WEBVTT 00:00:00.000 --> 00:00:02.899 Most of us treat AI like a high -tech magic eight 00:00:02.899 --> 00:00:05.960 ball. Oh, totally. We ask a simple question, 00:00:06.099 --> 00:00:08.980 and we get a generic answer back. Exactly. You've 00:00:08.980 --> 00:00:11.619 probably seen the viral trends, like asking an 00:00:11.619 --> 00:00:14.259 LLM to analyze your whole chat history. It's 00:00:14.259 --> 00:00:17.300 fun, yeah. But it is just the very, very tip 00:00:17.300 --> 00:00:19.440 of the iceberg for what these tools can actually 00:00:19.440 --> 00:00:22.359 do for you. OK, so let's unpack this. This deep 00:00:22.359 --> 00:00:25.000 dive is about the truth of advanced prompt engineering. 00:00:25.529 --> 00:00:28.109 We're moving beyond just talking to the machine. 00:00:28.410 --> 00:00:30.309 We're learning how to build structured systems 00:00:30.309 --> 00:00:32.409 within it. And the secret here, it isn't really 00:00:32.409 --> 00:00:35.170 about which model you use. No, not at all. It's 00:00:35.170 --> 00:00:37.570 about how you talk to it. It's how you architect 00:00:37.570 --> 00:00:40.189 the input. We've pulled back the curtain on the 00:00:40.189 --> 00:00:42.850 systems used by researchers at places like Google, 00:00:43.170 --> 00:00:45.649 OpenAI, and Anthropic. And we've broken them 00:00:45.649 --> 00:00:49.729 down into 10 surprisingly simple but really powerful 00:00:49.729 --> 00:00:51.750 techniques. This is basically your blueprint 00:00:51.750 --> 00:00:55.270 to go from being an amateur to... Well, an AI 00:00:55.270 --> 00:00:57.670 architect. And it all starts with one essential 00:00:57.670 --> 00:01:00.850 mindset shift. You have to stop thinking of the 00:01:00.850 --> 00:01:03.670 AI as a human assistant you chat with. And start 00:01:03.670 --> 00:01:05.909 treating it like a complex simulator you direct. 00:01:06.250 --> 00:01:09.010 Here's the core truth that really changes everything. 00:01:09.750 --> 00:01:13.670 The AI does not think. Not in the way a person 00:01:13.670 --> 00:01:15.829 does. It doesn't have an opinion. Right. It's 00:01:15.829 --> 00:01:18.790 a highly sophisticated probability engine. A 00:01:18.790 --> 00:01:21.510 simulator that has, you know, processed basically 00:01:21.510 --> 00:01:24.260 the entire internet. Because it's read everything, 00:01:24.640 --> 00:01:27.219 it knows how a Nobel Prize -winning physicist 00:01:27.219 --> 00:01:29.540 talks. And it knows how a sarcastic teenager 00:01:29.540 --> 00:01:32.680 talks. It has a library of a million different 00:01:32.680 --> 00:01:35.579 masks it can wear. The moment most users fail 00:01:35.579 --> 00:01:37.819 is when they ask for its opinion without setting 00:01:37.819 --> 00:01:40.280 the stage. Yeah, if you don't tell it which mask 00:01:40.280 --> 00:01:44.799 to put on, it just defaults to this boring, generalized, 00:01:45.140 --> 00:01:47.099 super safe average of everything it's ever read. 00:01:47.140 --> 00:01:49.409 That's the blandest possible result. OK, let 00:01:49.409 --> 00:01:51.430 me use a quick analogy to really nail this down. 00:01:51.769 --> 00:01:53.430 You wouldn't walk into a garage and just ask 00:01:53.430 --> 00:01:56.049 the first person you see for complex car repair 00:01:56.049 --> 00:01:58.250 advice. No, of course not. You'd find the master 00:01:58.250 --> 00:02:00.069 mechanics. You have to act like a movie director. 00:02:00.450 --> 00:02:03.150 Set the scene. Don't ask a random stranger. Tell 00:02:03.150 --> 00:02:06.090 the AI who to simulate. You are now a master 00:02:06.090 --> 00:02:08.430 mechanic with 20 years of experience who only 00:02:08.430 --> 00:02:11.710 works on vintage European sports cars. So connecting 00:02:11.710 --> 00:02:14.250 this to the bigger picture, what happens if we 00:02:14.250 --> 00:02:17.009 don't set that precise stage? Well, the model 00:02:17.009 --> 00:02:19.629 just pulls from that average data. You get an 00:02:19.629 --> 00:02:23.210 output that's generalized, average, and probably 00:02:23.210 --> 00:02:26.030 useless for what you actually need. And that 00:02:26.030 --> 00:02:28.050 leads perfectly into our first technique, which 00:02:28.050 --> 00:02:30.810 is persona adoption. Right. Most people kind 00:02:30.810 --> 00:02:32.629 of do this, but they do it wrong. They're just 00:02:32.629 --> 00:02:35.870 too vague. We need deep specificity, right? If 00:02:35.870 --> 00:02:38.650 you just say, you are a coder, the AI has no 00:02:38.650 --> 00:02:40.949 idea what that means. Is it a web developer? 00:02:41.389 --> 00:02:43.889 A machine learning expert? Exactly. That lack 00:02:43.889 --> 00:02:47.590 of semantic density is fatal to getting good 00:02:47.590 --> 00:02:50.150 output. Instead, you have to be really specific. 00:02:50.370 --> 00:02:52.770 Say, you are a senior data engineer with 10 years 00:02:52.770 --> 00:02:55.710 of experience in Python, specializing in cloud 00:02:55.710 --> 00:02:58.349 infrastructure and security. That level of detail, 00:02:58.610 --> 00:03:01.189 it changes everything. The AI's vocabulary, its 00:03:01.189 --> 00:03:04.189 logic structure, it's forced to access deeper, 00:03:04.210 --> 00:03:06.469 more relevant parts of its knowledge. Just think 00:03:06.469 --> 00:03:08.449 about the difference. A bad prompt is, write 00:03:08.449 --> 00:03:11.030 a blog post about coffee. It's generic. It's 00:03:11.030 --> 00:03:14.770 boring. A good prompt is, act as a world -class 00:03:14.770 --> 00:03:18.009 barista running a cafe in southern Italy. Write 00:03:18.009 --> 00:03:21.370 300 words on the art of making espresso. Focus 00:03:21.370 --> 00:03:24.530 on the aroma, the ritual, the crema. So does 00:03:24.530 --> 00:03:26.490 that specificity just pull specific words or 00:03:26.490 --> 00:03:28.550 is it deeper than that? It's much deeper. The 00:03:28.550 --> 00:03:31.229 AI accesses specific parts of its training data 00:03:31.229 --> 00:03:33.650 related to that expert character and their logic. 00:03:33.969 --> 00:03:36.189 Okay, let's talk about the single biggest problem 00:03:36.189 --> 00:03:40.169 with these models. Hallucinations. Factual errors. 00:03:40.409 --> 00:03:43.979 Yes. So we have to talk about The Chain of Verification, 00:03:44.159 --> 00:03:46.719 or COVE. This is a technique from Google researchers. 00:03:46.939 --> 00:03:49.419 What's so cool about this is that COVE forces 00:03:49.419 --> 00:03:52.199 the AI to check its own work before it shows 00:03:52.199 --> 00:03:54.400 the result to you. It's like an internal quality 00:03:54.400 --> 00:03:56.280 control system you build right into the prompt. 00:03:56.379 --> 00:03:58.639 It has a four -step process, all in one prompt. 00:03:58.879 --> 00:04:01.659 First, it generates an initial answer. Then second, 00:04:01.740 --> 00:04:03.879 it generates a list of questions to test if that 00:04:03.879 --> 00:04:06.479 answer is actually true. Third, it answers its 00:04:06.479 --> 00:04:08.280 own questions. It's literally fact -checking 00:04:08.280 --> 00:04:10.500 itself. And then finally, it fixes the original 00:04:10.500 --> 00:04:12.979 answer based on that check. It's brilliant. So 00:04:12.979 --> 00:04:14.860 if you ask it to explain the fall of the Roman 00:04:14.860 --> 00:04:17.839 Empire, it might ask itself, did I get the dates 00:04:17.839 --> 00:04:21.759 for Diocletian's reforms right? Or did I correctly 00:04:21.759 --> 00:04:24.879 attribute the Visigoths invasion? So how does 00:04:24.879 --> 00:04:27.519 this actually save the user time versus just 00:04:27.519 --> 00:04:30.399 doing manual checks? It ensures accuracy and 00:04:30.399 --> 00:04:33.040 detail from the start, fixing potential errors 00:04:33.040 --> 00:04:35.600 internally before you even see them. It's about 00:04:35.600 --> 00:04:38.620 trust. Next up, Anthropic found something really 00:04:38.620 --> 00:04:41.860 counterintuitive. Sometimes showing the AI what 00:04:41.860 --> 00:04:44.259 not to do is just as powerful as showing it what 00:04:44.259 --> 00:04:46.480 to do. This is called few shot with negative 00:04:46.480 --> 00:04:48.779 examples. We all know a few shot, right? You 00:04:48.779 --> 00:04:50.519 give it a good example of what you want. But 00:04:50.519 --> 00:04:52.759 if you pair that good example with a bad one, 00:04:53.019 --> 00:04:55.040 and this is the key, you explain why it's bad. 00:04:55.389 --> 00:04:58.329 the AI learns its boundaries way faster. You 00:04:58.329 --> 00:05:00.589 know, I still wrestle with prompt drift myself, 00:05:00.689 --> 00:05:03.889 especially when the AI starts sounding too robotic 00:05:03.889 --> 00:05:06.610 or like overly excited. Yeah, it gets all the 00:05:06.610 --> 00:05:08.709 exclamation points out. It's like it loses its 00:05:08.709 --> 00:05:10.610 personality after a while. This technique is 00:05:10.610 --> 00:05:12.569 the best way to fix that. You're giving it red 00:05:12.569 --> 00:05:14.990 flags. So look at this template. Good example. 00:05:15.329 --> 00:05:18.310 Five insightful ways to save money today. Bad 00:05:18.310 --> 00:05:22.250 example. Save money now. Urgent. And then you 00:05:22.250 --> 00:05:25.170 explain why it's bad. uses all caps, sounds like 00:05:25.170 --> 00:05:28.250 spam, lacks authority. So what's the biggest 00:05:28.250 --> 00:05:30.810 benefit for someone who just hates that robotic 00:05:30.810 --> 00:05:33.550 or, you know, over -the -top tone? Providing 00:05:33.550 --> 00:05:36.389 those bad examples really helps the AI learn 00:05:36.389 --> 00:05:39.529 boundaries and avoid that robotic or overly excited 00:05:39.529 --> 00:05:42.290 style. Okay, our next set of techniques is all 00:05:42.290 --> 00:05:46.110 about forcing the AI to slow down. If you rush 00:05:46.110 --> 00:05:49.480 the model, it takes shortcuts. It guesses. We 00:05:49.480 --> 00:05:53.019 want complex, layered thinking. OpenAI uses something 00:05:53.019 --> 00:05:55.040 called the Structured Thinking Protocol for this. 00:05:55.699 --> 00:05:57.899 You force the model to think in designed layers 00:05:57.899 --> 00:05:59.839 instead of just jumping to an answer. You have 00:05:59.839 --> 00:06:02.519 to segment its thought process. So, layer one, 00:06:02.800 --> 00:06:05.319 understand the goal. Layer two, analyze the variables. 00:06:05.439 --> 00:06:08.180 Layer three, strategize the approach. And only 00:06:08.180 --> 00:06:11.000 then, layer four, execute the final output. You'd 00:06:11.000 --> 00:06:12.920 use this for really difficult decisions, like 00:06:12.920 --> 00:06:14.920 should you buy or rent a house? Exactly. And 00:06:14.920 --> 00:06:17.769 the fifth technique. from Google DeepMind tackles 00:06:17.769 --> 00:06:20.829 overconfidence. The AI always sounds 100 % sure 00:06:20.829 --> 00:06:23.029 of itself. Even when it's just wrong. That's 00:06:23.029 --> 00:06:24.829 where confidence -weighted prompting comes in. 00:06:25.290 --> 00:06:27.790 You can ask the AI to rate its own confidence 00:06:27.790 --> 00:06:31.430 from 0 to 100%. And here's the trick. You tell 00:06:31.430 --> 00:06:33.949 it that if its confidence is less than, say, 00:06:34.029 --> 00:06:37.149 80 %? It has to provide an alternative answer 00:06:37.149 --> 00:06:40.149 or state its assumptions. This is a total game 00:06:40.149 --> 00:06:42.709 changer for unreliable questions, like what were 00:06:42.709 --> 00:06:44.910 the average temperatures in London in the year 00:06:44.910 --> 00:06:48.329 1650? It'll give an answer, but maybe say confidence, 00:06:48.649 --> 00:06:52.089 65 % based on limited historical records. So 00:06:52.089 --> 00:06:54.629 what's the core danger of the AI always sounding 00:06:54.629 --> 00:06:57.170 so certain? The user might rely on a low -confidence 00:06:57.170 --> 00:06:59.509 answer without understanding the uncertainty 00:06:59.509 --> 00:07:03.310 or the assumptions the model made. Welcome back 00:07:03.310 --> 00:07:06.009 to the Deep Dive. We're moving from foundational 00:07:06.009 --> 00:07:08.850 control to maximizing relevance and quality. 00:07:09.490 --> 00:07:11.709 So advanced prompt engineering is all about control. 00:07:12.209 --> 00:07:14.870 This next technique from entropic context injection 00:07:14.870 --> 00:07:17.709 with boundaries is like setting up a digital 00:07:17.709 --> 00:07:19.850 knowledge fence. This is so important if you're 00:07:19.850 --> 00:07:21.829 dealing with specific information. You paste 00:07:21.829 --> 00:07:24.370 in your text, a user manual, a resume, whatever, 00:07:24.589 --> 00:07:27.009 and you tell the AI. Only use information from 00:07:27.009 --> 00:07:29.110 the context below. If the answer isn't there, 00:07:29.329 --> 00:07:32.230 say insufficient information. It guarantees the 00:07:32.230 --> 00:07:34.879 AI stays on topic. It prevents it from pulling 00:07:34.879 --> 00:07:37.720 in random stuff from the internet that might 00:07:37.720 --> 00:07:40.180 contradict your source. Absolutely crucial for 00:07:40.180 --> 00:07:42.220 customer support, where a wrong guess could be 00:07:42.220 --> 00:07:44.620 a disaster. And then there's iterative refinement. 00:07:45.060 --> 00:07:46.980 No human gets it perfect on the first draft. 00:07:47.100 --> 00:07:49.519 And neither does the AI. So OpenAI uses this 00:07:49.519 --> 00:07:51.959 loop where you build an editor right into the 00:07:51.959 --> 00:07:54.560 prompt. You ask it to write a draft, then immediately 00:07:54.560 --> 00:07:57.199 critique itself, and then rewrite it based on 00:07:57.199 --> 00:07:59.819 that critique. Whoa. Imagine scaling that to 00:07:59.819 --> 00:08:02.860 a billion queries. Having an editor built right 00:08:02.860 --> 00:08:05.560 into every single draft. The quality jump from 00:08:05.560 --> 00:08:07.939 iteration one to iteration three is just huge. 00:08:08.100 --> 00:08:11.519 So what is the key outcome difference between 00:08:11.519 --> 00:08:14.480 iteration one and iteration three? After that 00:08:14.480 --> 00:08:17.079 self -critique loop, the writing becomes dramatically 00:08:17.079 --> 00:08:19.759 sharper and honestly more human sounding. For 00:08:19.759 --> 00:08:21.779 technique eight, Google brain researchers found 00:08:21.779 --> 00:08:24.399 we have to flip the rules. It's called constraint 00:08:24.399 --> 00:08:26.879 first prompting. Right. Historically, we put 00:08:26.879 --> 00:08:29.079 the rules at the end. We say summarize this article, 00:08:29.279 --> 00:08:31.649 make it funny in under 500 words. But by the 00:08:31.649 --> 00:08:34.190 time the AI gets to the constraints, it's already 00:08:34.190 --> 00:08:36.690 started planning the answer. The plan is already 00:08:36.690 --> 00:08:40.110 in motion. You have to flip it. List your hard 00:08:40.110 --> 00:08:43.870 constraints first. Must be under 200 words. Must 00:08:43.870 --> 00:08:46.830 not use the word delve. Then you list your soft 00:08:46.830 --> 00:08:49.830 preferences, like use a funny tone. Why is it 00:08:49.830 --> 00:08:51.649 so much better to put the constraints at the 00:08:51.649 --> 00:08:53.789 very start of the prompt? Because when the AI 00:08:53.789 --> 00:08:56.730 knows the rules first, it plans the entire output 00:08:56.730 --> 00:08:59.250 to fit those precise rules right from the beginning. 00:08:59.509 --> 00:09:02.330 Okay, technique nine. Multi -perspective prompting 00:09:02.330 --> 00:09:05.730 is inspired by Anthropic's work on reducing bias. 00:09:06.110 --> 00:09:08.389 Instead of asking for one answer, you ask for 00:09:08.389 --> 00:09:10.649 three different perspectives on a topic. This 00:09:10.649 --> 00:09:13.210 forces the model to explore the full semantic 00:09:13.210 --> 00:09:15.850 space, which gives you a smarter, fairer, more 00:09:15.850 --> 00:09:18.450 balanced answer. So instead of analyze remote 00:09:18.450 --> 00:09:20.870 work, you'd ask it to analyze it from three angles. 00:09:21.309 --> 00:09:23.950 Perspective one, the employee focus on happiness 00:09:23.950 --> 00:09:27.590 and costs. Perspective 2, the boss focus on productivity 00:09:27.590 --> 00:09:30.370 and culture. And perspective 3, the environment 00:09:30.370 --> 00:09:33.690 focus on carbon footprint. And only then do you 00:09:33.690 --> 00:09:36.070 ask it to synthesize a recommendation. It's like 00:09:36.070 --> 00:09:38.309 adversarial planning. It's just brilliant. And 00:09:38.309 --> 00:09:41.970 finally, technique 10, meta -prompting. This 00:09:41.970 --> 00:09:44.210 is the nuclear option. Yeah. This is what the 00:09:44.210 --> 00:09:46.809 red teams at OpenAI use when they need the absolute 00:09:46.809 --> 00:09:49.629 best output for a really complex task. It's genius 00:09:49.629 --> 00:09:51.809 because it solves the main problem we have as 00:09:51.809 --> 00:09:54.929 users. We often don't know how to ask for what 00:09:54.929 --> 00:09:57.470 we want. But the AI knows what input it needs 00:09:57.470 --> 00:10:00.110 to give the best result. So with meta -prompting, 00:10:00.250 --> 00:10:02.610 you ask the AI to write the perfect prompt for 00:10:02.610 --> 00:10:05.309 you. The template is clean. You state your goal, 00:10:05.470 --> 00:10:08.509 like, I need to accomplish X. Then you tell the 00:10:08.509 --> 00:10:11.409 AI, analyze my goal, write the single perfect 00:10:11.409 --> 00:10:14.470 prompt to achieve it. Then execute that perfect 00:10:14.470 --> 00:10:16.940 prompt you just wrote. I use this for a complex 00:10:16.940 --> 00:10:19.279 legal disclaimer. I am not a lawyer. The prompt 00:10:19.279 --> 00:10:21.519 the AI wrote for itself was 10 times better than 00:10:21.519 --> 00:10:23.019 anything I could have come up with. So is this 00:10:23.019 --> 00:10:25.159 basically like hiring a free prompt engineer 00:10:25.159 --> 00:10:27.860 that knows exactly how the LLM works? Absolutely. 00:10:28.179 --> 00:10:30.700 It leverages the AI's knowledge of its own internal 00:10:30.700 --> 00:10:33.299 systems to create the optimal input. So what 00:10:33.299 --> 00:10:35.539 does this all mean? The big takeaway is simple. 00:10:35.720 --> 00:10:38.559 Stop asking simple questions. Start building 00:10:38.559 --> 00:10:40.799 simulators. If you're going to change just a 00:10:40.799 --> 00:10:43.659 few things, stop using those generic you questions. 00:10:43.820 --> 00:10:46.639 Start assigning highly specific personas and 00:10:46.639 --> 00:10:48.799 use chain of verification for anything that involves 00:10:48.799 --> 00:10:51.580 facts. The gap between a beginner and an expert 00:10:51.580 --> 00:10:54.620 here isn't genius. It's just knowing how to talk 00:10:54.620 --> 00:10:57.509 to the machine. You now have the manual. So go 00:10:57.509 --> 00:10:59.570 try one of these. I'd say start with constraint 00:10:59.570 --> 00:11:02.210 -first prompting on your next email. You will 00:11:02.210 --> 00:11:04.389 see an immediate jump in quality. And consider 00:11:04.389 --> 00:11:06.830 this as you start building your simulators. If 00:11:06.830 --> 00:11:09.389 the AI performs better when simulating a single 00:11:09.389 --> 00:11:12.509 20 -year expert. What happens if you force it 00:11:12.509 --> 00:11:15.190 to simulate a committee of experts who are intentionally 00:11:15.190 --> 00:11:17.610 designed to argue with each other before they 00:11:17.610 --> 00:11:18.370 reach a consensus?