WEBVTT 00:00:00.000 --> 00:00:03.120 So picture this. You spend an entire evening, 00:00:03.180 --> 00:00:06.200 like, carefully typing out the absolute perfect 00:00:06.200 --> 00:00:08.599 detailed question for Claude. Oh yeah, we've 00:00:08.599 --> 00:00:10.939 all been there. Right. And you expect this sharp, 00:00:10.939 --> 00:00:14.519 brilliant answer back. But what you actually 00:00:14.519 --> 00:00:17.339 get is a response so painfully safe and generic 00:00:17.339 --> 00:00:19.480 that you basically just end up rewriting it yourself 00:00:19.480 --> 00:00:22.320 anyway. It is incredibly frustrating. And honestly, 00:00:22.480 --> 00:00:23.940 it makes it so easy to just throw your hands 00:00:23.940 --> 00:00:27.059 up, you know? You just assume the AI isn't as 00:00:27.059 --> 00:00:29.629 capable as the hype suggests. Welcome to this 00:00:29.629 --> 00:00:32.369 deep dive. Today we are looking at the actual 00:00:32.369 --> 00:00:34.750 mechanics of why that happens. Exactly. We are 00:00:34.750 --> 00:00:37.030 unpacking a source text on mastering the art 00:00:37.030 --> 00:00:39.590 of the Claude prompt. And our mission today is 00:00:39.590 --> 00:00:42.409 to figure out how treating Claude less like a 00:00:42.409 --> 00:00:44.890 magic eight ball and more like a real computational 00:00:44.890 --> 00:00:47.929 system completely changes your output. Right, 00:00:47.929 --> 00:00:50.369 because we are moving way past the basic how 00:00:50.369 --> 00:00:53.189 to write a prompt advice. We are looking at the 00:00:53.189 --> 00:00:55.969 literal mechanics of how a large language model 00:00:55.969 --> 00:00:58.789 processes your text. Yeah. And this happens before 00:00:58.789 --> 00:01:01.149 it even begins to generate a single word. We're 00:01:01.149 --> 00:01:04.329 going to explore six highly specific methods 00:01:04.329 --> 00:01:07.030 today. Everything from establishing contextual 00:01:07.030 --> 00:01:09.769 boundaries all the way up to triggering deep 00:01:09.769 --> 00:01:12.810 adaptive reasoning. Which is where things get 00:01:12.810 --> 00:01:16.689 really fascinating. Because if your input lacks... 00:01:16.510 --> 00:01:19.829 structure, and clear parameters, it will always 00:01:19.829 --> 00:01:22.709 default to a safe, generic average. Yeah, that 00:01:22.709 --> 00:01:24.769 makes sense. It's just a mathematical certainty, 00:01:25.109 --> 00:01:26.849 regardless of how powerful the model actually 00:01:26.849 --> 00:01:29.530 is. So I want to start at the foundation, the 00:01:29.530 --> 00:01:32.230 very bottom layer of how a model constructs a 00:01:32.230 --> 00:01:34.310 response. Which really comes down to identity. 00:01:34.670 --> 00:01:36.829 Right. Before Claude can answer effectively, 00:01:37.049 --> 00:01:39.530 it needs to know who it is. And I have to make 00:01:39.530 --> 00:01:41.530 a vulnerable admission right up front here. Oh. 00:01:41.950 --> 00:01:44.310 Let's hear it. I still wrestle with prompt drift 00:01:44.310 --> 00:01:46.829 myself, expecting the AI to just know what I 00:01:46.829 --> 00:01:48.569 want. Yeah, well, you are definitely not alone 00:01:48.569 --> 00:01:52.030 in that. It is human nature to anthropomorphize 00:01:52.030 --> 00:01:54.810 these systems. People skip the role assignment 00:01:54.810 --> 00:01:57.030 step constantly. They treat the chat box like 00:01:57.030 --> 00:01:58.989 a basic search engine and just type in a raw 00:01:58.989 --> 00:02:01.430 question. And the source text gives a highly 00:02:01.430 --> 00:02:04.709 practical example of this. Think about a plain 00:02:04.709 --> 00:02:08.050 request, like explain how a balance sheet works. 00:02:08.310 --> 00:02:11.289 Right. The model looks at that. and it pulls 00:02:11.289 --> 00:02:14.250 from the mathematical average of every time a 00:02:14.250 --> 00:02:16.770 balance sheet is mentioned in its training data. 00:02:17.009 --> 00:02:20.330 So you just get a completely accurate but perfectly 00:02:20.330 --> 00:02:22.889 dry answer. Exactly. It reads just like a textbook 00:02:22.889 --> 00:02:26.050 or a Wikipedia page. But then you take that exact 00:02:26.050 --> 00:02:29.169 same request and wrap it inside a specific persona. 00:02:29.550 --> 00:02:32.879 You write You are a CFO with 20 years of experience 00:02:32.879 --> 00:02:35.719 explaining financial concepts to non -finance 00:02:35.719 --> 00:02:37.960 executives. And then you say, explain how a balance 00:02:37.960 --> 00:02:40.699 sheet works. Right. The raw facts stay exactly 00:02:40.699 --> 00:02:43.219 the same. But because you apply that CFO rule, 00:02:43.520 --> 00:02:46.280 the output completely transforms. It really does. 00:02:46.319 --> 00:02:48.360 It stops sounding like an encyclopedia and starts 00:02:48.360 --> 00:02:50.599 reading like a practical executive summary. Think 00:02:50.599 --> 00:02:53.000 of Claude as a really smart new hire on their 00:02:53.000 --> 00:02:55.319 first day. The skill is there. It just needs 00:02:55.319 --> 00:02:57.259 direction. That's a great way to look at it. 00:02:57.319 --> 00:03:00.259 Because of how the model retrieves information, 00:03:00.919 --> 00:03:03.219 adding that one sentence dramatically shifts 00:03:03.219 --> 00:03:05.319 the probability distribution of the words it 00:03:05.319 --> 00:03:08.599 will choose next. Why does adding a fictional 00:03:08.599 --> 00:03:11.379 role change the actual factual data that comes 00:03:11.379 --> 00:03:13.560 back? Well, it doesn't change the facts. It changes 00:03:13.560 --> 00:03:16.740 the retrieval pathway. It filters the vast training 00:03:16.740 --> 00:03:19.819 data through a specific contextual lens. It gives 00:03:19.819 --> 00:03:23.219 the AI a specific lens to filter the information 00:03:23.219 --> 00:03:26.319 through. Exactly. You are shrinking the universe 00:03:26.319 --> 00:03:29.400 of possible answers down to what a seasoned executive 00:03:29.400 --> 00:03:31.620 would actually say. And the text mentions a great 00:03:31.620 --> 00:03:34.560 tip here. You can save these roles inside a Claude 00:03:34.560 --> 00:03:36.870 project or in your custom instructions. Yes. 00:03:37.689 --> 00:03:39.650 That way, they apply automatically to every new 00:03:39.650 --> 00:03:41.669 chat. You don't have to type it out every single 00:03:41.669 --> 00:03:46.370 day. But even if Claude knows it is a CFO, if 00:03:46.370 --> 00:03:49.250 you slide a messy, unorganized stack of papers 00:03:49.250 --> 00:03:51.969 across its desk, it's still going to fail quietly. 00:03:52.210 --> 00:03:54.770 Oh, absolutely. Which brings us to the next structural 00:03:54.770 --> 00:03:57.310 method. Once it knows who it is, it needs to 00:03:57.310 --> 00:03:59.289 know exactly what you were handing it. The problem 00:03:59.289 --> 00:04:01.810 is mashing background information, the actual 00:04:01.810 --> 00:04:04.810 task, and the formatting into one massive text 00:04:04.810 --> 00:04:07.449 block. People do this all the time. And the model 00:04:07.449 --> 00:04:09.669 is just left to guess where the context ends 00:04:09.669 --> 00:04:12.550 and the instructions begin. Even a great CFO 00:04:12.550 --> 00:04:15.370 persona will fail if the data is a mess. So we 00:04:15.370 --> 00:04:18.569 need to look at XML tags to solve this. And I'll 00:04:18.569 --> 00:04:20.550 define XML tags for you real quick. Please do. 00:04:20.800 --> 00:04:23.339 They are labels that create clear boundaries 00:04:23.339 --> 00:04:25.779 around different parts of your text. That's perfectly 00:04:25.779 --> 00:04:29.420 put. Anthropic's own documentation heavily recommends 00:04:29.420 --> 00:04:32.220 using these tags. You wrap your background info 00:04:32.220 --> 00:04:35.839 inside a tag, literally labeled context. And 00:04:35.839 --> 00:04:38.379 you put the task inside an instructions tag? 00:04:38.480 --> 00:04:40.860 Right. You can even nest them, like putting individual 00:04:40.860 --> 00:04:43.819 documents inside a larger document section. The 00:04:43.819 --> 00:04:46.399 source gives a really specific example involving 00:04:46.399 --> 00:04:49.339 Q2 SAS data. Oh yeah, the revenue report. Yeah. 00:04:49.579 --> 00:04:51.639 The messy version is just a brain dump. Like, 00:04:51.639 --> 00:04:54.759 here is Q2 data, revenue grew 12%, churn increased, 00:04:55.019 --> 00:04:57.500 hiring slowed down, analyze this. When you feed 00:04:57.500 --> 00:04:59.980 a language model a dense block of text like that, 00:05:00.439 --> 00:05:02.759 its attention mechanism just gets diluted. It 00:05:02.759 --> 00:05:04.620 is trying to weigh the importance of all those 00:05:04.620 --> 00:05:07.259 words simultaneously. Exactly. But when you use 00:05:07.259 --> 00:05:10.279 tags, the AI reads the purpose of each section 00:05:10.279 --> 00:05:13.439 immediately. The Q2 data sits cleanly inside 00:05:13.439 --> 00:05:15.959 a context bracket. It feels like stacking Lego 00:05:15.959 --> 00:05:17.959 blocks of data instead of just tossing them in 00:05:17.959 --> 00:05:21.600 a messy pile. That is a great analogy. It completely 00:05:21.600 --> 00:05:24.160 changes how the architecture processes the prompt. 00:05:24.300 --> 00:05:27.399 So does the model actually process tags differently 00:05:27.399 --> 00:05:30.579 than regular punctuation? It does. It has been 00:05:30.579 --> 00:05:33.240 trained to parse them as structural markers, 00:05:33.259 --> 00:05:36.800 which significantly reduces ambiguity. The tags 00:05:36.800 --> 00:05:39.420 act as literal walls, stopping the instructions 00:05:39.420 --> 00:05:41.649 from blurring together. Yes, exactly. They are 00:05:41.649 --> 00:05:44.569 structural walls. Okay, so we have our CFO persona 00:05:44.569 --> 00:05:47.430 and their workspace is neatly organized with 00:05:47.430 --> 00:05:51.529 XML tags. But we still have a problem. Interpretation. 00:05:51.709 --> 00:05:54.009 Right. Even neatly separated written instructions 00:05:54.009 --> 00:05:56.449 leave way too much room for interpretation. They 00:05:56.449 --> 00:05:58.730 really do. You might write a rule like write 00:05:58.730 --> 00:06:01.290 in a casual but professional tone and keep it 00:06:01.290 --> 00:06:03.930 under 100 words. What does that actually mean 00:06:03.930 --> 00:06:06.850 to an AI? Exactly. It lands differently every 00:06:06.850 --> 00:06:09.230 single time. It is just guessing what those abstract 00:06:09.230 --> 00:06:11.589 rules look like. So the source recommends showing 00:06:11.589 --> 00:06:14.769 examples over instructions. This is method three. 00:06:14.910 --> 00:06:17.670 And it is so effective. Anthropic recommends 00:06:17.670 --> 00:06:19.899 showing three to five examples wrapped inside 00:06:19.899 --> 00:06:22.800 an example tag to remove the guesswork. Let's 00:06:22.800 --> 00:06:25.000 look at the noise canceling headphones task from 00:06:25.000 --> 00:06:27.519 the text. OK, yeah. You need a product description. 00:06:27.720 --> 00:06:29.160 Right. If you just give instructions, you're 00:06:29.160 --> 00:06:31.399 writing things like keep the tone light, don't 00:06:31.399 --> 00:06:34.480 be too salesy, avoid long sentences. Which is 00:06:34.480 --> 00:06:37.459 just a minefield for the AI. Yeah. But if you 00:06:37.459 --> 00:06:40.360 just give Claude a direct pattern to match, it 00:06:40.360 --> 00:06:43.100 is infinitely better than explaining what it 00:06:43.100 --> 00:06:46.360 should avoid. A positive example carries so much 00:06:46.360 --> 00:06:48.500 more weight than a long list of restrictions. 00:06:48.660 --> 00:06:51.800 Cool. So why is showing a positive example so 00:06:51.800 --> 00:06:53.879 much stronger than giving a list of negative 00:06:53.879 --> 00:06:56.459 constraints? Well, pattern matching is the core 00:06:56.459 --> 00:06:59.160 strength of language models. They are built to 00:06:59.160 --> 00:07:02.040 identify and complete patterns naturally. Showing 00:07:02.040 --> 00:07:04.360 it what works is simply faster than eliminating 00:07:04.360 --> 00:07:07.560 what doesn't. Exactly. It takes a lot of processing 00:07:07.560 --> 00:07:11.120 overhead for an AI to navigate complex logical 00:07:11.120 --> 00:07:13.699 exclusions. Just show it what you want. meet. 00:07:14.040 --> 00:07:16.459 But what happens when the task gets so complex 00:07:16.459 --> 00:07:18.959 that even great examples aren't enough? That 00:07:18.959 --> 00:07:21.019 is when you hit a processing ceiling. Right. 00:07:21.720 --> 00:07:25.939 Which brings us to method four. Breaking tasks 00:07:25.939 --> 00:07:28.939 into a prompt chain. When you have four different 00:07:28.939 --> 00:07:31.300 jobs competing in one prompt -like research, 00:07:31.660 --> 00:07:34.939 analyze, draft, and format, each job gets less 00:07:34.939 --> 00:07:37.639 processing focus. So you have to split the work. 00:07:37.740 --> 00:07:40.399 Explain how the chain works in practice. Well, 00:07:40.420 --> 00:07:43.980 prompt one, extracts, say ten specific findings 00:07:43.980 --> 00:07:46.459 from your data. Okay. Then Prop 2 takes those 00:07:46.459 --> 00:07:48.660 findings and groups them into three business 00:07:48.660 --> 00:07:52.060 themes. And finally, Prop 3 turns that into a 00:07:52.060 --> 00:07:53.899 management report with action recommendations. 00:07:54.240 --> 00:07:56.360 And the source mentions a self -correction pattern 00:07:56.360 --> 00:07:58.480 at the end. Yeah, you ask it to generate a draft. 00:07:58.750 --> 00:08:01.410 review it against your criteria, and then refine 00:08:01.410 --> 00:08:03.149 it. I do want to point out some nuance here, 00:08:03.589 --> 00:08:05.750 though. Anthropic's current guidance has actually 00:08:05.750 --> 00:08:08.470 changed on this. It has, yeah. Because newer 00:08:08.470 --> 00:08:11.170 models have adaptive thinking that handles a 00:08:11.170 --> 00:08:13.449 lot of this internally now. So since newer models 00:08:13.449 --> 00:08:15.610 think adaptively, is manual chaining becoming 00:08:15.610 --> 00:08:19.189 obsolete? It's not obsolete, no. It's just repurposed. 00:08:19.750 --> 00:08:22.170 Manual chaining is best reserved now for inspecting 00:08:22.170 --> 00:08:24.910 intermediate steps or enforcing a strict pipeline. 00:08:25.069 --> 00:08:28.129 Not obsolete, just shifting to quality control 00:08:28.129 --> 00:08:31.209 for specific multi -step pipelines. Exactly. 00:08:31.470 --> 00:08:33.570 It's for when you really need to audit the AI's 00:08:33.570 --> 00:08:35.929 work step by step. Speaking of the model's internal 00:08:35.929 --> 00:08:39.129 processing, that brings us to method five, adaptive 00:08:39.129 --> 00:08:42.850 thinking for accuracy. This is huge. Some tasks 00:08:42.850 --> 00:08:45.929 demand a profound level of accuracy, where you 00:08:45.929 --> 00:08:48.169 literally need the AI to slow down and think. 00:08:48.269 --> 00:08:50.809 Right. The text explains two controls for this. 00:08:51.210 --> 00:08:53.529 The effort setting, which goes from low to max. 00:08:53.850 --> 00:08:56.490 and the thinking toggle. Everyday tasks like 00:08:56.490 --> 00:08:59.470 drafting short emails just need speed. You keep 00:08:59.470 --> 00:09:01.970 it on low effort. But high stakes tasks like 00:09:01.970 --> 00:09:04.529 financial analysis or strategic decisions, those 00:09:04.529 --> 00:09:06.750 need serious depth. Let's dive into the course 00:09:06.750 --> 00:09:08.970 pricing example to show this. Yeah, contrast 00:09:08.970 --> 00:09:11.870 a simple prompt like suggest a price for my course 00:09:11.870 --> 00:09:14.289 with a deeply constrained one. Right, the simple 00:09:14.289 --> 00:09:16.769 one just gives you a generic guess. But the constrained 00:09:16.769 --> 00:09:19.470 prompt asks Quad to weigh monthly retention, 00:09:19.789 --> 00:09:22.769 churn percentages, and customer acquisition cost. 00:09:22.919 --> 00:09:26.179 And it asks it to calculate a trade -off between 00:09:26.179 --> 00:09:29.080 three -month cash flow and one -year total profit. 00:09:29.159 --> 00:09:33.259 Exactly. Whoa. Imagine it weighing those real 00:09:33.259 --> 00:09:35.820 -world constraints perfectly before it even types 00:09:35.820 --> 00:09:38.259 a word. It's incredible. It basically stops being 00:09:38.259 --> 00:09:40.539 a text generator and becomes a strategic modeling 00:09:40.539 --> 00:09:44.500 engine. But how do we recognize in our own daily 00:09:44.500 --> 00:09:47.059 work when a task actually warrants this extra 00:09:47.059 --> 00:09:50.080 computing time? You use the doubt test. When 00:09:50.080 --> 00:09:52.500 a fast response sounds right, but leaves you 00:09:52.500 --> 00:09:54.240 doubting if it holds up to scrutiny. Because 00:09:54.240 --> 00:09:56.500 a wrong answer costs more than waiting for deeper 00:09:56.500 --> 00:09:59.120 reasoning. Exactly. You weigh the cost of a wrong 00:09:59.120 --> 00:10:01.559 answer against the value of your time. Look for 00:10:01.559 --> 00:10:04.120 doubt. If a generic answer costs you money, turn 00:10:04.120 --> 00:10:06.600 it on. Perfect rule of thumb. We are going to 00:10:06.600 --> 00:10:09.340 take a quick break for our sponsor. Sponsor. 00:10:09.779 --> 00:10:13.340 Welcome back. So deep reasoning solves the tough 00:10:13.340 --> 00:10:16.399 calculations. But what about simple frustrating 00:10:16.399 --> 00:10:18.480 mistakes that just keep happening week after 00:10:18.480 --> 00:10:21.559 week? That is method six. Fix the prompt system. 00:10:21.629 --> 00:10:24.409 not just the output. A cloud prompt should perform 00:10:24.409 --> 00:10:27.009 better 30 days from now than it does today. It 00:10:27.009 --> 00:10:29.230 really should. There's a great story in the text 00:10:29.230 --> 00:10:31.909 about a weekly customer revenue report. Oh, right. 00:10:31.990 --> 00:10:34.149 For three weeks in a row, it came back missing 00:10:34.149 --> 00:10:37.490 a month over month comparison. Yeah. And the 00:10:37.490 --> 00:10:40.009 user had to fix it manually every single time. 00:10:40.309 --> 00:10:42.970 They kept tweaking the final text. So what was 00:10:42.970 --> 00:10:46.320 the fix? Stop changing the output. Add a constraint 00:10:46.320 --> 00:10:49.039 directly to the prompt. They added, always compare 00:10:49.039 --> 00:10:51.679 this month's numbers to last month's and state 00:10:51.679 --> 00:10:54.580 the percentage change. Exactly. And by placing 00:10:54.580 --> 00:10:57.659 this fix inside a Claude project, it applies 00:10:57.659 --> 00:11:00.440 automatically to all future chats. It removes 00:11:00.440 --> 00:11:02.980 the repeated mistake right at its source. Yep. 00:11:03.120 --> 00:11:05.860 No more manual edits. So why is human nature 00:11:05.860 --> 00:11:08.519 so resistant to fixing the prompt instead of 00:11:08.519 --> 00:11:11.480 just tweaking the output? Honestly, editing text 00:11:11.480 --> 00:11:14.460 feels like tangible progress. Tweaking a final 00:11:14.460 --> 00:11:17.480 draft feels fast, while debugging a prompt feels 00:11:17.480 --> 00:11:19.700 like coding. Because tweaking the final text 00:11:19.700 --> 00:11:22.679 feels faster in the moment. Exactly. But it traps 00:11:22.679 --> 00:11:25.539 you in a cycle of endless rework. We have covered 00:11:25.539 --> 00:11:28.960 a massive amount of ground today, and I want 00:11:28.960 --> 00:11:31.179 to synthesize the overarching core philosophy 00:11:31.179 --> 00:11:33.200 here. Yeah, let's bring it all together. A truly 00:11:33.200 --> 00:11:35.980 strong Claude prompt doesn't come from one clever 00:11:35.980 --> 00:11:39.220 hack or a magic sentence. Not at all. It is a 00:11:39.220 --> 00:11:41.679 connected system of habits. You're setting roles, 00:11:42.039 --> 00:11:44.740 defining boundaries with tags, leading with positive 00:11:44.740 --> 00:11:47.759 examples, chaining complex steps, applying adaptive 00:11:47.759 --> 00:11:50.080 thinking, and relentlessly debugging the prompt 00:11:50.080 --> 00:11:53.919 itself. Two secs silence. If you stop treating 00:11:53.919 --> 00:11:56.360 AI like a magic search box and started treating 00:11:56.360 --> 00:11:58.639 it like a highly capable colleague who just needs 00:11:58.639 --> 00:12:01.059 a clean desk, a clear role, and a good onboarding, 00:12:01.419 --> 00:12:03.639 How much time could you actually buy back next 00:12:03.639 --> 00:12:06.120 month? It is a profound shift in perspective. 00:12:06.279 --> 00:12:07.980 It really is something to think about. Thank 00:12:07.980 --> 00:12:10.320 you for joining us for this Leap Dive. Stay curious 00:12:10.320 --> 00:12:11.519 and we will talk to you next time.