WEBVTT 00:00:00.000 --> 00:00:01.940 So if you've ever talked to an AI like Claude 00:00:01.940 --> 00:00:03.960 and you've asked for something deep, something 00:00:03.960 --> 00:00:06.740 professional, but what you get back just feels, 00:00:06.860 --> 00:00:09.640 I don't know, generic. A little flat. Yeah, flat 00:00:09.640 --> 00:00:12.119 or even just wrong. It can feel like you've just 00:00:12.119 --> 00:00:14.560 hit a brick wall. That frustration is so common. 00:00:14.679 --> 00:00:16.500 But what we're learning is that the wall isn't 00:00:16.500 --> 00:00:19.839 really the model's power. It's the quality of 00:00:19.839 --> 00:00:21.500 the conversation we're starting. We're whispering 00:00:21.500 --> 00:00:24.940 at it. Exactly. We're giving these vague instructions 00:00:24.940 --> 00:00:28.239 to this massive system. And then we act surprised 00:00:28.239 --> 00:00:31.660 when... The output is just, you know, high -tech 00:00:31.660 --> 00:00:34.479 mush. Well, that ends today. We know that you, 00:00:34.579 --> 00:00:36.840 the learner, are always looking for that shortcut, 00:00:37.000 --> 00:00:39.619 that efficient path to getting high -value knowledge 00:00:39.619 --> 00:00:43.340 and mastering AI communication. That is the path. 00:00:43.479 --> 00:00:45.619 It's the ultimate shortcut. So today we're doing 00:00:45.619 --> 00:00:48.399 a deep dive into the playbook. We've taken Anthropic's 00:00:48.399 --> 00:00:50.679 own guidance, their 10 golden rules for crafting 00:00:50.679 --> 00:00:53.100 really effective prompts for models like Claude 00:00:53.100 --> 00:00:56.090 Opus and Sonnet. Our mission today is to move 00:00:56.090 --> 00:00:59.070 you from just being a passive user of AI to becoming 00:00:59.070 --> 00:01:01.810 a precision orchestrator. Yeah, and we're going 00:01:01.810 --> 00:01:03.750 to structure this as a clear path. We've got 00:01:03.750 --> 00:01:06.310 three pillars, okay? First, the foundational 00:01:06.310 --> 00:01:09.549 rules for control and intent. Second, how to 00:01:09.549 --> 00:01:12.170 structure execution for big projects. And then 00:01:12.170 --> 00:01:14.409 finally, the advanced tricks that the top 1 % 00:01:14.409 --> 00:01:16.769 are using. Before we jump in, there's one term 00:01:16.769 --> 00:01:19.670 we should probably define. Right. Few -shot prompting. 00:01:20.079 --> 00:01:21.959 What is that? It's actually really simple. It 00:01:21.959 --> 00:01:24.700 just means giving the AI examples. You show Claude 00:01:24.700 --> 00:01:27.099 a sample of the perfect output you want, the 00:01:27.099 --> 00:01:30.140 tone, the format, and you just tell it, do this. 00:01:30.319 --> 00:01:33.019 Got it. Mimic this exactly. Exactly. Okay, so 00:01:33.019 --> 00:01:34.959 let's start with that foundation. Rule number 00:01:34.959 --> 00:01:40.069 one is to be, and I like this wording. painfully, 00:01:40.069 --> 00:01:42.829 obsessively specific. Painfully is the key word. 00:01:43.049 --> 00:01:46.310 This isn't just a good habit. It's about understanding 00:01:46.310 --> 00:01:48.469 how literal these machines are. If you leave 00:01:48.469 --> 00:01:50.909 any room for ambiguity, the AI has to guess. 00:01:51.069 --> 00:01:53.579 And guessing is where hallucinations start. It's 00:01:53.579 --> 00:01:56.120 where you get bad output. Quad is perfectly, 00:01:56.219 --> 00:01:58.379 almost unnervingly literal. You have to close 00:01:58.379 --> 00:02:01.200 every single loophole. The analogy they use is 00:02:01.200 --> 00:02:03.980 great. Don't just ask for a pizza. Right. You'll 00:02:03.980 --> 00:02:06.280 get who knows what. You have to say a large, 00:02:06.280 --> 00:02:09.039 thin crust pepperoni pizza with extra cheese 00:02:09.039 --> 00:02:12.439 cut into squares. And when you apply that same 00:02:12.439 --> 00:02:15.759 logic. to high stakes work, that's where the 00:02:15.759 --> 00:02:18.080 power is. Don't just say write a Python function. 00:02:18.340 --> 00:02:20.340 Specify everything. Write a Python function for 00:02:20.340 --> 00:02:22.340 a data analyst audience. Make sure all the doc 00:02:22.340 --> 00:02:25.159 strings follow PEP8 standards. And I need three 00:02:25.159 --> 00:02:28.400 specific examples for edge cases. There's no 00:02:28.400 --> 00:02:30.580 room for randomness there. That leads right into 00:02:30.580 --> 00:02:34.060 rule two, which is tell the AI why you want it. 00:02:34.080 --> 00:02:36.219 The intent principle. And this is transformative. 00:02:36.479 --> 00:02:39.800 A command tells the AI what to do. Intent tells 00:02:39.800 --> 00:02:41.979 it why. And that turns a command into a mission. 00:02:42.300 --> 00:02:45.699 Oh, okay. The why is the secret sauce. If I just 00:02:45.699 --> 00:02:48.500 say, summarize this legal document, you might 00:02:48.500 --> 00:02:50.360 focus on everything equally, the footnotes, the 00:02:50.360 --> 00:02:53.000 header. Right. But if I say, summarize this for 00:02:53.000 --> 00:02:55.360 a non -lawyer executive who needs to grasp key 00:02:55.360 --> 00:02:58.659 risks before a meeting in one hour, now the AI 00:02:58.659 --> 00:03:00.699 knows. Now it knows that anything that isn't 00:03:00.699 --> 00:03:02.800 a key risk is just noise. Exactly. It changes 00:03:02.800 --> 00:03:04.879 its entire approach. It becomes a specialized 00:03:04.879 --> 00:03:07.229 risk analyst, not just a summarizer. And that 00:03:07.229 --> 00:03:09.370 priority shift is where rule three comes in. 00:03:09.509 --> 00:03:12.810 Your examples are the unquestionable truth. This 00:03:12.810 --> 00:03:15.370 is that few shot prompting inaction. When you 00:03:15.370 --> 00:03:17.750 give Claude a sample, it treats it like sacred 00:03:17.750 --> 00:03:20.759 text. It will mimic that style, that structure, 00:03:20.840 --> 00:03:23.680 with total fidelity. That's so powerful. You 00:03:23.680 --> 00:03:25.460 don't have to try and describe an abstract tone 00:03:25.460 --> 00:03:27.860 like professional but approachable. No, you just 00:03:27.860 --> 00:03:29.500 show it. You just give it an example of what 00:03:29.500 --> 00:03:31.419 that looks like. Yeah, the best practice is to 00:03:31.419 --> 00:03:34.539 keep a folder of your own best work, your best 00:03:34.539 --> 00:03:37.000 emails, your best reports. And when you need 00:03:37.000 --> 00:03:39.319 something new, you just say, here's the tone 00:03:39.319 --> 00:03:42.250 I need. Now apply it to this new request. It's 00:03:42.250 --> 00:03:45.129 the fastest way to clone your own voice. So by 00:03:45.129 --> 00:03:48.169 supplying that intent, how does the model fundamentally 00:03:48.169 --> 00:03:50.990 change its approach? Intent lets it filter content. 00:03:51.810 --> 00:03:54.229 It prioritizes the important stuff like risks 00:03:54.229 --> 00:03:57.150 over the low -value stuff like jargon. Okay, 00:03:57.189 --> 00:03:59.469 so if those first three rules are about defining 00:03:59.469 --> 00:04:01.949 the mission, what happens when you have a long, 00:04:02.009 --> 00:04:04.389 complex task? Right. Now we're shifting from 00:04:04.389 --> 00:04:08.270 being a basic user to someone who actually structures 00:04:08.270 --> 00:04:10.150 the entire workflow. This is where you become 00:04:10.150 --> 00:04:12.770 the power user. Exactly. And it starts with rule 00:04:12.770 --> 00:04:16.290 four. Big projects need small, strategic checkpoints. 00:04:16.670 --> 00:04:18.709 Oh, this is a big one. Trying to get a 50 -page 00:04:18.709 --> 00:04:21.589 report in one shot. It's such a common mistake. 00:04:21.769 --> 00:04:24.810 It just overwhelms the model. And the result 00:04:24.810 --> 00:04:27.790 is almost always this incoherent blob where the 00:04:27.790 --> 00:04:30.050 beginning has no idea what the end is doing. 00:04:30.350 --> 00:04:32.529 I've done this. I tried to write a 10 -page market 00:04:32.529 --> 00:04:34.970 analysis and it spent, I don't know, 40 % of 00:04:34.970 --> 00:04:37.870 its time on this huge, irrelevant intro about 00:04:37.870 --> 00:04:39.870 the history of market analysis. That's such a 00:04:39.870 --> 00:04:42.350 relatable failure. Yeah. The fix is to become 00:04:42.350 --> 00:04:44.800 the project manager. You force what's called 00:04:44.800 --> 00:04:47.379 a chain of thought approach. So step by step. 00:04:47.459 --> 00:04:49.939 Step by step. First, just generate the outline. 00:04:50.420 --> 00:04:52.620 That's it. You approve it. Then you say, okay, 00:04:52.680 --> 00:04:54.860 now write the introduction based only on that 00:04:54.860 --> 00:04:56.819 outline. Then chapter one, you course correct 00:04:56.819 --> 00:04:59.480 at every step. That way you catch a fatal flaw 00:04:59.480 --> 00:05:02.360 early instead of after an hour of wasted generation. 00:05:02.920 --> 00:05:05.519 Precisely. Which brings us to rule five, which 00:05:05.519 --> 00:05:09.699 is about tone. Demand action, not advice. Be 00:05:09.699 --> 00:05:12.680 direct. Be very direct. Use an optive command. 00:05:13.180 --> 00:05:14.939 You're pushing it into execution mode, so you 00:05:14.939 --> 00:05:16.959 want to avoid all that softening language. But 00:05:16.959 --> 00:05:19.600 what if you're using it for, say, client communication? 00:05:20.160 --> 00:05:22.379 You don't want to sound abrasive. The key is 00:05:22.379 --> 00:05:24.420 that the command needs to be clear, not rude. 00:05:25.259 --> 00:05:27.779 The passive way is asking, what do you think 00:05:27.779 --> 00:05:30.720 of this Python code? The active, precise way 00:05:30.720 --> 00:05:33.160 is, refactor this code to improve performance 00:05:33.160 --> 00:05:36.379 and readability. Return only the refactored code. 00:05:37.720 --> 00:05:40.120 The command itself dictates the format and the 00:05:40.120 --> 00:05:43.600 goal. No fluff needed. Right. Clarity is kindness 00:05:43.600 --> 00:05:47.699 to the AI, which leads to rule sex, specifying 00:05:47.699 --> 00:05:50.220 the what, not the whatnot. Controlling the format. 00:05:50.420 --> 00:05:53.569 This is so critical for automation. Telling an 00:05:53.569 --> 00:05:56.870 AI what not to do is actually really vague. Phrases 00:05:56.870 --> 00:05:59.870 like don't use markdown or don't use bullet points 00:05:59.870 --> 00:06:02.129 get ignored all the time. You know, this is my 00:06:02.129 --> 00:06:04.589 vulnerable admission for the day. I still wrestle 00:06:04.589 --> 00:06:06.649 with this. I use negative commands and it just 00:06:06.649 --> 00:06:08.509 seems to ignore them sometimes. It spits out 00:06:08.509 --> 00:06:10.029 the bullet points anyway. And that's exactly 00:06:10.029 --> 00:06:12.689 why. The model prioritizes the positive act of 00:06:12.689 --> 00:06:15.670 generating text over remembering a negative constraint. 00:06:15.930 --> 00:06:19.470 So the fix is positive instruction. Always. Instead 00:06:19.470 --> 00:06:22.170 of what to avoid, define the exact structured 00:06:22.170 --> 00:06:25.410 output you want. For instance, output the final 00:06:25.410 --> 00:06:28.269 answer as a single JSON object with two keys, 00:06:28.490 --> 00:06:31.209 summary and key points, which must be an array 00:06:31.209 --> 00:06:35.209 of strings. That is unambiguous. For the learner, 00:06:35.329 --> 00:06:37.250 what is the biggest risk of using a negative 00:06:37.250 --> 00:06:40.779 command like don't use X? They're vague. The 00:06:40.779 --> 00:06:43.300 model often ignores them and does the exact thing 00:06:43.300 --> 00:06:45.319 you want it to avoid. So we've got the foundation. 00:06:45.319 --> 00:06:47.540 We've got the structure. These last four rules, 00:06:47.680 --> 00:06:49.439 these are the advanced techniques. Yeah, this 00:06:49.439 --> 00:06:51.000 is where the magic happens. This is what the 00:06:51.000 --> 00:06:52.800 expert prompters are doing. Okay, what's rule 00:06:52.800 --> 00:06:55.620 seven? Rule seven is to use simple labels, specifically 00:06:55.620 --> 00:06:58.980 XML tags, to steer behavior. Most people just 00:06:58.980 --> 00:07:01.120 throw everything into the prompt. The instructions, 00:07:01.300 --> 00:07:03.740 the context, the examples, it's all one big blob. 00:07:04.019 --> 00:07:06.459 And Claude sees it that way too. Right. Until 00:07:06.459 --> 00:07:09.420 you use these markers. Tags like instructions 00:07:09.420 --> 00:07:13.139 or document act like highlighters for the AI. 00:07:13.339 --> 00:07:15.740 It helps it cleanly separate the different parts 00:07:15.740 --> 00:07:17.459 of your prompt. So you're creating structural 00:07:17.459 --> 00:07:20.000 blocks for it. Exactly. You wrap your source 00:07:20.000 --> 00:07:22.399 material in a document tag, your instructions 00:07:22.399 --> 00:07:25.800 in an instructions tag. This has a huge effect 00:07:25.800 --> 00:07:27.879 on how the model parses everything. It gets much 00:07:27.879 --> 00:07:30.279 cleaner, much more predictable. Whoa. So you 00:07:30.279 --> 00:07:33.579 could use tags to manage, like... Dozens of different 00:07:33.579 --> 00:07:36.500 documents in a single prompt for a huge comparative 00:07:36.500 --> 00:07:39.120 analysis. Now you're getting it. It's like stacking 00:07:39.120 --> 00:07:42.040 Lego blocks of data. It's a game changer for 00:07:42.040 --> 00:07:44.480 managing these massive context windows. That 00:07:44.480 --> 00:07:47.540 makes so much sense. And then rule eight is for 00:07:47.540 --> 00:07:50.639 any kind of sequential work. If you're using 00:07:50.639 --> 00:07:53.199 an agent workflow, you have to announce it. You 00:07:53.199 --> 00:07:55.339 mean for tasks that take multiple steps? Yeah, 00:07:55.379 --> 00:07:57.220 any back and forth. You have to tell the AI that 00:07:57.220 --> 00:08:00.649 the job is ongoing. Otherwise, its context could 00:08:00.649 --> 00:08:02.810 just reset between turns. You're basically giving 00:08:02.810 --> 00:08:05.129 it short -term memory. So just a single sentence, 00:08:05.230 --> 00:08:07.329 like you are an agent in a multi -turn workflow. 00:08:07.689 --> 00:08:10.670 That's it. That one line forces it to track its 00:08:10.670 --> 00:08:13.029 own progress and saves you from repeating yourself 00:08:13.029 --> 00:08:15.949 over and over. Okay. Rule nine is about external 00:08:15.949 --> 00:08:19.149 tools, right? Like a calculator. Yes. And the 00:08:19.149 --> 00:08:23.589 guidance here is don't scream must. If you command 00:08:23.589 --> 00:08:25.990 it, you must use the calculator for all math. 00:08:26.290 --> 00:08:29.050 You get clumsy results. It'll call the calculator 00:08:29.050 --> 00:08:31.990 for 2 plus 2. Which just slows everything down. 00:08:32.090 --> 00:08:34.309 You want it to use judgment. Precisely. So you 00:08:34.309 --> 00:08:37.070 use softer guiding language. You should use the 00:08:37.070 --> 00:08:39.289 tool when you need to perform complex calculations. 00:08:39.649 --> 00:08:42.710 This lets Claude exercise some judgment and only 00:08:42.710 --> 00:08:45.090 use the tool when it's actually helpful. It's 00:08:45.090 --> 00:08:47.649 a much smarter outcome. Much smarter. And finally, 00:08:47.750 --> 00:08:49.789 rule 10 is one of the more fascinating ones. 00:08:49.970 --> 00:08:52.879 If your output feels sloppy. Avoid the word think. 00:08:53.360 --> 00:08:56.480 Really? Avoid think? It sounds like a superstition, 00:08:56.500 --> 00:08:58.639 I know, but the testing on this is consistent. 00:08:59.179 --> 00:09:02.500 Asking the model to think seems to invite speculation, 00:09:02.840 --> 00:09:05.080 maybe shallower reasoning. So what do you use 00:09:05.080 --> 00:09:08.620 instead? Stronger, action -oriented words. Consider, 00:09:08.799 --> 00:09:11.559 analyze, evaluate. You're demanding a more structured 00:09:11.559 --> 00:09:14.600 step -by -step process. It's like analyze forces 00:09:14.600 --> 00:09:17.100 it to show its work, whereas think lets it just 00:09:17.100 --> 00:09:19.220 reassociate. That's a perfect way to describe 00:09:19.220 --> 00:09:21.779 it. It shows that good prompting is as much about 00:09:21.779 --> 00:09:24.659 empirical observation as it is about perfect 00:09:24.659 --> 00:09:27.720 logic. Beyond the prompt itself, what changes 00:09:27.720 --> 00:09:31.139 when the model expects an agent workflow? It 00:09:31.139 --> 00:09:33.320 starts tracking progress and preparing for the 00:09:33.320 --> 00:09:36.259 next step. It avoids looping or restarting from 00:09:36.259 --> 00:09:39.179 scratch. When you tie all of these rules together, 00:09:39.360 --> 00:09:42.600 the big idea that emerges is that the AI is, 00:09:42.740 --> 00:09:46.549 at its core, intensely literal. And it really 00:09:46.549 --> 00:09:48.429 wants to be helpful. You're moving away from 00:09:48.429 --> 00:09:50.330 treating it like a search engine. Exactly. You 00:09:50.330 --> 00:09:52.149 stop whispering at it and you start treating 00:09:52.149 --> 00:09:55.350 it like a highly specialized, hyper -competent 00:09:55.350 --> 00:09:57.789 employee. So the core transformation for you, 00:09:57.889 --> 00:10:00.990 the listener, is to really embrace being the 00:10:00.990 --> 00:10:04.029 director. Claude is the actor and it's just waiting 00:10:04.029 --> 00:10:06.690 for precise instructions. And when you give it 00:10:06.690 --> 00:10:08.990 that specificity, that intent, and that clear 00:10:08.990 --> 00:10:12.299 structure. the results shift almost immediately. 00:10:12.580 --> 00:10:15.120 From generic to really professional grade output. 00:10:15.360 --> 00:10:17.759 So here's our call to action. We encourage you 00:10:17.759 --> 00:10:20.120 to open Clog right now or whatever model you're 00:10:20.120 --> 00:10:23.440 using and just apply rule number one, hyper -specificity, 00:10:23.440 --> 00:10:26.519 and rule number two, the intent principle. You're 00:10:26.519 --> 00:10:28.299 going to see a difference right away. An immediate 00:10:28.299 --> 00:10:30.159 difference in the quality of what you get back. 00:10:30.480 --> 00:10:32.639 And maybe here's a final thought to mull over. 00:10:33.440 --> 00:10:35.840 If we need this level of precision, this much 00:10:35.840 --> 00:10:39.179 intent and structural clarity to communicate 00:10:39.179 --> 00:10:41.899 well with an AI, what does that tell us about 00:10:41.899 --> 00:10:44.139 the clarity we owe to the other complex humans 00:10:44.139 --> 00:10:46.919 in our lives? Something to think about. Yeah. 00:10:47.519 --> 00:10:49.600 As you write your next prompt. Thank you for 00:10:49.600 --> 00:10:51.500 sharing your source material for this deep dive 00:10:51.500 --> 00:10:53.759 into the art of AI orchestration. We'll catch 00:10:53.759 --> 00:10:54.179 you next time.