WEBVTT 00:00:00.000 --> 00:00:03.319 You add, like, 200 words to an AI prompt, hoping 00:00:03.319 --> 00:00:06.960 it gets smarter. Beat. But what if it just needed 00:00:06.960 --> 00:00:11.140 one single sentence to pause and ask you what 00:00:11.140 --> 00:00:13.119 you actually mean? Yeah, it completely shifts 00:00:13.119 --> 00:00:15.199 the dynamic. I mean, you move away from just 00:00:15.199 --> 00:00:17.760 dictating to a machine, right? Right. And instead, 00:00:17.920 --> 00:00:20.079 you force it into a dialogue before it generates 00:00:20.079 --> 00:00:22.820 a single word. Welcome to today's deep dive. 00:00:23.480 --> 00:00:26.079 We're dissecting the clarifying question method 00:00:26.079 --> 00:00:28.320 today. It's a game changer, really. It is. It's 00:00:28.320 --> 00:00:30.800 this remarkably simple technique that basically 00:00:30.800 --> 00:00:35.119 replaces those massive paragraph long super prompts 00:00:35.119 --> 00:00:37.119 we've all been trying to write. Yeah, no fluff. 00:00:37.299 --> 00:00:39.899 Just the exact mechanics of why this works. Exactly. 00:00:40.380 --> 00:00:41.960 So our roadmap for today, we're going to unpack 00:00:41.960 --> 00:00:44.560 that specific sentence that changes AI behavior. 00:00:44.700 --> 00:00:47.240 We'll dissect a real -world landscaping failure 00:00:47.240 --> 00:00:50.619 to really see the psychology of missing context. 00:00:50.659 --> 00:00:53.140 That's a fascinating example, too. Oh, it really 00:00:53.140 --> 00:00:55.340 is. And then we'll tailor this method across 00:00:55.340 --> 00:00:58.479 four specific use cases, and finally, build a 00:00:58.479 --> 00:01:00.689 five -step repeatable workflow. Because it's 00:01:00.689 --> 00:01:03.189 a powerful shift in perspective. We're moving 00:01:03.189 --> 00:01:06.049 from just throwing data at a wall to actually 00:01:06.049 --> 00:01:08.489 building a structural foundation. Right. Because 00:01:08.489 --> 00:01:11.569 before we can fix bad AI outputs, we kind of 00:01:11.569 --> 00:01:13.450 have to admit why they happen in the first place. 00:01:13.530 --> 00:01:16.049 Yeah. That's the hard part. It is. I mean, I'll 00:01:16.049 --> 00:01:18.689 admit, I still wrestle with prompt drift myself 00:01:18.689 --> 00:01:23.370 to sex silence. I catch myself assuming the machine 00:01:23.370 --> 00:01:26.870 just knows. We all do it. Right. Mostly because 00:01:27.049 --> 00:01:30.030 The context is so clear in my own head. I write 00:01:30.030 --> 00:01:32.329 what feels like a perfect explanation, and the 00:01:32.329 --> 00:01:35.670 result is just completely generic. Well, that's 00:01:35.670 --> 00:01:38.030 the trap of the curse of knowledge. You know 00:01:38.030 --> 00:01:40.329 exactly who the content is for. You know the 00:01:40.329 --> 00:01:43.989 tone, the success metrics, the subtle nuances. 00:01:44.250 --> 00:01:46.129 But none of that crucial information actually 00:01:46.129 --> 00:01:49.489 gets typed out. Precisely. The AI is essentially 00:01:49.489 --> 00:01:52.790 operating blind, so it's forced to rely on broad 00:01:52.790 --> 00:01:55.750 statistical averages just to fill in your gaps. 00:01:56.329 --> 00:01:57.989 ordering a steak without saying how you want 00:01:57.989 --> 00:02:00.269 it cooked. Oh, that's a perfect way to put it. 00:02:00.370 --> 00:02:02.030 You know, the chef has to guess. You can't really 00:02:02.030 --> 00:02:04.810 get mad when it arrives well done because you 00:02:04.810 --> 00:02:07.650 left out the main variable. Exactly. And the 00:02:07.650 --> 00:02:10.110 fix for this is incredibly straightforward. Just 00:02:10.110 --> 00:02:12.490 the one sentence. Yeah, you just add one specific 00:02:12.490 --> 00:02:14.530 sentence to the very end of whatever you're asking. 00:02:14.930 --> 00:02:18.009 You write, ask me three to five clarifying questions 00:02:18.009 --> 00:02:20.990 so you are 99 % certain of what I want you to 00:02:20.990 --> 00:02:23.819 do. Wait, so. Adding that one sentence actually 00:02:23.819 --> 00:02:26.159 stops it from hallucinating? It does. That sounds 00:02:26.159 --> 00:02:28.800 almost too easy. Does giving the AI permission 00:02:28.800 --> 00:02:31.599 to ask questions just make the model lazy? It 00:02:31.599 --> 00:02:34.039 actually forces the exact opposite. It triggers 00:02:34.039 --> 00:02:38.569 an evaluation step. Okay, how so? Well... Before 00:02:38.569 --> 00:02:41.110 it generates a response, the model has to scan 00:02:41.110 --> 00:02:43.469 your prompt against its training weights for 00:02:43.469 --> 00:02:46.949 that specific task. It actively looks for low 00:02:46.949 --> 00:02:49.810 -confidence variables. So instead of hallucinating 00:02:49.810 --> 00:02:52.930 an answer to bridge a logical gap, it surfaces 00:02:52.930 --> 00:02:55.830 that gap to you as a question. So it stops the 00:02:55.830 --> 00:02:58.590 guessing and forces rigorous analysis. Precisely. 00:02:58.590 --> 00:03:01.949 And if you look at OpenAI's own GPT 5 .5 prompting 00:03:01.949 --> 00:03:04.629 guide, it explicitly confirms this behavior. 00:03:04.689 --> 00:03:06.969 Right. The models are optimized to perform when 00:03:06.969 --> 00:03:09.240 desired. outcomes and constraints are mathematically 00:03:09.240 --> 00:03:12.639 defined. So take a simple request, like asking 00:03:12.639 --> 00:03:15.819 for a LinkedIn post about AI automation. Yeah, 00:03:16.000 --> 00:03:18.780 normally it just spits out a generic, enthusiastic 00:03:18.780 --> 00:03:22.180 post with way too many emojis. We've all seen 00:03:22.180 --> 00:03:24.580 those. They are painfully obvious. Right. But 00:03:24.580 --> 00:03:27.479 with that clarifying sentence, it stops. It actually 00:03:27.479 --> 00:03:30.639 asks you questions. Yeah. It asks, is your audience 00:03:30.639 --> 00:03:33.699 technical founders or general marketers? It asks, 00:03:34.020 --> 00:03:36.719 are we aiming for a contrarian take or an educational 00:03:36.719 --> 00:03:40.400 tone? What is the specific call to action, things 00:03:40.400 --> 00:03:42.340 like that? Exactly. It's pulling the implicit 00:03:42.340 --> 00:03:44.979 context out of your head and making it explicit 00:03:44.979 --> 00:03:48.039 text. And honestly, half the time, those questions 00:03:48.039 --> 00:03:50.180 make you realize you didn't actually know the 00:03:50.180 --> 00:03:52.699 answers yourself yet? Oh, absolutely. It clarifies 00:03:52.699 --> 00:03:55.319 your own thinking first. So since we know the 00:03:55.319 --> 00:03:57.020 AI is going to take our literal instructions 00:03:57.020 --> 00:03:59.460 and run wild with them, let's look at what happens 00:03:59.460 --> 00:04:01.780 when an intelligent guess is completely wrong. 00:04:01.919 --> 00:04:04.439 Yes, the landscaping example. Right. I was looking 00:04:04.439 --> 00:04:06.900 at this incredibly clear case study earlier. 00:04:06.919 --> 00:04:09.860 and it perfectly illustrates this failure. It's 00:04:09.860 --> 00:04:12.500 a brilliant look at the gap between instruction 00:04:12.500 --> 00:04:15.580 and intent. So imagine a landscaping company 00:04:15.580 --> 00:04:18.660 needing portfolio material. They want to show 00:04:18.660 --> 00:04:21.399 potential clients the incredible value of their 00:04:21.399 --> 00:04:23.939 renovation work. Right, so the user writes a 00:04:23.939 --> 00:04:27.639 very standard, reasonable prompt. They ask Gemini 00:04:27.639 --> 00:04:30.779 to create a side -by -side image showing a neglected 00:04:30.779 --> 00:04:33.920 lawn next to a healthy, well -maintained lawn. 00:04:34.120 --> 00:04:36.420 And Gemini renders a visually stunning image 00:04:36.420 --> 00:04:39.259 almost immediately. Technically, the output is 00:04:39.259 --> 00:04:42.019 flawless. Yeah, you have a dead, overgrown lawn 00:04:42.019 --> 00:04:44.459 on the left. You have a lush, manicured lawn 00:04:44.459 --> 00:04:46.720 on the right. But the image is totally useless 00:04:46.720 --> 00:04:48.899 for a portfolio. Completely useless. Because 00:04:48.899 --> 00:04:51.060 it generated two completely different properties. 00:04:51.139 --> 00:04:53.709 Right. The user wanted a temporal transformation. 00:04:54.129 --> 00:04:56.629 They wanted a before and after of the same house 00:04:56.629 --> 00:05:00.050 to show the impact of their work. OK, but I have 00:05:00.050 --> 00:05:02.009 to play devil's advocate here for a second. Sure. 00:05:02.290 --> 00:05:04.930 Shouldn't a sophisticated model just use common 00:05:04.930 --> 00:05:07.769 sense to know it's for a landscaping portfolio? 00:05:07.970 --> 00:05:09.870 Well, that's the fundamental illusion of AI. 00:05:10.050 --> 00:05:12.569 It doesn't have common sense. It just has parameters. 00:05:13.189 --> 00:05:16.149 Exactly. A model is just the underlying brain 00:05:16.149 --> 00:05:20.180 powering the AI assistant. It maps tokens. The 00:05:20.180 --> 00:05:22.959 model maps the phrase side by side and neglected 00:05:22.959 --> 00:05:26.939 next to healthy to spatial comparison data. Spatial, 00:05:27.019 --> 00:05:30.160 not temporal. Right. It doesn't map those words 00:05:30.160 --> 00:05:32.740 to temporal data like a time lapse unless you 00:05:32.740 --> 00:05:34.980 explicitly instruct it to bridge those two states 00:05:34.980 --> 00:05:37.879 across time. Right. And? Exactly. It just follows 00:05:37.879 --> 00:05:40.779 the math. Creating two distinct properties was 00:05:40.779 --> 00:05:43.699 highly logical. It was a mathematically sound 00:05:43.699 --> 00:05:46.519 interpretation of the exact text provided. Which 00:05:46.519 --> 00:05:49.079 is wild. How often do we penalize these tools 00:05:49.079 --> 00:05:51.220 for our own emissions? Oh, we do it constantly. 00:05:51.319 --> 00:05:53.720 We get frustrated and blame the tool. But the 00:05:53.720 --> 00:05:56.060 machine only knows the words you actually typed. 00:05:56.259 --> 00:05:58.620 The context of same property, different timeline, 00:05:59.079 --> 00:06:01.399 existed purely in the user's mind. We basically 00:06:01.399 --> 00:06:03.439 blame the algorithm for our own missing detail. 00:06:03.519 --> 00:06:06.860 We really do. So clarifying questions act as 00:06:06.860 --> 00:06:09.240 like an insurance policy against your own blind 00:06:09.240 --> 00:06:11.560 spots. Exactly. They catch the gap between what 00:06:11.560 --> 00:06:13.720 you typed and what you actually meant. So if 00:06:13.720 --> 00:06:16.639 the problem is missing details... The questions 00:06:16.639 --> 00:06:18.800 we wanted to ask are going to be wildly different 00:06:18.800 --> 00:06:20.920 depending on the task. Absolutely. A business 00:06:20.920 --> 00:06:23.220 plan needs different guardrails than a landscape 00:06:23.220 --> 00:06:26.100 photo, which means we need to adapt this method 00:06:26.100 --> 00:06:28.560 based on the tool we're using. Yeah, you can't 00:06:28.560 --> 00:06:31.319 use a universal set of questions. The context 00:06:31.319 --> 00:06:33.860 required for a Python script is totally useless 00:06:33.860 --> 00:06:36.720 for a marketing email. Right. So the source provides 00:06:36.720 --> 00:06:39.420 four distinct versions of this approach to cover 00:06:39.420 --> 00:06:43.240 your basis. Whoa, beat. Imagine scaling this 00:06:43.240 --> 00:06:46.680 to a billion queries. the amount of saved computing 00:06:46.680 --> 00:06:49.680 power is staggering. Oh, it's a massive leap 00:06:49.680 --> 00:06:52.759 in global efficiency just from preventing bad 00:06:52.759 --> 00:06:55.329 drafts. Let's break down those four specific 00:06:55.329 --> 00:06:57.730 use cases, starting with the basic version. Sure, 00:06:57.769 --> 00:06:59.949 so this is for your everyday low -stakes tasks. 00:07:00.170 --> 00:07:02.629 For an everyday task, I assume I don't need a 00:07:02.629 --> 00:07:05.350 massive interrogation from the AI? No, not at 00:07:05.350 --> 00:07:07.449 all. I just tell it my topic and let it figure 00:07:07.449 --> 00:07:09.949 out the main gaps? Exactly. You ask it to find 00:07:09.949 --> 00:07:12.509 the missing info and ask three to five questions? 00:07:12.889 --> 00:07:15.009 It's looking for basic constraints in what a 00:07:15.009 --> 00:07:17.509 successful outcome looks like. Okay, that prevents 00:07:17.509 --> 00:07:20.490 it from relying on a fixed generic checklist. 00:07:20.670 --> 00:07:23.790 Right. But what about high -stakes tasks? The 00:07:23.790 --> 00:07:25.910 second version is the business version, and it 00:07:25.910 --> 00:07:28.129 specifically highlights Claude as the best tool 00:07:28.129 --> 00:07:30.649 for this. Yeah, Claude is fantastic for this. 00:07:30.970 --> 00:07:34.129 Why is that? I mean, I know we can use chat GPT, 00:07:34.170 --> 00:07:36.769 but what makes Claude's architecture better for 00:07:36.769 --> 00:07:39.420 business logic? Well... Claude's underlying system 00:07:39.420 --> 00:07:42.480 prompts make it highly attuned to risk and constraints. 00:07:42.620 --> 00:07:46.240 It tends to explore complex, interconnected factors 00:07:46.240 --> 00:07:49.420 before making broad recommendations. So the business 00:07:49.420 --> 00:07:51.819 version prompt leans into that. It's asking about 00:07:51.819 --> 00:07:53.939 business objectives, metrics, and timelines. 00:07:54.139 --> 00:07:57.300 Yes. You instruct Claude to focus on resources 00:07:57.300 --> 00:07:59.600 and constraints before it attempts to build a 00:07:59.600 --> 00:08:02.139 strategic plan. That really elevates the output 00:08:02.139 --> 00:08:05.379 from just generic advice to actual strategic 00:08:05.379 --> 00:08:08.500 consulting. It really does. version is the image 00:08:08.500 --> 00:08:11.360 version, where Gemini really shines. Why Gemini? 00:08:11.560 --> 00:08:14.379 Because image generation is so unforgiving. A 00:08:14.379 --> 00:08:17.100 single misunderstood word changes the entire 00:08:17.100 --> 00:08:19.839 composition. Oh for sure. Gemini's multimodal 00:08:19.839 --> 00:08:22.860 integration allows it to be highly specific about 00:08:22.860 --> 00:08:26.279 visual nuance and composition. So what specific 00:08:26.279 --> 00:08:29.339 context is the image version looking for? It 00:08:29.339 --> 00:08:31.899 asks about the visual style, the audience, and 00:08:31.899 --> 00:08:35.059 the branding. But most importantly, it asks about 00:08:35.059 --> 00:08:38.230 negative constraints. Like what to avoid. Exactly. 00:08:38.730 --> 00:08:41.029 It asks what visual elements should be strictly 00:08:41.029 --> 00:08:44.049 avoided in the final render. That is huge. Knowing 00:08:44.049 --> 00:08:46.190 what not to do is usually more important than 00:08:46.190 --> 00:08:49.330 what to do in design. Oh, 100%. Finally, we have 00:08:49.330 --> 00:08:52.809 the content creator version. And ChatGPT is flagged 00:08:52.809 --> 00:08:56.059 as the strongest tool here. Yeah, because ChatGPT 00:08:56.059 --> 00:08:58.860 is trained heavily on adapting writing styles 00:08:58.860 --> 00:09:01.460 and conversational tones. Contact creation isn't 00:09:01.460 --> 00:09:03.639 just about the information, it's about the delivery. 00:09:03.779 --> 00:09:06.639 Right, it's about pacing. So ChatGPT asks about 00:09:06.639 --> 00:09:09.200 the specific platform. It looks for your hook 00:09:09.200 --> 00:09:11.759 angle. It clarifies the call to action. Exactly. 00:09:11.980 --> 00:09:14.279 It asks what the reader should think, feel, or 00:09:14.279 --> 00:09:16.820 do after reading it. Can we mix these tools? 00:09:17.559 --> 00:09:20.419 Say, use Claude's business logic to write a Gemini 00:09:20.419 --> 00:09:22.950 image prompt. Yes, tool stacking is an elite 00:09:22.950 --> 00:09:26.070 workflow. You use Claude's analytical power to 00:09:26.070 --> 00:09:28.610 interrogate your business needs. And build a 00:09:28.610 --> 00:09:31.529 comprehensive brief. Exactly. Then you hand that 00:09:31.529 --> 00:09:34.409 highly clarified brief over to Gemini for the 00:09:34.409 --> 00:09:36.549 actual generation. So different tools require 00:09:36.549 --> 00:09:38.490 completely different types of context clues. 00:09:38.850 --> 00:09:41.889 Exactly. You match the required context to the 00:09:41.889 --> 00:09:45.110 specific model's inherent strengths. Sponsor. 00:09:45.490 --> 00:09:47.750 This deep dive is supported by our sponsors, 00:09:48.049 --> 00:09:50.129 who help keep the lights on and the curiosity 00:09:50.129 --> 00:09:52.870 flowing. So having the theoretical prompts is 00:09:52.870 --> 00:09:55.250 great, but let's talk about the actual friction 00:09:55.250 --> 00:09:57.529 of doing this every day. Right, because a technique 00:09:57.529 --> 00:10:00.350 is only useful if you actually use it. Exactly. 00:10:00.590 --> 00:10:02.470 We want to turn this into a frictionless habit. 00:10:02.850 --> 00:10:04.929 There's a really sharp five -step workflow for 00:10:04.929 --> 00:10:07.250 this. You need a systematic way to start with 00:10:07.250 --> 00:10:10.250 a rough idea and refine it quickly. Step one 00:10:10.250 --> 00:10:12.370 is just the brain dump, right? You write a rough 00:10:12.370 --> 00:10:14.549 prompt defining the basic task. Yeah, you don't 00:10:14.549 --> 00:10:16.710 stress about getting it perfect. Just write a 00:10:16.710 --> 00:10:19.299 newsletter about productivity. Simple. Step 2 00:10:19.299 --> 00:10:21.779 is where you add the clarifying line, but there's 00:10:21.779 --> 00:10:25.539 a crucial detail here. Yes, very crucial. You 00:10:25.539 --> 00:10:28.659 must explicitly tell the AI not to use a generic 00:10:28.659 --> 00:10:30.940 checklist. Because if you don't constrain the 00:10:30.940 --> 00:10:33.639 questions, it's going to ask you 20 useless things 00:10:33.639 --> 00:10:36.559 about font choices. Exactly. You have to instruct 00:10:36.559 --> 00:10:39.220 it to prioritize only the questions with the 00:10:39.220 --> 00:10:42.179 highest impact on this specific task. This is 00:10:42.179 --> 00:10:45.320 exactly like stacking Lego blocks of data. You 00:10:45.320 --> 00:10:47.220 build the foundation before you put the roof 00:10:47.220 --> 00:10:49.840 on. That is a perfect analogy. You define the 00:10:49.840 --> 00:10:53.539 base. Which brings us to step three, answering 00:10:53.539 --> 00:10:55.519 the questions. And people tend to overthink this 00:10:55.519 --> 00:10:57.179 part, don't they? Oh, they feel like they need 00:10:57.179 --> 00:10:59.799 to write an essay in response. My biggest tip 00:10:59.799 --> 00:11:03.080 here is to keep it incredibly brief. Just punchy 00:11:03.080 --> 00:11:05.279 bullet points. Right. If it asks about audience, 00:11:05.480 --> 00:11:08.559 just type small business owners. Tone. Professional. 00:11:08.919 --> 00:11:12.259 You are just feeding it raw variables. Step four 00:11:12.259 --> 00:11:15.250 is critical for quality control. You don't just 00:11:15.250 --> 00:11:17.809 let the AI run wild and generate the draft yet. 00:11:17.990 --> 00:11:20.669 No, you ask it to create the final output only 00:11:20.669 --> 00:11:23.070 after confirming the brief. Right, in like three 00:11:23.070 --> 00:11:25.649 to five bullet points. And identifying any remaining 00:11:25.649 --> 00:11:27.769 assumptions it plans to make. This is where you 00:11:27.769 --> 00:11:30.889 catch that landscaping error. If the AI summarizes 00:11:30.889 --> 00:11:33.429 the brief and says, assuming you want two different 00:11:33.429 --> 00:11:35.750 houses, you stop it right there. You correct 00:11:35.750 --> 00:11:38.470 the temporal mapping before it renders a useless 00:11:38.470 --> 00:11:41.149 image. It keeps the output completely anchored. 00:11:41.250 --> 00:11:44.299 And finally. Step five is preserving that effort. 00:11:44.840 --> 00:11:48.320 You save the best version for reuse. Yeah, using 00:11:48.320 --> 00:11:51.440 tools like Text Blaze for browsers or Raycast 00:11:51.440 --> 00:11:53.620 Snippets if you're on a Mac. You just type a 00:11:53.620 --> 00:11:55.799 short trigger, it drops your customized clarifying 00:11:55.799 --> 00:11:58.980 prompt in, and you're off. It becomes automatic. 00:11:59.279 --> 00:12:01.600 Doesn't adding all these steps actually slow 00:12:01.600 --> 00:12:04.279 down our daily workflow? I hear that a lot, but 00:12:04.279 --> 00:12:07.299 no. It takes an extra 60 seconds up front, but 00:12:07.299 --> 00:12:09.919 it completely eliminates the half hour you usually 00:12:09.919 --> 00:12:12.620 spend. regenerating bad drafts. Taking a minute 00:12:12.620 --> 00:12:15.480 now saves hours of rewriting later. It's an upfront 00:12:15.480 --> 00:12:17.980 investment in structural integrity. This brings 00:12:17.980 --> 00:12:20.980 us to our big idea recap. We've covered the mechanics, 00:12:21.299 --> 00:12:24.279 the psychology, and the workflow. And the ultimate 00:12:24.279 --> 00:12:26.720 takeaway here completely changes how you view 00:12:26.720 --> 00:12:29.759 these tools. AI prompt quality isn't about word 00:12:29.759 --> 00:12:32.799 count, it's about context. Exactly. When you 00:12:32.799 --> 00:12:35.360 stop demanding immediate answers and start inviting 00:12:35.360 --> 00:12:38.659 a dialogue, the technology transforms. It stops 00:12:38.659 --> 00:12:41.000 being a basic search engine. And it becomes a 00:12:41.000 --> 00:12:45.059 highly capable collaborative partner. A few focused 00:12:45.059 --> 00:12:47.220 surgical questions will improve your results 00:12:47.220 --> 00:12:49.779 far more than agonizing over perfect adjectives. 00:12:50.000 --> 00:12:52.399 Absolutely. We want you to try this exact sentence 00:12:52.399 --> 00:12:55.460 on your next task today. Just copy it, paste 00:12:55.460 --> 00:12:57.379 it at the end of whatever you're asking the AI 00:12:57.379 --> 00:13:00.889 to do. Watch it pause, evaluate your logic, and 00:13:00.889 --> 00:13:02.970 engage with you. It's a small mechanical change 00:13:02.970 --> 00:13:04.970 that yields a massive difference. He really is. 00:13:05.110 --> 00:13:07.789 If pausing to ask three clarifying questions 00:13:07.789 --> 00:13:10.529 drastically improves how a machine understands 00:13:10.529 --> 00:13:14.210 us, two sec silence. Imagine what it could do 00:13:14.210 --> 00:13:15.929 for how we communicate with each other.