WEBVTT 00:00:00.000 --> 00:00:03.080 So if you're using powerful AI like Claude, just 00:00:03.080 --> 00:00:05.080 to ask, you know, write me an email or summarize 00:00:05.080 --> 00:00:09.119 this article, that's fine. But really think about 00:00:09.119 --> 00:00:11.119 it this way. It's like you have this incredibly 00:00:11.119 --> 00:00:14.039 powerful supercomputer, maybe something like 00:00:14.039 --> 00:00:16.760 NASA would use, right? And you're basically just 00:00:16.760 --> 00:00:19.079 using it to, I don't know, balance your checkbook. 00:00:19.820 --> 00:00:22.859 You're leaving almost all of its real power completely 00:00:22.859 --> 00:00:27.010 untapped. So today, our deep dive is all about 00:00:27.010 --> 00:00:29.469 advanced Claude AI prompt engineering. We're 00:00:29.469 --> 00:00:31.609 going way beyond that basic chat box interaction. 00:00:31.730 --> 00:00:33.149 We're going to show you how to stop thinking 00:00:33.149 --> 00:00:35.130 of Claude as just a general assistant and start 00:00:35.130 --> 00:00:37.469 turning it into, well... a specialized employee, 00:00:37.609 --> 00:00:39.390 like a super smart team member that's permanently 00:00:39.390 --> 00:00:41.810 trained for your specific business needs or your 00:00:41.810 --> 00:00:44.670 most complex workflows. Exactly. And the real 00:00:44.670 --> 00:00:47.189 secret here, the thing most people miss, isn't 00:00:47.189 --> 00:00:50.090 some, you know, magic keyword. It's actually 00:00:50.090 --> 00:00:53.329 about designing the system underneath. It's about 00:00:53.329 --> 00:00:56.310 understanding how to give the AI a kind of long 00:00:56.310 --> 00:00:59.509 term memory operational rules that actually stick 00:00:59.509 --> 00:01:02.490 around. This deep dive is really your fast track 00:01:02.490 --> 00:01:05.269 to getting a huge advantage. It can totally transform 00:01:05.269 --> 00:01:08.290 the quality. the speed, the relevance of what 00:01:08.290 --> 00:01:09.989 the AI gives you, to the point where it feels 00:01:09.989 --> 00:01:13.049 less like using a tool and more like you've actually 00:01:13.049 --> 00:01:15.129 outsourced some critical thinking. OK, great. 00:01:15.250 --> 00:01:17.590 Let's unpack that. Because before we get into 00:01:17.590 --> 00:01:20.590 the really technical setup, the code that kind 00:01:20.590 --> 00:01:23.530 of installs these AI employees, we need to get 00:01:23.530 --> 00:01:25.430 the basics right. We need a sort of shared language 00:01:25.430 --> 00:01:27.530 for instructing any AI. I mean, our sources really 00:01:27.530 --> 00:01:30.049 emphasize that even those complex prompt system 00:01:30.049 --> 00:01:32.750 will just fall flat if the fundamentals aren't 00:01:32.730 --> 00:01:34.709 Rock solid. Yeah, think of it like onboarding 00:01:34.709 --> 00:01:36.670 a new team member, right? If you just give them 00:01:36.670 --> 00:01:39.030 a really vague job description and send them 00:01:39.030 --> 00:01:41.510 off, you're going to get fuzzy, probably unhelpful 00:01:41.510 --> 00:01:43.670 results. You have to be crystal clear about the 00:01:43.670 --> 00:01:45.989 job parameters and that clarity while it really 00:01:45.989 --> 00:01:48.049 boils down to five essential components. OK, 00:01:48.150 --> 00:01:51.489 first one, roll. And you're saying don't be shy 00:01:51.489 --> 00:01:54.909 here. It's not just be a writer. It's more like 00:01:54.909 --> 00:01:58.370 you are a senior copywriter specializing in persuasive 00:01:58.370 --> 00:02:01.310 B2B tech briefs or, like the example we'll get 00:02:01.310 --> 00:02:04.269 to, director of content and data strategy. Precisely. 00:02:04.310 --> 00:02:07.090 The role sets the expertise, the tone, the knowledge 00:02:07.090 --> 00:02:09.449 base it draws from. It's fundamental. Then number 00:02:09.449 --> 00:02:13.250 two is audience. Why is that so critical? Because 00:02:13.250 --> 00:02:15.590 it dictates everything about the style and complexity. 00:02:15.789 --> 00:02:17.870 I mean, are you writing for your CEO? They probably 00:02:17.870 --> 00:02:20.189 need high -level bullets. Or are you explaining 00:02:20.189 --> 00:02:22.590 something complex to a total beginner customer? 00:02:22.949 --> 00:02:24.830 That's a completely different approach. If you 00:02:24.830 --> 00:02:27.469 don't define the audience, the AI just defaults 00:02:27.469 --> 00:02:31.210 to this bland, generic voice that honestly satisfies 00:02:31.210 --> 00:02:33.909 no one. Right, makes sense. Pillar three is task. 00:02:34.090 --> 00:02:36.270 And this is where the hyper -specificity comes 00:02:36.270 --> 00:02:38.889 in. Yes, exactly. We need to move past the vague 00:02:38.889 --> 00:02:41.009 stuff like, write an ad. To something really 00:02:41.009 --> 00:02:43.509 actionable. Like, write three Facebook ads using 00:02:43.509 --> 00:02:45.810 these specific customer pain points, each headline 00:02:45.810 --> 00:02:48.830 under 10 words, body copy under 100 words. That 00:02:48.830 --> 00:02:52.550 level of detail. The AI just can't handle ambiguity 00:02:52.550 --> 00:02:55.569 well. It needs clear instructions. Okay. Fourth 00:02:55.569 --> 00:02:58.259 component. Context, basically, what background 00:02:58.259 --> 00:03:00.819 information does the AI need to actually do the 00:03:00.819 --> 00:03:03.460 task? This could be product details, maybe customer 00:03:03.460 --> 00:03:06.860 profiles, competitor analysis, or, like we'll 00:03:06.860 --> 00:03:09.099 see in the advanced cases, pointing it to specific 00:03:09.099 --> 00:03:11.379 data files it needs to read. You have to provide 00:03:11.379 --> 00:03:14.400 the necessary world knowledge. Got it. And finally, 00:03:14.680 --> 00:03:18.080 number five. format. Yeah, don't just assume 00:03:18.080 --> 00:03:20.400 it knows how you want the output. If you want 00:03:20.400 --> 00:03:22.960 a markdown table, you need to say output as a 00:03:22.960 --> 00:03:25.419 markdown table. If you need a numbered list of 00:03:25.419 --> 00:03:28.560 steps, specify output as numbered bullet points. 00:03:28.680 --> 00:03:32.740 Be explicit. So roll, audience, task, context, 00:03:32.860 --> 00:03:35.539 and format. You consistently nail those five 00:03:35.539 --> 00:03:37.159 things. You're already way ahead of probably 00:03:37.159 --> 00:03:39.379 90 % of people using these tools. It makes a 00:03:39.379 --> 00:03:41.580 huge difference. OK, so those five pillars cover 00:03:41.580 --> 00:03:43.419 the what and the how of giving instructions. 00:03:43.800 --> 00:03:46.900 But now the really interesting part, the where. 00:03:47.539 --> 00:03:50.000 Where do we actually put these detailed specialized 00:03:50.000 --> 00:03:52.340 instructions so Claude remembers them every single 00:03:52.340 --> 00:03:54.680 time? This is that long -term memory idea, right? 00:03:54.979 --> 00:03:57.479 Right. And this is where things get, maybe sound 00:03:57.479 --> 00:03:59.419 a bit technical, but it's actually surprisingly 00:03:59.419 --> 00:04:01.960 straightforward to set up. These really advanced 00:04:01.960 --> 00:04:04.379 workflows we're about to discuss, they need a 00:04:04.379 --> 00:04:06.500 specific technical foundation just to work at 00:04:06.500 --> 00:04:09.289 all. And there are two key things you absolutely 00:04:09.289 --> 00:04:11.969 need according to the sources. First, you have 00:04:11.969 --> 00:04:15.250 to be using the Cloud desktop app on your computer. 00:04:16.069 --> 00:04:17.970 Apparently these powerful system prompt features 00:04:17.970 --> 00:04:20.850 just don't work reliably or maybe at all on the 00:04:20.850 --> 00:04:22.930 regular website version. That's the first crucial 00:04:22.930 --> 00:04:25.569 piece. And the second is you need to create a 00:04:25.569 --> 00:04:29.009 new project within that desktop app. Okay, so 00:04:29.009 --> 00:04:32.560 a project is like... a dedicated workspace. Exactly. 00:04:32.699 --> 00:04:35.100 Think of it as the dedicated office or container 00:04:35.100 --> 00:04:38.300 for your specialized AI employee. So if you're 00:04:38.300 --> 00:04:40.160 building that content strategist, you might name 00:04:40.160 --> 00:04:42.579 the project content strategist or, you know, 00:04:42.779 --> 00:04:45.579 Athena 2 .0, like in our example. But the magic 00:04:45.579 --> 00:04:47.399 part, the bit that really gives it that long 00:04:47.399 --> 00:04:50.120 -term memory and makes it feel trained, is the 00:04:50.120 --> 00:04:53.300 system prompt. Oh, okay. So this is the big shift. 00:04:53.459 --> 00:04:56.439 All that detailed code, the massive block of 00:04:56.439 --> 00:04:58.540 instructions and rules we'll look at. You're 00:04:58.540 --> 00:05:00.560 not pasting that into the normal chat box where 00:05:00.560 --> 00:05:03.079 you type everyday questions. Oh, definitely not. 00:05:03.139 --> 00:05:06.079 You paste that entire block of code into the 00:05:06.079 --> 00:05:08.899 project's settings. There's a specific, usually 00:05:08.899 --> 00:05:13.060 quite large text box labeled system prompt, or 00:05:13.060 --> 00:05:15.170 maybe just instructions. That's where it goes. 00:05:15.410 --> 00:05:17.189 And putting it there makes it permanent for that 00:05:17.189 --> 00:05:19.750 project. Yes. The system prompt basically acts 00:05:19.750 --> 00:05:22.490 like the AI's core programming or its constitution 00:05:22.490 --> 00:05:25.050 for that specific project. Because of where it's 00:05:25.050 --> 00:05:27.649 placed, every instruction inside it gets the 00:05:27.649 --> 00:05:30.230 highest priority. Claude automatically refers 00:05:30.230 --> 00:05:32.329 back to it every single time you start a new 00:05:32.329 --> 00:05:35.110 conversation within that project. It's the difference 00:05:35.110 --> 00:05:37.870 between constantly reminding a freelancer about 00:05:37.870 --> 00:05:40.670 the job specs versus giving a full -time employee 00:05:40.670 --> 00:05:43.029 a detailed handbook they have to follow. Okay, 00:05:43.170 --> 00:05:45.509 that makes a lot more sense. Let's see this in 00:05:45.509 --> 00:05:47.870 action then. Let's look at the first installed 00:05:47.870 --> 00:05:51.230 employee, this concept called Athena 2 .0. The 00:05:51.230 --> 00:05:53.930 goal here is to build a data -driven business 00:05:53.930 --> 00:05:57.449 brain, right? Turning Claude into an expert that 00:05:57.449 --> 00:05:59.350 basically knows everything about your company 00:05:59.350 --> 00:06:01.529 because it's linked to your documents and goals. 00:06:01.689 --> 00:06:05.240 Right. So we install Athena using a very specific 00:06:05.240 --> 00:06:08.720 system prompt. Her role is defined as Director 00:06:08.720 --> 00:06:11.339 of Content and Data Strategy for the sustainable 00:06:11.339 --> 00:06:14.680 fashion brand ZanStyle. But look closely at the 00:06:14.680 --> 00:06:17.060 core directive in the prompt. Your job is to 00:06:17.060 --> 00:06:20.360 make decisions using data, not guesses. Always 00:06:20.360 --> 00:06:22.360 follow the brand style and focus on business 00:06:22.360 --> 00:06:25.079 results. That immediately tells it. Data first, 00:06:25.180 --> 00:06:27.620 opinion second, business outcomes matter most. 00:06:27.800 --> 00:06:29.800 And the workflow part of the prompt seems really 00:06:29.800 --> 00:06:31.480 structured. It looks like Athena has to follow 00:06:31.480 --> 00:06:33.399 seven specific steps before it even gives an 00:06:33.399 --> 00:06:35.720 answer. Yes, mandatory steps. This isn't just 00:06:35.720 --> 00:06:38.040 a suggestion in the prompt. It uses internal 00:06:38.040 --> 00:06:40.939 tags and structure to force a specific step -by 00:06:40.939 --> 00:06:42.879 -step reasoning process. It has to follow them. 00:06:43.040 --> 00:06:44.860 What are some of those steps? Well, it starts 00:06:44.860 --> 00:06:47.420 with things like analyze request and find unclear 00:06:47.420 --> 00:06:50.579 points. But then crucially, step three forces 00:06:50.579 --> 00:06:52.740 it. If there are unclear points in your request, 00:06:52.779 --> 00:06:56.360 it must... ask for clarification. Just one or 00:06:56.360 --> 00:06:58.600 two short, specific questions to make sure it 00:06:58.600 --> 00:07:01.339 understands before it dives in. Okay, hold on. 00:07:01.519 --> 00:07:04.079 If I'm trying to automate things, making ask 00:07:04.079 --> 00:07:06.560 me questions, doesn't that slow things down? 00:07:06.920 --> 00:07:09.480 I kind of want the answer now. That's a really 00:07:09.480 --> 00:07:11.620 common reaction, but it's thinking about speed 00:07:11.620 --> 00:07:14.699 over accuracy. This step is the difference between 00:07:14.699 --> 00:07:17.519 basic automation and genuine competence. Think 00:07:17.519 --> 00:07:19.980 about it. Asking one quick clarifying question 00:07:19.980 --> 00:07:22.759 upfront prevents the AI from spending time generating 00:07:22.759 --> 00:07:25.660 a long, detailed response based on a completely 00:07:25.660 --> 00:07:28.079 wrong assumption. It avoids that garbage in and 00:07:28.079 --> 00:07:30.660 garbage out problem. By forcing clarification, 00:07:30.740 --> 00:07:33.040 you guarantee the final output is actually relevant, 00:07:33.220 --> 00:07:35.680 which saves you potentially hours of rework later. 00:07:35.860 --> 00:07:38.100 OK. Yeah, see the logic there. Better to clarify 00:07:38.100 --> 00:07:40.620 than redo. So once it understands, it finds the 00:07:40.620 --> 00:07:42.949 documents it needs, starts the work. But then 00:07:42.949 --> 00:07:45.170 there's a quality check. Absolutely critical. 00:07:45.509 --> 00:07:48.310 Step six forces the AI to check your work against 00:07:48.310 --> 00:07:49.990 something called the self -correction checklist 00:07:49.990 --> 00:07:52.649 that's also defined in the system prompt. And 00:07:52.649 --> 00:07:54.730 that checklist is where you define what good 00:07:54.730 --> 00:07:57.189 looks like. Things like, does the writing style 00:07:57.189 --> 00:08:00.089 match the brand guide? Is every conclusion supported 00:08:00.089 --> 00:08:03.430 by data from a file? Are the suggestions specific 00:08:03.430 --> 00:08:06.990 and actionable? This makes the AI pause and evaluate 00:08:06.990 --> 00:08:10.009 its own output before showing it to you. It dramatically 00:08:10.009 --> 00:08:12.850 boosts the reliability. Wow, okay. And there's 00:08:12.850 --> 00:08:15.029 one more piece to Athena, something about being 00:08:15.029 --> 00:08:17.910 proactive. Yeah, the proactive analysis protocol. 00:08:18.069 --> 00:08:19.850 This part of the instruction tells Athena that, 00:08:20.129 --> 00:08:22.250 say, once a week it has to automatically scan 00:08:22.250 --> 00:08:25.129 the relevant data files. find something interesting, 00:08:25.350 --> 00:08:27.269 unusual, or maybe problematic that you didn't 00:08:27.269 --> 00:08:29.509 ask about, and report it. So it's designed to 00:08:29.509 --> 00:08:31.389 surface insights you might have missed. Exactly. 00:08:31.629 --> 00:08:33.570 It's programmed to be curious about the data 00:08:33.570 --> 00:08:36.289 that's moving towards real specialized intelligence, 00:08:36.490 --> 00:08:38.649 not just reacting to commands. All right. Super 00:08:38.649 --> 00:08:41.250 powerful. Let's pivot to the second use case. 00:08:41.610 --> 00:08:44.090 This one seems more focused on personal productivity 00:08:44.090 --> 00:08:48.529 and maybe even mood. Zenith, the personal performance 00:08:48.529 --> 00:08:51.409 coach. Yeah, Zenith is fascinating. It applies 00:08:51.409 --> 00:08:53.789 principles from behavioral science directly to 00:08:53.789 --> 00:08:57.090 the AI's operation. We know, like, real productivity 00:08:57.090 --> 00:08:59.950 isn't always about just pushing harder. Sometimes 00:08:59.950 --> 00:09:02.610 the smart move is to ease off or switch to a 00:09:02.610 --> 00:09:05.830 different kind of task. So Zenith acts as a combination 00:09:05.830 --> 00:09:08.250 of an executive assistant and a personal performance 00:09:08.250 --> 00:09:10.669 coach. And it starts every single morning by 00:09:10.669 --> 00:09:13.350 asking you, good morning, what is your energy 00:09:13.350 --> 00:09:16.179 level today from 110? And based on your answer, 00:09:16.340 --> 00:09:18.899 it changes what it suggests. Precisely. That's 00:09:18.899 --> 00:09:21.340 the core logic. If you say your energy is low, 00:09:21.679 --> 00:09:24.100 maybe a 1 to 3 zenith programming tells it to 00:09:24.100 --> 00:09:26.799 shift gears. It'll suggest simple, low -effort 00:09:26.799 --> 00:09:29.240 tasks, things like clearing out your inbox, maybe 00:09:29.240 --> 00:09:31.299 organizing some files. And it does it with a 00:09:31.299 --> 00:09:33.919 gentle tone, using warmer, softer emojis, like 00:09:33.919 --> 00:09:36.639 a little plant sprout, or a coffee cup, or a 00:09:36.639 --> 00:09:38.639 heart CEO. It's trying to conserve your mental 00:09:38.639 --> 00:09:40.759 energy. OK, and if you're feeling high energy, 00:09:40.879 --> 00:09:42.980 like an 8 to 10? Then the prom directs it to 00:09:42.980 --> 00:09:45.889 do the opposite. It suggests diving into complex, 00:09:45.889 --> 00:09:48.210 creative work that needs focus and brain power. 00:09:48.750 --> 00:09:51.090 And the tone shifts too, using more energetic, 00:09:51.230 --> 00:09:54.009 powerful emojis like a rocket, or a fire -fo, 00:09:54.309 --> 00:09:57.490 or sparkles you. It's designed to match the cognitive 00:09:57.490 --> 00:09:59.940 load to your current capacity. That's quite clever. 00:10:00.379 --> 00:10:01.919 Is there a safety mechanism built in? What if 00:10:01.919 --> 00:10:04.899 you're just completely overloaded? Yes, absolutely 00:10:04.899 --> 00:10:07.820 essential. There is a specific stress protocol 00:10:07.820 --> 00:10:10.440 built in. If you type words like overwhelmed, 00:10:10.820 --> 00:10:14.360 swamped, stressed, or similar, Zenith's instructions 00:10:14.360 --> 00:10:16.740 mandate that it must immediately stop whatever 00:10:16.740 --> 00:10:19.639 else it was doing. It first offers just a brief, 00:10:19.940 --> 00:10:22.059 comforting acknowledgement. Then its only goal 00:10:22.059 --> 00:10:24.679 is to help you identify one single, very small, 00:10:24.779 --> 00:10:27.419 manageable thing you can do next. It's hard -coded 00:10:27.419 --> 00:10:29.299 to break that feeling of panic and just help 00:10:29.299 --> 00:10:31.460 you regain a tiny bit of control. And it connects 00:10:31.460 --> 00:10:33.399 to other tools, too. I guess that's something 00:10:33.399 --> 00:10:35.899 about Google Calendar. Yeah, that's a great example 00:10:35.899 --> 00:10:38.519 of using a hard -coded constraint for well -being. 00:10:39.379 --> 00:10:41.519 Zenith can be configured to access your Google 00:10:41.519 --> 00:10:44.679 Calendar. And a non -negotiable rule programmed 00:10:44.679 --> 00:10:46.840 into its system prompt is that it must always 00:10:46.840 --> 00:10:49.279 ensure there's at least 15 minutes of break time 00:10:49.279 --> 00:10:51.419 scheduled between any back -to -back meetings 00:10:51.419 --> 00:10:54.659 it helps arrange. That's brilliant. A small rule. 00:10:54.879 --> 00:10:57.899 programmed once that prevents that constant meeting 00:10:57.899 --> 00:11:00.740 burnout forever. Exactly, little systemic fixes. 00:11:00.960 --> 00:11:03.220 Okay, so we've seen Athena for deep business 00:11:03.220 --> 00:11:05.820 data and Zenith for personal focus and energy 00:11:05.820 --> 00:11:07.919 management. Let's quickly touch on the other 00:11:07.919 --> 00:11:10.120 three examples mentioned in the sources just 00:11:10.120 --> 00:11:12.039 to give people a sense of the sheer range of 00:11:12.039 --> 00:11:14.179 specialized employees you could build with this 00:11:14.179 --> 00:11:17.250 approach. Sure. First up was Nexus, positioned 00:11:17.250 --> 00:11:20.830 as a data analyst specifically for, say, community 00:11:20.830 --> 00:11:23.769 management contexts. Think analyzing chat logs 00:11:23.769 --> 00:11:26.070 from a Discord server or comments on a forum, 00:11:26.149 --> 00:11:28.490 like for a Dittar Pro Hub example. And its key 00:11:28.490 --> 00:11:30.509 role is about data integrity, right? Yep. Right. 00:11:30.889 --> 00:11:34.269 To ensure trustworthiness, any advice or conclusion 00:11:34.269 --> 00:11:37.190 Nexus offers must be based on evidence it finds 00:11:37.190 --> 00:11:39.830 in at least two different data files or sources. 00:11:40.139 --> 00:11:43.480 This stops it from overreacting to, like, one 00:11:43.480 --> 00:11:46.259 angry comment. It forces it to look for patterns. 00:11:46.480 --> 00:11:48.840 And it does more than just count comments. Yeah, 00:11:48.840 --> 00:11:52.019 it can do sentiment analysis, calculate the percentage 00:11:52.019 --> 00:11:54.879 of positive, negative, neutral comments in text 00:11:54.879 --> 00:11:57.700 files. But more importantly, it's programmed 00:11:57.700 --> 00:12:00.340 to dig deeper, to look for the deepest reason 00:12:00.340 --> 00:12:03.100 behind issues. So it might correlate, say, the 00:12:03.100 --> 00:12:05.299 launch of a difficult new course with a sudden 00:12:05.299 --> 00:12:07.500 increase in members leaving the community, suggesting 00:12:07.500 --> 00:12:09.440 the course difficulty might be the real problem. 00:12:09.679 --> 00:12:12.720 OK, interesting. Then there was Creator, the 00:12:12.720 --> 00:12:14.940 web app developer. This one sounds ambitious. 00:12:15.279 --> 00:12:17.259 It's focused on using Claude's artifacts feature 00:12:17.259 --> 00:12:19.779 to build small interactive tools directly in 00:12:19.779 --> 00:12:22.279 the chat. Things like quizzes, calculators, simple 00:12:22.279 --> 00:12:24.860 dashboards. Artifacts. That's the thing where 00:12:24.860 --> 00:12:27.460 Claude actually generates working code, like 00:12:27.460 --> 00:12:29.779 a mini app, right there in the conversation window. 00:12:30.000 --> 00:12:32.580 Exactly. So you get a functional tool instantly, 00:12:32.960 --> 00:12:34.460 potentially without you needing to write any 00:12:34.460 --> 00:12:37.039 code yourself. And the example was an opportunity 00:12:37.039 --> 00:12:39.679 cost calculator. What was the advanced part there? 00:12:40.120 --> 00:12:42.399 The optimization built into its system prompt. 00:12:42.840 --> 00:12:45.639 Creator was instructed to always create two versions 00:12:45.639 --> 00:12:47.419 of the lead capture button for the calculator. 00:12:48.139 --> 00:12:50.500 Maybe one button says download detailed analysis 00:12:50.500 --> 00:12:53.470 and the other says get time -saving tips. And 00:12:53.470 --> 00:12:55.429 then it has to randomly show one version or the 00:12:55.429 --> 00:12:57.570 other to different users. It essentially becomes 00:12:57.570 --> 00:13:00.210 an AI programmer that also includes built -in 00:13:00.210 --> 00:13:02.850 AD testing to constantly optimize itself for 00:13:02.850 --> 00:13:05.350 better results, like lead generation. Plus, it 00:13:05.350 --> 00:13:07.429 had instructions about ensuring accessibility 00:13:07.429 --> 00:13:10.750 in the code it generates. Wow. OK. An AI dev 00:13:10.750 --> 00:13:12.990 and tester rolled into one. And the last one 00:13:12.990 --> 00:13:15.860 was Synthesizer. What's its specialty? Synthesizer 00:13:15.860 --> 00:13:18.440 is the strategic information specialist. It's 00:13:18.440 --> 00:13:20.620 designed for high level, high stakes preparation 00:13:20.620 --> 00:13:23.019 work. The example was generating a crucial key 00:13:23.019 --> 00:13:25.500 for product strategy report. And its main job 00:13:25.500 --> 00:13:28.000 involves handling lots of information. Massive 00:13:28.000 --> 00:13:31.379 amounts. Its task is to read and analyze all 00:13:31.379 --> 00:13:33.679 relevant documents from potentially multiple 00:13:33.679 --> 00:13:36.720 disparate sources. Imagine pointing it to three 00:13:36.720 --> 00:13:39.440 different Google Drive folders, one with engineering 00:13:39.440 --> 00:13:42.080 notes, one with marketing campaign results, and 00:13:42.080 --> 00:13:44.789 one with customer support ticket summaries. How 00:13:44.789 --> 00:13:46.529 does it decide what's important out of all that? 00:13:46.649 --> 00:13:49.269 That's its core strategic filter. The system 00:13:49.269 --> 00:13:51.549 prompt instructs synthesizer to only prioritize 00:13:51.549 --> 00:13:54.190 and highlight ideas, future requests, or problems 00:13:54.190 --> 00:13:56.090 if they are mentioned independently in at least 00:13:56.090 --> 00:13:58.269 two out of the three specified data sources. 00:13:58.769 --> 00:14:00.929 This automatically surfaces recurring themes 00:14:00.929 --> 00:14:02.909 and points of cross -departmental agreement, 00:14:03.350 --> 00:14:05.710 filtering out isolated comments or niche requests. 00:14:06.049 --> 00:14:08.450 It finds the consensus. And it doesn't just stop 00:14:08.450 --> 00:14:10.610 at the report, right? It automates the next steps. 00:14:10.840 --> 00:14:13.179 Yes, that's the final piece of its workflow. 00:14:13.820 --> 00:14:15.740 Based on the strategic priorities identified 00:14:15.740 --> 00:14:19.200 in its report, Synthesizer is programmed to automatically 00:14:19.200 --> 00:14:21.980 create corresponding tasks in a project management 00:14:21.980 --> 00:14:25.379 tool like Notion. And it even assigns those tasks 00:14:25.379 --> 00:14:28.120 to the correct team leads. Maybe the UX lead 00:14:28.120 --> 00:14:30.700 gets the research tasks, the engineering lead 00:14:30.700 --> 00:14:33.320 gets the technical implementation tasks, product 00:14:33.320 --> 00:14:35.759 marketing gets the go -to -market tasks. It connects 00:14:35.759 --> 00:14:38.840 the analysis directly to action. OK, these examples. 00:14:39.529 --> 00:14:43.049 Athena, Zenith, Nexus, Creator, Synthesizer, 00:14:43.710 --> 00:14:45.830 they really paint a picture of what's possible. 00:14:45.909 --> 00:14:48.230 It's way beyond just asking simple questions. 00:14:48.470 --> 00:14:50.429 Absolutely. So before we wrap this up, maybe 00:14:50.429 --> 00:14:52.110 we should touch on the overall philosophy here 00:14:52.110 --> 00:14:53.970 and also some common mistakes people make when 00:14:53.970 --> 00:14:56.789 trying this. Good idea. So key tips first. What's 00:14:56.789 --> 00:14:59.149 the mindset shift? I think the absolute number 00:14:59.149 --> 00:15:02.230 one thing is start thinking of the AI as a system, 00:15:02.350 --> 00:15:04.669 not just a discrete tool you use occasionally. 00:15:04.850 --> 00:15:07.389 You're actually designing an operational workflow, 00:15:07.429 --> 00:15:10.519 a pipeline. Right. And related to that, don't 00:15:10.519 --> 00:15:13.120 try to build one AI that does everything. That 00:15:13.120 --> 00:15:15.980 Swiss Army knife approach usually fails. Instead, 00:15:16.360 --> 00:15:19.320 create many different highly specialized employees 00:15:19.320 --> 00:15:22.899 or projects. Athena for data, Zenith for focus, 00:15:23.200 --> 00:15:25.840 Nexus for community insights. Each one should 00:15:25.840 --> 00:15:28.440 have a laser sharp focus. And crucially, this 00:15:28.440 --> 00:15:30.960 isn't a one and done setup. You have to iterate. 00:15:31.100 --> 00:15:33.519 Test your prompts, see what results you get, 00:15:33.919 --> 00:15:36.000 measure them against your goals, and then tweak 00:15:36.000 --> 00:15:37.960 and refine the system prompt until it's working 00:15:37.960 --> 00:15:41.080 perfectly for your specific needs. It's a continuous 00:15:41.080 --> 00:15:44.100 improvement loop. OK, so build systems, specialize, 00:15:44.379 --> 00:15:46.639 iterate. What about the common pitfalls? Where 00:15:46.639 --> 00:15:48.600 do people usually go wrong with these advanced 00:15:48.600 --> 00:15:50.860 prompts? Well, the first one, maybe ironically, 00:15:51.159 --> 00:15:53.679 is still giving unclear instructions, even within 00:15:53.679 --> 00:15:56.340 a long, complex prompt. You can have thousands 00:15:56.340 --> 00:15:59.059 of words in that system prompt. But if key parts 00:15:59.059 --> 00:16:01.519 lack specificity, like if you don't clearly define 00:16:01.519 --> 00:16:03.820 the target aesthetic for that fashion brand or 00:16:03.820 --> 00:16:06.159 the precise goal of synthesizer's report, the 00:16:06.159 --> 00:16:08.539 whole thing can still underperform or fail its 00:16:08.539 --> 00:16:11.480 main objective. Clarity remains king. Makes sense. 00:16:11.980 --> 00:16:15.340 What else? Second big mistake. Not telling the 00:16:15.340 --> 00:16:18.379 AI what not to do. Constraints are just as vital 00:16:18.379 --> 00:16:20.980 as positive commands. You need to proactively 00:16:20.980 --> 00:16:24.399 add negative rules. Things like, do not use tired 00:16:24.399 --> 00:16:27.259 marketing jargon, like world class solution or 00:16:27.259 --> 00:16:30.399 game changing, or avoid using passive voice in 00:16:30.399 --> 00:16:32.720 your summaries. These boundaries are essential 00:16:32.720 --> 00:16:35.379 for shaping the output quality and style. Right, 00:16:35.440 --> 00:16:37.419 defining the don'ts is as important as the dos. 00:16:37.860 --> 00:16:40.120 And the third mistake. Forgetting to tell what 00:16:40.120 --> 00:16:43.600 to do when it encounters errors. A really robust 00:16:43.600 --> 00:16:46.340 system, like a competent employee, needs clear 00:16:46.340 --> 00:16:49.440 instructions for handling failure. Remember Athena 00:16:49.440 --> 00:16:52.259 2 .0's rule. If you cannot find a file, tell 00:16:52.259 --> 00:16:54.279 me right away and ask for the correct path. Do 00:16:54.279 --> 00:16:56.639 not continue by yourself. You have to build in 00:16:56.639 --> 00:16:59.279 that kind of explicit error handling. Otherwise, 00:16:59.399 --> 00:17:01.299 the system might just try to guess or hallucinate 00:17:01.299 --> 00:17:03.480 its way through a problem, leading to bad outputs. 00:17:03.720 --> 00:17:06.079 Okay, unclear instructions, forgetting negative 00:17:06.079 --> 00:17:09.329 constraints, and no error handling plan. Got 00:17:09.329 --> 00:17:11.650 it. So pulling this all together, what's the 00:17:11.650 --> 00:17:13.569 big picture here? What do we really learn today? 00:17:13.950 --> 00:17:16.490 I think the core realization is that the true 00:17:16.490 --> 00:17:18.890 power of these advanced AIs isn't just locked 00:17:18.890 --> 00:17:21.750 inside the model's raw intelligence. It's unlocked 00:17:21.750 --> 00:17:24.150 by how effectively you design the instructions. 00:17:24.670 --> 00:17:27.009 You've essentially learned the secret that separates 00:17:27.009 --> 00:17:30.230 basic users from, well... AI architects. You're 00:17:30.230 --> 00:17:32.630 moving from just using the AI to actively designing 00:17:32.630 --> 00:17:35.049 and training your own bespoke digital team. Yeah, 00:17:35.049 --> 00:17:37.049 you become the designer of the system. Exactly. 00:17:37.089 --> 00:17:39.730 And the big takeaway is systemic thinking. By 00:17:39.730 --> 00:17:42.329 investing, say, a few focused hours today to 00:17:42.329 --> 00:17:44.470 build one of these custom project based systems, 00:17:44.529 --> 00:17:47.910 installing an Athena or a Zenith using the system 00:17:47.910 --> 00:17:50.289 prompt, you're setting up an automated engine. 00:17:50.430 --> 00:17:52.470 It's a self correcting machine that can potentially 00:17:52.470 --> 00:17:54.470 save you hundreds of hours down the line. It 00:17:54.470 --> 00:17:56.529 frees up your valuable time and mental energy 00:17:56.529 --> 00:17:59.670 for the truly high level stuff. strategic thinking, 00:18:00.230 --> 00:18:03.009 complex human interactions, genuine creativity. 00:18:03.369 --> 00:18:04.970 Okay, so here's a final thought for you, the 00:18:04.970 --> 00:18:07.569 listener, to really mull over. We've explored 00:18:07.569 --> 00:18:11.250 how to build one of these highly capable AI employees 00:18:11.250 --> 00:18:15.029 today. Now, really consider the immense competitive 00:18:15.029 --> 00:18:17.089 advantage you could gain when you have an entire 00:18:17.089 --> 00:18:20.630 team of these specialists. Imagine Athena managing 00:18:20.630 --> 00:18:23.490 your business data, Zenith optimizing your daily 00:18:23.490 --> 00:18:26.470 focus, Nexus constantly analyzing your customer 00:18:26.470 --> 00:18:30.140 feedback. all working tirelessly, 24 -7, following 00:18:30.140 --> 00:18:32.819 your exact rules. This isn't just some neat tech 00:18:32.819 --> 00:18:35.220 trick. It represents a fundamentally new layer 00:18:35.220 --> 00:18:37.180 of operational capability, a new way to work. 00:18:37.259 --> 00:18:39.380 It's automated competence. So the challenge is, 00:18:39.900 --> 00:18:42.500 pick one critical workflow in your job or business, 00:18:42.740 --> 00:18:44.519 get the Claw Desktop app if you haven't already, 00:18:45.099 --> 00:18:47.380 and start designing and building your very first 00:18:47.380 --> 00:18:49.259 AI executive today. See what happens.