WEBVTT 00:00:00.000 --> 00:00:02.980 Welcome to the deep dive today. We're tackling 00:00:02.980 --> 00:00:06.500 something pretty fundamental I think is the skill 00:00:06.500 --> 00:00:08.900 of well writing code actually being replaced 00:00:08.900 --> 00:00:12.000 by just writing a clear email Or something like 00:00:12.000 --> 00:00:14.199 it. It's that feeling right that creative friction. 00:00:14.199 --> 00:00:17.359 You get this this brilliant idea for an app Maybe 00:00:17.359 --> 00:00:19.500 it's simple. Maybe it's a whole business But 00:00:19.500 --> 00:00:22.039 then you think about learning JavaScript Python 00:00:22.039 --> 00:00:24.800 dealing with all the setup and it just stops 00:00:24.800 --> 00:00:27.379 you cold Yeah, that huge technical wall that 00:00:27.379 --> 00:00:30.589 barrier that used to mean months, maybe years 00:00:30.589 --> 00:00:32.689 of learning. Our source material today suggests 00:00:32.689 --> 00:00:36.509 that barrier. It's mostly gone thanks to generative 00:00:36.509 --> 00:00:39.509 AI. Exactly. And this big shift, people are starting 00:00:39.509 --> 00:00:42.210 to call it vibe coding. It's more about focusing 00:00:42.210 --> 00:00:44.490 on the feeling, the function you want, rather 00:00:44.490 --> 00:00:48.020 than getting lost in confusing code syntax. So 00:00:48.020 --> 00:00:50.240 our mission today is to really unpack how that 00:00:50.240 --> 00:00:52.340 works. We'll define vibe coding, look at the 00:00:52.340 --> 00:00:54.399 tools you need from simple chat interfaces right 00:00:54.399 --> 00:00:57.219 up to these powerful AI agents, and we'll break 00:00:57.219 --> 00:01:00.039 down the specific process, what the sources are 00:01:00.039 --> 00:01:02.020 calling the golden prompt. Yeah, and we'll dive 00:01:02.020 --> 00:01:04.920 into the money side too. New models this opens 00:01:04.920 --> 00:01:08.540 up. And maybe most importantly, we'll hit those 00:01:08.540 --> 00:01:11.000 crucial reality checks, things you really need 00:01:11.000 --> 00:01:13.299 to know before you jump in. Okay, let's unpack 00:01:13.299 --> 00:01:17.879 this idea then. Vibe coding. Before AI, Building 00:01:17.879 --> 00:01:20.739 software felt, I don't know, like building a 00:01:20.739 --> 00:01:23.120 house all by yourself. You're mixing the cement, 00:01:23.359 --> 00:01:25.299 laying every single brick, checking the plumbing, 00:01:25.459 --> 00:01:27.640 and if you mess up one tiny detail, the whole 00:01:27.640 --> 00:01:30.129 thing could be shaky. Totally. Vibe coding flips 00:01:30.129 --> 00:01:32.049 that role completely. You're not the one laying 00:01:32.049 --> 00:01:34.209 bricks anymore, you're the architect, the visionary. 00:01:34.310 --> 00:01:36.549 You still have control, absolutely, but you're 00:01:36.549 --> 00:01:39.709 builders. They're this super fast, hyper -efficient 00:01:39.709 --> 00:01:42.269 AI team working around the clock. So you just 00:01:42.269 --> 00:01:44.290 tell the team, okay, I need a wall here, big 00:01:44.290 --> 00:01:47.230 window, facing west for the sunset, and the AI 00:01:47.230 --> 00:01:49.650 handles the framing, the materials, putting the 00:01:49.650 --> 00:01:51.650 glass, and all that detail. Your focus is just 00:01:51.650 --> 00:01:53.790 on the final vision. And what's really fascinating 00:01:53.790 --> 00:01:56.390 is why this works now. It's not just slightly 00:01:56.390 --> 00:02:00.870 better code. The tech behind tools like ChatGPT, 00:02:01.109 --> 00:02:02.969 it's gotten incredibly good at understanding 00:02:02.969 --> 00:02:06.890 context, exponentially better. A few years back, 00:02:07.329 --> 00:02:09.729 AI could maybe write tiny bits of code, often 00:02:09.729 --> 00:02:11.889 buggy, isolated snippets. Right, snippets that 00:02:11.889 --> 00:02:13.750 didn't know about anything else, yeah, no state. 00:02:14.150 --> 00:02:17.550 Precisely. Now these models can handle huge complex 00:02:17.550 --> 00:02:21.030 instructions. Think like a whole floor plan, 00:02:21.229 --> 00:02:23.800 not just a door frame. And crucially, they remember 00:02:23.800 --> 00:02:26.080 context. They know what was built last week, 00:02:26.159 --> 00:02:28.300 how the new roof needs to connect. This means 00:02:28.300 --> 00:02:30.539 they manage complex dependencies, keep track 00:02:30.539 --> 00:02:33.539 of the app's state across lots of requests. Ah, 00:02:33.599 --> 00:02:36.280 that deep context memory. That's the key, then, 00:02:36.360 --> 00:02:38.819 to speeding things up so much from months down 00:02:38.819 --> 00:02:40.800 to days. It really compresses the whole creation 00:02:40.800 --> 00:02:43.659 time scale dramatically. Let's people try ideas 00:02:43.659 --> 00:02:45.639 that were just too expensive or time -consuming 00:02:45.639 --> 00:02:48.150 before. Right. Lots more experimentation is possible 00:02:48.150 --> 00:02:51.289 now. OK, so if the AI is the builder doing the 00:02:51.289 --> 00:02:54.150 heavy lifting, what's the absolute first thing 00:02:54.150 --> 00:02:57.050 the architect, the human, has to nail down? The 00:02:57.050 --> 00:02:59.370 vision. It has to be crystal clear before you 00:02:59.370 --> 00:03:01.509 even touch a tool. Leaving from that vision, 00:03:01.590 --> 00:03:04.550 let's talk tools. For someone just starting out, 00:03:04.750 --> 00:03:08.090 just curious, the sources say, start simple. 00:03:08.460 --> 00:03:10.960 Yeah, stick with the standard chat tools first, 00:03:11.319 --> 00:03:14.719 like your basic chat GPT or Claude. They're great 00:03:14.719 --> 00:03:17.120 for, say, simple calculators or getting the code 00:03:17.120 --> 00:03:19.680 for a basic one -page website, even simple text 00:03:19.680 --> 00:03:22.460 games. You can copy the HTML and CSS they give 00:03:22.460 --> 00:03:24.520 you, open it straight in your browser. It feels 00:03:24.520 --> 00:03:27.020 kind of magical. But there's a clear limit, right? 00:03:27.159 --> 00:03:29.900 That code is static. If you need users to log 00:03:29.900 --> 00:03:32.539 in or you need to store data permanently or connect 00:03:32.539 --> 00:03:35.699 to some financial API... These basic tools fall 00:03:35.699 --> 00:03:37.960 short. Exactly. They can write the blueprint, 00:03:38.280 --> 00:03:40.580 but they can't run the factory, as you put it. 00:03:40.780 --> 00:03:43.300 For building real, dynamic products, apps that 00:03:43.300 --> 00:03:46.280 handle user data, you got to step up to AI agents. 00:03:46.340 --> 00:03:48.680 AI agents, OK. Which raises the question, what's 00:03:48.680 --> 00:03:50.840 the real difference? Well, agents don't just 00:03:50.840 --> 00:03:53.719 write code. They also run it. They execute commands, 00:03:53.979 --> 00:03:56.099 fix errors automatically, and handle the tricky 00:03:56.099 --> 00:03:58.430 stuff of putting it online. So an agent isn't 00:03:58.430 --> 00:04:01.210 just spitting out text. It's like software using 00:04:01.210 --> 00:04:04.669 other software, a mini CEO for the build. That's 00:04:04.669 --> 00:04:06.469 a good way to think about it. Yeah. It has an 00:04:06.469 --> 00:04:09.729 execution environment. It can use APIs, manage 00:04:09.729 --> 00:04:12.330 secret keys, handle those compressed configuration 00:04:12.330 --> 00:04:15.110 files. All the stuff that used to trip up beginners 00:04:15.110 --> 00:04:17.550 constantly, it abstracts away the infrastructure 00:04:17.550 --> 00:04:19.810 pain. Which lowers the barrier to entry, but 00:04:19.810 --> 00:04:22.670 also raises the ceiling of what you can actually 00:04:22.670 --> 00:04:24.949 build. Definitely. And we're seeing platforms 00:04:24.949 --> 00:04:27.089 emerge to help with this. Replit's a big one. 00:04:27.470 --> 00:04:29.449 Its agent feature handles tough things like setting 00:04:29.449 --> 00:04:32.569 up databases, environment stuff. Bolt .new, as 00:04:32.569 --> 00:04:34.129 mentioned, is great for web apps, runs right 00:04:34.129 --> 00:04:37.009 in the browser, super fast prototyping. And lovable 00:04:37.009 --> 00:04:39.230 gets a shout out for making things look really 00:04:39.230 --> 00:04:41.769 modern and polished, aesthetically pleasing. 00:04:42.310 --> 00:04:45.089 without needing design skills. What's the biggest 00:04:45.089 --> 00:04:47.430 sort of inherent difficulty when you move from 00:04:47.430 --> 00:04:50.170 just chatting with an AI to using one of these 00:04:50.170 --> 00:04:53.490 more sophisticated agents? Agents handle complex 00:04:53.490 --> 00:04:55.629 operational tasks, like actually running and 00:04:55.629 --> 00:04:58.209 debugging code in real time. OK, let's get into 00:04:58.209 --> 00:05:01.230 the core process. The sources lay it out in stages, 00:05:01.449 --> 00:05:04.750 phase one. The blueprint. They're really strict 00:05:04.750 --> 00:05:06.470 about this. Before you write a single prompt, 00:05:06.750 --> 00:05:08.930 spend 10 minutes just writing down exactly what 00:05:08.930 --> 00:05:10.970 the app should do. Yeah. because most people 00:05:10.970 --> 00:05:13.649 fail by being vague. They just type, make a water 00:05:13.649 --> 00:05:16.829 app. That's like a guaranteed way to get prompt 00:05:16.829 --> 00:05:19.410 drift where the AI just wanders off. You need 00:05:19.410 --> 00:05:22.069 absolute clarity. Take their hydro habits tracker 00:05:22.069 --> 00:05:25.670 example to find the audience. Desk workers forgetting 00:05:25.670 --> 00:05:28.870 water. The one core function, clicking a button 00:05:28.870 --> 00:05:31.910 to add water intake. Nice to haves, like sharing 00:05:31.910 --> 00:05:34.430 with friends. Clearly mark those for later. Right. 00:05:34.589 --> 00:05:36.750 Nail the core first. Then, once that blueprint's 00:05:36.750 --> 00:05:38.589 solid, you write the instruction, the golden 00:05:38.589 --> 00:05:41.290 prompt. And there's a formula they suggest. Role, 00:05:41.350 --> 00:05:43.569 goal, features, and rules. It forces clarity. 00:05:43.889 --> 00:05:45.930 For HydroHabits, the role isn't just developer. 00:05:46.310 --> 00:05:49.029 It's expert web developer, specialized in mobile 00:05:49.029 --> 00:05:51.550 -first, high -conversion apps. Super specific. 00:05:51.850 --> 00:05:54.829 The goal. Build a water intake tracker. Three 00:05:54.829 --> 00:05:58.139 liter, daily target. No ambiguity. And the features 00:05:58.139 --> 00:06:01.040 get really specific too, almost visual. Simple 00:06:01.040 --> 00:06:03.740 login, circular progress bar, quick add buttons 00:06:03.740 --> 00:06:07.360 for 250 milliliter, 500 milliliters, 1L, a street 00:06:07.360 --> 00:06:10.000 counter, and that satisfying confetti animation 00:06:10.000 --> 00:06:13.279 when you hit the goal, details matter. And rules 00:06:13.279 --> 00:06:15.860 and style are just as vital. Use light blues 00:06:15.860 --> 00:06:18.420 white, keep it clean, minimal, make sure it works 00:06:18.420 --> 00:06:21.769 perfectly on a phone, stop the AI guessing. I 00:06:21.769 --> 00:06:23.449 have to admit, I still wrestle with prompt drift 00:06:23.449 --> 00:06:25.610 myself sometimes. You know, you start clear. 00:06:25.990 --> 00:06:28.389 But after a few changes, the AI totally forgets 00:06:28.389 --> 00:06:30.569 the original point and starts building, well, 00:06:30.790 --> 00:06:33.230 something else entirely. This level of detail 00:06:33.230 --> 00:06:36.089 up front feels absolutely necessary now. It really 00:06:36.089 --> 00:06:38.149 is. It's like the contract with the AI. Yeah. 00:06:38.250 --> 00:06:41.230 And it'll build that first version fast. Sources 00:06:41.230 --> 00:06:43.649 say it's usually about 80 % right. Maybe the 00:06:43.649 --> 00:06:45.750 buttons are too small, colors slightly off, that 00:06:45.750 --> 00:06:47.029 kind of thing. Which leads straight into phase 00:06:47.029 --> 00:06:49.490 three. This is the vibe part, right? You talk 00:06:49.490 --> 00:06:51.410 to the agent like a human assistant molding it. 00:06:51.529 --> 00:06:53.629 You say, hey, those buttons, too small, make 00:06:53.629 --> 00:06:56.189 them 50 % bigger. Or that blue's too harsh, try 00:06:56.189 --> 00:06:58.889 a softer sky blue. Exactly. It's a conversation. 00:06:59.069 --> 00:07:01.149 You keep going back and forth, refining details, 00:07:01.389 --> 00:07:03.389 shaping the app until you're happy. But isn't 00:07:03.389 --> 00:07:05.949 that last 20 %? Isn't that where the real complexity 00:07:05.949 --> 00:07:09.269 lies? If I still meet a human to fix tricky bugs 00:07:09.269 --> 00:07:12.329 or device issues, have we really removed the 00:07:12.329 --> 00:07:14.829 barrier or just shifted it? That's a critical 00:07:14.829 --> 00:07:17.439 point. The barrier's been massively low. From 00:07:17.439 --> 00:07:20.699 100 % build down to maybe the final 20 % polish. 00:07:21.100 --> 00:07:24.240 The sources are clear. AI gets you 0 to 80 % 00:07:24.240 --> 00:07:26.720 minutes. That's the revolution. But yeah, the 00:07:26.720 --> 00:07:28.459 human architect still needs to guide that last 00:07:28.459 --> 00:07:31.079 mile, refine it. And when it is ready, going 00:07:31.079 --> 00:07:34.779 live seems almost anti -climactic. These modern 00:07:34.779 --> 00:07:36.959 tools, Replet, Bolt, they just have a deploy 00:07:36.959 --> 00:07:39.459 button. Pretty much. Click it. wait a bit while 00:07:39.459 --> 00:07:41.720 the agent sorts out server stuff, and boom, you 00:07:41.720 --> 00:07:44.300 get a real URL to share. A live product, just 00:07:44.300 --> 00:07:46.500 like that. Just like that. OK, but how often 00:07:46.500 --> 00:07:48.980 does the AI just completely whiff it on the first 00:07:48.980 --> 00:07:50.939 try? Like, totally miss the mark, forcing you 00:07:50.939 --> 00:07:52.800 to start over. It's usually about 80 % right. 00:07:53.639 --> 00:07:56.819 That last 20 % needs that targeted human refinement. 00:07:56.959 --> 00:08:00.939 Yeah. Sponsor reads. OK, so because you can build 00:08:00.939 --> 00:08:03.540 things, maybe 10 times faster, prototype quickly. 00:08:04.360 --> 00:08:06.519 You just, you get more shots on goal, right? 00:08:06.560 --> 00:08:08.620 More chances to succeed. And this changes the 00:08:08.620 --> 00:08:11.100 financial models available. Absolutely. The big 00:08:11.100 --> 00:08:13.920 one this speed enables is the Microsoft model. 00:08:14.519 --> 00:08:16.980 Sauce being software as a service. A Microsoft 00:08:16.980 --> 00:08:20.699 is just a small, really niche app. Solves one 00:08:20.699 --> 00:08:23.410 tiny problem for a very specific group. Like 00:08:23.410 --> 00:08:26.129 a special invoice generator just for freelance 00:08:26.129 --> 00:08:27.790 writers where that meeting note cleaner turning 00:08:27.790 --> 00:08:30.430 messy text into a PDF. Exactly. Things like that. 00:08:30.470 --> 00:08:32.649 And you might only charge five, maybe $10 a month. 00:08:32.730 --> 00:08:34.990 You don't need a million users. If you find just 00:08:34.990 --> 00:08:38.429 100 people paying $5 a month, that's $500 monthly 00:08:38.429 --> 00:08:40.690 recurring revenue. for something you potentially 00:08:40.690 --> 00:08:43.429 built in a weekend. Whoa. Okay, imagine scaling 00:08:43.429 --> 00:08:47.669 a really solid Microsoft built that fast to hundreds, 00:08:47.870 --> 00:08:50.269 maybe thousands of users. That whole rapid iteration 00:08:50.269 --> 00:08:52.070 and growth potential, that feels fundamentally 00:08:52.070 --> 00:08:55.289 new. Very exciting. Yeah. It is. The sources 00:08:55.289 --> 00:08:57.049 point to a couple of other paths too. One is 00:08:57.049 --> 00:09:00.090 freelancing. You take someone else's idea, business 00:09:00.090 --> 00:09:02.090 owners maybe, run it through your agent tools. 00:09:02.250 --> 00:09:04.090 You charge a premium for that result. Basically, 00:09:04.210 --> 00:09:05.649 you're selling your prompting skill. And the 00:09:05.649 --> 00:09:07.490 third one, which I think gets overlooked, is 00:09:07.490 --> 00:09:09.970 just saving money or time at your existing job. 00:09:10.250 --> 00:09:13.190 Building small internal tools. Automating some 00:09:13.190 --> 00:09:15.929 repetitive tasks like that PDF cleaner example. 00:09:16.309 --> 00:09:19.009 Maybe it saves your team two hours a week. That's 00:09:19.009 --> 00:09:21.309 real value back to the company, even if you never 00:09:21.309 --> 00:09:23.669 sell the tool itself. Definitely. Building those 00:09:23.669 --> 00:09:25.850 internal efficiencies, often the quickest way 00:09:25.850 --> 00:09:28.470 to show immediate value with these tools in a 00:09:28.470 --> 00:09:30.929 workplace. OK, this all sounds amazing, like 00:09:30.929 --> 00:09:34.610 pure gold, really, a tech dream. But every revolution 00:09:34.610 --> 00:09:36.690 has its fine print, doesn't it? The caution tape. 00:09:36.990 --> 00:09:39.570 Yeah. Before anyone rushes off to build their 00:09:39.570 --> 00:09:42.490 Microsoft empire, we need those reality checks 00:09:42.490 --> 00:09:44.690 from the sources. Right. Reality check number 00:09:44.690 --> 00:09:47.690 one, the last 10 % is hard. We touched on this. 00:09:48.049 --> 00:09:51.929 AI gets you, say, 90 % there super fast. But 00:09:51.929 --> 00:09:54.970 that last 10 % Specific bugs making it work on 00:09:54.970 --> 00:09:58.149 all devices, weird edge cases, that takes human 00:09:58.149 --> 00:10:00.450 patience. Sometimes serious debugging. Yeah, 00:10:00.470 --> 00:10:02.649 you might spend an hour chasing one line of code 00:10:02.649 --> 00:10:05.690 the AI wrote, even if the 999 lines before it 00:10:05.690 --> 00:10:08.230 were perfect. Sometimes you just have to delete 00:10:08.230 --> 00:10:10.169 the last change and explain the problem differently 00:10:10.169 --> 00:10:12.450 to the agent, right, to kind of unstick it. Exactly. 00:10:12.710 --> 00:10:16.840 Reality check two. Security. This is huge. It's 00:10:16.840 --> 00:10:19.440 explicitly your responsibility. The AI wants 00:10:19.440 --> 00:10:22.899 to help you to complete the task. It's not automatically 00:10:22.899 --> 00:10:25.779 building like Fort Knox for your logins. So if 00:10:25.779 --> 00:10:27.600 you're handling sensitive stuff, credit cards, 00:10:27.879 --> 00:10:29.759 personal info, you have to be incredibly careful. 00:10:30.059 --> 00:10:32.899 For payments, the advice is clear. Use trusted 00:10:32.899 --> 00:10:35.519 external tools. Stripe, lemon squeezy, things 00:10:35.519 --> 00:10:37.519 like that. Don't try to build your own payment 00:10:37.519 --> 00:10:40.399 system from scratch with AI. Just don't. Way 00:10:40.399 --> 00:10:42.899 too risky. Way too risky. And reality check three, 00:10:43.259 --> 00:10:46.330 cost. Building complex stuff, especially with 00:10:46.330 --> 00:10:48.750 powerful agents using lots of computation that 00:10:48.750 --> 00:10:51.649 uses credits or tokens. It costs money. Learning 00:10:51.649 --> 00:10:53.950 in simple prototypes, usually cheap or even free 00:10:53.950 --> 00:10:56.389 to start. But if your app gets high traffic or 00:10:56.389 --> 00:10:58.809 it's really complex, your bill will go up. Always 00:10:58.809 --> 00:11:01.220 check the pricing pages first. When should someone 00:11:01.220 --> 00:11:03.399 who's stuck in that iteration phase going back 00:11:03.399 --> 00:11:05.799 and forth just give up on that thread and start 00:11:05.799 --> 00:11:08.700 fresh with a new golden prompt? If the AI sees 00:11:08.700 --> 00:11:11.600 you stuck in a loop on one specific bug, like 00:11:11.600 --> 00:11:13.679 the same air keeps popping up no matter how you 00:11:13.679 --> 00:11:16.379 explain the fix. Maybe step back, try explaining 00:11:16.379 --> 00:11:19.320 the fix differently, or simplify the goal for 00:11:19.320 --> 00:11:21.299 that step. Yeah. That often breaks this cycle. 00:11:21.700 --> 00:11:25.279 So stepping back, what does this all mean? Big 00:11:25.279 --> 00:11:27.960 picture. It feels like we're in a really special 00:11:27.960 --> 00:11:30.620 moment in time. That massive wall between just 00:11:30.620 --> 00:11:33.080 having an idea and actually holding it, using 00:11:33.080 --> 00:11:35.740 it, it's almost gone. It really has. You don't 00:11:35.740 --> 00:11:37.799 need permission anymore. In the same way, you 00:11:37.799 --> 00:11:39.700 don't necessarily need that computer science 00:11:39.700 --> 00:11:41.519 degree to start building real things, even things 00:11:41.519 --> 00:11:44.100 that make money. That's incredibly democratizing. 00:11:44.340 --> 00:11:46.919 The core skill seems to have shifted entirely. 00:11:47.200 --> 00:11:50.179 From execution, the sheer labor of typing code, 00:11:50.259 --> 00:11:53.360 to clarity, the ability to think clearly and 00:11:53.360 --> 00:11:55.720 write a precise, thoughtful prompt. That's the 00:11:55.720 --> 00:11:58.159 valuable skill now. Yeah. The new currency is 00:11:58.159 --> 00:12:02.720 just curiosity and a willingness to try things 00:12:02.720 --> 00:12:04.960 out. To stop seeing software as this arcane, 00:12:05.100 --> 00:12:08.080 complex thing and start thinking of it more like 00:12:08.080 --> 00:12:10.600 molding clay, but with language. My best advice, 00:12:10.820 --> 00:12:13.120 echoing the sources. Stop just reading about 00:12:13.120 --> 00:12:15.159 it. Go build something silly today. Seriously. 00:12:15.360 --> 00:12:17.360 Build a button that makes a funny noise. Build 00:12:17.360 --> 00:12:19.779 a page that just shows random cat photos. Just 00:12:19.779 --> 00:12:21.720 start vibing with the tools to see what happens. 00:12:21.799 --> 00:12:24.460 And think about this. If you can get functional 00:12:24.460 --> 00:12:27.899 software from just a text description, what really 00:12:27.899 --> 00:12:30.480 complex, maybe multidisciplinary problem, something 00:12:30.480 --> 00:12:33.960 totally out of reach a year ago, is now solvable 00:12:33.960 --> 00:12:36.799 by the average curious person. That is a challenging 00:12:36.799 --> 00:12:38.580 thought and a wonderful place to leave it.