WEBVTT 00:00:00.000 --> 00:00:03.140 The difference between amazing AI output and 00:00:03.140 --> 00:00:06.900 that frustratingly generic robotic nonsense isn't 00:00:06.900 --> 00:00:08.519 actually the model you're using. You could be 00:00:08.519 --> 00:00:11.720 on the most advanced LLM available. The key is 00:00:11.720 --> 00:00:14.439 simply how you ask the question. We have to start 00:00:14.439 --> 00:00:17.019 treating our AI like a digital magic eight ball 00:00:17.019 --> 00:00:19.780 that just gives these quick, shallow answers. 00:00:20.219 --> 00:00:22.879 Welcome to the Deep Dive. Today we're unpacking 00:00:22.879 --> 00:00:25.940 a truly comprehensive guide. It details Anthropic's 00:00:25.940 --> 00:00:28.859 10 advanced rules for prompting Claude Opus 4 00:00:28.859 --> 00:00:31.829 .5. And this deep dive is customized for you, 00:00:31.929 --> 00:00:34.329 focusing on the mechanics of getting higher quality, 00:00:34.549 --> 00:00:37.310 more reliable work. It's a fascinating look under 00:00:37.310 --> 00:00:39.289 the hood. I think most of us are only tapping, 00:00:39.509 --> 00:00:41.369 what, maybe two -thirds of the model's brain 00:00:41.369 --> 00:00:43.369 power because our prompts are just lazy. We're 00:00:43.369 --> 00:00:45.530 asking for coffee when we should be handing the 00:00:45.530 --> 00:00:48.530 AI a detailed multi -step recipe. Exactly. The 00:00:48.530 --> 00:00:50.390 mission here is simple. We're moving past these 00:00:50.390 --> 00:00:52.350 vague, single -line requests and learning how 00:00:52.350 --> 00:00:54.549 to turn a large language model into a strategic 00:00:54.549 --> 00:00:56.880 thinking partner. This is all about applying 00:00:56.880 --> 00:01:00.799 clear context, specific constraints, and structured 00:01:00.799 --> 00:01:03.380 prompting to really 10x the quality of the insights 00:01:03.380 --> 00:01:05.640 you get back. So we've broken this guide into 00:01:05.640 --> 00:01:09.019 four core themes. First, we'll look at adopting 00:01:09.019 --> 00:01:12.180 a genuinely collaborative mindset. Then, we'll 00:01:12.180 --> 00:01:13.840 cover how to control the creative boundaries 00:01:13.840 --> 00:01:16.829 of the AI. After that, we'll delve into why breaking 00:01:16.829 --> 00:01:20.069 massive tasks down always works best. And finally, 00:01:20.170 --> 00:01:23.349 we'll analyze the secret power phrases that actually 00:01:23.349 --> 00:01:26.090 unlock Claude's deeper, more complex reasoning 00:01:26.090 --> 00:01:28.170 abilities. OK, let's start with that mindset 00:01:28.170 --> 00:01:30.950 shift. The first two rules, they deal directly 00:01:30.950 --> 00:01:33.370 with how you approach the AI. And frankly, this 00:01:33.370 --> 00:01:35.629 is where most users fail. They're either overly 00:01:35.629 --> 00:01:38.870 polite or, you know, on the other hand. too bossy, 00:01:38.870 --> 00:01:41.209 demanding things instantly. Neither tone works 00:01:41.209 --> 00:01:43.709 well with these advanced models. Right. We often 00:01:43.709 --> 00:01:45.849 forget that these are trained on human conversation, 00:01:46.189 --> 00:01:48.489 which means tone matters, but maybe not in the 00:01:48.489 --> 00:01:50.750 way we think. Rule one says we have to treat 00:01:50.750 --> 00:01:53.069 the AI like a high -performing teammate, not 00:01:53.069 --> 00:01:56.000 a servant. That's the distinction. The core requirement 00:01:56.000 --> 00:01:59.640 is always clarity and directness, but it's got 00:01:59.640 --> 00:02:01.659 to be delivered in a friendly, cooperative tone. 00:02:01.859 --> 00:02:04.099 If you barked an order at a new team member, 00:02:04.239 --> 00:02:06.519 you wouldn't get their best thinking. But at 00:02:06.519 --> 00:02:09.620 the same time, treating the AI as if it requires 00:02:09.620 --> 00:02:12.960 constant thank yous or flowery language just 00:02:12.960 --> 00:02:15.599 adds unnecessary tokens and noise to the prompt. 00:02:15.840 --> 00:02:18.840 So if we're correcting a document, the shift 00:02:18.840 --> 00:02:21.680 is away from just saying, correct this. Instead, 00:02:21.840 --> 00:02:24.340 we frame it as a goal, something like, please 00:02:24.340 --> 00:02:26.379 review this draft for grammatical errors and 00:02:26.379 --> 00:02:28.419 suggest three distinct ways I can modify the 00:02:28.419 --> 00:02:30.240 language to sound more professional and confident 00:02:30.240 --> 00:02:32.919 for a board meeting. That gives it the context 00:02:32.919 --> 00:02:35.180 and the tone. And that leads perfectly into rule 00:02:35.180 --> 00:02:37.580 two, which is the principle of explicitness. 00:02:37.939 --> 00:02:41.219 If collaboration is the attitude, then details 00:02:41.219 --> 00:02:44.189 are truly your superpower. If you don't explicitly 00:02:44.189 --> 00:02:46.610 specify what you want, you basically guarantee 00:02:46.610 --> 00:02:48.909 you'll get the bland, generic answers. The fluff. 00:02:49.210 --> 00:02:51.030 That unusable content that everyone complains 00:02:51.030 --> 00:02:53.270 about. Exactly. That's the key takeaway right 00:02:53.270 --> 00:02:55.889 there. Our source material emphasizes that you 00:02:55.889 --> 00:02:58.969 have to specify three crucial things. Quantity, 00:02:58.990 --> 00:03:02.110 topic, and audience. Don't just ask for some 00:03:02.110 --> 00:03:05.270 blog post ideas. That gives the AI infinite possibilities, 00:03:05.409 --> 00:03:08.370 so it just chooses the safest, most generic path. 00:03:08.919 --> 00:03:12.139 But if you specify, generate 10 compelling blog 00:03:12.139 --> 00:03:14.659 post titles about remote work's unexpected impact 00:03:14.659 --> 00:03:17.340 on urban planning, targeted specifically at city 00:03:17.340 --> 00:03:19.879 officials and real estate developers, well, suddenly 00:03:19.879 --> 00:03:22.360 the scope is narrow. The goal is clear and the 00:03:22.360 --> 00:03:25.139 output quality just skyrockets. And here's where 00:03:25.139 --> 00:03:27.939 the analysis gets deep. Using strong action verbs 00:03:27.939 --> 00:03:31.479 like generate, analyze, compare, critique, that 00:03:31.479 --> 00:03:34.020 isn't just a stylistic choice. It actually forces 00:03:34.020 --> 00:03:37.139 the model into a specific operational mode. Correct. 00:03:37.379 --> 00:03:40.180 It tells the model precisely what cognitive function 00:03:40.180 --> 00:03:43.180 to perform. Asking the AI to analyze a financial 00:03:43.180 --> 00:03:45.599 document forces it to use different, more detailed 00:03:45.599 --> 00:03:47.580 reasoning pathways than if you just asked it 00:03:47.580 --> 00:03:49.699 to write about the document. It moves the AI 00:03:49.699 --> 00:03:52.400 from summary mode. to an active operational mode. 00:03:52.520 --> 00:03:55.319 So if clarity is everything, how critical is 00:03:55.319 --> 00:03:57.599 it to define the specific action we want it to 00:03:57.599 --> 00:04:01.080 take? Action verbs move the AI from summary mode 00:04:01.080 --> 00:04:03.819 to operational mode. Let's move to how we control 00:04:03.819 --> 00:04:06.120 that output itself. Because getting the model 00:04:06.120 --> 00:04:08.560 to think hard is one thing. Getting it to produce 00:04:08.560 --> 00:04:11.280 something usable is another. Rule three is all 00:04:11.280 --> 00:04:13.639 about defining the boundaries. This is the paradox, 00:04:13.840 --> 00:04:17.220 right? That creativity actually thrives on constraints. 00:04:17.769 --> 00:04:19.769 When you give the AI an open -ended request, 00:04:20.069 --> 00:04:22.449 it just gets lost in all the possibilities. No 00:04:22.449 --> 00:04:26.069 structure equals chaos and, honestly, predictability. 00:04:26.350 --> 00:04:28.209 I love the analogy from the source material. 00:04:28.470 --> 00:04:31.189 It's like putting a fence around a garden. Without 00:04:31.189 --> 00:04:33.910 limits, the output just goes wild, and it often 00:04:33.910 --> 00:04:36.470 defaults to the most common words. Constraints 00:04:36.470 --> 00:04:38.569 force it to search deeper in its knowledge space. 00:04:39.009 --> 00:04:41.930 Yeah, think about this prompt. Write a 500 -word 00:04:41.930 --> 00:04:44.430 short story about a robot detective on Mars in 00:04:44.430 --> 00:04:46.550 the style of Raymond Chandler, but do not use 00:04:46.550 --> 00:04:50.189 the word cyber. red planet, or metallic. That 00:04:50.189 --> 00:04:52.509 is a deliberate constraint. That's fascinating. 00:04:52.769 --> 00:04:55.470 So limiting the AI's options, especially by banning 00:04:55.470 --> 00:04:58.449 common words, forces it to think harder and be 00:04:58.449 --> 00:05:01.069 more specific. It has to make these non -generic 00:05:01.069 --> 00:05:02.970 creative decisions because you've kind of tied 00:05:02.970 --> 00:05:05.769 one hand behind its back. It can't lean on those 00:05:05.769 --> 00:05:08.870 high probability phrases. And speaking of structure, 00:05:09.029 --> 00:05:12.079 rule five demands structured output. We need 00:05:12.079 --> 00:05:15.779 to actively stop accepting those long, undifferentiated 00:05:15.779 --> 00:05:18.600 walls of text that are hard to scan or integrate 00:05:18.600 --> 00:05:21.100 into other workflows. This is where we shift 00:05:21.100 --> 00:05:23.500 the responsibility for presentation onto the 00:05:23.500 --> 00:05:25.860 AI. Claude defaults to paragraphs because that's 00:05:25.860 --> 00:05:28.139 the most common text structure. But you can and 00:05:28.139 --> 00:05:30.560 you must request precise formats. We're talking 00:05:30.560 --> 00:05:33.860 about markdown tables, JSON or CSV. Right. And 00:05:33.860 --> 00:05:36.500 for listeners who might not know, JSON or JavaScript 00:05:36.500 --> 00:05:39.079 Object Notation is just a clean, standardized 00:05:39.079 --> 00:05:41.620 format that machines and programmers use to exchange 00:05:41.620 --> 00:05:44.019 structured data. Requesting data this way is 00:05:44.019 --> 00:05:46.360 incredibly efficient. The practical application 00:05:46.360 --> 00:05:49.199 is huge. So instead of asking for information 00:05:49.199 --> 00:05:52.139 about the last three Apollo missions, 15, 16, 00:05:52.199 --> 00:05:55.180 and 17 in PROS, you request a structured table, 00:05:55.339 --> 00:05:57.980 include their launch dates, the crew, and a key 00:05:57.980 --> 00:06:00.399 scientific achievement. And what you get back 00:06:00.399 --> 00:06:03.100 is immediately functional, sortable data. But 00:06:03.100 --> 00:06:06.160 why does requesting a format like JSON or a table 00:06:06.160 --> 00:06:08.879 instantly elevate the quality of the output? 00:06:09.220 --> 00:06:11.540 Structured formats eliminate guessing and make 00:06:11.540 --> 00:06:13.779 the output immediately usable. That makes perfect 00:06:13.779 --> 00:06:16.740 sense. Now, this leads us nicely to the iterative 00:06:16.740 --> 00:06:19.459 workflow. Our sources are insistent that trying 00:06:19.459 --> 00:06:22.000 to get a perfect final draft in one shot is a 00:06:22.000 --> 00:06:24.800 fundamental trap. It's unrealistic and always 00:06:24.800 --> 00:06:27.420 frustrating. So rule four demands we start with 00:06:27.420 --> 00:06:29.439 an exploratory draft. Yeah, we're moving past 00:06:29.439 --> 00:06:31.379 that all or nothing approach. The process has 00:06:31.379 --> 00:06:33.699 to be broken down. Always ask for an outline 00:06:33.699 --> 00:06:36.060 or a plan first. This lets you course correct 00:06:36.060 --> 00:06:38.800 early before the model burns a ton of time and 00:06:38.800 --> 00:06:41.379 tokens on a 5 ,000 word draft that missed the 00:06:41.379 --> 00:06:44.050 mark on section two. I still wrestle with prompt 00:06:44.050 --> 00:06:47.910 drift myself. Beat. For those unfamiliar, that's 00:06:47.910 --> 00:06:50.889 when the AI slowly forgets your original constraints 00:06:50.889 --> 00:06:53.110 or the core topic as the conversation gets longer. 00:06:53.370 --> 00:06:55.790 It's incredibly painful when you realize late 00:06:55.790 --> 00:06:58.449 in the process that a major assumption was wrong 00:06:58.449 --> 00:07:01.370 and you have to just throw everything away. Getting 00:07:01.370 --> 00:07:03.430 that outline right prevents that catastrophe. 00:07:03.810 --> 00:07:05.589 And if you apply that microstep thinking on a 00:07:05.589 --> 00:07:08.689 larger scale, you get rule 10, divide and conquer. 00:07:09.089 --> 00:07:11.949 Complex tasks overwhelm both the user and the 00:07:11.949 --> 00:07:14.560 model. you have to act like a conductor. breaking 00:07:14.560 --> 00:07:17.220 the entire symphony into controllable, smaller 00:07:17.220 --> 00:07:19.339 sequential stages. Let's use the business plan 00:07:19.339 --> 00:07:21.000 example. You never ask for the whole thing at 00:07:21.000 --> 00:07:23.180 once. You first request the table of contents, 00:07:23.399 --> 00:07:25.180 then you prompt for the executive summary, then 00:07:25.180 --> 00:07:27.779 the market analysis, and only then do you ask 00:07:27.779 --> 00:07:30.019 the AI to synthesize and check for contradictions 00:07:30.019 --> 00:07:32.360 between the sections. The efficiency boost is 00:07:32.360 --> 00:07:35.360 just dramatic. You maintain quality control at 00:07:35.360 --> 00:07:37.939 every single small step. You're reviewing the 00:07:37.939 --> 00:07:39.720 integrity of the brick, you know, before you 00:07:39.720 --> 00:07:42.259 let it become part of the wall. It prevents costly 00:07:42.259 --> 00:07:45.569 rework later. When dealing with huge documents, 00:07:45.769 --> 00:07:48.050 what is the single biggest benefit of the divide 00:07:48.050 --> 00:07:51.389 and conquer approach? Breaking tasks into chunks 00:07:51.389 --> 00:07:54.149 ensures high quality control at every stage. 00:07:54.550 --> 00:07:56.670 Okay, let's pivot now to what I think is the 00:07:56.670 --> 00:07:59.269 most exciting section. Rules six through nine. 00:07:59.629 --> 00:08:02.069 This is the toolkit for unlocking the model's 00:08:02.069 --> 00:08:04.829 deepest reasoning capabilities. Rule six is simple, 00:08:04.889 --> 00:08:08.240 but it's transformative. Explain the why. Context 00:08:08.240 --> 00:08:10.660 is king. The AI needs to understand the underlying 00:08:10.660 --> 00:08:13.160 purpose, the ultimate goal behind your request 00:08:13.160 --> 00:08:15.899 to deliver truly tailored results. It's the difference 00:08:15.899 --> 00:08:18.100 between a tool and a partner. Right. If you just 00:08:18.100 --> 00:08:20.040 ask for five coffee slogans, you get the most 00:08:20.040 --> 00:08:23.199 common answers, great coffee. But if you explain 00:08:23.199 --> 00:08:25.500 that the beans are ethically sourced, the target 00:08:25.500 --> 00:08:27.959 audience is environmentally conscious millennials, 00:08:28.199 --> 00:08:31.779 and the medium is a specific Instagram ad, well, 00:08:31.819 --> 00:08:34.519 then everything changes. That context shifts 00:08:34.519 --> 00:08:37.740 the entire output from generic marketing to tailored 00:08:37.740 --> 00:08:40.639 high value branding. It narrows the vocabulary 00:08:40.639 --> 00:08:43.600 and the concepts the AI even considers. And rule 00:08:43.600 --> 00:08:46.519 seven is about controlling brevity versus verbosity. 00:08:47.000 --> 00:08:49.320 Claude has no idea if you want a detailed dissertation 00:08:49.320 --> 00:08:51.659 or a bullet point summary unless you tell it. 00:08:51.740 --> 00:08:54.080 So you have to explicitly set the level of detail 00:08:54.080 --> 00:08:56.720 up front. We see this when comparing reasoning 00:08:56.720 --> 00:08:59.320 levels. Take a single topic, like photosynthesis. 00:08:59.580 --> 00:09:02.200 You can request deep thinking by stating, explain 00:09:02.200 --> 00:09:04.720 photosynthesis in detail for a university -level 00:09:04.720 --> 00:09:07.279 biology student, making sure to include complex 00:09:07.279 --> 00:09:11.080 concepts like C3 and C4 pathways. Or you simplify 00:09:11.080 --> 00:09:13.480 drastically. Explain photosynthesis like I'm 00:09:13.480 --> 00:09:15.980 five years old, using only analogies about baking 00:09:15.980 --> 00:09:18.539 and sunlight. The third level of control, rule 00:09:18.539 --> 00:09:20.740 eight, is providing a scaffold. This is where 00:09:20.740 --> 00:09:22.500 templates become your secret weapon because they 00:09:22.500 --> 00:09:25.210 enforce consistency. Stop asking Claude to summarize 00:09:25.210 --> 00:09:27.429 an article generally. That just leads to inconsistent 00:09:27.429 --> 00:09:30.269 links and formats. Instead, you give Claude a 00:09:30.269 --> 00:09:33.110 precise format, a summary template that explicitly 00:09:33.110 --> 00:09:36.929 states, main thesis, one sentence, key supporting 00:09:36.929 --> 00:09:39.870 points, three bullet points, and a concluding 00:09:39.870 --> 00:09:43.149 insight, one sentence. The AI has to fill in 00:09:43.149 --> 00:09:45.750 the blanks consistently every time, making its 00:09:45.750 --> 00:09:48.759 output easily ingestible. And finally, rule nine, 00:09:49.000 --> 00:09:52.320 use power phrases and expert personas. These 00:09:52.320 --> 00:09:54.779 are the immediate triggers that unlock Claude's 00:09:54.779 --> 00:09:57.139 higher level reasoning and access a much more 00:09:57.139 --> 00:09:59.740 sophisticated vocabulary. The power phrase toolkit 00:09:59.740 --> 00:10:02.620 is essential for advanced users. You use think 00:10:02.620 --> 00:10:05.639 step by step to explicitly force reasoning. It 00:10:05.639 --> 00:10:07.980 makes the AI lay out its internal logic before 00:10:07.980 --> 00:10:10.500 giving the final answer. You use critique your 00:10:10.500 --> 00:10:13.200 own response to trigger self -correction and 00:10:13.200 --> 00:10:15.480 iterative quality assurance. And the persona 00:10:15.480 --> 00:10:18.240 rule. Adopt the persona of an expert in field. 00:10:18.379 --> 00:10:20.779 It primes its domain knowledge. You're telling 00:10:20.779 --> 00:10:22.700 the model which part of its massive data set 00:10:22.700 --> 00:10:24.559 to privilege. Act as a leading constitutional 00:10:24.559 --> 00:10:26.980 lawyer is just so much better than tell me about 00:10:26.980 --> 00:10:30.720 the First Amendment. Whoa. Imagine scaling this 00:10:30.720 --> 00:10:33.440 iterative self -critiquing process where the 00:10:33.440 --> 00:10:35.879 model checks its own work across a million queries. 00:10:36.460 --> 00:10:39.659 That is how we unlock true, reliable expertise 00:10:39.659 --> 00:10:42.460 from these systems. You are leveraging the model's 00:10:42.460 --> 00:10:45.179 ability to self -audit. Which of the power phrases 00:10:45.179 --> 00:10:48.379 is most effective at forcing the AI to slow down 00:10:48.379 --> 00:10:51.820 its processing? Think step -by -step forces the 00:10:51.820 --> 00:10:54.720 model to show its actual reasoning. So to summarize 00:10:54.720 --> 00:10:57.220 the big idea for you, these rules are simple, 00:10:57.320 --> 00:10:59.820 but they're powerful. Be explicit in your brief, 00:11:00.000 --> 00:11:02.500 set deliberate creative boundaries, iterate in 00:11:02.500 --> 00:11:05.240 small stages, and use specific linguistic triggers. 00:11:05.500 --> 00:11:07.519 These aren't just minor tips. They are battle 00:11:07.519 --> 00:11:09.440 -tested techniques designed to eliminate generic 00:11:09.440 --> 00:11:11.769 output. Let's run one quick synthesis example 00:11:11.769 --> 00:11:14.110 to show how these rules stack up. Compare the 00:11:14.110 --> 00:11:16.029 vague request, tell me about stoicism, which 00:11:16.029 --> 00:11:18.450 gives you a generic introductory paragraph versus 00:11:18.450 --> 00:11:21.470 the correct way to prompt. The right prompt stacks 00:11:21.470 --> 00:11:23.909 six of these rules together. First, we use the 00:11:23.909 --> 00:11:26.129 persona rule, act as a university professor of 00:11:26.129 --> 00:11:28.190 ancient philosophy. Then we explain the why, 00:11:28.350 --> 00:11:30.789 the context. I am preparing a one -hour introductory 00:11:30.789 --> 00:11:33.269 lecture for first -year students with no prior 00:11:33.269 --> 00:11:36.899 knowledge. Next, we use divide and conquer. First, 00:11:37.039 --> 00:11:39.360 create a lecture outline with three main balanced 00:11:39.360 --> 00:11:42.100 sections. We constrain the scope. The outline 00:11:42.100 --> 00:11:44.600 must have a clear introduction, a body focused 00:11:44.600 --> 00:11:46.879 on practical application and a strong conclusion. 00:11:47.159 --> 00:11:50.139 Then we demand structured output. Please format 00:11:50.139 --> 00:11:52.500 this as a nested bulleted list using markdown 00:11:52.500 --> 00:11:55.580 headings. And finally, we use explicitness. For 00:11:55.580 --> 00:11:58.019 each major body point, include a key stoic figure 00:11:58.019 --> 00:12:00.659 such as Seneca and one core actionable idea. 00:12:01.080 --> 00:12:03.399 And the result is immediate. It's structured, 00:12:03.500 --> 00:12:05.460 it's audience -appropriate content, a professional 00:12:05.460 --> 00:12:08.039 lecture outline ready to teach from. The difference 00:12:08.039 --> 00:12:10.740 between that unusable AI slop and high -quality 00:12:10.740 --> 00:12:13.279 insight is solely in the 30 seconds you took 00:12:13.279 --> 00:12:16.220 to craft that powerful prompt. Claude is an extremely 00:12:16.220 --> 00:12:18.460 sharp tool, but it's only as effective as the 00:12:18.460 --> 00:12:20.600 person wielding it. You need to take those few 00:12:20.600 --> 00:12:23.429 extra seconds to set your brief. Add context, 00:12:23.570 --> 00:12:25.409 set boundaries, and deploy those power phrases. 00:12:25.769 --> 00:12:28.269 That's what turns a chat experience into a strategic 00:12:28.269 --> 00:12:30.830 partnership. So think about one complex information 00:12:30.830 --> 00:12:33.669 -gathering task you usually rush through in one 00:12:33.669 --> 00:12:36.509 prompt. How would breaking it into three smaller, 00:12:36.690 --> 00:12:39.490 constrained, and persona -driven prompts change 00:12:39.490 --> 00:12:42.129 the outcome? That's your next deep dive. Thank 00:12:42.129 --> 00:12:43.570 you for joining us for this deep dive.