WEBVTT 00:00:00.000 --> 00:00:01.960 You know, a lot of us have had this exact feeling. 00:00:02.720 --> 00:00:05.179 You open up a powerful AI tool, something like 00:00:05.179 --> 00:00:08.140 ChatGPT or Claude, you type in a question, maybe 00:00:08.140 --> 00:00:10.599 something complex, something important, and the 00:00:10.599 --> 00:00:12.279 answer that comes back is just, well, it's flat. 00:00:13.570 --> 00:00:15.470 Generic. Yeah, it's generic. It doesn't sound 00:00:15.470 --> 00:00:17.070 like your company It definitely doesn't sound 00:00:17.070 --> 00:00:18.910 like you hear this all the time and people blame 00:00:18.910 --> 00:00:21.230 the tech They think the models themselves just 00:00:21.230 --> 00:00:24.129 don't have real insight But the core discovery 00:00:24.129 --> 00:00:25.789 from the sources you shared about the art of 00:00:25.789 --> 00:00:28.429 asking is it's something totally different Okay, 00:00:28.570 --> 00:00:31.230 let's unpack that because the sources are suggesting 00:00:31.230 --> 00:00:34.210 the problem isn't the AI's intelligence at all 00:00:34.210 --> 00:00:37.049 It's our interaction style. It's that we're talking 00:00:37.049 --> 00:00:40.030 to this incredibly powerful tool like it's a 00:00:40.030 --> 00:00:42.939 search bar Exactly. Welcome to the deep dive. 00:00:43.359 --> 00:00:45.039 If you treat AI like Google, you're going to 00:00:45.039 --> 00:00:47.420 get generalized Google results. We have to start 00:00:47.420 --> 00:00:50.060 treating it like a specialized teammate. I mean, 00:00:50.079 --> 00:00:53.500 think of it like this. It's an eager, brilliant, 00:00:54.140 --> 00:00:58.020 super fast intern who is just desperate to please 00:00:58.020 --> 00:01:01.479 you. But. And this is the key part, an intern 00:01:01.479 --> 00:01:04.400 who absolutely hates saying, I don't know. That 00:01:04.400 --> 00:01:07.879 eager intern analogy feels so right. So our mission 00:01:07.879 --> 00:01:10.840 today is to distill this massive shift in approach. 00:01:11.140 --> 00:01:13.599 We're moving from treating AI like an oracle 00:01:13.599 --> 00:01:17.480 to treating it as this highly capable but kind 00:01:17.480 --> 00:01:19.719 of naive collaborator. And we're going to dive 00:01:19.719 --> 00:01:22.659 into five simple, really powerful techniques 00:01:22.659 --> 00:01:25.299 that experts use to get intelligent, personalized, 00:01:25.739 --> 00:01:28.439 and most importantly, reliably accurate results. 00:01:28.840 --> 00:01:31.799 This is all about something we call context engineering. 00:01:32.079 --> 00:01:34.719 Let's start right there. What is the fundamental 00:01:34.719 --> 00:01:36.700 problem we're trying to solve with this idea 00:01:36.700 --> 00:01:38.810 of context engineering? It's that eager interim 00:01:38.810 --> 00:01:41.769 problem you just mentioned. The AI is fundamentally 00:01:41.769 --> 00:01:44.209 programmed for helpfulness. So in its training, 00:01:44.450 --> 00:01:47.430 providing an answer, any answer, is prioritized 00:01:47.430 --> 00:01:49.969 way above accuracy or even honesty. So if you 00:01:49.969 --> 00:01:52.650 give it vague instructions, it has to guess the 00:01:52.650 --> 00:01:54.829 context because it just can't tolerate silence. 00:01:54.950 --> 00:01:57.430 It has to respond. And that guessing game, that's 00:01:57.430 --> 00:01:59.409 what leads to the biggest headache in using AI 00:01:59.409 --> 00:02:01.730 today. That's right. If the AI doesn't have the 00:02:01.730 --> 00:02:03.790 context that's in your head, it just makes things 00:02:03.790 --> 00:02:06.010 up. I mean, that's the literal definition of 00:02:06.010 --> 00:02:08.810 hallucinations. The AI is generating fiction 00:02:08.810 --> 00:02:10.870 because it's trying so hard to please you with 00:02:10.870 --> 00:02:13.770 a response. It just, it fills in the gaps. So 00:02:13.770 --> 00:02:16.370 if prompt engineering sounds a little too technical 00:02:16.370 --> 00:02:18.909 for people, we should think of context engineering 00:02:18.909 --> 00:02:22.569 as just bridging that gap. Bridging the gap between 00:02:22.569 --> 00:02:25.030 what's in our brain and what the AI needs to 00:02:25.030 --> 00:02:27.990 know to actually succeed. Precisely. If I walk 00:02:27.990 --> 00:02:30.610 up to you and just say, write a budget, you have 00:02:30.610 --> 00:02:33.250 to know who it's for, what the goal is, the time 00:02:33.250 --> 00:02:35.590 frame. Context engineering is just making sure 00:02:35.590 --> 00:02:37.750 you transfer all those constraints to the AI 00:02:37.750 --> 00:02:40.909 before it starts writing. So if that eager intern 00:02:40.909 --> 00:02:44.139 is so prone to guessing when we're vague, What's 00:02:44.139 --> 00:02:46.840 the biggest risk of providing insufficient context? 00:02:47.099 --> 00:02:49.699 Guesswork directly leads to hallucinations, because 00:02:49.699 --> 00:02:52.719 the AI will simply invent details to satisfy 00:02:52.719 --> 00:02:55.460 the request. That makes perfect sense. Okay, 00:02:55.759 --> 00:02:57.539 let's get into the first technique, which seems 00:02:57.539 --> 00:03:00.460 to directly address the speed and, frankly, the 00:03:00.460 --> 00:03:02.840 recklessness of this intern. Chain of thought 00:03:02.840 --> 00:03:06.349 reasoning. This is so critical. When an AI generates 00:03:06.349 --> 00:03:09.370 a response, it's not planning a full essay first. 00:03:09.669 --> 00:03:12.949 It's predicting one single word at a time based 00:03:12.949 --> 00:03:14.969 on the words that came before it. It's like speaking 00:03:14.969 --> 00:03:17.710 without really thinking deeply. Right. It's like 00:03:17.710 --> 00:03:19.530 a student just shouting out the first answer 00:03:19.530 --> 00:03:22.030 that pops into their head for a hard math problem. 00:03:22.110 --> 00:03:24.770 And that's often the wrong one. We need to force 00:03:24.770 --> 00:03:27.259 it to be more deliberate. We have to force the 00:03:27.259 --> 00:03:30.199 AI to show its work. And you don't need PODE 00:03:30.199 --> 00:03:33.159 for this. You just need to add one magic sentence 00:03:33.159 --> 00:03:35.819 to your instructions. Before you answer, please 00:03:35.819 --> 00:03:37.639 walk me through your thought process step by 00:03:37.639 --> 00:03:39.780 step. And I find this fascinating because the 00:03:39.780 --> 00:03:42.379 sources noted that making the AI write down its 00:03:42.379 --> 00:03:45.080 plan actually makes the final output smarter. 00:03:45.659 --> 00:03:48.039 It literally uses those planning words to predict 00:03:48.039 --> 00:03:50.240 better words that come after. It's like it's 00:03:50.240 --> 00:03:52.300 building a higher quality scaffold for its own 00:03:52.300 --> 00:03:54.610 answer. It really does. Let's use a real -life 00:03:54.610 --> 00:03:57.449 example. Planning a budget trip to Japan. A bad 00:03:57.449 --> 00:04:00.030 prompt is just, plan a five -day trip to Tokyo. 00:04:00.189 --> 00:04:03.090 Make it cheap. The AI will just spit out a generic 00:04:03.090 --> 00:04:05.250 itinerary. You'll get expensive hotels, some 00:04:05.250 --> 00:04:08.430 tourist traps. It's useless. But the good prompt, 00:04:08.669 --> 00:04:11.189 the chain of thought prompt, specifies the details 00:04:11.189 --> 00:04:13.490 and then demands that planning phase. Something 00:04:13.490 --> 00:04:16.620 like... I need a five -day itinerary for Tokyo 00:04:16.620 --> 00:04:19.420 on a tight budget. Before you create the itinerary, 00:04:19.639 --> 00:04:21.500 please think through the constraints step -by 00:04:21.500 --> 00:04:24.660 -step. Prioritize low -cost transport, research 00:04:24.660 --> 00:04:27.740 cheap but high -quality food, and only suggest 00:04:27.740 --> 00:04:30.379 free activities. Then the AI starts by saying, 00:04:30.459 --> 00:04:33.240 okay, my first constraint is the budget, so I'll 00:04:33.240 --> 00:04:35.959 prioritize subway passes and skip the high -speed 00:04:35.959 --> 00:04:38.620 rail. For meals, I'll look at convenience stores. 00:04:39.319 --> 00:04:41.259 That thought process locks in the constraints, 00:04:41.339 --> 00:04:43.680 so the final itinerary is something you can actually 00:04:43.680 --> 00:04:46.040 use. And here's a bit of a vulnerable admission. 00:04:46.500 --> 00:04:49.120 I still wrestle with prompt drift myself. You 00:04:49.120 --> 00:04:51.160 know, when the final itinerary looks good, it's 00:04:51.160 --> 00:04:53.399 just so easy to skip reading that initial thought 00:04:53.399 --> 00:04:55.019 process. Yeah. But we have to read the thinking 00:04:55.019 --> 00:04:56.899 part, too, right? Oh, absolutely. Because if 00:04:56.899 --> 00:04:59.839 the AI makes a bad assumption in its planning, 00:05:00.160 --> 00:05:02.240 like, say, it assumes you're flying into an airport 00:05:02.240 --> 00:05:04.420 that you're not, you have to catch that flawed 00:05:04.420 --> 00:05:06.620 premise. Right. If you only look at the final 00:05:06.620 --> 00:05:09.139 answer, you've completely missed the core problem. 00:05:09.259 --> 00:05:11.639 It allows us to course correct before the AI 00:05:11.639 --> 00:05:15.079 commits fully to a bad direction. Yes. It addresses 00:05:15.079 --> 00:05:17.720 that fundamental flaw of speed over precision. 00:05:17.959 --> 00:05:20.560 Okay, so if chain of thought helps us improve 00:05:20.560 --> 00:05:23.819 the AI's logic, the next big challenge is making 00:05:23.819 --> 00:05:27.519 the AI sound like a person. Specifically, like 00:05:27.519 --> 00:05:30.660 you. This brings us to technique two, few -shot 00:05:30.660 --> 00:05:33.720 prompting. And this one tackles that really common 00:05:33.720 --> 00:05:36.819 mistake of using weak adjectives. I see people 00:05:36.819 --> 00:05:39.899 type, write a professional, engaging, funny email. 00:05:40.170 --> 00:05:42.470 Those words are subjective. They're basically 00:05:42.470 --> 00:05:44.629 meaningless to an AI that's trying to match a 00:05:44.629 --> 00:05:47.550 specific human voice. Adjectives are weak. Examples 00:05:47.550 --> 00:05:50.509 are strong. The AI is a phenomenal copycat machine. 00:05:50.569 --> 00:05:52.589 It's just brilliant at pattern matching. If you 00:05:52.589 --> 00:05:54.290 want it to write like you, you have to give it 00:05:54.290 --> 00:05:56.170 the data, the patterns of what you sound like. 00:05:56.430 --> 00:05:58.850 And the technique is surprisingly simple. You 00:05:58.850 --> 00:06:00.790 just find three to five examples of your own 00:06:00.790 --> 00:06:03.230 successful writing, emails that got a great reply, 00:06:03.529 --> 00:06:05.310 short reports you were proud of, and you paste 00:06:05.310 --> 00:06:07.449 them right into the chat. And the instruction 00:06:07.449 --> 00:06:11.569 is key. You ask the AI to Analyze these examples, 00:06:11.889 --> 00:06:14.290 then write a new email using the same style, 00:06:14.550 --> 00:06:17.230 tone, and format. It's like it's using your own 00:06:17.230 --> 00:06:19.850 work as a microtraining set. Let's use that example 00:06:19.850 --> 00:06:22.449 of writing a statistical topology letter. A bad 00:06:22.449 --> 00:06:25.089 prompt is just, write an apology email. Be polite. 00:06:25.290 --> 00:06:28.550 The result. Dear valued customer, we are deeply 00:06:28.550 --> 00:06:30.889 sorry for the inconvenience. It's so robotic. 00:06:30.970 --> 00:06:33.269 It just alienates people. But the good prompts 00:06:33.269 --> 00:06:35.569 gives it short, friendly examples. Something 00:06:35.569 --> 00:06:37.329 like, hey, Sarah, I saw your order was late. 00:06:37.389 --> 00:06:39.709 I am so sorry about that. And the AI picks up 00:06:39.709 --> 00:06:41.490 on the short sentences, the casual greeting, 00:06:41.529 --> 00:06:43.970 and it avoids all that corporate jargon. The 00:06:43.970 --> 00:06:46.509 apology it writes is personal and actually effective. 00:06:46.750 --> 00:06:48.610 OK, but let me ask a practical question here. 00:06:49.069 --> 00:06:51.170 If I paced in three or five paragraphs of my 00:06:51.170 --> 00:06:52.970 best writing, doesn't that make the prompt kind 00:06:52.970 --> 00:06:55.399 of massive? I mean, does that get slow or expensive 00:06:55.399 --> 00:06:58.120 to run every single time? That's a great point, 00:06:58.240 --> 00:07:00.360 especially if you're hitting your context window 00:07:00.360 --> 00:07:03.040 limit. But remember, the examples don't have 00:07:03.040 --> 00:07:05.319 to be long. They just need to be representative. 00:07:05.920 --> 00:07:08.839 A few strong short paragraphs are usually enough 00:07:08.839 --> 00:07:11.560 for the AI to pick up on your linguistic fingerprint, 00:07:11.959 --> 00:07:13.920 your sentence structure, your vocabulary, your 00:07:13.920 --> 00:07:16.920 rhythm. The value of getting a perfect output 00:07:16.920 --> 00:07:19.639 really outweighs that slight increase in the 00:07:19.639 --> 00:07:21.850 input data. And what about the opposite? Can 00:07:21.850 --> 00:07:23.930 you show it what you don't want? Absolutely. 00:07:24.149 --> 00:07:26.589 That's the advanced tip here. Using a negative 00:07:26.589 --> 00:07:29.269 example, you paste in a piece of writing you 00:07:29.269 --> 00:07:32.069 hate, maybe a super stiff formal corporate email, 00:07:32.430 --> 00:07:35.050 and you explicitly tell the AI, do not write 00:07:35.050 --> 00:07:37.250 like this, avoid this style. It learns just as 00:07:37.250 --> 00:07:39.850 quickly what to avoid. So how many examples are 00:07:39.850 --> 00:07:41.910 generally effective for the AI to pick up on 00:07:41.910 --> 00:07:44.589 a specific style or voice? We're aiming for three 00:07:44.589 --> 00:07:46.970 to five strong examples of your best work. That's 00:07:46.970 --> 00:07:49.730 a tangible number. Oh. OK, moving to technique 00:07:49.730 --> 00:07:52.769 three. Reverse prompting. This feels vital for 00:07:52.769 --> 00:07:55.189 those times when, as the user, you don't even 00:07:55.189 --> 00:07:56.769 know what information is important to share. 00:07:57.129 --> 00:07:59.170 This happens all the time with health and safety. 00:07:59.769 --> 00:08:02.769 Remember the eager intern. If you ask for a workout 00:08:02.769 --> 00:08:05.089 plan but you forget to mention an old me injury 00:08:05.089 --> 00:08:07.850 or that you only have dumbbells at home, the 00:08:07.850 --> 00:08:10.709 AI just guesses. And it might give you this generic, 00:08:11.149 --> 00:08:12.829 intense plan that could actually hurt you because 00:08:12.829 --> 00:08:15.910 it's missing that vital context. So we flip the 00:08:15.910 --> 00:08:18.720 script. Instead of us giving vague instructions, 00:08:19.220 --> 00:08:22.000 we force the AI to become the smart interviewer. 00:08:22.300 --> 00:08:25.220 The key sentence is so simple. Before you generate 00:08:25.220 --> 00:08:27.600 the response, please ask me any questions you 00:08:27.600 --> 00:08:30.720 need to know to do the best job possible. This 00:08:30.720 --> 00:08:33.860 forces it out of guessing mode and into information 00:08:33.860 --> 00:08:36.440 gathering mode. We can use that gardening scenario 00:08:36.440 --> 00:08:38.799 from the sources. A bad prompt is just, tell 00:08:38.799 --> 00:08:41.649 me what to plant in my garden. The AI guesses 00:08:41.649 --> 00:08:43.730 you're in a temperate zone and recommends tomatoes, 00:08:44.330 --> 00:08:46.110 but you live in a cold climate where they'll 00:08:46.110 --> 00:08:48.669 die instantly. Right, but the good prompt makes 00:08:48.669 --> 00:08:52.289 the AI ask five clarifying questions about your 00:08:52.289 --> 00:08:54.690 location, hours of sunlight, your maintenance 00:08:54.690 --> 00:08:57.429 level, whether you want edible plants. You answer 00:08:57.429 --> 00:08:59.250 those five questions and the plan it gives you 00:08:59.250 --> 00:09:01.429 is perfectly tailored and will actually succeed. 00:09:01.629 --> 00:09:03.370 So it seems like it saves a lot of time. You 00:09:03.370 --> 00:09:05.330 just bypass several rounds of bad answers. It 00:09:05.330 --> 00:09:07.769 absolutely does. It moves straight to an informed 00:09:07.769 --> 00:09:10.129 answer, which increases the speed of execution 00:09:10.129 --> 00:09:13.409 by eliminating all those useless revisions. Technique 00:09:13.409 --> 00:09:16.809 4 is incredibly powerful. And it goes back to 00:09:16.809 --> 00:09:19.950 the AI's massive training data. Let's use the 00:09:19.950 --> 00:09:24.309 library analogy. If AI is this huge digital library 00:09:24.309 --> 00:09:26.750 with every book in the world, from comic books 00:09:26.750 --> 00:09:30.710 to PhD textbooks, a generic question gives you 00:09:30.710 --> 00:09:32.529 the average of everything. And we don't want 00:09:32.529 --> 00:09:34.450 the average answer. We want the expert answer. 00:09:34.850 --> 00:09:36.769 So when you assign a role, you're telling the 00:09:36.769 --> 00:09:39.929 AI which section of the library to look in. You're 00:09:39.929 --> 00:09:42.309 telling it to ignore the blogs and focus only 00:09:42.309 --> 00:09:44.769 on the specialized high authority texts for that 00:09:44.769 --> 00:09:47.220 role. The structure is simple. but really effective. 00:09:47.740 --> 00:09:50.480 You are an expert job title with 20 years of 00:09:50.480 --> 00:09:52.820 experience. You think like a famous person. This 00:09:52.820 --> 00:09:55.600 immediately filters the knowledge the AI uses. 00:09:55.860 --> 00:09:58.059 Let's say you're checking a resume. A bad prompt 00:09:58.059 --> 00:10:01.100 is just, check my resume for mistakes. You'll 00:10:01.100 --> 00:10:03.139 get basic spell check, maybe some grammar notes. 00:10:03.460 --> 00:10:05.960 That's the average answer. But the good role 00:10:05.960 --> 00:10:09.379 assigned prompt is, you are a strict hiring manager 00:10:09.379 --> 00:10:13.379 at a top tech company. You reject 99 % of resumes. 00:10:13.879 --> 00:10:17.990 Tell me why you would reject mine. The AI completely 00:10:17.990 --> 00:10:21.110 changes its persona. It will ignore the typos 00:10:21.110 --> 00:10:22.929 and tell you your bullet points are too vague, 00:10:23.129 --> 00:10:25.610 that you lack quantifiable achievements, that 00:10:25.610 --> 00:10:27.570 your structure isn't optimized for automated 00:10:27.570 --> 00:10:30.730 tracking systems. That kind of high -stakes feedback 00:10:30.730 --> 00:10:33.250 is invaluable. And you can assign almost any 00:10:33.250 --> 00:10:36.200 persona. a tough negotiation expert, certified 00:10:36.200 --> 00:10:38.779 financial planner, even a friendly kindergarten 00:10:38.779 --> 00:10:41.399 teacher to simplify a really complicated topic. 00:10:41.860 --> 00:10:44.279 Oh, and just imagine the potential there. When 00:10:44.279 --> 00:10:46.620 you can instantly tap into the synthesized knowledge 00:10:46.620 --> 00:10:49.139 of a billion textbooks just by assigning a job 00:10:49.139 --> 00:10:52.100 title, you're accessing true, focused expertise 00:10:52.100 --> 00:10:54.600 instantly. You're bypassing all the noise. So 00:10:54.600 --> 00:10:56.659 what is the critical difference between the generic 00:10:56.659 --> 00:10:59.379 answer and the role assigned answer? Role Assignment 00:10:59.379 --> 00:11:02.200 accesses specific, high -quality knowledge instead 00:11:02.200 --> 00:11:04.519 of providing an average of all possible information. 00:11:04.899 --> 00:11:07.559 Okay, finally we get to Technique 5, Role Playing. 00:11:08.279 --> 00:11:10.379 And this feels like the ultimate application 00:11:10.379 --> 00:11:15.100 of Persona, turning the AI into a conversation 00:11:15.100 --> 00:11:17.840 simulator. Right. Pilots use flight simulators 00:11:17.840 --> 00:11:19.980 to crash planes safely until they know exactly 00:11:19.980 --> 00:11:23.360 how to handle turbulence. We can use AI as a 00:11:23.360 --> 00:11:25.440 conversation simulator to practice high -stakes 00:11:25.440 --> 00:11:28.679 talks, asking for a raise, delivering bad news, 00:11:28.879 --> 00:11:30.820 so you don't mess up the real interaction. The 00:11:30.820 --> 00:11:33.039 sources outline a really powerful three -window 00:11:33.039 --> 00:11:35.580 system for this. Let's use that example of the 00:11:35.580 --> 00:11:38.279 tough rent negotiation with our fictional landlord, 00:11:38.480 --> 00:11:41.159 Mr. Smith. Okay, so step A happens in chat one, 00:11:41.480 --> 00:11:44.000 the profiler. You build the character, you say, 00:11:44.320 --> 00:11:46.419 Mr. Smith is stubborn, he talks loudly, he's 00:11:46.419 --> 00:11:48.639 60 years old, and he always complains about property 00:11:48.639 --> 00:11:51.679 taxes. You use the AI to really flesh out that 00:11:51.679 --> 00:11:53.700 psychological profile. And then step B is chat 00:11:53.700 --> 00:11:56.580 two, the simulator. You paste that detailed description 00:11:56.580 --> 00:11:59.139 in, you set the scene, and you start the conversation. 00:11:59.259 --> 00:12:01.759 You are Mr. Smith. Stay in character. Ring ring. 00:12:02.120 --> 00:12:04.519 And when you make your offer, the AI, acting 00:12:04.519 --> 00:12:06.639 as Mr. Smith, will argue back. It'll challenge 00:12:06.639 --> 00:12:08.799 your assumptions, it'll maintain that stubborn 00:12:08.799 --> 00:12:11.120 persona. This is where you get to crash safely. 00:12:11.320 --> 00:12:14.320 You can try five different opening lines. You 00:12:14.320 --> 00:12:16.720 learn what makes him angry, what diffuses him, 00:12:16.820 --> 00:12:19.200 what objections he's going to raise. And then 00:12:19.200 --> 00:12:21.799 step C is the key. You integrate a technique 00:12:21.799 --> 00:12:25.399 we mentioned earlier. It's chat three, the coach 00:12:25.399 --> 00:12:28.820 and the Russian judge. You copy the entire transcript 00:12:28.820 --> 00:12:31.159 of your negotiation attempt and you paste it 00:12:31.159 --> 00:12:33.980 into a new chat window. Right. And now you assign 00:12:33.980 --> 00:12:38.379 a new specific role. You are a world -class negotiation 00:12:38.379 --> 00:12:41.539 coach. You must grade my performance, but you're 00:12:41.539 --> 00:12:44.879 a Russian judge. do not be polite, grade me on 00:12:44.879 --> 00:12:48.080 a scale of 1 to 10, and be brutal. Why is that 00:12:48.080 --> 00:12:50.279 separate coach step necessary rather than just 00:12:50.279 --> 00:12:52.580 asking the simulator for feedback? It ensures 00:12:52.580 --> 00:12:55.460 the analysis is objective, separating the character 00:12:55.460 --> 00:12:58.460 simulation from the performance review. Got it. 00:12:58.639 --> 00:13:00.840 And because the AI is explicitly told not to 00:13:00.840 --> 00:13:03.059 be nice, it will point out the logical holes, 00:13:03.139 --> 00:13:05.299 the emotional traps you fell into. That truth 00:13:05.299 --> 00:13:07.440 is way more helpful than just simple encouragement. 00:13:07.659 --> 00:13:10.100 Exactly. We usually only get one shot at a hard 00:13:10.100 --> 00:13:12.250 conversation. With this method, you can... Practice 00:13:12.250 --> 00:13:14.429 10 times until you're calm, you're prepared, 00:13:14.750 --> 00:13:16.649 and you're ready for the real call. Okay, to 00:13:16.649 --> 00:13:19.190 wrap this deep dive up, let's hit one final technical 00:13:19.190 --> 00:13:22.669 tip. This is the new chat rule. It's so simple, 00:13:22.750 --> 00:13:25.289 but it's crucial. Keep your workspace clean. 00:13:25.580 --> 00:13:28.720 The AI relies on the previous conversation history 00:13:28.720 --> 00:13:31.480 for context. If you change the topic, just start 00:13:31.480 --> 00:13:33.539 a fresh chat window immediately. Yeah, we've 00:13:33.539 --> 00:13:36.120 all done it. We ruin a perfectly built persona 00:13:36.120 --> 00:13:39.279 by asking an unrelated question. You don't want 00:13:39.279 --> 00:13:42.100 to break your serious CEO negotiations simulator 00:13:42.100 --> 00:13:44.059 by asking it about your fantasy football picks 00:13:44.059 --> 00:13:46.610 in the same thread. A long, messy history just 00:13:46.610 --> 00:13:48.870 confuses the AI. It's like using one notebook 00:13:48.870 --> 00:13:52.049 for math, history, and cooking. One focused task, 00:13:52.309 --> 00:13:55.330 one clean chat. That makes sense. OK, let's quickly 00:13:55.330 --> 00:13:57.669 summarize these five techniques we covered. One, 00:13:58.129 --> 00:14:00.669 chain of thought. Force the AI to think step 00:14:00.669 --> 00:14:03.889 by step for smarter, more logical outputs. Two, 00:14:04.549 --> 00:14:07.629 few -shot prompting. Provide three to five examples 00:14:07.629 --> 00:14:09.789 of your work for perfect style and tone matching. 00:14:10.169 --> 00:14:13.519 Three, Reverse prompting. Ask the AI to interview 00:14:13.519 --> 00:14:16.559 you to get that critical, perfect context. Four, 00:14:16.980 --> 00:14:19.919 assign a role. Make the AI an expert to access 00:14:19.919 --> 00:14:23.279 high quality, focused advice. And five, role 00:14:23.279 --> 00:14:25.779 playing. Use that three window system and the 00:14:25.779 --> 00:14:28.120 Russian judge to simulate and practice high stakes 00:14:28.120 --> 00:14:30.919 conversations safely. The core idea really is 00:14:30.919 --> 00:14:33.419 the same across all of these. AI is the best 00:14:33.419 --> 00:14:35.059 teammate you've ever had, but it only shines 00:14:35.059 --> 00:14:37.340 when you give it precise instructions. So stop 00:14:37.340 --> 00:14:39.299 talking to it like an omniscient database. Start 00:14:39.299 --> 00:14:41.720 talking to it like the smart, eager intern it 00:14:41.720 --> 00:14:44.039 is. An intern who needs clear constraints and 00:14:44.039 --> 00:14:46.039 boundaries. And our challenge to you is simple. 00:14:46.320 --> 00:14:48.179 Don't try to master all five techniques today. 00:14:48.299 --> 00:14:50.679 Just pick one. Maybe use FewShop for an important 00:14:50.679 --> 00:14:52.759 email you have to write this afternoon. Or try 00:14:52.759 --> 00:14:54.559 Chain of Thought for a complex decision you're 00:14:54.559 --> 00:14:56.600 trying to map out. Just try one right now. And 00:14:56.600 --> 00:14:58.379 as a final provocative thought for you to explore 00:14:58.379 --> 00:15:01.419 this week, if we have to explicitly tell the 00:15:01.419 --> 00:15:04.279 AI to be brutal and honest, how does that fundamental 00:15:04.279 --> 00:15:06.580 AI tendency, the need to be helpful over being 00:15:06.580 --> 00:15:09.519 truthful, reflect broader human behavioral tendencies 00:15:09.519 --> 00:15:10.840 in high pressure situations?