WEBVTT 00:00:00.000 --> 00:00:02.399 Typing the words build the thing into an AI feels 00:00:02.399 --> 00:00:05.759 like absolute magic, doesn't it? Oh, yeah, completely. 00:00:06.139 --> 00:00:08.460 But it's actually a trap. It is a massive trap. 00:00:08.599 --> 00:00:11.519 Honestly, it's the absolute fastest way to manufacture 00:00:11.519 --> 00:00:15.539 software bugs. And you do it with total, unwavering 00:00:15.539 --> 00:00:19.160 confidence. Welcome to the Deep Dive. We're unpacking 00:00:19.160 --> 00:00:22.039 something incredibly fundamental today. Beat. 00:00:22.539 --> 00:00:26.059 It is April 2026. Hard to believe, right? Right. 00:00:26.179 --> 00:00:28.620 And we are looking at a brilliant article by 00:00:28.620 --> 00:00:31.280 Max Ann today. Yeah, we're exploring the official 00:00:31.280 --> 00:00:34.380 developer workflow for Claude Code. This system 00:00:34.380 --> 00:00:36.520 was built by an engineer named Boris Cherny. 00:00:36.659 --> 00:00:39.039 We're really guiding you through a massive paradigm 00:00:39.039 --> 00:00:41.219 shift today. We are moving far away from the 00:00:41.219 --> 00:00:43.600 era of clever prompting. Exactly right. I mean, 00:00:43.600 --> 00:00:46.100 clever prompting is completely dead. Yeah. we're 00:00:46.100 --> 00:00:48.500 finally entering the era of architectural discipline 00:00:48.500 --> 00:00:51.740 and we'll walk you through five specific actionable 00:00:51.740 --> 00:00:55.240 steps today these steps turn your ai from a disposable 00:00:55.240 --> 00:00:58.240 chatbot into a custom operating system it's just 00:00:58.240 --> 00:01:00.340 a complete reframing of how you interact with 00:01:00.340 --> 00:01:03.039 the machine you aren't just giving it tasks anymore 00:01:03.039 --> 00:01:05.840 you're building an environment for it to succeed 00:01:05.840 --> 00:01:08.219 so let's talk about the blueprints you know you 00:01:08.219 --> 00:01:10.719 can't build a solid architecture without looking 00:01:10.719 --> 00:01:13.500 at them Right. And that's exactly where the first 00:01:13.500 --> 00:01:16.180 major failure happens. The biggest mistake developers 00:01:16.180 --> 00:01:18.719 make actually happens before they write a single 00:01:18.719 --> 00:01:21.560 line. Yeah. It's a trap we call reactive coding. 00:01:21.719 --> 00:01:23.799 And reactive coding feels like you're making 00:01:23.799 --> 00:01:26.180 real progress. You jump right into the terminal. 00:01:26.359 --> 00:01:28.859 You start generating code immediately. Because 00:01:28.859 --> 00:01:31.659 it's fun. Right. But you're really just creating 00:01:31.659 --> 00:01:34.019 a massive mess for yourself later. You end up 00:01:34.019 --> 00:01:37.459 reacting to a cascade of issues. You're not planning 00:01:37.459 --> 00:01:40.530 for them at all. Beat. That brings us to our 00:01:40.530 --> 00:01:43.829 first major shift. This is a big one. Yeah. Boris 00:01:43.829 --> 00:01:47.189 Cherny starts 80 % of his sessions in plan mode. 00:01:47.370 --> 00:01:49.650 80%. That number should really stop you in your 00:01:49.650 --> 00:01:51.790 tracks. Variously. He is the engineer who actually 00:01:51.790 --> 00:01:54.530 built the tool, and he spends almost all his 00:01:54.530 --> 00:01:57.810 time planning. To access this mode. You simply 00:01:57.810 --> 00:02:01.349 hit shift and tab twice. It forces the strategic 00:02:01.349 --> 00:02:04.129 thinking to happen before the actual building 00:02:04.129 --> 00:02:07.469 starts. Plan mode uses multi -model reasoning 00:02:07.469 --> 00:02:10.969 to actually verify your logic. Let's unpack that 00:02:10.969 --> 00:02:13.469 for a second because... Multi -model reasoning 00:02:13.469 --> 00:02:16.110 sounds incredibly dense. Oh, sure. It just means 00:02:16.110 --> 00:02:17.969 having different AI models check each other's 00:02:17.969 --> 00:02:20.669 work. Got it. One model drafts a plan, and another 00:02:20.669 --> 00:02:22.969 model tries to find the holes in it. So it aggressively 00:02:22.969 --> 00:02:25.509 stops you from rushing ahead blindly. Exactly. 00:02:25.509 --> 00:02:28.530 You use a very specific setup prompt here. You 00:02:28.530 --> 00:02:30.870 have Claude ask you about the fundamental problem 00:02:30.870 --> 00:02:33.629 you're solving. You meticulously define the intended 00:02:33.629 --> 00:02:37.129 user. You establish the specific success metrics. 00:02:37.449 --> 00:02:40.110 You figure out exactly what victory looks like. 00:02:40.229 --> 00:02:43.189 And critically, you define the anti -goals. You 00:02:43.189 --> 00:02:46.050 tell the AI what it should explicitly avoid doing. 00:02:46.490 --> 00:02:49.509 Two sec silence. I'm looking at this setup and, 00:02:49.550 --> 00:02:53.229 you know, I get setting goals. Right. But the 00:02:53.229 --> 00:02:57.129 intense focus on anti -goals feels slightly counterintuitive. 00:02:57.169 --> 00:02:59.550 Why is defining an anti -goal sometimes more 00:02:59.550 --> 00:03:01.930 important than defining the actual goal? Because 00:03:01.930 --> 00:03:03.969 it prevents confident hallucination. Which means 00:03:03.969 --> 00:03:08.050 the AI smoothly generating completely wrong but 00:03:08.050 --> 00:03:11.289 plausible answers. Exactly. Large language models 00:03:11.289 --> 00:03:13.610 are fundamentally people pleaser. Yeah, they 00:03:13.610 --> 00:03:15.650 really are. If you don't set strict negative 00:03:15.650 --> 00:03:18.530 boundaries, the AI just fills in the blanks. 00:03:18.530 --> 00:03:21.050 It guesses what you want and it usually guesses 00:03:21.050 --> 00:03:24.110 wrong. It creates code that looks perfect but 00:03:24.110 --> 00:03:27.909 fails structurally. Right. Think about assembling 00:03:27.909 --> 00:03:30.919 a complex piece of furniture. If you don't look 00:03:30.919 --> 00:03:32.780 at the instructions, you might build something 00:03:32.780 --> 00:03:35.300 resembling a chair. Sure. But the moment you 00:03:35.300 --> 00:03:38.319 sit on it, the entire thing collapses. The anti 00:03:38.319 --> 00:03:40.300 -goals are the instructions telling you which 00:03:40.300 --> 00:03:42.840 screws absolutely do not belong together. Right. 00:03:42.919 --> 00:03:45.520 Telling it what to avoid prevents expensive and 00:03:45.520 --> 00:03:48.080 frustrating wrong turns. You map out the entire 00:03:48.080 --> 00:03:50.159 minefield before you walk through it. It saves 00:03:50.159 --> 00:03:52.620 you hours of expensive refactoring later. So 00:03:52.620 --> 00:03:55.159 we've carefully mapped the logic. We have a solid 00:03:55.159 --> 00:03:58.099 plan. Now we need some basic ground rules for 00:03:58.099 --> 00:04:00.860 the environment. Beat. This leads us directly 00:04:00.860 --> 00:04:04.520 to the Claude .md file. Yeah, this is basically 00:04:04.520 --> 00:04:07.379 Claude's standing cheat sheet. The model reads 00:04:07.379 --> 00:04:09.900 this specific file at the start of every single 00:04:09.900 --> 00:04:12.099 chat. You might think you should give the AI 00:04:12.099 --> 00:04:15.280 a massive legal document, but the real secret 00:04:15.280 --> 00:04:19.410 here is ruthless minimalism. People tend to treat 00:04:19.410 --> 00:04:22.389 this file like a dusty storage unit. They hoard 00:04:22.389 --> 00:04:25.589 rules. They really do. They create this massive, 00:04:25.689 --> 00:04:29.009 overly anxious rulebook. They throw every single 00:04:29.009 --> 00:04:31.189 edge case they've ever encountered into it. And 00:04:31.189 --> 00:04:33.930 that directly causes instruction drift. Which 00:04:33.930 --> 00:04:36.449 is simply the AI getting lost in too many contradictory 00:04:36.449 --> 00:04:39.410 rules. The model only has so much attention to 00:04:39.410 --> 00:04:42.529 give. Right. If you feed it a 500 line markdown 00:04:42.529 --> 00:04:45.829 file of strict rules, the token waiting gets 00:04:45.829 --> 00:04:48.110 completely diluted. It loses focus. Exactly. 00:04:48.230 --> 00:04:50.329 It starts ignoring the core task because it's 00:04:50.329 --> 00:04:52.750 terrified of violating some obscure formatting 00:04:52.750 --> 00:04:54.709 rule from six months ago. The elite developer 00:04:54.709 --> 00:04:57.990 workflow in 2026 approaches this entirely differently. 00:04:58.209 --> 00:05:00.569 Yeah. You must keep this file under 100 lines. 00:05:00.750 --> 00:05:03.360 Under 100 lines. That requires incredible. discipline. 00:05:03.680 --> 00:05:06.339 Yeah, it does. It forces you to only include 00:05:06.339 --> 00:05:09.480 the absolute most critical project DNA. Two sec 00:05:09.480 --> 00:05:12.699 silence. I have to admit something vulnerable 00:05:12.699 --> 00:05:15.579 here. Oh. I still wrestle with prompt drift myself. 00:05:15.879 --> 00:05:19.259 Oh yeah. I get anxious when a model fails. Yeah. 00:05:19.339 --> 00:05:21.639 And I just start endlessly adding rules to the 00:05:21.639 --> 00:05:24.439 file. It's a really hard habit to break. We all 00:05:24.439 --> 00:05:27.180 do that. It is completely human nature. Right. 00:05:27.259 --> 00:05:29.399 We want to feel in control of the machine's output. 00:05:30.139 --> 00:05:32.560 When something breaks, our instinct is to build 00:05:32.560 --> 00:05:36.639 a taller fence. But in this new era, taller fences 00:05:36.639 --> 00:05:40.060 just block the AI's view. The article suggests 00:05:40.060 --> 00:05:42.639 a brilliant cleanup prompt tactic to fix this. 00:05:42.720 --> 00:05:46.040 You actively ask Claude to review the file. and 00:05:46.040 --> 00:05:48.879 delete outdated or conflicting information. Or 00:05:48.879 --> 00:05:51.480 you go nuclear and just burn it down completely. 00:05:51.699 --> 00:05:53.899 Just delete it. Yeah, you delete the entire file 00:05:53.899 --> 00:05:56.319 to remove legacy friction. But wait, if we just 00:05:56.319 --> 00:05:58.600 delete the file, aren't we losing the fundamental 00:05:58.600 --> 00:06:01.420 project DNA we've worked so hard on? Not at all. 00:06:01.560 --> 00:06:03.759 You have to remember that these underlying models 00:06:03.759 --> 00:06:06.540 improve constantly over time. Oh, true. What 00:06:06.540 --> 00:06:09.420 used to require strict, explicit rules is now 00:06:09.420 --> 00:06:11.759 handled entirely naturally by the model's base 00:06:11.759 --> 00:06:13.980 intelligence. The AI just understands standard 00:06:13.980 --> 00:06:16.490 best practices better now. Exactly. Storing old 00:06:16.490 --> 00:06:19.230 rules doesn't add clarity anymore. It just adds 00:06:19.230 --> 00:06:21.629 friction. It's like leaving training wheels on 00:06:21.629 --> 00:06:24.290 a motorcycle. It actively prevents the machine 00:06:24.290 --> 00:06:27.250 from leaning into the curves. So fewer rules 00:06:27.250 --> 00:06:30.310 actually force better AI alignment and less confusion. 00:06:30.610 --> 00:06:33.129 Spot on. You want a beautifully sharp tool, not 00:06:33.129 --> 00:06:35.750 a bloated corporate policy manual. Okay, let's 00:06:35.750 --> 00:06:37.410 look at where we are. We've planned the core 00:06:37.410 --> 00:06:40.050 code. We've burned down the rulebook and set 00:06:40.050 --> 00:06:44.189 lean boundaries. Now, how do you actually trust 00:06:44.189 --> 00:06:46.970 the output it gives you? You don't. You absolutely 00:06:46.970 --> 00:06:49.709 never blindly trust the output. You force the 00:06:49.709 --> 00:06:51.949 AI to check its own work. This is the concept 00:06:51.949 --> 00:06:55.050 of self -verifying loops. Yeah. Honestly. It's 00:06:55.050 --> 00:06:57.550 quietly the most powerful idea in this entire 00:06:57.550 --> 00:07:00.410 article. It is brilliant because it mimics senior 00:07:00.410 --> 00:07:03.250 engineering behavior. You give Claude a tangible 00:07:03.250 --> 00:07:05.930 tool to inspect its own output. You give it direct 00:07:05.930 --> 00:07:08.129 access to a browser or a test runner. Right. 00:07:08.189 --> 00:07:10.089 You don't just say, write this function. You 00:07:10.089 --> 00:07:12.350 say, write this function, run the test suite, 00:07:12.529 --> 00:07:15.029 read the error logs, and fix it until it passes. 00:07:15.290 --> 00:07:18.009 And this seemingly simple feedback loop creates 00:07:18.009 --> 00:07:22.050 a massive 3x lift in output quality. A three 00:07:22.050 --> 00:07:25.769 times quality improvement. It fundamentally shifts 00:07:25.769 --> 00:07:28.850 the AI from being a passive writer to an active 00:07:28.850 --> 00:07:32.829 participant. Whoa. Just imagine the AI catching 00:07:32.829 --> 00:07:35.670 its own bugs in the dark before it even reaches 00:07:35.670 --> 00:07:37.949 your terminal. That's genuinely incredible. It 00:07:37.949 --> 00:07:39.709 completely changes your relationship with the 00:07:39.709 --> 00:07:41.990 machine. Yeah. It's the difference between reviewing 00:07:41.990 --> 00:07:45.209 a messy first draft and approving a polished 00:07:45.209 --> 00:07:48.009 final product. You implement this by using a 00:07:48.009 --> 00:07:50.660 reusable cleanup prompt. You explicitly tell 00:07:50.660 --> 00:07:53.060 Claude to recheck all the work it just completed. 00:07:53.199 --> 00:07:55.439 You have it independently ensure that all best 00:07:55.439 --> 00:07:58.220 practices were followed. You force it to confirm 00:07:58.220 --> 00:08:00.980 that no weird structural complications were introduced. 00:08:01.220 --> 00:08:03.319 I'm thinking about edge cases here, though. Okay. 00:08:03.500 --> 00:08:05.319 What if you're working on something abstract 00:08:05.319 --> 00:08:08.079 where there isn't a simple test runner? or a 00:08:08.079 --> 00:08:10.779 clear visual layout to inspect. That is actually 00:08:10.779 --> 00:08:14.480 a massive glaring red flag. Really? Yeah. If 00:08:14.480 --> 00:08:16.680 there is no clear way to mechanically check the 00:08:16.680 --> 00:08:18.939 result, you failed step one. You mean the plan 00:08:18.939 --> 00:08:21.959 mode phase? Exactly. It means the initial task 00:08:21.959 --> 00:08:24.620 was not framed clearly enough. Oh, I see. If 00:08:24.620 --> 00:08:27.779 success cannot be measured or verified by a tool, 00:08:27.959 --> 00:08:30.639 then the AI is just guessing. And we already 00:08:30.639 --> 00:08:33.799 know how dangerous guessing is. Right. If the 00:08:33.799 --> 00:08:36.419 AI can't check its work... Your initial goal 00:08:36.419 --> 00:08:39.659 wasn't clearly defined. Precisely. Verification 00:08:39.659 --> 00:08:42.539 has to be baked into the actual setup. It cannot 00:08:42.539 --> 00:08:46.179 be an afterthought you tack on at the end. Sponsor. 00:08:46.639 --> 00:08:50.120 So your AI is now self -verifying its own beautiful 00:08:50.120 --> 00:08:52.960 work. Things are moving fast. But really fast. 00:08:53.179 --> 00:08:55.740 Because it's moving fast, you might be extremely 00:08:55.740 --> 00:08:58.100 tempted to run five sessions at once. Oh, sure. 00:08:58.299 --> 00:09:00.539 You want to speed things up even more. But here's 00:09:00.539 --> 00:09:02.830 exactly why you shouldn't do that. Running parallel 00:09:02.830 --> 00:09:05.909 sessions on the exact same file is a catastrophic 00:09:05.909 --> 00:09:08.950 mistake. You end up with what Boris calls context 00:09:08.950 --> 00:09:12.169 collision, which is simply AI confusion from 00:09:12.169 --> 00:09:14.830 multiple sessions editing identical files. Think 00:09:14.830 --> 00:09:16.350 of it this way. It's like handing two separate 00:09:16.350 --> 00:09:18.750 contractors the exact same half -finished kitchen 00:09:18.750 --> 00:09:21.289 remodel. Okay, that sounds bad already. You blindfold 00:09:21.289 --> 00:09:23.950 them both, and you do not tell either one what 00:09:23.950 --> 00:09:26.129 the other is doing. So one is installing cabinets 00:09:26.129 --> 00:09:28.029 while the other is tearing down the wall. Instead 00:09:28.029 --> 00:09:30.690 of accelerated progress, you just get overlapping 00:09:30.690 --> 00:09:33.559 chaos. Right, and you are the one who ends up 00:09:33.559 --> 00:09:36.100 untangling those contradictory code merges at 00:09:36.100 --> 00:09:40.259 11 p .m. Ouch. The AI models overwrite each other's 00:09:40.259 --> 00:09:42.620 logic because they don't share a real -time brain. 00:09:42.899 --> 00:09:46.940 The elegant solution here is to use fresh, independent 00:09:46.940 --> 00:09:50.120 sessions, but you strictly only use them for 00:09:50.120 --> 00:09:53.100 distinct silos. Maintaining clean context is 00:09:53.100 --> 00:09:56.340 a developer superpower. A clean, focused context 00:09:56.340 --> 00:09:59.669 window spots incredibly simple solutions. those 00:09:59.669 --> 00:10:02.409 solutions get completely lost in crowded noisy 00:10:02.409 --> 00:10:05.509 chats so you might dedicate one isolated session 00:10:05.509 --> 00:10:08.590 entirely to documentation you use another separate 00:10:08.590 --> 00:10:10.710 session for the core logic right and another 00:10:10.710 --> 00:10:13.409 strictly for edge cases you isolate the variables 00:10:13.409 --> 00:10:15.789 but wait where exactly is the line between a 00:10:15.789 --> 00:10:18.450 distinct silo and just another part of the exact 00:10:18.450 --> 00:10:21.059 same task It all comes down to the final product 00:10:21.059 --> 00:10:23.559 destination. If the two sessions are trying to 00:10:23.559 --> 00:10:26.740 edit the exact same code block, that is a collision 00:10:26.740 --> 00:10:29.480 waiting to happen. Makes sense. But if one session 00:10:29.480 --> 00:10:32.220 is writing the core script and the other is writing 00:10:32.220 --> 00:10:35.379 a separate readme file, that is a perfectly distinct 00:10:35.379 --> 00:10:38.100 silo. Keep their jobs completely separate and 00:10:38.100 --> 00:10:40.139 they won't step on each other's toes. Divide 00:10:40.139 --> 00:10:43.419 and conquer. But keep the territories completely 00:10:43.419 --> 00:10:46.559 segregated. That brings us beautifully to our 00:10:46.559 --> 00:10:50.389 final major shift. This is the ultimate overarching 00:10:50.389 --> 00:10:54.549 goal of the 2026 workflow. We are finally moving 00:10:54.549 --> 00:10:58.480 from a manual day -to -day grind. to long -term 00:10:58.480 --> 00:11:01.340 skill -based leverage. Most people still use 00:11:01.340 --> 00:11:04.700 AI as a disposable conversational chatbot. They 00:11:04.700 --> 00:11:06.639 write a long prompt, they get a decent answer, 00:11:06.740 --> 00:11:09.120 they close the window. Yeah. And they start completely 00:11:09.120 --> 00:11:11.440 over from scratch tomorrow. It keeps you eternally 00:11:11.440 --> 00:11:13.480 stuck in manual mode. You're just pulling levers 00:11:13.480 --> 00:11:15.799 every single day. Exactly. The real paradigm 00:11:15.799 --> 00:11:18.679 shift is turning your repeat work into a repeatable 00:11:18.679 --> 00:11:22.009 skill. Think about basketball. A prompt is telling 00:11:22.009 --> 00:11:24.149 a player exactly how to dribble the ball down 00:11:24.149 --> 00:11:26.970 the court every single time. A skill is just 00:11:26.970 --> 00:11:29.289 calling a designated, heavily practiced play 00:11:29.289 --> 00:11:31.529 from the sideline. That is a brilliant way to 00:11:31.529 --> 00:11:34.409 picture it. You define the complex skill once 00:11:34.409 --> 00:11:38.029 and you securely store it. And then the AI fundamentally 00:11:38.029 --> 00:11:41.230 inherits your specific project DNA every single 00:11:41.230 --> 00:11:43.970 time it wakes up. Wow. It instantly knows exactly 00:11:43.970 --> 00:11:46.710 how we do things around here. It knows the coding 00:11:46.710 --> 00:11:49.129 standards, the formatting quirks, the preferred 00:11:49.129 --> 00:11:51.190 libraries. You're building infrastructure, not 00:11:51.190 --> 00:11:53.870 just sending text messages. Beat. But you also 00:11:53.870 --> 00:11:56.570 never want to bet against the model itself. Never. 00:11:56.610 --> 00:11:58.750 Betting against the model's trajectory is the 00:11:58.750 --> 00:12:01.049 worst strategic mistake you can make right now. 00:12:01.419 --> 00:12:04.480 The article stresses this heavily. Do not spend 00:12:04.480 --> 00:12:08.139 countless hours optimizing tiny, fragile prompt 00:12:08.139 --> 00:12:10.700 tricks for today's specific limitations. Because 00:12:10.700 --> 00:12:14.200 underlying AI models are improving at a staggering 00:12:14.200 --> 00:12:16.460 rate right now. Those little quirks and limitations 00:12:16.460 --> 00:12:19.019 you're hacking around, they will simply vanish 00:12:19.019 --> 00:12:21.860 in the very next model update. You should invest 00:12:21.860 --> 00:12:24.100 your energy in durable systems, better context 00:12:24.100 --> 00:12:26.799 management, and tighter iteration loops instead. 00:12:27.059 --> 00:12:28.500 You want to build architecture that actually 00:12:28.500 --> 00:12:31.799 survives the next major AI release. But if the 00:12:31.799 --> 00:12:34.779 underlying models are improving so rapidly, won't 00:12:34.779 --> 00:12:37.860 our documented skills become obsolete just as 00:12:37.860 --> 00:12:41.059 fast as our fragile prompts? Not if you build 00:12:41.059 --> 00:12:43.399 them correctly. Okay. Good systems and clear 00:12:43.399 --> 00:12:46.019 logical instructions actually scale beautifully 00:12:46.019 --> 00:12:48.000 with the model's intelligence. That makes sense. 00:12:48.100 --> 00:12:50.320 When the model gets smarter, your system just 00:12:50.320 --> 00:12:53.299 runs faster and makes fewer errors. Clever prompt 00:12:53.299 --> 00:12:55.200 hacks, on the other hand, just break immediately 00:12:55.200 --> 00:12:57.860 when the model updates. Build structured workflows 00:12:57.860 --> 00:13:00.960 that scale rather than relying on fragile, temporary 00:13:00.960 --> 00:13:03.740 prompt hacks. Yes, you are literally building 00:13:03.740 --> 00:13:06.159 an operating system. You are not compiling a 00:13:06.159 --> 00:13:08.440 collection of magic spells. Let's synthesize 00:13:08.440 --> 00:13:10.879 this entire journey for a moment. Mastering clawed 00:13:10.879 --> 00:13:13.200 code isn't about being a wizard engineer with 00:13:13.200 --> 00:13:15.210 secret knowledge. It's really not about having 00:13:15.210 --> 00:13:18.169 the secret magical list of prompts saved in a 00:13:18.169 --> 00:13:20.350 document. It is entirely about architectural 00:13:20.350 --> 00:13:23.570 discipline. You absolutely must plan the logic 00:13:23.570 --> 00:13:26.049 before you start building. Keep your standing 00:13:26.049 --> 00:13:29.350 rules incredibly lean and minimal. Verify all 00:13:29.350 --> 00:13:31.450 of your outputs with automated self -checking 00:13:31.450 --> 00:13:34.250 loops. Silo your parallel tasks strictly to avoid 00:13:34.250 --> 00:13:37.929 nasty context collisions. And finally, systemize 00:13:37.929 --> 00:13:40.350 your repeat manual work into highly durable skills. 00:13:40.750 --> 00:13:44.299 Less wizard. more operating system. It is remarkably 00:13:44.299 --> 00:13:47.000 simple. It isn't flashy at all, but it actually 00:13:47.000 --> 00:13:50.519 works reliably at scale. So here is your actionable 00:13:50.519 --> 00:13:53.100 call to action for today. Go into your project 00:13:53.100 --> 00:13:56.919 and delete your bloated Cleland E .MD file right 00:13:56.919 --> 00:13:59.559 now. Seriously, just burn it down. Start completely 00:13:59.559 --> 00:14:01.940 fresh. And try starting your very next development 00:14:01.940 --> 00:14:04.720 project in plan mode. Let the models argue about 00:14:04.720 --> 00:14:07.409 your logic before you build. Two sec silence. 00:14:07.909 --> 00:14:10.169 I want to leave you with one final, slightly 00:14:10.169 --> 00:14:12.570 provocative thought today. Something to really 00:14:12.570 --> 00:14:14.870 chew on while you work. If the gold standard 00:14:14.870 --> 00:14:17.590 right now is manually systemizing your repeat 00:14:17.590 --> 00:14:20.769 work into these stored skills, or wait, sorry, 00:14:20.889 --> 00:14:22.830 into these stored skills, what happens in the 00:14:22.830 --> 00:14:25.269 very near future? What happens when the AI starts 00:14:25.269 --> 00:14:27.590 silently analyzing your daily terminal behavior 00:14:27.590 --> 00:14:30.330 to suggest and build its own skills without you 00:14:30.330 --> 00:14:32.769 even asking? Oh, it is absolutely coming. The 00:14:32.769 --> 00:14:34.870 machine will inevitably learn exactly how you 00:14:34.870 --> 00:14:37.059 work just by watching you. We started today by 00:14:37.059 --> 00:14:39.659 saying build the thing is a dangerous trap. Someday, 00:14:39.679 --> 00:14:42.360 very soon, the AI might already know exactly 00:14:42.360 --> 00:14:44.379 how to build the thing before you even type a 00:14:44.379 --> 00:14:46.419 single word. OTRO music.