WEBVTT 00:00:00.000 --> 00:00:02.560 The digital world is absolutely buzzing about 00:00:02.560 --> 00:00:05.780 GPT -5 right now. Some people are hailing it 00:00:05.780 --> 00:00:09.900 as this monumental leap forward for AI, truly 00:00:09.900 --> 00:00:12.259 transformative. Yeah, it's incredible, no doubt. 00:00:12.500 --> 00:00:15.279 But there's also this undercurrent I'm seeing, 00:00:15.400 --> 00:00:18.359 a bit of frustration for many users. Frustration? 00:00:18.460 --> 00:00:20.579 How so? Well, it's been like, you know, we've 00:00:20.579 --> 00:00:22.739 just bought this Formula One race car. It's an 00:00:22.739 --> 00:00:24.620 engineering marvel, capable of amazing speed 00:00:24.620 --> 00:00:27.129 and precision. OK. But we're still trying to 00:00:27.129 --> 00:00:29.269 drive it like our old automatic sedan. We're 00:00:29.269 --> 00:00:30.949 just not using it to its full potential yet. 00:00:31.129 --> 00:00:35.149 That's a perfect analogy and a really insightful 00:00:35.149 --> 00:00:38.030 way to frame it, actually. So if you felt that 00:00:38.030 --> 00:00:40.670 GPT -5's responses have been maybe inconsistent 00:00:40.670 --> 00:00:42.689 or you're not quite getting those breakthrough 00:00:42.689 --> 00:00:45.100 results you expected, well, you are definitely 00:00:45.100 --> 00:00:47.859 not alone. This deep dive is really for you. 00:00:47.939 --> 00:00:49.659 We've pulled together a whole stack of sources, 00:00:49.939 --> 00:00:52.359 everything from early access tester insights 00:00:52.359 --> 00:00:56.200 to the official guidance to break down 11 proven 00:00:56.200 --> 00:00:58.880 prompting techniques. Our mission today really 00:00:58.880 --> 00:01:02.380 is to transform your interactions with GPT -5. 00:01:02.799 --> 00:01:05.700 We want to move them away from being a game of 00:01:05.700 --> 00:01:08.390 chance. where you're just hoping for a good outcome 00:01:08.390 --> 00:01:11.010 into something more like precision engineering. 00:01:11.569 --> 00:01:13.530 So we'll start with the foundational stuff, the 00:01:13.530 --> 00:01:15.750 pillars that kind of redefine how we interact 00:01:15.750 --> 00:01:17.950 with it. Then we'll move to some more advanced 00:01:17.950 --> 00:01:21.290 strategies and finally wrap up with how you can 00:01:21.290 --> 00:01:24.200 combine these for... really world -class results. 00:01:24.760 --> 00:01:27.819 OK, sounds good. Let's unpack this first concept 00:01:27.819 --> 00:01:30.620 then. GPT -5 isn't just another update, is it? 00:01:30.680 --> 00:01:33.159 It feels like a genuine paradigm shift. Totally. 00:01:33.379 --> 00:01:35.799 OpenAI engineered this model for what they're 00:01:35.799 --> 00:01:38.819 calling surgical instruction following. What 00:01:38.819 --> 00:01:41.319 does that actually mean in practice? Well, what's 00:01:41.319 --> 00:01:43.060 fascinating here, and this isn't just marketing 00:01:43.060 --> 00:01:46.620 fluff, right? It means GPT -5 adheres strictly 00:01:46.620 --> 00:01:49.879 almost like literally to what you say. It doesn't 00:01:49.879 --> 00:01:52.909 make assumptions or try too hard to infer your 00:01:52.909 --> 00:01:56.329 intent like older models might have. And a prominent 00:01:56.329 --> 00:01:59.349 AI researcher, someone who got early access, 00:01:59.730 --> 00:02:01.489 put pretty starkly, they said, prompts don't 00:02:01.489 --> 00:02:03.950 just influence results anymore. They make or 00:02:03.950 --> 00:02:06.909 break them entirely. That's a powerful statement. 00:02:07.129 --> 00:02:10.349 It suggests the bar for precision has been raised 00:02:10.349 --> 00:02:13.969 significantly. It really has. And this strict 00:02:13.969 --> 00:02:17.169 adherence, it's a huge departure from the more 00:02:17.169 --> 00:02:20.129 forgiving AI tools we've kind of gotten used 00:02:20.129 --> 00:02:23.129 to, hasn't it? Absolutely. Previous models often 00:02:23.129 --> 00:02:26.009 tried to understand vague conversational requests, 00:02:26.610 --> 00:02:27.870 sort of fill in the blanks for you. All right, 00:02:27.990 --> 00:02:30.530 it tried to guess what you meant. Exactly. GPT 00:02:30.530 --> 00:02:32.710 -5, though, it brings back the art of prompt 00:02:32.710 --> 00:02:35.189 engineering, makes it an absolutely essential 00:02:35.189 --> 00:02:38.400 skill again. So when we talk about surgical instruction, 00:02:38.719 --> 00:02:41.360 how does that fundamental shift actually change 00:02:41.360 --> 00:02:44.680 our day -to -day interaction with the AI? Well, 00:02:44.800 --> 00:02:47.439 it means the era of just casual conversational 00:02:47.439 --> 00:02:49.840 prompting is kind of over for serious use cases. 00:02:50.539 --> 00:02:52.080 Instead of thinking of it as a chat partner, 00:02:52.800 --> 00:02:55.599 you need to think of GPT -5 as this highly capable 00:02:55.599 --> 00:02:58.719 but incredibly literal expert. Like a specialist. 00:02:59.099 --> 00:03:01.319 Precisely. You wouldn't give a surgeon vague 00:03:01.319 --> 00:03:04.460 instructions, right? You give precise ones. This 00:03:04.460 --> 00:03:07.020 model demands that exact same level of rigor. 00:03:07.280 --> 00:03:09.159 Okay, that makes sense. Treat it like the precision 00:03:09.159 --> 00:03:12.580 instrument it is. You got it. So our first foundational 00:03:12.580 --> 00:03:15.639 pillar in this new world, it sounds incredibly 00:03:15.639 --> 00:03:19.120 simple, but it's remarkably impactful. You need 00:03:19.120 --> 00:03:22.719 to explicitly tell GPT -5 to think more. Just 00:03:22.719 --> 00:03:25.060 tell it to think harder. Yeah, exactly. Phrases 00:03:25.060 --> 00:03:27.240 that might sound almost, I don't know, whimsical, 00:03:27.400 --> 00:03:29.879 like, take a deep breath and think through this 00:03:29.879 --> 00:03:32.639 step by step, or analyze this request from first 00:03:32.639 --> 00:03:36.020 principles before formulating a response. They 00:03:36.020 --> 00:03:38.060 actually work. And not just a little bit. It's 00:03:38.060 --> 00:03:40.400 not about being polite. It's about directing 00:03:40.400 --> 00:03:42.460 the model's internal process. OK, so why does 00:03:42.460 --> 00:03:44.900 that work? Why does explicitly asking for deep 00:03:44.900 --> 00:03:47.560 thought actually make the model perform better? 00:03:48.000 --> 00:03:50.039 It seems to be about allocating computational 00:03:50.039 --> 00:03:53.039 budget. Think of it like this. These large models, 00:03:53.039 --> 00:03:56.719 they operate on a finite budget of processing 00:03:56.719 --> 00:03:59.159 power and internal steps for each response. Right. 00:03:59.280 --> 00:04:02.039 Limited resources. Yeah. Simple, straightforward 00:04:02.039 --> 00:04:05.099 instructions to get a fast, often kind of superficial 00:04:05.099 --> 00:04:08.840 processing path. But when you demand deeper thought, 00:04:09.199 --> 00:04:11.539 you're telling the model, hey, engage your more 00:04:11.539 --> 00:04:14.949 complex reasoning pathways. Ah, okay. It's like 00:04:14.949 --> 00:04:16.810 giving a chess engine more time to think, you 00:04:16.810 --> 00:04:19.329 know. It explores way more possibilities internally 00:04:19.329 --> 00:04:21.310 before it commits to an output. You're basically 00:04:21.310 --> 00:04:24.810 pushing it past the easy, obvious first answer. 00:04:24.939 --> 00:04:27.959 So you're directing it to use more of its underlying 00:04:27.959 --> 00:04:30.459 intelligence to really stretch its capabilities. 00:04:30.579 --> 00:04:33.100 That's interesting. It is. And one experienced 00:04:33.100 --> 00:04:36.019 prompt engineer even recommended an ultra think 00:04:36.019 --> 00:04:38.779 protocol for really critical tasks. An example 00:04:38.779 --> 00:04:41.100 was something like, engage in a deep thinking 00:04:41.100 --> 00:04:43.660 process for a minimum of three minutes. Your 00:04:43.660 --> 00:04:46.680 goal is to produce a world class output. Wow, 00:04:46.819 --> 00:04:48.920 three minutes. That's pretty intense for an AI 00:04:48.920 --> 00:04:50.879 response time. It signals the importance and 00:04:50.879 --> 00:04:53.339 complexity, right? It tells the model to really 00:04:53.339 --> 00:04:55.579 allocate resources. Okay, building on that idea 00:04:55.579 --> 00:04:58.480 of internal processing, another critical technique 00:04:58.480 --> 00:05:01.060 is guiding GPT -5 through a structured planning 00:05:01.060 --> 00:05:03.639 phase before it even starts generating the final 00:05:03.639 --> 00:05:06.819 response. Like an architect creating a meticulous 00:05:06.819 --> 00:05:09.980 blueprint before any construction starts, that's 00:05:09.980 --> 00:05:12.180 the level of upfront planning we should aim for. 00:05:12.579 --> 00:05:15.139 Right, because without a plan, the model can 00:05:15.139 --> 00:05:18.339 easily miss crucial steps or maybe jump to conclusions 00:05:18.339 --> 00:05:22.139 or just deliver a disorganized response. Researchers 00:05:22.139 --> 00:05:24.699 have found that models like GPT -5, the ones 00:05:24.699 --> 00:05:27.680 with strong instruction following skills, they 00:05:27.680 --> 00:05:31.560 really excel when you give them a metacognitive 00:05:31.560 --> 00:05:33.939 framework. Metacognitive framework, like thinking 00:05:33.939 --> 00:05:37.160 about thinking. Exactly. It's basically a blueprint 00:05:37.160 --> 00:05:39.779 or a plan. for how it should think about the 00:05:39.779 --> 00:05:42.699 task before it even begins generating the actual 00:05:42.699 --> 00:05:45.019 answer. You're front loading the intelligence. 00:05:45.220 --> 00:05:47.529 That makes a lot of sense. Planning isn't just 00:05:47.529 --> 00:05:50.250 for humans anymore. Apparently not. Guiding the 00:05:50.250 --> 00:05:52.829 AI through a plan like this just helps prevent 00:05:52.829 --> 00:05:55.850 disorganized kind of haphazard output. So an 00:05:55.850 --> 00:05:57.949 effective structure would involve, what, breaking 00:05:57.949 --> 00:05:59.970 down the request? Yeah, break it down into core 00:05:59.970 --> 00:06:02.850 components, maybe identify information gaps it 00:06:02.850 --> 00:06:05.949 needs to fill, propose a clear execution strategy, 00:06:06.269 --> 00:06:08.790 and then this is important, define specific success 00:06:08.790 --> 00:06:11.540 criteria. You can even ask it to present that 00:06:11.540 --> 00:06:14.060 plan to you for approval first. Oh, interesting. 00:06:14.160 --> 00:06:16.160 Like a check -in. Yeah. Gives you a chance to 00:06:16.160 --> 00:06:18.120 course correct before it burns through its computational 00:06:18.120 --> 00:06:20.060 budget on an output you might not even want. 00:06:20.100 --> 00:06:23.579 It's smart. For complex multi -stage tasks, this 00:06:23.579 --> 00:06:25.899 really is like building an internal mental to 00:06:25.899 --> 00:06:29.300 -do list for the AI. This structured approach 00:06:29.300 --> 00:06:32.379 forces a more deliberate, logical progression. 00:06:32.779 --> 00:06:36.199 It moves it from just improvising to actually 00:06:36.199 --> 00:06:38.939 strategic execution. Strategic execution from 00:06:38.939 --> 00:06:41.920 an AI. OK. Now, this next one, the principle 00:06:41.920 --> 00:06:45.019 of unambiguous specificity. I would argue this 00:06:45.019 --> 00:06:47.300 is the single most important rule to master with 00:06:47.300 --> 00:06:51.040 GPT -5. OK. Be relentlessly explicit about absolutely 00:06:51.040 --> 00:06:54.420 everything. vagueness in this new era, it's truly 00:06:54.420 --> 00:06:56.600 it's kryptonite. And this is where that surgical 00:06:56.600 --> 00:06:59.279 precision really bites, isn't it? Totally. Earlier 00:06:59.279 --> 00:07:02.180 models like GPT -4, maybe they inferred context 00:07:02.180 --> 00:07:04.259 or read between the lines a bit. They tried. 00:07:04.500 --> 00:07:07.680 But GPT -5 interprets instructions with, like 00:07:07.680 --> 00:07:11.000 you said, near literal precision. OpenAI's own 00:07:11.000 --> 00:07:13.399 documentation confirms this, calling its precision 00:07:13.399 --> 00:07:16.259 a double -edged sword. Right. It's incredibly 00:07:16.259 --> 00:07:19.519 powerful. But it demands this unprecedented level 00:07:19.519 --> 00:07:22.439 of clarity and exactness from us, the users. 00:07:22.480 --> 00:07:24.819 So if you're not specific... It won't guess. 00:07:24.959 --> 00:07:26.899 It'll just follow the vague instruction, often 00:07:26.899 --> 00:07:30.160 leading to, well, frustration. How specific do 00:07:30.160 --> 00:07:33.220 we really need to get them for GPT -5 to understand? 00:07:34.100 --> 00:07:36.939 Extremely specific. You need to avoid all inference 00:07:36.939 --> 00:07:39.889 or ambiguity. Assume nothing is implied. Okay, 00:07:39.930 --> 00:07:42.410 give me an example. Like, tone. Right. You can't 00:07:42.410 --> 00:07:44.610 just say, write this in a friendly tone anymore. 00:07:44.709 --> 00:07:46.870 Doesn't work well. Instead, you need to define 00:07:46.870 --> 00:07:49.470 friendly. Okay. You might say something like, 00:07:49.810 --> 00:07:52.529 adopt the persona of a helpful, encouraging mentor 00:07:52.529 --> 00:07:55.370 who uses analogies to explain complex topics. 00:07:55.649 --> 00:07:57.689 The tone should be professional yet accessible. 00:07:57.930 --> 00:08:00.370 Avoiding jargon where possible. See the difference. 00:08:00.490 --> 00:08:02.490 Yeah, that's much more concrete. What about formatting? 00:08:02.829 --> 00:08:05.490 Same deal. Don't just ask for a blog post. Specify 00:08:05.490 --> 00:08:08.279 precisely. Generate a 1 ,500 word blog post, 00:08:08.620 --> 00:08:10.800 format it in Markdown. It needs an H1 title, 00:08:11.180 --> 00:08:13.879 three H2 subheadings, use bullet points to summarize 00:08:13.879 --> 00:08:16.620 key ideas, and wrap it up with a concluding paragraph 00:08:16.620 --> 00:08:18.839 title, final thoughts. Wow, OK. Detail, detail, 00:08:19.040 --> 00:08:21.720 detail. The more detail up front, the less back 00:08:21.720 --> 00:08:23.439 and forth and reprompting you'll need later. 00:08:23.740 --> 00:08:25.899 Saves time in the end. Right. Building on that 00:08:25.899 --> 00:08:29.029 need for specificity. You also need to structure 00:08:29.029 --> 00:08:32.590 your prompts really meticulously. Yes. GPT -5 00:08:32.590 --> 00:08:35.529 apparently delivers its best performance in response 00:08:35.529 --> 00:08:39.429 to well -architected requests. It's not just 00:08:39.429 --> 00:08:41.309 listing instructions. It's about the framework. 00:08:41.730 --> 00:08:44.529 Exactly. Think about that recent trend of JSON 00:08:44.529 --> 00:08:46.470 prompting. People are talking about it a lot. 00:08:46.509 --> 00:08:48.870 Yeah, I've seen that. Well, the thing is, it 00:08:48.870 --> 00:08:51.289 isn't about the JSON format itself being magical. 00:08:51.470 --> 00:08:54.129 You know? OK. It's about the structure it forces 00:08:54.129 --> 00:08:57.769 you to create. A clear, hierarchical way of organizing 00:08:57.769 --> 00:09:01.129 your request. That structure is the key. It compels 00:09:01.129 --> 00:09:03.250 the model to be more systematic. Is it enough 00:09:03.250 --> 00:09:05.750 just to list instructions, then? Or does the 00:09:05.750 --> 00:09:08.169 order and structure really matter for GPT -5? 00:09:08.230 --> 00:09:09.950 Oh, the order and the hierarchical structure 00:09:09.950 --> 00:09:12.490 are crucial. They make a huge difference in performance. 00:09:12.610 --> 00:09:14.909 OK, so structure is vital. Let's take an example. 00:09:15.110 --> 00:09:17.850 A vague prompt like, write a launch announcement 00:09:17.850 --> 00:09:20.679 email. How would you structure that? A well -structured 00:09:20.679 --> 00:09:23.220 version would break that down into explicit sections. 00:09:23.340 --> 00:09:27.039 You define the persona. You are the CEO of a 00:09:27.039 --> 00:09:29.980 tech startup. Right. The audience. Targeting 00:09:29.980 --> 00:09:32.379 early adopters familiar with our beta program. 00:09:33.019 --> 00:09:36.019 The core components include three potential subject 00:09:36.019 --> 00:09:39.200 lines, an opening hook focusing on user pain 00:09:39.200 --> 00:09:42.379 points, explain the solution at a clear call 00:09:42.379 --> 00:09:45.320 to action button text like get early access now. 00:09:45.419 --> 00:09:48.360 Very specific. And then constraints. Keep it 00:09:48.360 --> 00:09:51.399 under 400 words, avoid technical jargon, format 00:09:51.399 --> 00:09:54.419 as simple HTML suitable for email. That level 00:09:54.419 --> 00:09:57.320 of fine -grained detail really helps it nail 00:09:57.320 --> 00:09:59.169 your vision. That's a lot more involved than 00:09:59.169 --> 00:10:01.429 just asking for an email. It is. But it gets 00:10:01.429 --> 00:10:04.269 you closer to the target on the first try. Some 00:10:04.269 --> 00:10:06.350 advanced users even use something called a spec 00:10:06.350 --> 00:10:09.289 format for really complex prompts. Situation, 00:10:09.710 --> 00:10:12.330 purpose, execution, constraints. Spec? Yeah. 00:10:12.409 --> 00:10:14.429 It's all about providing that logical framework, 00:10:14.590 --> 00:10:16.370 like giving it a detailed project plan so it 00:10:16.370 --> 00:10:19.049 doesn't wander off. OK. Structure, specificity. 00:10:19.070 --> 00:10:21.190 What's next? This next one is really fascinating, 00:10:21.250 --> 00:10:24.960 I think. GPT -5's performance on complex reasoning 00:10:24.960 --> 00:10:28.419 tasks actually improves when it knows it's going 00:10:28.419 --> 00:10:30.919 to be required to explain its thought process. 00:10:31.080 --> 00:10:34.059 Wait, so asking the AI for its reasoning can 00:10:34.059 --> 00:10:36.340 actually make it smarter in the moment? It seems 00:10:36.340 --> 00:10:38.879 so. It forces a more coherent logical chain, 00:10:39.320 --> 00:10:41.620 improving the output quality. How does that work? 00:10:42.259 --> 00:10:45.419 Well, it forces the model to construct a more 00:10:45.419 --> 00:10:48.799 coherent logical chain. before it arrives at 00:10:48.799 --> 00:10:50.799 the final conclusion. It can't just jump to an 00:10:50.799 --> 00:10:53.100 answer. It has to build a path to it. And how 00:10:53.100 --> 00:10:55.059 do you trigger that? You just add a simple clause, 00:10:55.399 --> 00:10:58.000 like, before providing the final answer, begin 00:10:58.000 --> 00:11:00.919 your response with a section titled, My Reasoning. 00:11:01.379 --> 00:11:03.840 That seemingly small instruction can really elevate 00:11:03.840 --> 00:11:06.139 the quality of the output. And presumably you 00:11:06.139 --> 00:11:09.179 get to see that reasoning too. Exactly. This 00:11:09.179 --> 00:11:11.519 chain of thought prompting, as it's called, doesn't 00:11:11.519 --> 00:11:14.159 just give us a better output. It gives us invaluable 00:11:14.159 --> 00:11:16.820 insight into the model's mind, so to speak. You 00:11:16.820 --> 00:11:19.059 can see where it went wrong, maybe. Precisely. 00:11:19.379 --> 00:11:21.440 You can actually see where its logic might have 00:11:21.440 --> 00:11:24.039 gone astray, or where it maybe misunderstood 00:11:24.039 --> 00:11:26.820 a nuance. It's incredibly helpful for debugging 00:11:26.820 --> 00:11:29.360 your own prompts. That transparency is huge. 00:11:29.620 --> 00:11:33.279 It really is. I mean, honestly. Soft laugh. I 00:11:33.279 --> 00:11:35.240 still wrestle with prompt drift myself sometimes. 00:11:35.399 --> 00:11:38.509 Gain the output to stay on track. So seeing the 00:11:38.509 --> 00:11:40.809 model's actual reasoning laid out, that would 00:11:40.809 --> 00:11:42.929 be a game changer for debugging my own prompts, 00:11:43.250 --> 00:11:45.129 not just its output. Yeah, it's like having an 00:11:45.129 --> 00:11:48.350 x -ray into its cognitive process. You can pinpoint 00:11:48.350 --> 00:11:50.769 the logical flaw. That's incredibly valuable 00:11:50.769 --> 00:11:53.529 for refining requests and just understanding 00:11:53.529 --> 00:11:56.190 the model better, sponsor. All right, let's shift 00:11:56.190 --> 00:11:58.370 gears into some more advanced strategies now, 00:11:58.570 --> 00:12:02.750 because GPT -5 is so... literal conflicting instructions 00:12:02.750 --> 00:12:05.309 seem like they could be a real problem. Oh, absolutely. 00:12:05.549 --> 00:12:07.730 They can easily send it into a kind of computational 00:12:07.730 --> 00:12:10.269 loop, you know, wasting valuable cycles while 00:12:10.269 --> 00:12:12.649 it tries to reconcile what it sees as a paradox. 00:12:13.169 --> 00:12:15.169 Right. The stakes for clarity here are just much 00:12:15.169 --> 00:12:17.490 higher than with previous more forgiving models. 00:12:17.629 --> 00:12:20.330 So how often do we accidentally create contradictory 00:12:20.330 --> 00:12:22.690 rules for the A .I.? Probably more often than 00:12:22.690 --> 00:12:25.190 we realize. And the key to preventing this kind 00:12:25.190 --> 00:12:27.649 of confusion is that you really need to build 00:12:27.649 --> 00:12:31.399 in an explicit hierarchy. or an override condition 00:12:31.399 --> 00:12:33.919 for your rules. A hierarchy, like rule one beats 00:12:33.919 --> 00:12:36.580 rule two. Exactly. It struggles immensely when 00:12:36.580 --> 00:12:38.820 rules clash without a clear time marker. You 00:12:38.820 --> 00:12:40.600 have to assume it will take every instruction 00:12:40.600 --> 00:12:43.460 literally and won't know which one takes precedence 00:12:43.460 --> 00:12:45.779 unless you tell it. Okay. Can you give an example? 00:12:46.059 --> 00:12:49.000 Sure. Think of like a medical scheduling assistant 00:12:49.000 --> 00:12:52.240 AI. You might have instruction one. Never book 00:12:52.240 --> 00:12:54.720 an appointment without explicit patient consent. 00:12:55.360 --> 00:12:57.860 Makes sense. Right. Standard procedure. But then 00:12:57.860 --> 00:13:00.509 instruction two. Immediately auto -assign the 00:13:00.509 --> 00:13:02.929 earliest available slot for any incoming high 00:13:02.929 --> 00:13:05.809 -risk alerts. Ah, OK. I see the conflict. What 00:13:05.809 --> 00:13:07.970 happens in an emergency? Exactly. In a real -world 00:13:07.970 --> 00:13:10.149 emergency, those rules could clash. Which one 00:13:10.149 --> 00:13:12.370 does it follow? So you need to tell it. You have 00:13:12.370 --> 00:13:14.629 to write that protocol directly into the prompt, 00:13:14.809 --> 00:13:17.549 something like, primary rule. Never book without 00:13:17.549 --> 00:13:20.409 explicit patient consent. Emergency override 00:13:20.409 --> 00:13:23.950 condition. For any alert flagged code red, the 00:13:23.950 --> 00:13:26.940 primary rule is temporarily suspended. auto -find 00:13:26.940 --> 00:13:29.820 the earliest available same -day slot immediately. 00:13:29.960 --> 00:13:33.659 Got it. A clear tiebreaker. It seems obvious 00:13:33.659 --> 00:13:35.259 to us, but you have to spell it out. You have 00:13:35.259 --> 00:13:38.039 to spell out the if -then explicitly to avoid 00:13:38.039 --> 00:13:40.399 that internal conflict. OK, that makes sense. 00:13:40.600 --> 00:13:42.899 What other advanced tricks are there? Well, GPT 00:13:42.899 --> 00:13:45.919 -5 also has this really powerful emergent capability. 00:13:46.419 --> 00:13:48.779 It can actually critique and improve its own 00:13:48.779 --> 00:13:50.860 work. It can evaluate itself. Yeah, it's not 00:13:50.860 --> 00:13:52.860 just about tweaking things slightly. It's like 00:13:52.860 --> 00:13:55.759 a form of internal quality control. So the AI 00:13:55.759 --> 00:13:58.620 can essentially become its own editor and quality 00:13:58.620 --> 00:14:01.440 control. Pretty much, yeah. It drafts, critiques 00:14:01.440 --> 00:14:03.820 against criteria, and refines until it meets 00:14:03.820 --> 00:14:05.840 the standard you set. How do you leverage that? 00:14:06.059 --> 00:14:08.379 You instruct the model to first create its own 00:14:08.379 --> 00:14:10.500 evaluation rubric for the task. Create its own 00:14:10.500 --> 00:14:13.659 rubric. Yeah. And then iterate on its response 00:14:13.659 --> 00:14:17.159 until it meets that self -imposed standard. This 00:14:17.159 --> 00:14:19.399 is where you really see the precision engineering 00:14:19.399 --> 00:14:21.580 idea come alive. It's like it's building its 00:14:21.580 --> 00:14:23.559 own internal compass for what excellent looks 00:14:23.559 --> 00:14:26.399 like. OK, walk me through an example, maybe for 00:14:26.399 --> 00:14:28.340 a business strategy. Sure. You tell it something 00:14:28.340 --> 00:14:32.159 like, first, create an internal rubric detailing 00:14:32.159 --> 00:14:34.799 what constitutes a world class business strategy. 00:14:35.039 --> 00:14:38.080 Cover market analysis, competitive advantage, 00:14:38.539 --> 00:14:41.100 financial projections, and implementation plan. 00:14:41.399 --> 00:14:45.340 Okay, step one. Define success. Right. Then, 00:14:45.500 --> 00:14:47.679 generate a first draft of the strategy based 00:14:47.679 --> 00:14:50.220 on this input, provide input. Next, critically 00:14:50.220 --> 00:14:52.259 evaluate this draft against your own rubric. 00:14:52.700 --> 00:14:55.259 If any category doesn't meet the highest standard 00:14:55.259 --> 00:14:57.779 on a world class, discard the draft and start 00:14:57.779 --> 00:15:00.320 again, incorporating specific feedback from your 00:15:00.320 --> 00:15:03.559 self -critique. Wow. That's a loop. Draft, critique, 00:15:03.960 --> 00:15:06.299 refine. Exactly. This self -correction loop is 00:15:06.299 --> 00:15:09.059 incredibly powerful. It turns the AI into its 00:15:09.059 --> 00:15:12.179 own QA engineer. Combining self -evaluation and 00:15:12.179 --> 00:15:14.679 iteration, that could lead to exceptional results. 00:15:15.039 --> 00:15:17.059 It's like having a hyper -efficient, totally 00:15:17.059 --> 00:15:19.100 objective editor built right in. That's a good 00:15:19.100 --> 00:15:20.820 way to put it. All right. Here's a concept that 00:15:20.820 --> 00:15:23.580 sounds a bit mind -bending. You can actually 00:15:23.580 --> 00:15:26.360 ask GPT -5 to improve your own prompts. Yes. 00:15:26.519 --> 00:15:28.799 This is called meta -prompting. Metaprompting. 00:15:28.879 --> 00:15:31.919 It's like asking the AI to teach you how to talk 00:15:31.919 --> 00:15:34.620 to it more effectively. You leverage its expertise 00:15:34.620 --> 00:15:36.940 on itself. How does that work in practice? You 00:15:36.940 --> 00:15:40.299 basically set it up like this. You are a world 00:15:40.299 --> 00:15:43.659 -class prompt engineer, specializing in creating 00:15:43.659 --> 00:15:47.039 clear, concise, and highly effective instructions 00:15:47.039 --> 00:15:50.259 for advanced large language models like GPT -5. 00:15:50.360 --> 00:15:53.039 Give it a roll. Right. I will provide you with 00:15:53.039 --> 00:15:55.669 a prompt I have written. along with my intended 00:15:55.669 --> 00:15:58.169 goal and any issues I'm currently seeing in the 00:15:58.169 --> 00:16:00.929 output it produces. Your task is to analyze my 00:16:00.929 --> 00:16:03.789 prompt and rewrite it to be clearer, more structured, 00:16:04.090 --> 00:16:06.389 more explicit, and ultimately more effective 00:16:06.389 --> 00:16:08.529 at achieving my desired outcome. And then you 00:16:08.529 --> 00:16:10.250 feed it your prompt and explain the problems 00:16:10.250 --> 00:16:12.210 you're having. Exactly. You explain your goal, 00:16:12.529 --> 00:16:15.470 what you tried, and what went wrong. In GBT -5, 00:16:15.809 --> 00:16:17.870 using its own deep understanding of what makes 00:16:17.870 --> 00:16:20.789 a prompt effective for itself helps you communicate 00:16:20.789 --> 00:16:23.620 better. That creates a really powerful feedback 00:16:23.620 --> 00:16:26.600 loop. Could this be the fastest way for us to 00:16:26.600 --> 00:16:29.740 actually learn how to prompt GPT -5 better? Absolutely. 00:16:30.019 --> 00:16:32.860 I think it could be. It essentially uses the 00:16:32.860 --> 00:16:36.399 AI's own expertise to teach you the AI's native 00:16:36.399 --> 00:16:38.299 language, so to speak. It's a direct shortcut 00:16:38.299 --> 00:16:41.799 to mastering this new way of interacting. Using 00:16:41.799 --> 00:16:45.879 the AI to learn the AI. Very meta. OK, what else? 00:16:46.059 --> 00:16:48.740 Well, GBT -5 also has pretty robust, agentic 00:16:48.740 --> 00:16:51.519 capabilities. Agentic? What does that mean in 00:16:51.519 --> 00:16:54.200 plain English? It means it can perform complex, 00:16:54.440 --> 00:16:57.659 multi -step tasks more autonomously. It can act 00:16:57.659 --> 00:17:00.259 more like an intelligent agent managing a process 00:17:00.259 --> 00:17:02.659 rather than just responding to a single query. 00:17:02.779 --> 00:17:05.220 OK, it can manage workflows. To some extent, 00:17:05.400 --> 00:17:07.839 yes. And you can control this sophisticated behavior 00:17:07.839 --> 00:17:10.119 even through natural language instructions. For 00:17:10.119 --> 00:17:12.160 instance, you can specify its reasoning effort. 00:17:12.319 --> 00:17:14.380 Reasoning effort. Yeah, you might say something 00:17:14.380 --> 00:17:16.640 like, approach this problem with a high level 00:17:16.640 --> 00:17:19.119 of reasoning effort. Explore multiple potential 00:17:19.119 --> 00:17:21.400 solutions and consider nuanced implications before 00:17:21.400 --> 00:17:24.119 answering. You're telling it to really dig deep, 00:17:24.380 --> 00:17:26.619 use more of its processing power. So you can 00:17:26.619 --> 00:17:29.140 dial the thinking up or down. Kind of, yeah. 00:17:29.660 --> 00:17:33.140 And importantly, you can control verbosity independently 00:17:33.140 --> 00:17:35.940 of that reasoning effort. Verbosity. So how much 00:17:35.940 --> 00:17:38.779 it talks. Exactly. You can tell it. Conduct a 00:17:38.779 --> 00:17:41.079 deep and thorough analysis high reasoning effort, 00:17:41.559 --> 00:17:44.660 but summarize your final findings concisely in 00:17:44.660 --> 00:17:48.839 no more than 200 words, low verbosity. Ah, okay. 00:17:49.079 --> 00:17:51.460 So we can tell it to think deeply, but then just 00:17:51.460 --> 00:17:53.420 give us the short version. Yes. You're separating 00:17:53.420 --> 00:17:56.039 the thinking effort from the output length, which 00:17:56.039 --> 00:17:58.759 is a really powerful lever for efficiency and 00:17:58.759 --> 00:18:01.619 clarity. That separation of concerns feels very 00:18:01.619 --> 00:18:04.279 efficient. Get the deep work done, but give me 00:18:04.279 --> 00:18:06.910 the executive summary. Precisely. It lets you 00:18:06.910 --> 00:18:09.170 tailor the internal computational work to the 00:18:09.170 --> 00:18:11.690 task's complexity while still getting a streamlined, 00:18:12.029 --> 00:18:14.470 easy -to -digest response at the end. Very useful. 00:18:14.609 --> 00:18:16.390 OK, what about handling multiple things at once? 00:18:16.490 --> 00:18:18.950 Right. So for complex workflows, GPT -5 can actually 00:18:18.950 --> 00:18:21.170 handle multiple independent tasks simultaneously, 00:18:21.470 --> 00:18:23.329 in parallel. Simultaneously. Yeah, which is a 00:18:23.329 --> 00:18:26.089 huge time saver, provided, and this is the key 00:18:26.089 --> 00:18:28.829 part, that the tasks don't depend on each other's 00:18:28.829 --> 00:18:31.029 outputs. So they have to be truly separate jobs. 00:18:31.410 --> 00:18:33.230 Correct. You're not waiting for one thing to 00:18:33.230 --> 00:18:35.069 finish before the next one starts, if they're 00:18:35.069 --> 00:18:37.049 independent. Can you give an example of a prompt 00:18:37.049 --> 00:18:39.730 like that? Sure. Imagine a single prompt saying, 00:18:40.349 --> 00:18:42.390 perform the following three tasks in parallel. 00:18:43.369 --> 00:18:46.150 Task one, research and summarize the top five 00:18:46.150 --> 00:18:50.869 marketing trends for Q3 2025. Task two, write 00:18:50.869 --> 00:18:53.809 a 500 -word blog post introduction about AI's 00:18:53.809 --> 00:18:57.529 impact on small businesses. And task three. analyze 00:18:57.529 --> 00:19:00.529 the attached customer sentiment data, CSV, and 00:19:00.529 --> 00:19:02.990 provide a bulleted summary of positive and negative 00:19:02.990 --> 00:19:05.289 trends. All from one prompt running at the same 00:19:05.289 --> 00:19:07.490 time. That's the idea. All from one prompt, all 00:19:07.490 --> 00:19:10.150 processing concurrently. Whoa. Okay, imagine 00:19:10.150 --> 00:19:12.829 scaling that across an entire enterprise, thousands, 00:19:13.009 --> 00:19:14.990 maybe millions of queries running in parallel 00:19:14.990 --> 00:19:17.650 like that. Right. That's a serious loop in efficiency. 00:19:18.289 --> 00:19:20.730 That's astounding, actually. It really is. The 00:19:20.730 --> 00:19:22.970 potential for throughput and speed is massive. 00:19:23.160 --> 00:19:25.920 So when would this parallel processing feature 00:19:25.920 --> 00:19:29.079 be most useful, practically speaking? It's best 00:19:29.079 --> 00:19:31.619 for scenarios where you have truly distinct tasks 00:19:31.619 --> 00:19:34.140 that can run independently. Things like batch 00:19:34.140 --> 00:19:36.920 content generation on different topics, summarizing 00:19:36.920 --> 00:19:38.859 different documents, or maybe running initial 00:19:38.859 --> 00:19:42.380 research phases across multiple unrelated domains 00:19:42.380 --> 00:19:45.539 simultaneously. Any time the tasks don't rely 00:19:45.539 --> 00:19:47.380 on each other's results. Got it. Independent 00:19:47.380 --> 00:19:50.450 tasks running side by side. Exactly. And one 00:19:50.450 --> 00:19:54.450 more tool worth mentioning for really mission 00:19:54.450 --> 00:19:57.250 -critical applications, or maybe if you're developing 00:19:57.250 --> 00:20:00.950 against the OpenAI API pretty heavily, they actually 00:20:00.950 --> 00:20:04.170 offer a dedicated prompt optimizer tool. A tool 00:20:04.170 --> 00:20:06.809 from OpenAI specifically for prompts. Yeah, available 00:20:06.809 --> 00:20:09.390 via their developer platform. It programmatically 00:20:09.390 --> 00:20:12.190 analyzes your prompts and suggests concrete improvement. 00:20:12.369 --> 00:20:14.240 How does it do that? It often gives detailed 00:20:14.240 --> 00:20:16.319 explanations about why certain changes would 00:20:16.319 --> 00:20:19.319 be beneficial. It can spot ambiguities you missed, 00:20:19.880 --> 00:20:21.960 recommend better structuring, maybe even add 00:20:21.960 --> 00:20:23.740 validation steps you hadn't thought of. So it's 00:20:23.740 --> 00:20:25.599 kind of like having a prompt engineering coach 00:20:25.599 --> 00:20:28.400 built right into OpenAI's platform. Essentially, 00:20:28.799 --> 00:20:31.119 yeah. Automatically enhances your prompts for 00:20:31.119 --> 00:20:34.380 GPT -5 based on what OpenAI knows works best 00:20:34.380 --> 00:20:37.079 with their own model. It's like an AI co -pilot 00:20:37.079 --> 00:20:39.119 for your prompt engineering. Okay, interesting. 00:20:39.200 --> 00:20:41.980 So if I feed it something simple like, make a 00:20:41.980 --> 00:20:44.819 website about class and cars. It might come back 00:20:44.819 --> 00:20:47.440 with a much more robust, optimized version. Maybe 00:20:47.440 --> 00:20:49.480 one that includes a conceptual checklist for 00:20:49.480 --> 00:20:52.599 the model to follow. Or explicit aesthetic instructions 00:20:52.599 --> 00:20:56.559 like, emulate the visual style of mid -20th century 00:20:56.559 --> 00:20:59.940 automotive magazines. And maybe detailed validation 00:20:59.940 --> 00:21:02.940 steps to ensure the final output actually aligns 00:21:02.940 --> 00:21:05.240 with your vision. It basically turns your initial 00:21:05.240 --> 00:21:07.740 vague thought into a much more solid instruction 00:21:07.740 --> 00:21:10.579 set. Exactly. It helps bridge that gap between 00:21:10.579 --> 00:21:13.680 your idea and what the model needs to execute 00:21:13.680 --> 00:21:15.440 it well. All right. We've covered a lot of ground, 00:21:15.720 --> 00:21:17.660 from foundational pillars to advanced techniques. 00:21:18.079 --> 00:21:19.920 How do these all fit together? Well, the true 00:21:19.920 --> 00:21:22.160 power, I think, really comes when you start layering 00:21:22.160 --> 00:21:25.380 these techniques, combining them. Our sources 00:21:25.380 --> 00:21:27.680 actually provided an example of a master prompt 00:21:27.680 --> 00:21:30.160 for creating something like a social media content 00:21:30.160 --> 00:21:32.799 calendar. And it artfully combines several things 00:21:32.799 --> 00:21:36.670 we talked about, defining the AI's role. including 00:21:36.670 --> 00:21:39.950 explicit pre -planning steps, giving detailed 00:21:39.950 --> 00:21:42.390 execution instructions, using structural elements 00:21:42.390 --> 00:21:44.750 like sections and bullet points, and even building 00:21:44.750 --> 00:21:47.630 in quality assurance checks. It was like a comprehensive 00:21:47.630 --> 00:21:50.890 blueprint for getting high -quality output consistently. 00:21:51.490 --> 00:21:54.130 So it's about synergy, using multiple techniques 00:21:54.130 --> 00:21:57.720 in concert. Exactly. And just as important as 00:21:57.720 --> 00:22:00.019 using these techniques is avoiding the common 00:22:00.019 --> 00:22:02.440 mistakes that people fall into, especially coming 00:22:02.440 --> 00:22:05.160 from older models. Right. What are the big pitfalls 00:22:05.160 --> 00:22:07.950 to watch out for with GPT -5? Well, first is 00:22:07.950 --> 00:22:11.029 legacy prompting, just using your old GPT -3 00:22:11.029 --> 00:22:13.650 or GPT -4 prompts and expecting them to work 00:22:13.650 --> 00:22:15.470 the same. They likely won't be precise enough. 00:22:15.589 --> 00:22:17.390 OK. Need to update our prompts. Then there's 00:22:17.390 --> 00:22:20.190 implicit intent. Just assuming the model understands 00:22:20.190 --> 00:22:23.069 context you haven't explicitly provided. It probably 00:22:23.069 --> 00:22:25.769 doesn't. Be explicit. Got it. Structural apathy. 00:22:25.930 --> 00:22:28.390 Just throwing unstructured blocks of text at 00:22:28.390 --> 00:22:30.690 it and hoping for the best. It needs that architecture 00:22:30.690 --> 00:22:32.730 we discussed. Structure matter. And of course, 00:22:33.130 --> 00:22:35.960 logical contradictions. Without that clear tiebreaker 00:22:35.960 --> 00:22:37.799 rule we talked about, that can really confuse 00:22:37.799 --> 00:22:41.539 it. Need that hierarchy. And finally, just underutilizing 00:22:41.539 --> 00:22:44.039 its strengths. Not engaging these more advanced 00:22:44.039 --> 00:22:46.519 features like planning, self -critique, or controlling 00:22:46.519 --> 00:22:48.799 reasoning effort. You're leaving performance 00:22:48.799 --> 00:22:51.519 on the table. Okay, avoid those pitfalls. Anything 00:22:51.519 --> 00:22:54.819 else? One more big one. Stop using subjective 00:22:54.819 --> 00:22:59.180 language like good or nice. or professional without 00:22:59.180 --> 00:23:02.400 defining them. Define your terms. Yes. Instead 00:23:02.400 --> 00:23:05.539 of write a good blog post, say, write a blog 00:23:05.539 --> 00:23:08.640 post that is good by being data -driven, citing 00:23:08.640 --> 00:23:10.920 three academic sources, and providing actionable 00:23:10.920 --> 00:23:13.839 advice for the reader. Give it concrete, measurable 00:23:13.839 --> 00:23:16.299 criteria for what good means in this specific 00:23:16.299 --> 00:23:20.019 context. Translate subjective desires into objective 00:23:20.019 --> 00:23:22.390 instructions. You nailed it. That translation 00:23:22.390 --> 00:23:24.450 is key. That's really well put. It feels like 00:23:24.450 --> 00:23:27.250 the biggest mindset shift GPT -5 demands, then, 00:23:27.609 --> 00:23:30.069 is viewing it less like a casual conversational 00:23:30.069 --> 00:23:32.329 partner and more like a brilliant, extremely 00:23:32.329 --> 00:23:34.730 powerful, but highly specialized instrument. 00:23:34.809 --> 00:23:37.750 A tool that needs careful calibration and precise 00:23:37.750 --> 00:23:40.349 input. Exactly. It requires a new language of 00:23:40.349 --> 00:23:43.059 interaction, almost. So let's recap the big idea 00:23:43.059 --> 00:23:45.880 here. What's the main takeaway for someone trying 00:23:45.880 --> 00:23:48.640 to get the most out of GPT -5? What's the biggest 00:23:48.640 --> 00:23:51.660 mental hurdle we need to overcome to truly master 00:23:51.660 --> 00:23:54.339 GPT -5? I think it's shifting our view of it 00:23:54.339 --> 00:23:56.960 away from being a conversational partner and 00:23:56.960 --> 00:23:59.119 towards seeing it as a precision instrument. 00:23:59.500 --> 00:24:02.900 It demands rigor, clarity, surgical instructions. 00:24:03.200 --> 00:24:05.339 Yeah, it's not about casual chat anymore. It's 00:24:05.339 --> 00:24:07.619 about crafting those precise instructions, like 00:24:07.619 --> 00:24:10.099 an engineer designing for a very specific outcome. 00:24:10.200 --> 00:24:12.319 And that extra effort you put into designing 00:24:12.319 --> 00:24:15.880 a really superior prompt isn't just a chore, 00:24:16.039 --> 00:24:18.500 right? Not at all. It's the very act of unlocking 00:24:18.500 --> 00:24:21.779 the model's immense power. That upfront work 00:24:21.779 --> 00:24:24.339 leads directly to significantly higher quality, 00:24:24.779 --> 00:24:27.420 much greater consistency, and far more reliable 00:24:27.420 --> 00:24:29.880 outputs down the line. The reward for that initial 00:24:29.880 --> 00:24:32.799 effort is genuinely significant. It can transform 00:24:32.799 --> 00:24:34.660 what might feel like a frustrating interaction 00:24:34.660 --> 00:24:37.329 into something incredibly productive. Absolutely. 00:24:37.829 --> 00:24:40.190 And mastering these techniques really positions 00:24:40.190 --> 00:24:43.349 you at the leading edge of AI utilization. You're 00:24:43.349 --> 00:24:46.470 not just keeping pace. You're actively leveraging 00:24:46.470 --> 00:24:49.109 its transformative potential. Whatever work you 00:24:49.109 --> 00:24:52.930 do, this seems increasingly critical. So we really 00:24:52.930 --> 00:24:56.190 encourage everyone listening to experiment relentlessly 00:24:56.190 --> 00:24:57.990 with these techniques. Don't just take our word 00:24:57.990 --> 00:25:00.549 for it. Yeah, try them out. Adapt them to your 00:25:00.549 --> 00:25:03.289 unique use cases. See what works for you. and 00:25:03.289 --> 00:25:05.710 contribute to our collective understanding. The 00:25:05.710 --> 00:25:08.529 more we all refine our prompts, the more we discover 00:25:08.529 --> 00:25:11.009 what these amazing models are truly capable of. 00:25:11.150 --> 00:25:13.750 It's true. The landscape of AI is evolving so 00:25:13.750 --> 00:25:16.670 incredibly fast. Mastering tools like GPT -5, 00:25:16.849 --> 00:25:18.829 it's really becoming a core competency, isn't 00:25:18.829 --> 00:25:21.309 it? For productivity, for innovation. Essential. 00:25:21.529 --> 00:25:23.430 It's becoming essential for navigating the modern 00:25:23.430 --> 00:25:25.950 world, I think. So the final thought maybe is, 00:25:26.329 --> 00:25:29.349 what kind of... Intricate, perhaps world -changing 00:25:29.349 --> 00:25:32.690 questions can you now ask, knowing that GPT -5 00:25:32.690 --> 00:25:35.289 is listening with such surgical precision. Yeah, 00:25:35.390 --> 00:25:37.809 what becomes possible now? It's a pretty exciting 00:25:37.809 --> 00:25:40.089 time to be exploring this. It really is. Thank 00:25:40.089 --> 00:25:42.509 you for joining us on this deep dive today. My 00:25:42.509 --> 00:25:44.230 pleasure. Until next time, keep exploring.