WEBVTT 00:00:00.000 --> 00:00:03.919 Imagine constantly rewriting, say, a 500 word 00:00:03.919 --> 00:00:06.860 instruction manual over and over just to get 00:00:06.860 --> 00:00:09.439 one consistent result from your AI. Right. The 00:00:09.439 --> 00:00:11.919 real headache in a lot of modern AI work isn't 00:00:11.919 --> 00:00:14.160 about being creative, you know. Yeah. It's the 00:00:14.160 --> 00:00:16.079 repetition. It's just exhausting. Yeah. We're 00:00:16.079 --> 00:00:19.079 talking about serious prompt fatigue, that burnout 00:00:19.079 --> 00:00:22.500 feeling. But there's a fix now from a major platform. 00:00:22.539 --> 00:00:25.239 And the key thing is it's actually portable. 00:00:26.739 --> 00:00:29.760 Welcome to the deep dive. Today we're digging 00:00:29.760 --> 00:00:32.179 into the source material on this new AI feature 00:00:32.179 --> 00:00:35.640 they're calling Skills. And this isn't just like 00:00:35.640 --> 00:00:38.240 a minor shortcut. It feels more like a fundamental 00:00:38.240 --> 00:00:40.719 shift, moving away from writing those endless 00:00:40.719 --> 00:00:44.640 prompts towards building reusable, detailed rule 00:00:44.640 --> 00:00:46.840 books for your AI. So we're going to unpack what 00:00:46.840 --> 00:00:49.020 these skills really are, how they stack up against 00:00:49.020 --> 00:00:51.560 older methods like custom instructions, and importantly, 00:00:51.579 --> 00:00:53.359 how you can actually build them and then use 00:00:53.359 --> 00:00:55.619 them across different platforms. Think Claude, 00:00:55.740 --> 00:00:58.179 ChatGPT, maybe others down the line. Yeah, our 00:00:58.179 --> 00:01:00.000 goal today is pretty straightforward. Give you 00:01:00.000 --> 00:01:03.579 a clear map, a way to create these powerful portable 00:01:03.469 --> 00:01:05.950 abilities that could genuinely save you hours. 00:01:06.390 --> 00:01:08.250 Okay, let's unpack that starting point. Before 00:01:08.250 --> 00:01:11.370 this skills idea, pretty much everything relied 00:01:11.370 --> 00:01:13.909 on you crafting that perfect prompt, right? The 00:01:13.909 --> 00:01:16.549 big instruction you feed the AI. Totally. Need 00:01:16.549 --> 00:01:19.430 a financial analysis. Yeah. Okay, write a huge 00:01:19.430 --> 00:01:22.290 detailed prompt. Need a presentation draft. Another 00:01:22.290 --> 00:01:25.370 one. Complex job search plan. Yep, you guessed 00:01:25.370 --> 00:01:28.109 it. Another massive prompt. And doing this day 00:01:28.109 --> 00:01:31.239 in, day out. Yeah. It's mentally draining. You're 00:01:31.239 --> 00:01:33.560 basically tweaking similar instructions constantly. 00:01:34.040 --> 00:01:36.540 And that friction, it just makes hard work feel 00:01:36.540 --> 00:01:39.920 tiring and inefficient. The source used a good 00:01:39.920 --> 00:01:41.920 analogy. It's like having this incredibly smart 00:01:41.920 --> 00:01:44.219 assistant, but you still have to explain exactly 00:01:44.219 --> 00:01:46.340 how to make coffee, exactly how to organize files 00:01:46.340 --> 00:01:49.599 every single morning, that constant reset. It's 00:01:49.599 --> 00:01:52.480 taxing, even with a template. Precisely. But 00:01:52.480 --> 00:01:54.760 why is that constant re -explaining so mentally 00:01:54.760 --> 00:01:57.079 tough when the AI is so powerful? Well, it boils 00:01:57.079 --> 00:01:59.879 down to the mental load, right? Switching context 00:01:59.879 --> 00:02:01.920 and having to double check that every single 00:02:01.920 --> 00:02:04.200 instruction is perfect every single time. So 00:02:04.200 --> 00:02:07.180 what's really interesting is how a skill is defined 00:02:07.180 --> 00:02:09.599 here. It's basically a detailed rule book. It 00:02:09.599 --> 00:02:11.599 holds all the instructions, all the preferences, 00:02:12.240 --> 00:02:15.430 for one specific task. And you only build it 00:02:15.430 --> 00:02:17.949 once. And the key part, I think, is that the 00:02:17.949 --> 00:02:20.710 AI uses the right skill automatically when the 00:02:20.710 --> 00:02:23.229 topic comes up. You don't have to, like, tag 00:02:23.229 --> 00:02:26.590 it or remember to say use skill X. That auto 00:02:26.590 --> 00:02:29.069 detection feels like the real game changer. Oh, 00:02:29.110 --> 00:02:31.550 absolutely. I mean, I still wrestle with prompt 00:02:31.550 --> 00:02:34.449 drift myself sometimes, you know, where the AI 00:02:34.449 --> 00:02:36.789 kind of forgets the complex rules you gave it 00:02:36.789 --> 00:02:39.169 way back at the start of a long chat. This auto 00:02:39.169 --> 00:02:41.509 detection thing, that could be a huge step away 00:02:41.509 --> 00:02:43.389 from that whole problem. OK, so let's try to 00:02:43.389 --> 00:02:45.800 dist - The older methods, like custom instructions 00:02:45.800 --> 00:02:48.639 in ChatGPT, they're sort of like one general 00:02:48.639 --> 00:02:50.819 sticky note the AI always reads, right? Which 00:02:50.819 --> 00:02:53.659 can cause issues. Exactly. That's how you end 00:02:53.659 --> 00:02:56.620 up with your super formal business expert AI 00:02:56.620 --> 00:03:00.300 trying to write a love poem using corporate jargon. 00:03:00.729 --> 00:03:03.150 It doesn't know when not to apply the instruction. 00:03:03.909 --> 00:03:06.030 Skills, though, are different. They're like having 00:03:06.030 --> 00:03:09.250 a whole shelf of specialized books. The AI only 00:03:09.250 --> 00:03:12.090 grabs the Excel data analysis skill book when 00:03:12.090 --> 00:03:14.050 you actually start talking about spreadsheets 00:03:14.050 --> 00:03:16.629 and data. Which makes the output cleaner, more 00:03:16.629 --> 00:03:19.189 accurate. And you can have, what, maybe 20 of 00:03:19.189 --> 00:03:21.629 these specialized rule books ready to go. But 00:03:21.629 --> 00:03:24.490 these complex rules, they need a simple container, 00:03:24.610 --> 00:03:26.389 right? How are they packaged so they can be shared 00:03:26.389 --> 00:03:28.729 easily? Yeah, it's surprisingly simple. They're 00:03:28.729 --> 00:03:31.909 just plain text files, usually marked down, all 00:03:31.909 --> 00:03:34.449 bundled up together in a standard zip file. And 00:03:34.449 --> 00:03:36.289 the nice thing is, you don't need to be some 00:03:36.289 --> 00:03:38.629 kind of prompt engineering guru to get started. 00:03:38.849 --> 00:03:41.110 The AI actually guides you through making one. 00:03:41.430 --> 00:03:43.229 Yeah, the process starts pretty easily in the 00:03:43.229 --> 00:03:45.330 Claude settings. You just find skills or maybe 00:03:45.330 --> 00:03:47.370 capabilities, turn on custom skills, and you're 00:03:47.370 --> 00:03:50.030 good to go. Then you start a chat. Say you want 00:03:50.030 --> 00:03:52.889 to create a new skill. And this seems important. 00:03:53.449 --> 00:03:55.810 You ask Claude to act like an expert in building 00:03:55.810 --> 00:03:58.259 these things. Then it starts asking you the key 00:03:58.259 --> 00:04:01.439 questions. That's right. The AI will ask about 00:04:01.439 --> 00:04:03.580 the primary goal for the skill, the specific 00:04:03.580 --> 00:04:06.139 brand tone of voice you need and who your target 00:04:06.139 --> 00:04:08.500 audience is. It kind of forces you to get clear 00:04:08.500 --> 00:04:10.500 on the basics. The source material gave some 00:04:10.500 --> 00:04:13.099 good examples, like the brand voice was defined 00:04:13.099 --> 00:04:15.740 as a friendly, knowledgeable friend and there 00:04:15.740 --> 00:04:19.420 were ban words, never used cheap or by now or 00:04:19.420 --> 00:04:22.180 shocking sale. The main goal was engagement, 00:04:22.819 --> 00:04:25.360 explicitly not direct selling. And then Claude 00:04:25.360 --> 00:04:28.290 digs deeper. It asks platform -specific stuff, 00:04:28.470 --> 00:04:30.910 like what's the ideal length for a Facebook post 00:04:30.910 --> 00:04:33.490 using this skill? What's the hashtag strategy 00:04:33.490 --> 00:04:36.529 for Instagram? The right tone for TikTok. That 00:04:36.529 --> 00:04:38.670 back and forth really refines it. OK, but hold 00:04:38.670 --> 00:04:40.910 on. If I'm, say, a freelancer juggling maybe 00:04:40.910 --> 00:04:43.410 five or six different recurring tasks, is it 00:04:43.410 --> 00:04:45.170 really worth investing the time to spell out, 00:04:45.170 --> 00:04:48.490 like, 15 super detailed rules for every single 00:04:48.490 --> 00:04:50.810 platform within one skill? Well, that's kind 00:04:50.810 --> 00:04:53.290 of the point of asking the AI to act as the expert 00:04:53.290 --> 00:04:56.379 builder. It asks better, more targeted questions 00:04:56.379 --> 00:04:58.879 upfront. It helps you build that solid foundation 00:04:58.879 --> 00:05:01.160 correctly the first time, instead of you having 00:05:01.160 --> 00:05:03.420 to patch up mistakes later. Saves time in the 00:05:03.420 --> 00:05:04.759 long run. Okay, this is where it gets really 00:05:04.759 --> 00:05:06.519 practical. Let's look at the actual rules from 00:05:06.519 --> 00:05:08.439 the example skill and the storage material, the 00:05:08.439 --> 00:05:10.879 social media content plan skill. Yeah, this was 00:05:10.879 --> 00:05:13.220 cool. For Facebook, the rules were super specific, 00:05:13.680 --> 00:05:17.339 really sticky, length, strictly 150 to 200 words, 00:05:17.759 --> 00:05:21.730 title. always IN, all CPPS, and I have to include 00:05:21.730 --> 00:05:24.569 a main keyword, you must use bullet points. And 00:05:24.569 --> 00:05:27.089 the post absolutely has to end with one open 00:05:27.089 --> 00:05:29.170 -ended question to get comments going. Then for 00:05:29.170 --> 00:05:31.910 Instagram, totally different vibe. Focus on feelings, 00:05:32.069 --> 00:05:35.250 keep it short, under 50 words, period. And crucially, 00:05:35.430 --> 00:05:38.069 the rulebook demanded three specific brand hashtags, 00:05:38.529 --> 00:05:41.769 like hard -coded ones, plus five relevant trending 00:05:41.769 --> 00:05:45.269 hashtags. No deviation. And TikTok was fascinating. 00:05:45.420 --> 00:05:48.339 It broke it down into a ready -made 15 -second 00:05:48.339 --> 00:05:51.459 video script structure. Three scenes. Scene 1, 00:05:51.660 --> 00:05:54.540 seconds 0 to 3, that's your hook. Scene 2, seconds 00:05:54.540 --> 00:05:57.920 4 to 10, present the solution. Scene 3, 11 to 00:05:57.920 --> 00:06:01.000 15 seconds, call to action. It even asked for 00:06:01.000 --> 00:06:03.620 a suggested style of audio. Wow. So, okay, instead 00:06:03.620 --> 00:06:05.879 of writing this monster prompt explaining all 00:06:05.879 --> 00:06:07.699 those platform nuances every single time, you 00:06:07.699 --> 00:06:09.959 just say something simple like, topic, benefits 00:06:09.959 --> 00:06:12.279 of drinking enough water, use my social media 00:06:12.279 --> 00:06:14.259 skill, and boom, the skill handles all those 00:06:14.259 --> 00:06:17.279 detailed platform -specific rules automatically. 00:06:17.379 --> 00:06:19.399 That's the power of systemizing it, right? So, 00:06:19.459 --> 00:06:21.000 when you give that simple instruction, what's 00:06:21.000 --> 00:06:23.740 the key rule that stops the AI from accidentally 00:06:23.740 --> 00:06:26.000 using that short, punchy Instagram caption on 00:06:26.000 --> 00:06:28.290 Facebook? It's that strict length requirement. 00:06:28.550 --> 00:06:31.269 Under 50 words for Insta versus the mandatory 00:06:31.269 --> 00:06:34.370 150 to 200 for Facebook, the rule book kind of 00:06:34.370 --> 00:06:37.110 polices itself based on those constraints. Mid 00:06:37.110 --> 00:06:40.629 -roll sponsor, red placeholder. Now, the part 00:06:40.629 --> 00:06:42.769 that makes this whole concept potentially revolutionary, 00:06:42.769 --> 00:06:45.910 I think, is the portability. Because these skills 00:06:45.910 --> 00:06:48.189 are just structured text files inside a zip folder, 00:06:48.610 --> 00:06:50.769 you can build your rulebook in Claude and then 00:06:50.769 --> 00:06:53.410 just use it in ChatGPT. Exactly. You just upload 00:06:53.410 --> 00:06:56.230 that zip file to a version of ChatGPT that can 00:06:56.230 --> 00:06:58.209 read files, GPT -4 for instance, and then give 00:06:58.209 --> 00:07:00.389 it one critical instruction. Something like, 00:07:00.790 --> 00:07:02.970 act like an AI that follows 100 % of the rules 00:07:02.970 --> 00:07:05.069 inside the uploaded file. So you're essentially 00:07:05.069 --> 00:07:08.189 giving ChatGPT the exact same detailed training 00:07:08.189 --> 00:07:10.569 manual that your Claude skill uses. That means 00:07:10.569 --> 00:07:13.649 you can get consistent on -brand output across 00:07:13.649 --> 00:07:15.769 different AI platforms. You're not locked into 00:07:15.769 --> 00:07:18.029 one vendor. That's huge. Yeah, and there was 00:07:18.029 --> 00:07:19.990 another clever trick mentioned. You could actually 00:07:19.990 --> 00:07:23.750 use a different AI to sort of proofread or upgrade 00:07:23.750 --> 00:07:27.490 your skills rulebook. Use chat GPT to look for 00:07:27.490 --> 00:07:29.790 potential logic flaws in your clog -built skill. 00:07:30.189 --> 00:07:33.110 For example, it might spot that asking for a 00:07:33.110 --> 00:07:36.050 trending background sound for TikTok is maybe 00:07:36.050 --> 00:07:38.930 too vague. You know, what's trending changes 00:07:38.930 --> 00:07:40.850 constantly. And the fix was straightforward. 00:07:41.089 --> 00:07:43.230 Just update the rule. Instead of asking for the 00:07:43.230 --> 00:07:45.370 trending sound, tell it to suggest a type of 00:07:45.370 --> 00:07:48.170 music, maybe lo -fi or happy pop music, something 00:07:48.170 --> 00:07:50.870 the AI can actually deliver consistently. So 00:07:50.870 --> 00:07:52.509 this ability to just take the rulebook itself 00:07:52.509 --> 00:07:54.790 and move it around. How does that fundamentally 00:07:54.790 --> 00:07:57.649 change things for us as users? How does it loosen 00:07:57.649 --> 00:08:00.709 the grip of being tied to just one AI tool? Well, 00:08:00.850 --> 00:08:02.550 skills create these abilities that are truly 00:08:02.550 --> 00:08:05.610 portable and reusable. It shifts the intelligence 00:08:05.610 --> 00:08:07.970 or at least the specific operational rules away 00:08:07.970 --> 00:08:10.410 from the underlying AI model and puts it into 00:08:10.410 --> 00:08:12.990 your hands in a format you can share and control. 00:08:13.550 --> 00:08:16.389 Reduces that platform dependency. So the big 00:08:16.389 --> 00:08:18.970 question then is, when should you actually invest 00:08:18.970 --> 00:08:21.709 the time to build one of these skills? The source 00:08:21.709 --> 00:08:24.589 gave a pretty good rule of thumb. If a task is 00:08:24.589 --> 00:08:26.110 important enough that you'd normally write up 00:08:26.110 --> 00:08:28.649 a detailed training guide for a human, then yeah, 00:08:28.829 --> 00:08:31.110 make it a skill. Makes sense. So that covers 00:08:31.110 --> 00:08:34.190 things like recurring jobs, weekly reports, maybe. 00:08:34.190 --> 00:08:36.250 Yeah. Complex projects like putting together 00:08:36.250 --> 00:08:38.649 a strategic plan. And definitely anything with 00:08:38.649 --> 00:08:40.970 really specific rules you have to follow. Brand 00:08:40.970 --> 00:08:44.169 voice. legal disclaimers, data formatting, that 00:08:44.169 --> 00:08:46.289 kind of stuff. We saw some great specific examples, 00:08:46.429 --> 00:08:48.350 like a PowerPoint outline skill. It could be 00:08:48.350 --> 00:08:50.889 built to always use the five by five rule, you 00:08:50.889 --> 00:08:53.070 know, max five lines per slide, max five words 00:08:53.070 --> 00:08:55.649 per line, and maybe enforce a standard 10 slide 00:08:55.649 --> 00:08:59.289 structure every time, flawlessly. Or imagine 00:08:59.289 --> 00:09:02.730 a basic Excel data analysis skill. It could be 00:09:02.730 --> 00:09:05.750 trained to always follow a set process. First, 00:09:05.970 --> 00:09:08.429 understand the data columns. Second, suggest 00:09:08.429 --> 00:09:11.519 data cleaning steps. Third, run basic analysis 00:09:11.519 --> 00:09:15.240 averages, medians, counts. And fourth, only present 00:09:15.240 --> 00:09:17.919 the results using clear, formatted bullet points. 00:09:18.200 --> 00:09:20.590 No rambling. Whoa. Okay, when you start thinking 00:09:20.590 --> 00:09:22.570 about scaling that, the potential is pretty wild. 00:09:22.850 --> 00:09:24.889 The source hints at a future where maybe there's 00:09:24.889 --> 00:09:26.970 a marketplace. You could just download an expert 00:09:26.970 --> 00:09:29.509 created financial analysis skill and instantly 00:09:29.509 --> 00:09:31.409 plug that sophisticated system into whatever 00:09:31.409 --> 00:09:34.470 AI tool you prefer. Right. So let's say you build 00:09:34.470 --> 00:09:36.669 two foundational skills yourself. One for basic 00:09:36.669 --> 00:09:38.950 SEO analysis, maybe, and another for generating 00:09:38.950 --> 00:09:40.929 meeting summaries. What's the crucial difference 00:09:40.929 --> 00:09:42.649 in the outcome you're demanding from each of 00:09:42.649 --> 00:09:46.009 those skills? Ah, good question. The SEO skill 00:09:46.009 --> 00:09:50.009 would be laser focused on output format, structure, 00:09:50.269 --> 00:09:52.370 hitting keyword targets. Whereas the meeting 00:09:52.370 --> 00:09:54.950 summary skill is all about accurately capturing 00:09:54.950 --> 00:09:57.990 the final decisions made and critically the specific 00:09:57.990 --> 00:10:00.730 action items, who needs to do what and by when. 00:10:01.190 --> 00:10:03.990 Different goals, different rules. So wrapping 00:10:03.990 --> 00:10:06.990 this up. What's the big takeaway here? This deep 00:10:06.990 --> 00:10:09.610 dive really highlights a shift, doesn't it? Moving 00:10:09.610 --> 00:10:12.230 from constantly being in prompt writing mode 00:10:12.230 --> 00:10:15.169 to more of a proactive ability building mindset. 00:10:15.490 --> 00:10:17.690 Yeah, exactly. Skills let you offload that heavy 00:10:17.690 --> 00:10:19.809 lifting of explaining the context and the rules 00:10:19.809 --> 00:10:22.470 every single time, which means your actual prompts 00:10:22.470 --> 00:10:24.850 can become much shorter, much clearer. They just 00:10:24.850 --> 00:10:26.409 trigger the detailed rule book you already built. 00:10:26.629 --> 00:10:29.590 You're essentially building this library of reusable 00:10:29.590 --> 00:10:32.990 systems based intelligence, making complex, often 00:10:32.990 --> 00:10:34.799 repetitive professional to work feel simpler 00:10:34.799 --> 00:10:36.840 and way more efficient. But it's not about stopping 00:10:36.840 --> 00:10:38.340 prompting altogether, right? That's important. 00:10:38.740 --> 00:10:40.799 You still need to tell the AI what the specific 00:10:40.799 --> 00:10:43.500 job is for today. You provide the topic, the 00:10:43.500 --> 00:10:46.320 raw input. But the how is already baked into 00:10:46.320 --> 00:10:48.899 this skill. So if you start building this library 00:10:48.899 --> 00:10:51.940 now, maybe just focus on creating one simple 00:10:51.940 --> 00:10:55.000 skill this week for a task you repeat. Often 00:10:55.000 --> 00:10:57.440 you'll likely be way ahead of the curve. ahead 00:10:57.440 --> 00:10:59.799 of others still grinding out those long, complex 00:10:59.799 --> 00:11:02.240 prompts from scratch every time. Yeah, absolutely. 00:11:02.759 --> 00:11:04.100 So maybe the final thought to leave you with 00:11:04.100 --> 00:11:07.200 is this. Think about one core part of your job 00:11:07.200 --> 00:11:09.700 right now, something that relies heavily on repetition 00:11:09.700 --> 00:11:12.240 and following specific rules. How quickly could 00:11:12.240 --> 00:11:14.299 you automate the compliance part of that task, 00:11:14.700 --> 00:11:17.360 perfectly, every time, with a single portable 00:11:17.360 --> 00:11:17.720 skill?