WEBVTT 00:00:00.000 --> 00:00:02.540 Everyone is fighting the exact wrong war right 00:00:02.540 --> 00:00:04.960 now. Oh, absolutely. We constantly jump from 00:00:04.960 --> 00:00:08.820 Claude to ChatGPT to Gemini. Yeah, looking for 00:00:08.820 --> 00:00:11.140 the next big upgrade. Right. We keep chasing 00:00:11.140 --> 00:00:14.599 the illusion of a magic model. Beat. But the 00:00:14.599 --> 00:00:16.800 model is no longer the bottleneck. No, it really 00:00:16.800 --> 00:00:20.390 isn't. Welcome to this deep dive. Today, we explore 00:00:20.390 --> 00:00:23.469 what an agentic workflow actually is. It is completely 00:00:23.469 --> 00:00:25.289 going to change how you work. We will discover 00:00:25.289 --> 00:00:28.170 why cramming prompts absolutely destroys your 00:00:28.170 --> 00:00:32.409 output. And we will map the exact method to build 00:00:32.409 --> 00:00:36.149 daily skills. People still treat AI like a magic 00:00:36.149 --> 00:00:40.469 search box. And honestly, that is a massive mistake 00:00:40.469 --> 00:00:42.990 for your daily productivity. We really need to 00:00:42.990 --> 00:00:45.369 treat it like a brand new employee instead. Like 00:00:45.369 --> 00:00:48.479 onboarding a real person. Exactly. We are completely 00:00:48.479 --> 00:00:50.479 changing how we interact with these tools. Right, 00:00:50.640 --> 00:00:53.219 because an agentic workflow isn't just one megaprompt. 00:00:53.380 --> 00:00:55.820 No, not at all. It is the entire loop the agent 00:00:55.820 --> 00:00:58.399 runs through. It includes the specific context, 00:00:58.640 --> 00:01:01.000 the tools, and your hard rules. Yeah, and it 00:01:01.000 --> 00:01:03.259 even includes where the results get saved eventually. 00:01:03.340 --> 00:01:05.239 So it's a whole environment. Picture the model 00:01:05.239 --> 00:01:08.400 is just one machine on an assembly line. A perfect 00:01:08.400 --> 00:01:11.420 machine on a broken line still produces broken 00:01:11.420 --> 00:01:13.680 parts. That is a great way to look at it. You 00:01:13.680 --> 00:01:16.019 wouldn't hire a brilliant college grad and just 00:01:16.019 --> 00:01:19.120 walk out. Right. You would never just say, handle 00:01:19.120 --> 00:01:21.459 the whole company today. Exactly. You have to 00:01:21.459 --> 00:01:23.840 clearly define what good and bad looks like. 00:01:24.019 --> 00:01:26.819 But wait, these models have read the entire internet. 00:01:27.079 --> 00:01:30.000 Yeah, they have. Shouldn't that vast public knowledge 00:01:30.000 --> 00:01:33.219 be enough for basic tasks? Like, why do they 00:01:33.219 --> 00:01:36.319 need constant hand -holding for every little 00:01:36.319 --> 00:01:39.519 thing? because vague inputs will always yield 00:01:39.519 --> 00:01:43.140 incredibly vague outputs. These models just predict 00:01:43.140 --> 00:01:46.640 the next likely text chunk mathematically. They 00:01:46.640 --> 00:01:49.900 know general facts, but not your highly specific 00:01:49.900 --> 00:01:52.019 standards. They don't know your company's taste. 00:01:52.219 --> 00:01:54.140 Right. They do not actually think the way that 00:01:54.140 --> 00:01:57.780 humans think. The agentic workflow is the specific 00:01:57.780 --> 00:02:00.590 training plan for that new hire. Got it. So public 00:02:00.590 --> 00:02:02.569 knowledge isn't enough. It needs your specific 00:02:02.569 --> 00:02:04.670 training plan. That is exactly it. That makes 00:02:04.670 --> 00:02:06.670 total sense when you frame it like onboarding. 00:02:06.909 --> 00:02:10.330 But here is the problem. Since the AI desperately 00:02:10.330 --> 00:02:13.810 needs our instructions, we overdo it. We go way 00:02:13.810 --> 00:02:16.590 too far. Our natural instinct is to give it absolutely 00:02:16.590 --> 00:02:19.449 everything immediately. We just dump entire company 00:02:19.449 --> 00:02:21.810 manuals right into the chat window blindly. And 00:02:21.810 --> 00:02:24.530 that instinct completely breaks the machine over 00:02:24.530 --> 00:02:26.370 time. It totally does. We have to talk about 00:02:26.370 --> 00:02:29.150 the mechanics of the token window here. A token 00:02:29.150 --> 00:02:31.509 is basically just a tiny fragment of a word. 00:02:31.629 --> 00:02:34.509 Right. It is how the AI digests information. 00:02:34.849 --> 00:02:38.129 The AI processes text using these small mathematical 00:02:38.129 --> 00:02:42.569 chunks. Anthropix beta models are hitting 1 million 00:02:42.569 --> 00:02:44.930 tokens right now. Which is just a massive amount 00:02:44.930 --> 00:02:47.830 of data. It is. Most standard models give you 00:02:47.830 --> 00:02:50.729 200 ,000 tokens. But there's a huge problem with 00:02:50.729 --> 00:02:53.110 actually using all that space. Right. The recent 00:02:53.110 --> 00:02:55.770 research from Chroma highlights this perfectly. 00:02:55.889 --> 00:02:58.210 Yeah, the needle in a haystack test. Exactly. 00:02:58.909 --> 00:03:02.250 All 18 leading models drop in quality with longer 00:03:02.250 --> 00:03:04.930 contexts. They enter what the developer community 00:03:04.930 --> 00:03:08.610 calls the dumb zone. The dumb zone. That sounds 00:03:08.610 --> 00:03:10.770 incredibly frustrating to deal with. Oh, it is. 00:03:10.810 --> 00:03:14.629 It happens between 50 ,000 and 150 ,000 tokens. 00:03:15.810 --> 00:03:18.409 The attention mechanisms inside the neural network 00:03:18.409 --> 00:03:21.599 literally get diluted. So the agent basically 00:03:21.599 --> 00:03:23.740 just forgets your initial instruction. Right. 00:03:23.840 --> 00:03:26.560 It forgets and it gets incredibly sloppy. This 00:03:26.560 --> 00:03:29.500 degrading phenomenon is widely known as context 00:03:29.500 --> 00:03:32.919 rot today. I still wrestle with prompt drift 00:03:32.919 --> 00:03:36.169 myself. to sex silence. It happens to everyone. 00:03:36.530 --> 00:03:39.270 You ran a brilliant prompt with perfect safety 00:03:39.270 --> 00:03:42.909 guardrails. And then 20 messages later, the AI 00:03:42.909 --> 00:03:45.469 completely forgets those fundamental rules. Yes. 00:03:45.729 --> 00:03:47.830 An agent that is sharp at message five fails 00:03:47.830 --> 00:03:50.349 at 30. It just totally loses the plot. And this 00:03:50.349 --> 00:03:53.770 brings us directly to the 95 % rule. 95 % of 00:03:53.770 --> 00:03:56.689 users do not need giant instruction files. because 00:03:56.689 --> 00:03:59.789 it just wastes resources. Repeating known facts, 00:03:59.969 --> 00:04:02.530 like explaining what TypeScript is, just burns 00:04:02.530 --> 00:04:05.830 tokens. You can easily burn 140 ,000 tokens over 00:04:05.830 --> 00:04:08.370 20 turns. Wow, just on repetitive basics. You 00:04:08.370 --> 00:04:10.810 are paying precious context for information the 00:04:10.810 --> 00:04:13.669 agent already knows. So why do we instinctively 00:04:13.669 --> 00:04:16.800 react to an agent's failure by adding more? Like, 00:04:16.959 --> 00:04:19.500 why do we constantly cram more rules into the 00:04:19.500 --> 00:04:21.980 prompt? Because we mistakenly think that more 00:04:21.980 --> 00:04:25.339 text creates more safety. We think we are building 00:04:25.339 --> 00:04:28.360 better walls. In reality, your core signal just 00:04:28.360 --> 00:04:31.560 gets buried under useless noise. The AI simply 00:04:31.560 --> 00:04:34.079 loses the plot entirely when it gets overwhelmed. 00:04:34.379 --> 00:04:36.199 So we are just confusing it. You should only 00:04:36.199 --> 00:04:38.759 use large files for very specific things. Things 00:04:38.759 --> 00:04:42.079 like mandatory legal compliance language or strict 00:04:42.079 --> 00:04:45.350 brand voice guidelines. Makes sense. We think 00:04:45.350 --> 00:04:48.110 more text adds safety, but it just creates confusing 00:04:48.110 --> 00:04:50.490 noise. Exactly. It is honestly like shouting 00:04:50.490 --> 00:04:53.149 at someone in a crowded room. That is a perfect 00:04:53.149 --> 00:04:56.629 analogy. So if we cannot cram the context window, 00:04:56.709 --> 00:04:59.009 what do we do? Right. How are we supposed to 00:04:59.009 --> 00:05:01.470 give the AI the rules it actually needs? There 00:05:01.470 --> 00:05:03.670 has to be a better way. The elegant answer is 00:05:03.670 --> 00:05:05.790 to use something called skills. Right. A skill 00:05:05.790 --> 00:05:08.189 is just a small file with detailed instructions. 00:05:08.310 --> 00:05:10.430 Keep it super simple. It has a name, a short 00:05:10.430 --> 00:05:12.680 description, and an instruction block. And the 00:05:12.680 --> 00:05:15.220 secret here relies on a concept called progressive 00:05:15.220 --> 00:05:18.079 disclosure. Which Anthropic talked about recently. 00:05:18.240 --> 00:05:21.019 Right. The agent only loads the name and description 00:05:21.019 --> 00:05:23.480 into memory initially. It acts like a lightweight 00:05:23.480 --> 00:05:25.759 menu of available tools. Just a menu, not the 00:05:25.759 --> 00:05:28.060 whole meal. The heavy instruction body loads 00:05:28.060 --> 00:05:31.600 only if the agent actively needs it. The AI triggers 00:05:31.600 --> 00:05:34.310 a specific action. based on a semantic match. 00:05:34.629 --> 00:05:36.990 The token math on this mechanism is absolutely 00:05:36.990 --> 00:05:39.810 incredible. It really is wild. The metadata only 00:05:39.810 --> 00:05:43.009 costs about 50 tokens to scan initially, but 00:05:43.009 --> 00:05:46.170 the hidden instruction body might be 5 ,000 tokies 00:05:46.170 --> 00:05:49.329 long. This creates a nearly 18 to 1 token savings 00:05:49.329 --> 00:05:52.490 ratio for you. Whoa, imagine saving tens of thousands 00:05:52.490 --> 00:05:54.910 of tokens a week just hiding instructions. It 00:05:54.910 --> 00:05:57.490 adds up so fast. It is like stacking Lego blocks 00:05:57.490 --> 00:06:00.189 of data efficiently. It keeps the context window 00:06:00.189 --> 00:06:03.839 perfectly clean and high. focused? It is a massive 00:06:03.839 --> 00:06:06.160 upgrade for your daily workflow. For sure. But 00:06:06.160 --> 00:06:08.220 there's a huge warning we need to discuss right 00:06:08.220 --> 00:06:10.199 now. Okay what is the catch? You should never 00:06:10.199 --> 00:06:13.220 use random downloaded skills from the internet. 00:06:13.579 --> 00:06:15.759 Wait if I shouldn't use public skills aren't 00:06:15.759 --> 00:06:18.759 I stuck? Writing custom code for an agent sounds 00:06:18.759 --> 00:06:21.259 like a massive bottleneck like how do I actually 00:06:21.259 --> 00:06:23.540 create one without just guessing the code? Because 00:06:23.540 --> 00:06:26.379 public skills can contain highly dangerous prompt 00:06:26.379 --> 00:06:29.620 injections. Oh wow. They might have strange API 00:06:29.620 --> 00:06:32.689 endpoints hidden deep in the code. They could 00:06:32.689 --> 00:06:35.470 even contain malicious data exfiltration steps 00:06:35.470 --> 00:06:37.430 inside them. So they are stealing your data? 00:06:37.670 --> 00:06:40.050 You are essentially letting a total stranger 00:06:40.050 --> 00:06:42.209 write scripts for your agent. That sounds like 00:06:42.209 --> 00:06:45.089 a massive security risk for any company. It is. 00:06:45.589 --> 00:06:47.870 And even if they are totally safe, they were 00:06:47.870 --> 00:06:49.750 written for someone else. Right, for a different 00:06:49.750 --> 00:06:52.550 company's workflow. The seams will show up very 00:06:52.550 --> 00:06:55.350 fast in your own setup. So you read the public 00:06:55.350 --> 00:06:57.769 skills for structural inspiration only. Just 00:06:57.769 --> 00:06:59.730 to see how they did it. You borrow their general 00:06:59.730 --> 00:07:02.430 architecture, but you write the code yourself. 00:07:02.670 --> 00:07:05.069 You tailor it perfectly to your own proprietary 00:07:05.069 --> 00:07:07.209 data. Right, so borrow structure from public 00:07:07.209 --> 00:07:10.670 skills, but write the actual code yourself. Exactly. 00:07:10.930 --> 00:07:13.709 Protect your workflow. Midroll sponsor read. 00:07:13.920 --> 00:07:16.459 mid -roll sponsor read, content inserted here, 00:07:16.600 --> 00:07:19.139 and mid -roll sponsor read. So let's look at 00:07:19.139 --> 00:07:21.699 the exact step -by -step process of building 00:07:21.699 --> 00:07:23.519 one. Yes, let's get into the practical side. 00:07:23.639 --> 00:07:25.800 We are going to build a daily competitor tracking 00:07:25.800 --> 00:07:28.399 bot today. I really like this practical example 00:07:28.399 --> 00:07:31.939 from the outline. This bot checks five competitor 00:07:31.939 --> 00:07:34.439 websites every single morning. Automating the 00:07:34.439 --> 00:07:37.300 boring stuff. It looks at their pricing pages, 00:07:37.579 --> 00:07:40.759 blogs, change logs, and X accounts. Now if you 00:07:40.759 --> 00:07:43.379 use a lazy prompt, you just ask what is new. 00:07:43.470 --> 00:07:45.769 Which never works well. The agent will immediately 00:07:45.769 --> 00:07:48.149 fail and ask you confusing questions. It will 00:07:48.149 --> 00:07:51.370 ask about RSS readers, web scrapers, or visual 00:07:51.370 --> 00:07:53.870 comping. Right, because you gave it an incredibly 00:07:53.870 --> 00:07:56.050 vague objective. It doesn't know how to start. 00:07:56.170 --> 00:07:58.949 To fix this, you have to build a very clear spec. 00:07:59.269 --> 00:08:02.329 You need to add specific exclusion rules directly 00:08:02.329 --> 00:08:04.769 to the prompt. Give it actual boundaries. You 00:08:04.769 --> 00:08:08.610 tell it to only flag if a price moves 10%. Or 00:08:08.610 --> 00:08:11.629 flag it if a brand new paid tier suddenly appears. 00:08:11.649 --> 00:08:14.990 Yes. Or if an ex post has over 200 likes. Exactly. 00:08:15.269 --> 00:08:17.550 Then you tell it to save that output to a Notion 00:08:17.550 --> 00:08:21.500 page. Keep it organized. But here is the absolute 00:08:21.500 --> 00:08:25.759 golden rule of this entire process. You must 00:08:25.759 --> 00:08:28.759 run the workflow completely manually at first. 00:08:29.139 --> 00:08:31.759 You have to actually prove the logic works in 00:08:31.759 --> 00:08:35.110 reality. Yes. Do not skip this step. You wait 00:08:35.110 --> 00:08:38.509 for two completely clean back -to -back manual 00:08:38.509 --> 00:08:40.889 runs. Just to be sure it wasn't a fluke. Only 00:08:40.889 --> 00:08:43.570 when it succeeds twice do you save it as a skill. 00:08:43.710 --> 00:08:45.409 And you don't even write the skill description 00:08:45.409 --> 00:08:47.889 yourself. Well, you let the AI do it. You ask 00:08:47.889 --> 00:08:50.309 the agent to review those two successful runs. 00:08:50.649 --> 00:08:52.970 Then it writes its own skill description in under 00:08:52.970 --> 00:08:55.710 30 words. That ensures the skill is born from 00:08:55.710 --> 00:08:58.809 a real success. It is not just a blind guess 00:08:58.809 --> 00:09:02.590 from a human. It uses its own language. The AI 00:09:02.590 --> 00:09:04.649 knows exactly what trigger words will activate 00:09:04.649 --> 00:09:07.309 it best. That is the only kind of skill worth 00:09:07.309 --> 00:09:09.450 keeping around. But what happens when reality 00:09:09.450 --> 00:09:11.889 actually sets in on a Tuesday? Oh, it always 00:09:11.889 --> 00:09:13.950 does. Like a website completely changes this 00:09:13.950 --> 00:09:16.590 layout without any warning, or an API throws 00:09:16.590 --> 00:09:19.570 a random 503 error during a run, the skill is 00:09:19.570 --> 00:09:21.289 going to trip and break eventually. Of course 00:09:21.289 --> 00:09:23.210 it will break when the real world shifts. So 00:09:23.210 --> 00:09:26.090 what do you do? When it trips, you don't just 00:09:26.090 --> 00:09:28.490 rewrite it from scratch. You use what we call 00:09:28.490 --> 00:09:30.870 recursive improvement loops. Okay, how does that 00:09:30.870 --> 00:09:34.120 work? You ask the AI why it broke, and you fix 00:09:34.120 --> 00:09:36.899 it together. Ah. So we use recursive improvement 00:09:36.899 --> 00:09:39.259 loops to fix and update it together. Exactly. 00:09:39.440 --> 00:09:42.100 You iterate on the errors. But once you get one 00:09:42.100 --> 00:09:44.960 skill working, a dangerous temptation kicks in. 00:09:45.019 --> 00:09:47.179 Oh, I know what you were going to say. People 00:09:47.179 --> 00:09:51.000 immediately want to build a massive complex network 00:09:51.000 --> 00:09:53.340 of agents. I know everyone listening right now 00:09:53.340 --> 00:09:55.759 immediately wants to build swarms. You want 15 00:09:55.759 --> 00:09:57.799 sub -agents talking to each other with a maze 00:09:57.799 --> 00:09:59.500 of arrows. It looks so cool on a whiteboard. 00:09:59.799 --> 00:10:02.620 But you are saying that is a massive trap for 00:10:02.620 --> 00:10:04.720 beginners. It is the trap of wanting to look 00:10:04.720 --> 00:10:07.679 impressive online. You have to build a hut before 00:10:07.679 --> 00:10:10.080 you build a cathedral. Start small. Stick to 00:10:10.080 --> 00:10:12.639 one main agent and maybe three to five skills. 00:10:13.799 --> 00:10:16.559 You should only scale that system when your productivity 00:10:16.559 --> 00:10:19.000 truly demands it. Let's talk about using your 00:10:19.000 --> 00:10:21.960 actual code as context. Because the agent can 00:10:21.960 --> 00:10:24.019 read things straight from your code base. Right. 00:10:24.139 --> 00:10:27.200 It sees Next .js, SuperBase, and Tailwind automatically. 00:10:27.320 --> 00:10:29.919 It knows what they are. Do not waste your precious 00:10:29.919 --> 00:10:32.639 tokens explaining those public frameworks. You 00:10:32.639 --> 00:10:35.460 should be using clean starter templates instead 00:10:35.460 --> 00:10:38.639 of long explanations. There is a really great 00:10:38.639 --> 00:10:40.779 screening question for this context. You just 00:10:40.779 --> 00:10:43.440 ask yourself, can the AI find this information 00:10:43.440 --> 00:10:46.299 publicly? That is the perfect filter. If it can 00:10:46.299 --> 00:10:49.419 find it on the open web, Skip adding it. Don't 00:10:49.419 --> 00:10:52.039 repeat the internet. You only add the things 00:10:52.039 --> 00:10:55.659 the AI absolutely cannot find publicly. Things 00:10:55.659 --> 00:10:58.940 like your highly specific customer personas or 00:10:58.940 --> 00:11:01.899 pricing logic. Exactly. Or the specific things 00:11:01.899 --> 00:11:05.220 your company likes to say no to. That taste and 00:11:05.220 --> 00:11:08.080 constraint is highly proprietary to your business. 00:11:08.220 --> 00:11:10.679 Those are the high value pieces of context you 00:11:10.679 --> 00:11:13.500 have to protect. But isn't it safer to just remind 00:11:13.500 --> 00:11:15.740 the agent what framework we use just in case 00:11:15.740 --> 00:11:18.600 it somehow forgets the syntax during a long task? 00:11:18.919 --> 00:11:21.700 No, reminding it just clutters the precious context 00:11:21.700 --> 00:11:24.379 window unnecessarily. It takes up that valuable 00:11:24.379 --> 00:11:26.360 space. It pushes you right back toward the dumb 00:11:26.360 --> 00:11:28.820 zone we discussed earlier. Keep it out if public, 00:11:29.379 --> 00:11:31.779 included if proprietary. I see. Skip it in public. 00:11:31.899 --> 00:11:34.059 Keep it only if it is completely proprietary. 00:11:34.460 --> 00:11:37.960 We now have a blueprint for a lean, simple workflow. 00:11:38.139 --> 00:11:40.200 Which is great. But you need to know about the 00:11:40.200 --> 00:11:43.879 painful reality of the first 14 days. Let's share 00:11:43.879 --> 00:11:46.740 that skeleton template example for a weekly revenue 00:11:46.740 --> 00:11:49.500 review. It pulls your financial data directly 00:11:49.500 --> 00:11:52.019 from Striken Notion. Super useful automation. 00:11:52.340 --> 00:11:54.620 It calculates week -over -week deltas and flags 00:11:54.620 --> 00:11:58.639 15 % moves. Then it saves a clean five -line 00:11:58.639 --> 00:12:01.700 summary for your team. It sounds great, but getting 00:12:01.700 --> 00:12:03.740 this to work means hitting the two -week wall. 00:12:03.899 --> 00:12:06.240 The two -week wall. The first 10 to 14 days are 00:12:06.240 --> 00:12:09.220 gonna honestly hurt. You are gonna spend way 00:12:09.220 --> 00:12:12.679 more time fixing errors than producing. The automated 00:12:12.679 --> 00:12:15.159 system feels significantly worse than just doing 00:12:15.159 --> 00:12:17.639 it manually. It is super frustrating at first. 00:12:17.720 --> 00:12:20.539 Yeah. But around day 14, you finally hit the 00:12:20.539 --> 00:12:23.320 turning point, the recursive loops pay off, and 00:12:23.320 --> 00:12:25.139 the skills stop breaking. Everything just starts 00:12:25.139 --> 00:12:27.419 clicking. A reporting task that used to take 00:12:27.419 --> 00:12:30.019 an hour now takes 10 minutes. And this creates 00:12:30.019 --> 00:12:32.480 a massive career advantage for you over others. 00:12:33.059 --> 00:12:36.139 People using tight, agentic workflows are going 00:12:36.139 --> 00:12:38.299 to do the work of three people. They really are. 00:12:38.539 --> 00:12:41.669 A solo founder. with five sharp skills becomes 00:12:41.669 --> 00:12:45.129 incredibly fast. They will easily outrun a 10 00:12:45.129 --> 00:12:47.549 -person team that is still arguing about tools. 00:12:48.129 --> 00:12:50.710 That leverage is entirely real, but you have 00:12:50.710 --> 00:12:53.370 to actually earn it. Is this initial friction 00:12:53.370 --> 00:12:56.070 the real reason people mistakenly blame the AI 00:12:56.070 --> 00:12:58.490 models? Like, is this why they jump to a different 00:12:58.490 --> 00:13:00.870 provider the second something fails? Absolutely. 00:13:01.350 --> 00:13:03.570 They blame the machine because the initial training 00:13:03.570 --> 00:13:05.690 is hard. They don't want to do the work. They 00:13:05.690 --> 00:13:07.809 quit at the wall instead of pushing through the 00:13:07.809 --> 00:13:10.779 friction. They refuse to pay the initial training 00:13:10.779 --> 00:13:14.000 cost for their new employee. Exactly. They quit 00:13:14.000 --> 00:13:15.919 at the two -week wall instead of paying the training 00:13:15.919 --> 00:13:18.559 costs. The big takeaway here is that the model 00:13:18.559 --> 00:13:21.139 was never the bottleneck. It was always our process. 00:13:21.340 --> 00:13:24.299 Less context, clearer rules, and smaller starts 00:13:24.299 --> 00:13:27.200 are the absolute keys. They are what actually 00:13:27.200 --> 00:13:30.100 create a true agentic workflow for you. Keep 00:13:30.100 --> 00:13:33.039 it lean. Do it by hand, get two clean rends, 00:13:33.059 --> 00:13:35.399 and make it a skill. Thank you for joining us 00:13:35.399 --> 00:13:37.120 on this deep dive today. I want to leave you 00:13:37.120 --> 00:13:40.320 with a final thought before we go. Beat. Think 00:13:40.320 --> 00:13:42.659 about a repetitive task you do every single morning. 00:13:42.700 --> 00:13:45.600 We all have them. What are the unspoken, invisible 00:13:45.600 --> 00:13:47.940 rules sitting in your head right now? The ones 00:13:47.940 --> 00:13:50.159 you have never explicitly written down for another 00:13:50.159 --> 00:13:53.360 human? Let alone an AI. How could an AI ever 00:13:53.360 --> 00:13:55.879 guess them if you don't write them down? We used 00:13:55.879 --> 00:13:58.240 to think the magic was always in the model, but 00:13:58.240 --> 00:14:00.279 the real magic is making your invisible rules 00:14:00.279 --> 00:14:02.220 visible. OTRO music.