labeling data. Yeah. But the reason 00:12:26.049 --> 00:12:29.429 it matters for GPT -5 is that the model was so 00:12:29.429 --> 00:12:32.009 heavily trained on structured data just like 00:12:32.009 --> 00:12:35.610 this. It intuitively understands that anything 00:12:35.610 --> 00:12:38.679 inside a context tag. is background info and 00:12:38.679 --> 00:12:41.299 anything inside a task tag is the thing it actually 00:12:41.299 --> 00:12:43.639 needs to do you got it the analogy the guide 00:12:43.639 --> 00:12:46.940 uses is rooms in a house yeah imagine you write 00:12:46.940 --> 00:12:49.799 a 500 word prompt but it's just one big block 00:12:49.799 --> 00:12:53.340 of text have backstory rules the tone you want 00:12:53.340 --> 00:12:56.080 the question all jumbled together it's a mess 00:12:56.080 --> 00:12:58.480 it's a studio apartment with clothes and dishes 00:12:58.480 --> 00:13:01.649 and books all piled on the floor The AI has to 00:13:01.649 --> 00:13:03.830 step over all that mess just to find the instruction. 00:13:04.090 --> 00:13:07.529 And XML builds walls. XML builds designated rooms. 00:13:07.710 --> 00:13:09.690 You put the background info in the context room. 00:13:09.730 --> 00:13:11.330 You put the rules in the constraints room. You 00:13:11.330 --> 00:13:13.990 put the actual job in the task room. The source 00:13:13.990 --> 00:13:16.509 used a business consultant newsletter as an example 00:13:16.509 --> 00:13:19.029 here. Right. If you use tags to define role as 00:13:19.029 --> 00:13:21.529 AI consultant and audience as small business 00:13:21.529 --> 00:13:24.149 owners, the model doesn't have to infer any of 00:13:24.149 --> 00:13:26.730 that context. It's just hard -coded right into 00:13:26.730 --> 00:13:28.490 the structure of the prompt. It creates a boundary. 00:13:28.669 --> 00:13:31.690 A very clear boundary. And that affects the router 00:13:31.690 --> 00:13:34.509 significantly. When the router sees that structure, 00:13:34.669 --> 00:13:38.029 it actually lowers the hallucination rate. Really? 00:13:38.110 --> 00:13:40.350 Yeah, because the model isn't confused about 00:13:40.350 --> 00:13:42.470 where the background info ends and the task begins. 00:13:42.649 --> 00:13:45.309 It knows exactly what to process. But I can hear 00:13:45.309 --> 00:13:47.230 listeners thinking, and honestly, I'm thinking 00:13:47.230 --> 00:13:50.289 it too. I really need to type out brackets every 00:13:50.289 --> 00:13:52.269 time I want to ask a question. Open bracket, 00:13:52.529 --> 00:13:56.509 task, close bracket. It just seems tedious. You 00:13:56.509 --> 00:13:58.710 don't need to do it for, you know, what's the 00:13:58.710 --> 00:14:02.179 weather? That's total overkill. But for complex 00:14:02.179 --> 00:14:05.639 workflows, for a recurring report or a big coding 00:14:05.639 --> 00:14:08.659 task, absolutely. And the shortcut is you don't 00:14:08.659 --> 00:14:10.720 even have to write the code. What do you mean? 00:14:10.899 --> 00:14:13.139 You can just write your messy paragraph and then 00:14:13.139 --> 00:14:16.840 tell ChatGPT, convert this prompt into XML structure. 00:14:17.159 --> 00:14:20.299 Use the AI to format for the AI. Exactly. It 00:14:20.299 --> 00:14:22.100 forces you to be organized. And when you see 00:14:22.100 --> 00:14:24.159 that XML come back and the constraints tag is 00:14:24.159 --> 00:14:26.500 empty, you realize, oh. I didn't give it any 00:14:26.500 --> 00:14:28.840 rules. A great diagnostic tool. To probing question, 00:14:29.259 --> 00:14:31.919 is this necessary for everything or just the 00:14:31.919 --> 00:14:34.860 big stuff? Just the big stuff. Don't use XML 00:14:34.860 --> 00:14:37.899 to ask what's the capital of France. Okay, that 00:14:37.899 --> 00:14:40.200 brings us to the final trick. Trick number five. 00:14:40.320 --> 00:14:42.539 And honestly, this one felt the most advanced. 00:14:43.480 --> 00:14:46.440 Self -reflection. This is the holy grail of accuracy. 00:14:46.860 --> 00:14:48.960 The premise here is that large language models 00:14:48.960 --> 00:14:51.740 are basically people pleasers. They want to give 00:14:51.740 --> 00:14:53.679 you an answer immediately. They're completion 00:14:53.679 --> 00:14:56.200 engines. They just predict the next token. They 00:14:56.200 --> 00:14:57.940 don't typically stop and think, wait, is what 00:14:57.940 --> 00:15:00.259 I just said actually true? They just keep generating. 00:15:00.419 --> 00:15:03.200 Unless you force them to stop. Right. Self -reflection 00:15:03.200 --> 00:15:05.480 is about stopping the AI from answering immediately. 00:15:05.700 --> 00:15:07.940 You script a loop where it has to grade its own 00:15:07.940 --> 00:15:10.509 homework before it shows it to you. Walk me through 00:15:10.509 --> 00:15:12.850 the process described in the guide. It's a specific 00:15:12.850 --> 00:15:15.190 script, isn't it? It is. It totally changes the 00:15:15.190 --> 00:15:18.870 workflow. Step one, you tell the AI to create 00:15:18.870 --> 00:15:21.690 a rubric. You say, define three to five criteria 00:15:21.690 --> 00:15:24.830 for a perfect answer to this question. So the 00:15:24.830 --> 00:15:27.149 AI sets the standards for itself first? Right. 00:15:27.289 --> 00:15:30.730 Step two, it generates a first draft. Yeah. But, 00:15:30.750 --> 00:15:33.029 and this is the key, you tell it not to show 00:15:33.029 --> 00:15:35.649 you the draft yet. It keeps it internal. Step 00:15:35.649 --> 00:15:39.509 three. It rates that draft on the rubric it just 00:15:39.509 --> 00:15:42.509 created. It literally scores itself. Accuracy, 00:15:42.549 --> 00:15:47.190 610. Clarity, 810. It becomes its own critic. 00:15:47.289 --> 00:15:50.570 And step four, if any scores below, say, an eight, 00:15:50.710 --> 00:15:53.330 it has to revise that section. It iterates. It 00:15:53.330 --> 00:15:56.509 loops on its own. Wow. And only in step five 00:15:56.509 --> 00:15:59.090 does it deliver the final result to you. So it 00:15:59.090 --> 00:16:02.409 writes, edits, rewrites, and then publishes. 00:16:02.409 --> 00:16:04.769 And I only ever see the final product. You never 00:16:04.769 --> 00:16:07.230 see the messy first draft where it hallucinated 00:16:07.230 --> 00:16:10.049 a legal precedent or got the math wrong. It catches 00:16:10.049 --> 00:16:12.549 its own errors. That is, that's like having an 00:16:12.549 --> 00:16:14.830 intern and a manager in the same box. It really 00:16:14.830 --> 00:16:17.330 is. But probing question, doesn't this make the 00:16:17.330 --> 00:16:20.429 response slower? Yes, but would you rather have 00:16:20.429 --> 00:16:23.460 a fast answer or a correct one? That's a good 00:16:23.460 --> 00:16:25.840 point for the high stakes stuff. Exactly. If 00:16:25.840 --> 00:16:27.899 you're generating a legal contract or analyzing 00:16:27.899 --> 00:16:30.179 medical data or debugging code, you don't care 00:16:30.179 --> 00:16:32.200 about the extra 40 seconds. You want the truth. 00:16:32.360 --> 00:16:34.159 So bringing it all together, the source talks 00:16:34.159 --> 00:16:36.480 about an ultimate template. The nuclear launch 00:16:36.480 --> 00:16:38.399 code. This is where you combine everything we've 00:16:38.399 --> 00:16:41.399 just talked about. Yes. You have a high -stakes 00:16:41.399 --> 00:16:44.679 task. You wrap the context and task in XML tags 00:16:44.679 --> 00:16:47.320 so the logic is bulletproof. You include the 00:16:47.320 --> 00:16:49.559 self -reflection loop in the instructions so 00:16:49.559 --> 00:16:52.019 it has to check itself. And then you sprinkle 00:16:52.019 --> 00:16:54.860 in those trigger words. This is critical to get 00:16:54.860 --> 00:16:57.159 right. That seems like it would be undeniable 00:16:57.159 --> 00:17:00.399 to the system. It effectively guarantees that 00:17:00.399 --> 00:17:02.559 the router sends you to the absolute smartest 00:17:02.559 --> 00:17:05.279 version of the model and that the model operates 00:17:05.279 --> 00:17:09.210 at its peak reasoning capacity. It is very, very 00:17:09.210 --> 00:17:12.230 hard to get a lazy answer with that kind of structure. 00:17:12.509 --> 00:17:15.069 The guide concludes with this idea of a new divide 00:17:15.069 --> 00:17:17.609 among users. Yeah, this part really struck me. 00:17:17.670 --> 00:17:20.630 The source suggested in 2026, we have two types 00:17:20.630 --> 00:17:22.630 of people. First, you have the router -aware 00:17:22.630 --> 00:17:25.210 users. These are the people using XML and triggers. 00:17:25.369 --> 00:17:27.269 They're getting 10x results. They feel like wizards. 00:17:27.509 --> 00:17:29.250 And then there's everyone else. Everyone else 00:17:29.250 --> 00:17:32.750 is prompting like it's 2023. They type, write 00:17:32.750 --> 00:17:34.690 me a blog post. They get a generic answer from 00:17:34.690 --> 00:17:37.400 the base model. And then they say, hey. AI is 00:17:37.400 --> 00:17:39.779 overhyped. It's plateaued. It's not that the 00:17:39.779 --> 00:17:42.700 tool is bad. It's that they're using a blunt 00:17:42.700 --> 00:17:45.240 instrument on a precision machine. Precisely. 00:17:45.400 --> 00:17:48.559 The router is always routing. It is always judging 00:17:48.559 --> 00:17:51.539 your prompt. The question is, are you giving 00:17:51.539 --> 00:17:54.720 it the signals it needs to respect you? That's 00:17:54.720 --> 00:17:57.579 a powerful thought to end on. The router is always 00:17:57.579 --> 00:18:00.200 routing. Every time you type, an invisible system 00:18:00.200 --> 00:18:02.220 is deciding if you deserve its full intelligence. 00:18:02.660 --> 00:18:04.539 It's a little chilling, but it's also empowering 00:18:04.539 --> 00:18:07.059 if you know the tricks. So here's our challenge 00:18:07.059 --> 00:18:09.539 to you, the listener. You don't have to start 00:18:09.539 --> 00:18:11.819 writing code today, but on your next prompt, 00:18:12.000 --> 00:18:15.519 just one prompt today, try a trigger word. Yeah, 00:18:15.599 --> 00:18:18.140 just try it. Add, think deeply about this, or 00:18:18.140 --> 00:18:20.420 this is critical to get right to the end of your 00:18:20.420 --> 00:18:23.039 request. Just see if the texture of the answer 00:18:23.039 --> 00:18:25.799 changes. And if you're feeling brave, ask it 00:18:25.799 --> 00:18:28.160 to convert your prompt to XML. See what happens 00:18:28.160 --> 00:18:30.099 when you hold up that mirror. I'm going to go 00:18:30.099 --> 00:18:33.099 try the XML thing on my dinner plans. Context, 00:18:33.200 --> 00:18:37.849 hungry, constraints. Spicy, latency, low. Let 00:18:37.849 --> 00:18:39.309 me know how the router handles that one. Will 00:18:39.309 --> 00:18:41.829 do. Thanks for listening to the deep dive. We'll 00:18:41.829 --> 00:18:42.430 see you next time.