WEBVTT 00:00:00.000 --> 00:00:02.819 So if you've been using the latest AI models, 00:00:02.940 --> 00:00:05.799 maybe chat GPT -5, you might have noticed something 00:00:05.799 --> 00:00:08.400 a bit weird. Yeah, like you know you're using 00:00:08.400 --> 00:00:10.480 this incredibly powerful tool, right? State of 00:00:10.480 --> 00:00:13.539 the art. Exactly. But the results, they can be 00:00:13.539 --> 00:00:16.059 all over the place. Sometimes amazing, sometimes 00:00:16.059 --> 00:00:19.300 not so much. Right. You get this brilliant piece 00:00:19.300 --> 00:00:22.300 of analysis one minute, and then next the answer 00:00:22.300 --> 00:00:26.260 feels kind of lazy, almost worse than the older 00:00:26.260 --> 00:00:28.940 model sometimes. It's confusing, isn't it? You're 00:00:28.940 --> 00:00:31.339 paying for the premium model expecting top quality 00:00:31.339 --> 00:00:35.200 every time, but it dips. So why is this super 00:00:35.200 --> 00:00:38.509 smart AI acting well? Kind of inconsistent. It's 00:00:38.509 --> 00:00:40.689 a fair question, and it feels backwards. But 00:00:40.689 --> 00:00:43.149 the issue isn't really the AI model itself. It 00:00:43.149 --> 00:00:46.229 is powerful. OK. The problem comes from a couple 00:00:46.229 --> 00:00:48.689 of big changes OpenAI made in how the whole system 00:00:48.689 --> 00:00:50.149 worked. Basically, the way we used to write prompts, 00:00:50.390 --> 00:00:53.229 it's tripping up this new setup. Oh, OK. So our 00:00:53.229 --> 00:00:55.170 old habits are the problem. Let's dig into that, 00:00:55.189 --> 00:00:57.109 then. Yeah. The goal today is really to give 00:00:57.109 --> 00:00:59.710 you, the listener, a clear strategy to get back 00:00:59.710 --> 00:01:02.609 to those consistently high quality results. Definitely. 00:01:03.250 --> 00:01:05.189 We're going to unpack these two main changes. 00:01:05.430 --> 00:01:07.849 First, this invisible router thing, and also 00:01:07.849 --> 00:01:11.129 how the AI now follows instructions very strictly. 00:01:11.409 --> 00:01:14.969 And then we'll jump into five specific tips. 00:01:15.409 --> 00:01:17.930 We'll start simple with easy nudges and build 00:01:17.930 --> 00:01:19.810 up to something we call the perfection loop. 00:01:20.609 --> 00:01:23.469 It's a bit more effort, but wow, the results. 00:01:23.609 --> 00:01:25.469 All right, let's get into it. Segment one. Yeah. 00:01:25.950 --> 00:01:28.829 Why the old ways are failing. This invisible 00:01:28.829 --> 00:01:31.260 router. What's that about? I remember you used 00:01:31.260 --> 00:01:33.219 to be able to pick different models. You did, 00:01:33.359 --> 00:01:36.120 yeah. You could choose GPT -5, maybe a thinking 00:01:36.120 --> 00:01:38.260 version. You felt like you had control. Right. 00:01:38.439 --> 00:01:41.739 Now the interface seems simpler. But behind the 00:01:41.739 --> 00:01:43.719 curtain, when you send your prompt, it hits this 00:01:43.719 --> 00:01:47.480 invisible router first. Think of it like an automatic 00:01:47.480 --> 00:01:49.760 traffic cop for your request. A traffic cop? 00:01:50.140 --> 00:01:52.420 But if I'm paying for the best model, why does 00:01:52.420 --> 00:01:54.120 it need routing? Shouldn't it just go to the 00:01:54.120 --> 00:01:56.280 best one? That's the logical thought, yeah. But 00:01:56.280 --> 00:01:59.180 here's the catch, and where the user frustration 00:01:59.180 --> 00:02:01.579 often comes in, those really powerful models, 00:02:01.659 --> 00:02:03.920 they take more computer power, they cost more 00:02:03.920 --> 00:02:07.719 for OpenAI to run. Ah, so the router is maybe 00:02:07.719 --> 00:02:11.699 optimizing for cost. You got it. It's often engineered 00:02:11.699 --> 00:02:13.879 to find the most efficient path, which usually 00:02:13.879 --> 00:02:16.500 means the fastest and cheapest AI model available 00:02:16.500 --> 00:02:19.449 that might be able to answer the query. So it's 00:02:19.449 --> 00:02:22.689 defaulting to good enough to save resources, 00:02:23.310 --> 00:02:26.370 even if I wanted great. Precisely. And that good 00:02:26.370 --> 00:02:29.469 enough model gives you those, well, less impressive 00:02:29.469 --> 00:02:32.370 inconsistent answers sometimes. That's the hidden 00:02:32.370 --> 00:02:34.949 mechanical reason for the quality dips. OK, that 00:02:34.949 --> 00:02:37.150 makes a lot of sense, actually. So that's the 00:02:37.150 --> 00:02:39.810 router. What about the other big shift, the training? 00:02:40.009 --> 00:02:41.830 Right, the second piece is how it's trained. 00:02:42.270 --> 00:02:45.210 Chat GPT -5 got a lot of training focused on 00:02:45.210 --> 00:02:49.229 serving AI agents. These are programs, basically, 00:02:49.389 --> 00:02:51.490 that need instructions followed perfectly. No 00:02:51.490 --> 00:02:53.590 mistakes, no guessing. OK, so it got really good 00:02:53.590 --> 00:02:56.289 at following orders exactly. Super good. But 00:02:56.289 --> 00:02:58.770 the flip side for a regular user like you or 00:02:58.770 --> 00:03:01.669 me, if your prompt is a bit vague or you leave 00:03:01.669 --> 00:03:04.349 things out, It's much worse now at trying to 00:03:04.349 --> 00:03:05.830 figure out what you meant. It won't fill in the 00:03:05.830 --> 00:03:07.870 gaps like the older ones might have. Nope. It 00:03:07.870 --> 00:03:10.150 sticks rigidly to what you typed. No implied 00:03:10.150 --> 00:03:13.060 instructions. That sounds demanding. It can be. 00:03:13.280 --> 00:03:15.099 And honestly, I still mess this up sometimes. 00:03:15.180 --> 00:03:17.599 You know, I rush a prompt, make it too simple. 00:03:17.800 --> 00:03:20.800 Happens to the best of us. Right. Just last week, 00:03:20.879 --> 00:03:23.039 I asked for a quick summary and it gave me this 00:03:23.039 --> 00:03:25.560 long definition of terms I obviously already 00:03:25.560 --> 00:03:27.900 knew. Slight chuckle. Took me a few tries to 00:03:27.900 --> 00:03:29.860 get it right. It catches you out if you're not 00:03:29.860 --> 00:03:32.900 precise. OK, so the router pushes towards cheapness 00:03:32.900 --> 00:03:36.639 and the AI demands exact instructions. So the 00:03:36.639 --> 00:03:39.500 big question then, how do we make the router 00:03:39.500 --> 00:03:42.500 pick the smart expensive model when we need it. 00:03:42.620 --> 00:03:44.659 Ah, well, that's where our specific phrasing 00:03:44.659 --> 00:03:46.860 comes in. We have to signal it. Leading us nicely 00:03:46.860 --> 00:03:50.740 into the quick wins. Tip one. Tip one. Use router 00:03:50.740 --> 00:03:53.000 nudge phrases. This is probably the easiest thing 00:03:53.000 --> 00:03:56.080 you can do. Seriously low effort. OK. You just 00:03:56.080 --> 00:03:57.979 add a specific little phrase right at the end 00:03:57.979 --> 00:03:59.699 of your prompt. It's like a little flag for the 00:03:59.699 --> 00:04:02.560 router that says, hey, stop. This one needs actual 00:04:02.560 --> 00:04:04.860 thinking power. What are these magic words then? 00:04:05.180 --> 00:04:07.080 What tells the router to wake up the big guns? 00:04:07.439 --> 00:04:10.360 The ones that work consistently are phrases like, 00:04:10.800 --> 00:04:13.300 think carefully about this, or think deeply about 00:04:13.300 --> 00:04:15.419 this. Sometimes think hard about this works, 00:04:15.759 --> 00:04:18.980 too. Hmm. OK. Simple enough. Yeah. Using one 00:04:18.980 --> 00:04:21.360 of those basically forces the router to send 00:04:21.360 --> 00:04:23.720 your request to a more capable model, avoiding 00:04:23.720 --> 00:04:26.300 that quick, cheap default route we talked about. 00:04:26.480 --> 00:04:28.639 I can see that being useful. Like if I'm asking 00:04:28.639 --> 00:04:32.449 it to, say, outline pros and cons for a big decision, 00:04:32.810 --> 00:04:35.329 adding, think deeply about this, would probably 00:04:35.329 --> 00:04:38.050 get me a much better, more nuanced answer than 00:04:38.050 --> 00:04:40.290 just asking the question straight. Exactly. You 00:04:40.290 --> 00:04:41.990 should definitely use it for anything important. 00:04:42.189 --> 00:04:44.889 Business plans, you know, data analysis, drafting 00:04:44.889 --> 00:04:47.889 a critical email, things where quality really 00:04:47.889 --> 00:04:51.250 matters. Oh, and a quick heads up, don't use 00:04:51.250 --> 00:04:53.569 emotional language. Saying things like, this 00:04:53.569 --> 00:04:55.490 is really important to me doesn't seem to work. 00:04:55.730 --> 00:04:58.069 Right. It's a machine. It needs commands, not 00:04:58.069 --> 00:05:01.079 feelings. Precisely. Keep it clear. Keep it command 00:05:01.079 --> 00:05:03.779 -like. OK, so that's nudging the router. What 00:05:03.779 --> 00:05:06.060 about the output itself? Tip 2 is about length, 00:05:06.060 --> 00:05:08.199 right? Yeah, controlling the output length. Because 00:05:08.199 --> 00:05:10.879 the router also kind of defaults to shorter answers, 00:05:11.060 --> 00:05:13.040 again, to save processing time. So we need to 00:05:13.040 --> 00:05:15.120 be explicit about how much text we actually want. 00:05:15.240 --> 00:05:17.439 Makes sense. So for short stuff, like just the 00:05:17.439 --> 00:05:19.740 key points. Right. For a short output, you'd 00:05:19.740 --> 00:05:22.040 say something like, summarize the main points 00:05:22.040 --> 00:05:25.100 in under 100 words. Perfect for maybe a quick 00:05:25.100 --> 00:05:28.120 project update email or you know a tweet Okay, 00:05:28.319 --> 00:05:31.199 and if I need a bit more like the main points 00:05:31.199 --> 00:05:33.920 plus some background That's your medium output 00:05:33.920 --> 00:05:37.339 try asking it to explain this topic in about 00:05:37.339 --> 00:05:40.019 three to five short paragraphs That works well 00:05:40.019 --> 00:05:42.040 for explaining something like why your websites 00:05:42.040 --> 00:05:44.120 click through rate dropped Maybe to your team 00:05:44.120 --> 00:05:47.199 gives context and then the big stuff full reports 00:05:47.199 --> 00:05:50.800 articles detailed full output be specific provide 00:05:50.800 --> 00:05:53.519 a detailed and full analysis around 600, 800 00:05:53.519 --> 00:05:56.420 words, or even more, like write a comprehensive 00:05:56.420 --> 00:05:59.220 guide of about 1 ,200 words. That gets you the 00:05:59.220 --> 00:06:01.620 full document. Numbers seem key here. Specific 00:06:01.620 --> 00:06:03.879 word counts or paragraph counts. Absolutely. 00:06:04.199 --> 00:06:07.319 And a little pro tip here, use a text expander 00:06:07.319 --> 00:06:09.819 app. You can save these length commands like 00:06:09.819 --> 00:06:12.519 summarize under 100 words and just type a short 00:06:12.519 --> 00:06:15.300 code like S100 to insert the whole phrase. Saves 00:06:15.300 --> 00:06:18.350 a ton of time. Nice. Okay, so we can nudge the 00:06:18.350 --> 00:06:20.550 router and control the length, but what about 00:06:20.550 --> 00:06:22.610 improving the actual prompt before we send it? 00:06:22.689 --> 00:06:24.649 How do we make the prompt itself better? Great 00:06:24.649 --> 00:06:27.670 question. That leads us to using a meta prompt. 00:06:28.269 --> 00:06:30.750 We basically get the AI to help us write a better 00:06:30.750 --> 00:06:33.209 prompt for itself. Okay, tip three, the meta 00:06:33.209 --> 00:06:36.189 prompt optimizer. You're saying we can use the 00:06:36.189 --> 00:06:39.329 AI to improve our instructions for the AI. Exactly. 00:06:39.490 --> 00:06:41.850 It sounds a bit circular, maybe, but it works 00:06:41.850 --> 00:06:44.209 incredibly well. Most people don't know OpenAI 00:06:44.209 --> 00:06:46.870 has internal tools for this, but we can kind 00:06:46.870 --> 00:06:49.310 of replicate it with a specific type of prompt. 00:06:49.470 --> 00:06:52.889 A meta prompt. Yeah. Which is just a prompt about 00:06:52.889 --> 00:06:55.589 prompting. Yep. You're telling the AI to act 00:06:55.589 --> 00:06:58.709 like a prompt expert and fix your request. So 00:06:58.709 --> 00:07:01.649 wait, we're asking the AI to critique our prompting 00:07:01.649 --> 00:07:04.910 skills. Is that really necessary for, like, just 00:07:04.910 --> 00:07:07.569 asking for a blog post idea? Well, maybe not 00:07:07.569 --> 00:07:10.129 for every single query. But remember how strict 00:07:10.129 --> 00:07:12.170 it is now? Write the exact instruction following. 00:07:12.329 --> 00:07:14.430 By having it analyze your prompt first, you force 00:07:14.430 --> 00:07:17.290 it to clarify everything. You guarantee the instructions 00:07:17.290 --> 00:07:19.569 it finally acts on are super clear and detailed. 00:07:19.910 --> 00:07:21.970 The command is usually something like, you are 00:07:21.970 --> 00:07:24.509 an expert prompt engineer, analyze the weak points 00:07:24.509 --> 00:07:26.589 of my original prompt below, and then rewrite 00:07:26.589 --> 00:07:28.769 it to be much clearer and more effective. Hmm, 00:07:29.009 --> 00:07:32.430 okay. I can see how that would force clarity. 00:07:32.790 --> 00:07:35.170 Do you have an example? Sure, think about a vague 00:07:35.170 --> 00:07:38.449 request. Write a blog post about working from 00:07:38.449 --> 00:07:40.949 home effectively. Pretty standard prompt. Right. 00:07:41.310 --> 00:07:43.250 But the meta -prompt process might turn that 00:07:43.250 --> 00:07:45.949 into something much richer. The rewritten prompt 00:07:45.949 --> 00:07:48.689 would likely define the AI's role, like you are 00:07:48.689 --> 00:07:51.810 a productivity expert. Specify the target audience, 00:07:52.129 --> 00:07:54.490 remote workers new to the concept. Outline the 00:07:54.490 --> 00:07:57.790 required structure. Intro, three main tips with 00:07:57.790 --> 00:08:01.230 examples, conclusion, and set the tone, encouraging 00:08:01.230 --> 00:08:04.449 and practical. Wow, okay. That's way more specific. 00:08:04.610 --> 00:08:06.610 The difference in output quality must be huge. 00:08:06.810 --> 00:08:09.230 Night and day. That detailed prompt just gives 00:08:09.230 --> 00:08:11.269 the AI so much more to work with accurately. 00:08:11.649 --> 00:08:13.670 And if you start doing this meta -prompting regularly, 00:08:13.889 --> 00:08:15.529 I guess your prompts naturally start looking 00:08:15.529 --> 00:08:18.149 more structured. Which brings us to tip four. 00:08:18.490 --> 00:08:21.930 the XML sandwich. Exactly. You'll notice those 00:08:21.930 --> 00:08:24.110 improved prompts often use these angle brackets 00:08:24.110 --> 00:08:27.769 like this tag or context. These look like XML 00:08:27.769 --> 00:08:30.589 tags. OpenAI actually recommends this structure 00:08:30.589 --> 00:08:33.309 internally. It helps organize complex instructions. 00:08:33.830 --> 00:08:35.590 I like the analogy you used before thinking of 00:08:35.590 --> 00:08:38.350 them like clearly labeled boxes. Instead of just 00:08:38.350 --> 00:08:40.309 a messy paragraph of instructions, you're giving 00:08:40.309 --> 00:08:43.149 it distinct blocks of information. Audience, 00:08:43.309 --> 00:08:46.509 tone, goal. Precisely. It breaks it down logically. 00:08:47.070 --> 00:08:49.909 The AI processes structured information really 00:08:49.909 --> 00:08:52.139 well. Let's take an example. maybe asking for 00:08:52.139 --> 00:08:55.100 a study plan. Right. A bad prompt is just, make 00:08:55.100 --> 00:08:58.360 me an IELTS study plan, I need band 7. Vague. 00:08:58.980 --> 00:09:01.779 A good structured prompt would use tags. Current 00:09:01.779 --> 00:09:04.340 level band 6, current level, target score band, 00:09:04.679 --> 00:09:08.000 7 .5 target score, week area speaking fluency, 00:09:08.559 --> 00:09:11.259 complex grammar week areas, study time 10 hours 00:09:11.259 --> 00:09:13.960 week study time. Much clearer. Each piece of 00:09:13.960 --> 00:09:16.279 info is neatly packaged. And the result you get 00:09:16.279 --> 00:09:18.639 back reflects that clarity. Instead of generic 00:09:18.639 --> 00:09:20.919 advice, you'll likely get a detailed table, maybe 00:09:20.919 --> 00:09:23.259 in markdown format, with specific activities 00:09:23.259 --> 00:09:26.179 scheduled like practice part two cue cards, 30 00:09:26.179 --> 00:09:29.399 minutes, or shadow native speaker audio, 20 minutes. 00:09:29.580 --> 00:09:31.779 It becomes actionable. Yeah, that structure is 00:09:31.779 --> 00:09:34.100 powerful, especially I imagine if you're setting 00:09:34.100 --> 00:09:36.320 up custom instructions in chat GPT or building 00:09:36.320 --> 00:09:38.980 your own custom GPTs. Definitely essential for 00:09:38.980 --> 00:09:41.720 those. OK, so we've covered nudging, length, 00:09:42.659 --> 00:09:44.639 optimizing the prompt itself with meta prompts 00:09:44.639 --> 00:09:47.080 and structuring it with XML. This is all about 00:09:47.019 --> 00:09:50.679 crafting the input. But how do we kind of guarantee 00:09:50.679 --> 00:09:53.240 the output quality before we even see it? Ah, 00:09:53.240 --> 00:09:56.500 that's the final piece. We embed quality control 00:09:56.500 --> 00:09:59.799 inside the prompt itself. We make the AI check 00:09:59.799 --> 00:10:02.500 its own work. Interesting. Let's take a quick 00:10:02.500 --> 00:10:05.179 break and then dive into that final tip. Sounds 00:10:05.179 --> 00:10:08.240 good. Mid -roll sponsor read. We'll be right 00:10:08.240 --> 00:10:11.259 back after the break. All right, we're back. 00:10:11.440 --> 00:10:14.220 We were just talking about making the AI guarantee 00:10:14.220 --> 00:10:17.240 its own quality. What's the final tip? Tip five, 00:10:17.580 --> 00:10:19.639 the perfection loop. Now this one takes a bit 00:10:19.639 --> 00:10:20.980 more thought when you're writing the prompt, 00:10:21.039 --> 00:10:23.659 but the payoff is potentially huge. Okay, the 00:10:23.659 --> 00:10:25.720 perfection loop. How does it work? Instead of 00:10:25.720 --> 00:10:27.759 just asking for the output and hoping it's good, 00:10:27.980 --> 00:10:30.659 you first instruct the AI to define what a perfect 00:10:30.659 --> 00:10:32.879 result would even look like for this specific 00:10:32.879 --> 00:10:35.299 request. So you make it create its own checklist 00:10:35.299 --> 00:10:38.740 first, based on my goals. That's clever. Exactly. 00:10:38.899 --> 00:10:41.539 You tell it, okay, first... Figure out the criteria 00:10:41.539 --> 00:10:44.299 for a perfect answer here. Maybe that's uniqueness, 00:10:44.740 --> 00:10:47.679 clarity, engaging tone, fits the brand voice, 00:10:47.799 --> 00:10:50.679 whatever is relevant. Then you tell it, draft 00:10:50.679 --> 00:10:53.519 an answer. Then use your own checklist to grade 00:10:53.519 --> 00:10:56.279 that draft. Keep refining it internally until 00:10:56.279 --> 00:10:59.340 it scores a 10 out of 10. Only then show me the 00:10:59.340 --> 00:11:01.519 final result. Whoa. So it does the drafting, 00:11:01.840 --> 00:11:04.019 the critiquing, the editing all internally before 00:11:04.019 --> 00:11:06.820 I even see anything. Yep. You essentially push 00:11:06.820 --> 00:11:09.840 the quality control step. back onto the AI before 00:11:09.840 --> 00:11:11.559 it delivers. You only get the polished version. 00:11:11.740 --> 00:11:14.039 That's kind of amazing. So, for example, if I 00:11:14.039 --> 00:11:16.279 needed a content strategy. You could tell it. 00:11:16.659 --> 00:11:18.759 First, define the key elements of a successful 00:11:18.759 --> 00:11:22.120 Q4 content strategy for a B2B SaaS company. Make 00:11:22.120 --> 00:11:25.340 a checklist. Then, generate the strategy, review 00:11:25.340 --> 00:11:27.620 it against your checklist, revise it until it's 00:11:27.620 --> 00:11:30.419 perfect, and then show me the final 1010 strategy. 00:11:30.799 --> 00:11:32.639 It assesses itself on things like scalability, 00:11:32.759 --> 00:11:37.029 tone, alignment with goals. Or, for say, a YouTube 00:11:37.029 --> 00:11:40.570 script. Create a checklist for an engaging tutorial 00:11:40.570 --> 00:11:44.210 script. Hook, clear steps, visual cues, call 00:11:44.210 --> 00:11:47.149 to action, friendly tone. Write the script, grade 00:11:47.149 --> 00:11:50.929 it, rewrite it until perfect, then output. Wow, 00:11:51.190 --> 00:11:53.269 just thinking about that. Imagine scaling that 00:11:53.269 --> 00:11:56.090 kind of internal self -correction across, I don't 00:11:56.090 --> 00:11:58.230 know, millions of tasks. It's not just prompting 00:11:58.230 --> 00:12:00.480 anymore, that's like... managing an automated 00:12:00.480 --> 00:12:03.080 quality process. It really shifts the dynamic, 00:12:03.279 --> 00:12:05.000 doesn't it? It's best for those really complex, 00:12:05.200 --> 00:12:07.379 high -stakes tasks. You know, writing entire 00:12:07.379 --> 00:12:09.360 business plans, generating code that needs to 00:12:09.360 --> 00:12:12.440 work first time, crafting really long, important 00:12:12.440 --> 00:12:14.720 documents. Okay, so that's the fifth tip, the 00:12:14.720 --> 00:12:17.679 perfection loop. Now, the big takeaway here seems 00:12:17.679 --> 00:12:19.700 to be that these tips aren't really isolated 00:12:19.700 --> 00:12:22.159 tricks. Not at all. The real power comes when 00:12:22.159 --> 00:12:24.120 you start combining them. You layer them together 00:12:24.120 --> 00:12:26.259 to create what we sometimes call a super prompt. 00:12:26.399 --> 00:12:28.500 Right. Can we walk through building one, maybe 00:12:28.500 --> 00:12:30.399 for that project proposal example you mentioned, 00:12:30.779 --> 00:12:32.799 a fashion retail app? Sure. Perfect example. 00:12:33.120 --> 00:12:35.019 So first we'd start with structure. Tip four, 00:12:35.159 --> 00:12:37.860 we use those XML style tags. Okay, like company 00:12:37.860 --> 00:12:39.559 context, dot company context, problem statement, 00:12:39.679 --> 00:12:42.200 problem statement, and then maybe tags for each 00:12:42.200 --> 00:12:44.639 required section of the proposal. Section one 00:12:44.639 --> 00:12:46.679 introduction, section two solution, et cetera, 00:12:47.100 --> 00:12:49.700 laying it all out clearly. Exactly. Then we add 00:12:49.700 --> 00:12:53.480 tip two. Length control. We specify this proposal 00:12:53.480 --> 00:12:56.059 should be detailed and complete around 5 ,000, 00:12:56.059 --> 00:12:58.860 1 ,200 words. Give it a clear target. Got it. 00:12:59.120 --> 00:13:01.379 Structure, then length. What's next? Now we bring 00:13:01.379 --> 00:13:04.379 in the big one. Tip five, the perfection loop. 00:13:04.820 --> 00:13:07.679 We'd add instructions like, before writing, define 00:13:07.679 --> 00:13:09.899 an internal checklist for what makes a project 00:13:09.899 --> 00:13:12.559 proposal highly persuasive and likely to be approved. 00:13:13.200 --> 00:13:15.799 Grade your draft against this checklist. Refine 00:13:15.799 --> 00:13:18.019 it until it achieves a 10 -10 internal score, 00:13:18.320 --> 00:13:21.039 then present the final version. Okay, so structure, 00:13:21.220 --> 00:13:23.919 length, internal quality control. Are we missing 00:13:23.919 --> 00:13:26.340 anything? Just the final touch. Tip one, the 00:13:26.340 --> 00:13:29.059 router nudge. Right at the very end of the entire 00:13:29.059 --> 00:13:31.620 prompt we add, think very carefully about this. 00:13:31.860 --> 00:13:33.820 Ah, the signal to use the powerful model. Yep, 00:13:33.940 --> 00:13:35.840 so you see how it all stacks together. Structure 00:13:35.840 --> 00:13:37.879 provides clarity, length sets boundaries, the 00:13:37.879 --> 00:13:39.620 perfection loop ensures quality, and the nudge 00:13:39.620 --> 00:13:41.539 phrase makes sure the right brain is doing the 00:13:41.539 --> 00:13:45.279 work. That combination feels really robust. It 00:13:45.279 --> 00:13:47.659 addresses both the router issue and the strict 00:13:47.659 --> 00:13:50.480 instruction following. It gives the AI everything 00:13:50.480 --> 00:13:53.700 it needs to succeed based on its new rules. OK, 00:13:53.840 --> 00:13:56.080 so wrapping this all up, what's the main thing 00:13:56.080 --> 00:13:57.740 people should take away from this deep dive? 00:13:58.299 --> 00:14:00.299 I think the core idea is pretty straightforward. 00:14:01.039 --> 00:14:04.759 The era of just typing vague one -sentence prompts 00:14:04.759 --> 00:14:07.559 and hoping for the best. That's kind of over, 00:14:07.620 --> 00:14:09.700 at least for consistently high quality with models 00:14:09.700 --> 00:14:13.320 like GPT -5. Yes, the AI is more powerful. But 00:14:13.320 --> 00:14:15.960 that power comes with these new conditions, the 00:14:15.960 --> 00:14:18.799 invisible router, the need for absolute clarity. 00:14:19.080 --> 00:14:21.460 So our approach has to adapt. It's less about 00:14:21.460 --> 00:14:23.860 needing fancier tech and more about us being 00:14:23.860 --> 00:14:26.340 better communicators. More structured, now we 00:14:26.340 --> 00:14:29.259 ask. Exactly. Clear instructions, better organization. 00:14:29.419 --> 00:14:31.600 You can start small, you know, just add those 00:14:31.600 --> 00:14:33.539 router nudges and specify the length. That alone 00:14:33.539 --> 00:14:35.700 makes a difference. Right. Easy wins first. Then 00:14:35.700 --> 00:14:37.879 as you get comfortable, start playing with the 00:14:37.879 --> 00:14:40.500 meta prompts, use the XML structure for complex 00:14:40.500 --> 00:14:43.220 requests, and try out that perfection loop for 00:14:43.220 --> 00:14:45.779 really crucial tasks. It feels like these techniques 00:14:45.779 --> 00:14:48.360 aren't just about getting a better answer from 00:14:48.360 --> 00:14:50.960 the AI. They're also about building better processes. 00:14:51.240 --> 00:14:53.539 for working with the AI. That's a great way to 00:14:53.539 --> 00:14:56.019 put it. It's about setting up an effective partnership. 00:14:56.740 --> 00:14:59.360 You define the standards very clearly, and the 00:14:59.360 --> 00:15:03.059 AI has the power to meet them precisely. So here's 00:15:03.059 --> 00:15:04.799 a final thought to leave our listeners with, 00:15:05.159 --> 00:15:07.360 something to mull over. We've talked about how 00:15:07.360 --> 00:15:10.840 this AI is trained to follow exact commands perfectly 00:15:10.840 --> 00:15:14.539 now. Super literally. What do you think happens 00:15:14.539 --> 00:15:17.120 if you give it two instructions that are perfectly 00:15:17.120 --> 00:15:20.679 clear, perfectly structured? but they directly 00:15:20.679 --> 00:15:23.179 conflict with each other. What does forcing that 00:15:23.179 --> 00:15:26.399 kind of logical paradox make the system prioritize? 00:15:26.899 --> 00:15:29.080 Something to maybe experiment with. Ooh, that's 00:15:29.080 --> 00:15:32.340 a fascinating question. What breaks first? Definitely 00:15:32.340 --> 00:15:34.840 something to try. But for now, maybe just try 00:15:34.840 --> 00:15:36.860 combining those first couple of tips, the nudge 00:15:36.860 --> 00:15:38.759 and the length control, see what happens. Good 00:15:38.759 --> 00:15:40.200 starting point. Thanks for breaking all this 00:15:40.200 --> 00:15:42.320 down. My pleasure. Always fun talking about this 00:15:42.320 --> 00:15:42.460 stuff.