WEBVTT 00:00:00.000 --> 00:00:04.219 Welcome to the deep dive. OK, let's talk about 00:00:04.219 --> 00:00:06.179 that feeling. You know the one, right? Yeah. 00:00:06.240 --> 00:00:09.359 That mix of hope and, let's be honest, crushing 00:00:09.359 --> 00:00:11.619 disappointment when you get some AI -generated 00:00:11.619 --> 00:00:14.119 code. Oh, yeah. Looks perfect on the screen. 00:00:14.279 --> 00:00:17.000 Exactly. Looks absolutely perfect. But then the 00:00:17.000 --> 00:00:20.260 second you try to actually integrate it, it just 00:00:20.260 --> 00:00:22.440 crumbles. It falls apart. You end up just tossing 00:00:22.440 --> 00:00:25.519 these disconnected prompts at the AI, hoping 00:00:25.519 --> 00:00:27.780 something sticks. People are calling it vibe 00:00:27.780 --> 00:00:31.839 coding. Yeah, vibe coding. It's kitschy. It is. 00:00:32.259 --> 00:00:34.340 But look, today we're digging into something 00:00:34.340 --> 00:00:36.899 that feels pretty revolutionary, an approach 00:00:36.899 --> 00:00:38.979 that could genuinely change that frustration 00:00:38.979 --> 00:00:42.280 into, well, maybe phenomenal productivity. And 00:00:42.280 --> 00:00:43.960 that change, that whole approach, it's called 00:00:43.960 --> 00:00:46.579 context engineering. And you know, it's not just 00:00:46.579 --> 00:00:48.259 another buzzword that's going to vanish next 00:00:48.259 --> 00:00:50.359 month. Right. We see enough of those. We do. 00:00:50.479 --> 00:00:54.070 This is a, it's a systematic way, really. a methodology 00:00:54.070 --> 00:00:57.189 to dramatically improve AI coding results, basically 00:00:57.189 --> 00:00:59.909 by giving the AI the right information, precisely. 00:01:00.149 --> 00:01:01.810 That feels like such a key difference, doesn't 00:01:01.810 --> 00:01:04.150 it? It shifts the focus. It's less about the 00:01:04.150 --> 00:01:07.510 AI guessing and more about us, the developers, 00:01:08.189 --> 00:01:10.870 actually providing the clarity it needs. It's 00:01:10.870 --> 00:01:14.049 almost like we become architects for the AI. 00:01:14.250 --> 00:01:16.230 That's a great way to put it. Information architects. 00:01:16.310 --> 00:01:19.680 Yeah. So our mission for this deep dive is to 00:01:19.680 --> 00:01:21.900 really unpack how this works. We've got some 00:01:21.900 --> 00:01:25.560 great sources, including bits from context engineering, 00:01:25.939 --> 00:01:28.200 the future of AI assisted development to help 00:01:28.200 --> 00:01:30.599 guide us through this shift. Let's unpack this 00:01:30.599 --> 00:01:33.879 then. What's the fundamental problem? Why are 00:01:33.879 --> 00:01:36.319 developers struggling so much with AI code right 00:01:36.319 --> 00:01:38.700 now? Why doesn't it just work more often? Well, 00:01:38.700 --> 00:01:41.340 I think it really boils down to trust or maybe 00:01:41.340 --> 00:01:43.459 the lack of it. There was this recent study, 00:01:43.659 --> 00:01:45.640 Quoto did it, and it found something pretty alarming. 00:01:45.790 --> 00:01:49.409 Oh, what was that? 76 .4 % of developers just 00:01:49.409 --> 00:01:52.129 don't trust AI -generated code, not without a 00:01:52.129 --> 00:01:54.209 human looking it over carefully. Wow, over three 00:01:54.209 --> 00:01:56.590 quarters. That's huge. It is. And look, the core 00:01:56.590 --> 00:01:58.769 problem isn't really the AI model itself. These 00:01:58.769 --> 00:02:01.750 LLMs, they're incredibly smart, like brilliant 00:02:01.750 --> 00:02:04.629 interns. They have this vast encyclopedic knowledge, 00:02:04.989 --> 00:02:07.489 but they have zero practical experience with 00:02:07.489 --> 00:02:10.770 your project, your specific setup. That's a fantastic 00:02:10.770 --> 00:02:13.780 analogy. Brilliant interns, loads of theory, 00:02:14.039 --> 00:02:16.400 no clue about this actual job site. Exactly. 00:02:16.620 --> 00:02:20.360 They know general patterns, sure. But your project's 00:02:20.360 --> 00:02:22.819 architecture, the libraries you picked, your 00:02:22.819 --> 00:02:25.419 team's coding style, the business reason for 00:02:25.419 --> 00:02:27.080 the feature, they know none of that. Right. So 00:02:27.080 --> 00:02:28.979 that's the gap. So if they're these brilliant 00:02:28.979 --> 00:02:31.740 interns, what are they missing most when we just 00:02:31.740 --> 00:02:34.759 give them a simple, vague prompt? It's all about 00:02:34.759 --> 00:02:37.800 that context void. When you say something short 00:02:37.800 --> 00:02:41.229 like, Write me a user API with Express .js. Yeah, 00:02:41.250 --> 00:02:43.009 something simple like that. You're basically 00:02:43.009 --> 00:02:45.789 forcing the AI to guess. It has to guess, you 00:02:45.789 --> 00:02:47.789 know, your project's directory structure or using 00:02:47.789 --> 00:02:51.289 Mongo or Postgres. What about OS? Is it JWT OAuth 00:02:51.289 --> 00:02:53.590 2 .0? What are your logging standards, error 00:02:53.590 --> 00:02:56.710 handling, existing data models? All of it. It 00:02:56.710 --> 00:02:58.509 doesn't know any of that. It's completely in 00:02:58.509 --> 00:03:00.189 the dark on the stuff that actually matters for 00:03:00.189 --> 00:03:02.189 making the code work in your system. And without 00:03:02.189 --> 00:03:04.210 that specific info, it just spits out generic 00:03:04.210 --> 00:03:06.569 code, right? Based on common stuff it saw during 00:03:06.569 --> 00:03:10.020 training. Precisely. code that might technically 00:03:10.020 --> 00:03:12.900 run on its own, maybe, but it's almost guaranteed 00:03:12.900 --> 00:03:15.099 to be incompatible somehow with your existing 00:03:15.099 --> 00:03:18.020 project. Which leads to hours of painful debugging. 00:03:18.659 --> 00:03:20.780 Trying to figure out why this perfectly logical 00:03:20.780 --> 00:03:23.599 -looking code is breaking everything. Hours and 00:03:23.599 --> 00:03:27.340 hours, costly hours. And that's exactly why context 00:03:27.340 --> 00:03:30.539 engineering isn't just like a nice idea, it's 00:03:30.539 --> 00:03:33.580 becoming essential, a real game changer. Okay, 00:03:33.580 --> 00:03:35.759 so that brings us neatly to the term itself, 00:03:36.240 --> 00:03:38.689 context engineering. It's interesting. Andrzej 00:03:38.689 --> 00:03:40.669 Karpathy coined it, and he also gave us vibe 00:03:40.669 --> 00:03:42.669 coding, funnily enough. Yeah. Yeah. It covers 00:03:42.669 --> 00:03:44.669 both ends of the spectrum. Right. And Toby Lidke, 00:03:44.750 --> 00:03:46.909 the Shopify CEO, he put it really well. He said, 00:03:46.930 --> 00:03:49.389 I really like the term context engineering over 00:03:49.389 --> 00:03:52.030 prompt engineering. It describes the core skills 00:03:52.030 --> 00:03:54.610 better, the art of providing all the context 00:03:54.610 --> 00:03:56.990 for the task to be plausibly solvable by the 00:03:56.990 --> 00:03:59.229 large language model. That quote really does 00:03:59.229 --> 00:04:01.330 hit the nail on the head. It captures the essence 00:04:01.330 --> 00:04:03.990 perfectly. If you want to think about the difference, 00:04:04.389 --> 00:04:07.189 think about ordering food. OK. An analogy. I 00:04:07.189 --> 00:04:09.259 like it. Trumped engineering is like going to 00:04:09.259 --> 00:04:11.080 a chef and saying, cook me something nice with 00:04:11.080 --> 00:04:15.379 beef. OK. Vague could be anything. Exactly. Steak, 00:04:15.560 --> 00:04:19.199 stir fry, soup. Who knows? It depends entirely 00:04:19.199 --> 00:04:21.180 on the chef's assumptions about what you want. 00:04:21.379 --> 00:04:23.019 It's basically a shot in the dark. And how often 00:04:23.019 --> 00:04:25.410 does that work out perfectly? Probably not very 00:04:25.410 --> 00:04:27.589 often, so context engineering is the opposite. 00:04:27.790 --> 00:04:29.529 Completely. Context engineering is like handing 00:04:29.529 --> 00:04:32.449 the chef a super detailed recipe, right? Here's 00:04:32.449 --> 00:04:35.250 300 grams of Australian beef, cut exactly two 00:04:35.250 --> 00:04:37.550 centimeters thick, use the cast iron skillet, 00:04:37.649 --> 00:04:40.709 get it rip and hot, add rosemary, garlic, cook 00:04:40.709 --> 00:04:43.290 it medium rare, serve it with roasted asparagus. 00:04:43.490 --> 00:04:46.430 Ah, okay. No ambiguity there. The result is precisely 00:04:46.430 --> 00:04:48.509 what you asked for because you gave all the necessary 00:04:48.509 --> 00:04:51.569 details. Exactly. So if we want a full definition, 00:04:51.949 --> 00:04:54.120 context engineering is something like... The 00:04:54.120 --> 00:04:56.860 skill and art of strategically selecting, organizing, 00:04:57.120 --> 00:04:59.899 and managing the entire set of information and 00:04:59.899 --> 00:05:03.819 AI needs at each step to do complex tasks well 00:05:03.819 --> 00:05:06.220 efficiently and accurately. And the key is without 00:05:06.220 --> 00:05:08.319 overwhelming it with useless stuff or leaving 00:05:08.319 --> 00:05:10.860 out critical details. That's the art part. Finding 00:05:10.860 --> 00:05:13.300 that balance, feeding it the perfect information 00:05:13.300 --> 00:05:16.360 diet, you could say. OK, the perfect diet. So 00:05:16.360 --> 00:05:18.160 here's where it gets really interesting for me. 00:05:18.699 --> 00:05:21.959 How does having all this context actually make 00:05:21.959 --> 00:05:24.680 such a huge difference? What are the ingredients 00:05:24.680 --> 00:05:27.180 in this perfect diet you're talking about? Right, 00:05:27.279 --> 00:05:29.839 so context isn't just your prompt. It's everything 00:05:29.839 --> 00:05:32.660 the AI model gets to see before it starts generating 00:05:32.660 --> 00:05:34.819 anything. It's multi -layered. It's structured. 00:05:35.600 --> 00:05:38.819 It includes, well, quite a few things. Lay it 00:05:38.819 --> 00:05:41.199 on us. What's in the mix? OK, first, the current 00:05:41.199 --> 00:05:43.519 state of the project. That means things like 00:05:43.519 --> 00:05:45.879 your file tree structure, the actual content 00:05:45.879 --> 00:05:48.079 of relevant code files. Like the files it needs 00:05:48.079 --> 00:05:50.660 to modify or interact with. Exactly. And things 00:05:50.660 --> 00:05:53.839 like your package .json or requirements .txt 00:05:53.839 --> 00:05:56.660 so it knows your dependencies. Then there's historical 00:05:56.660 --> 00:05:59.019 information. Meaning? Previous turns in the same 00:05:59.019 --> 00:06:01.800 chat session. Or maybe bigger architectural decisions 00:06:01.800 --> 00:06:04.000 that were already made and documented somewhere 00:06:04.000 --> 00:06:06.360 in the project. Then of course your user prompt. 00:06:06.939 --> 00:06:09.149 But not just a one -liner. The detailed one we 00:06:09.149 --> 00:06:10.850 talked about with business and technical needs. 00:06:10.990 --> 00:06:14.769 Yes. Then crucially, available tools. This is 00:06:14.769 --> 00:06:17.110 where it gets powerful. You tell the AI about 00:06:17.110 --> 00:06:19.730 specific APIs or functions it's allowed to use. 00:06:20.029 --> 00:06:22.410 Ah, like function calling capabilities in newer 00:06:22.410 --> 00:06:25.230 models. Precisely. Things like a function to 00:06:25.230 --> 00:06:27.589 query your database or an API call to send an 00:06:27.589 --> 00:06:30.350 email. The AI can leverage these external tools. 00:06:30.970 --> 00:06:33.709 Then there's ARAGI instructions, retrieve a logmented 00:06:33.709 --> 00:06:36.529 generation. Guiding the AI on how to look stuff 00:06:36.529 --> 00:06:39.860 up. Yep. How it should search and use info from, 00:06:39.860 --> 00:06:43.019 say, external docs you provide official API docs, 00:06:43.259 --> 00:06:46.120 technical articles, your own internal wiki, and 00:06:46.120 --> 00:06:49.439 finally, long -term memory. Persistent rules. 00:06:49.600 --> 00:06:52.319 Kind of. Rules, preferences, standards you set 00:06:52.319 --> 00:06:54.759 up once, like, always use type hints in Python, 00:06:54.939 --> 00:06:58.139 or every public function needs unit tests. These 00:06:58.139 --> 00:07:00.819 apply to every single interaction. Wow, okay. 00:07:01.019 --> 00:07:03.100 That's... That's a lot more comprehensive than 00:07:03.100 --> 00:07:05.220 just a prompt. So by pulling all that together, 00:07:05.319 --> 00:07:07.879 the AI can actually cross reference things, connect 00:07:07.879 --> 00:07:10.040 the dots. Exactly. It can see the current code, 00:07:10.079 --> 00:07:12.139 check the standards file, look at the API docs 00:07:12.139 --> 00:07:14.420 you linked, understand the goal from your detailed 00:07:14.420 --> 00:07:16.720 prompt. It's not guessing anymore. It's working 00:07:16.720 --> 00:07:18.860 from a detailed blueprint. A blueprint. I like 00:07:18.860 --> 00:07:20.560 that. So if you compare the two approaches side 00:07:20.560 --> 00:07:23.339 by side, vibe coding versus context engineering, 00:07:23.980 --> 00:07:26.279 the difference must be stark. Oh, absolutely. 00:07:26.639 --> 00:07:29.339 Look at reliability. Vibe coding, pretty low. 00:07:29.879 --> 00:07:32.899 Context engineering. much higher debug time vibe 00:07:32.899 --> 00:07:35.439 coding often high context engineering way lower 00:07:35.439 --> 00:07:37.560 because errors are caught earlier or prevented 00:07:37.560 --> 00:07:40.620 scalability consistency vibe coding is hard to 00:07:40.620 --> 00:07:43.199 scale or make consistent context engineering 00:07:43.199 --> 00:07:45.819 high scalability because you can reuse templates 00:07:45.819 --> 00:07:48.560 and rules consistency is much better what about 00:07:48.560 --> 00:07:52.379 cost like token usage that's interesting vibe 00:07:52.379 --> 00:07:55.680 coding feels cheap per prompt, but the back and 00:07:55.680 --> 00:07:58.120 forth adds up fast. Context engineering might 00:07:58.120 --> 00:08:00.680 have a larger initial request, but it's often 00:08:00.680 --> 00:08:03.019 more optimal overall because you get it right, 00:08:03.259 --> 00:08:06.180 or much closer to right, the first time. Fewer 00:08:06.180 --> 00:08:08.399 iterations. Okay, so the theory makes a ton of 00:08:08.399 --> 00:08:11.379 sense. Blueprint, detailed info, better results. 00:08:11.879 --> 00:08:13.600 But the big question for everyone listening is 00:08:13.600 --> 00:08:15.680 probably, OK, great, but how do I actually do 00:08:15.680 --> 00:08:17.980 this? How do I stop vibe coding and start building 00:08:17.980 --> 00:08:19.779 with this blueprint? Let's get practical. Right. 00:08:19.860 --> 00:08:21.459 Let's walk through it. There's a great template 00:08:21.459 --> 00:08:23.699 out there inspired by Coal Midden that really 00:08:23.699 --> 00:08:26.000 demonstrates the principles. And remember, these 00:08:26.000 --> 00:08:29.220 ideas work with basically any capable AI model. 00:08:29.579 --> 00:08:32.080 OK, step by step. What's step one? Step one is 00:08:32.080 --> 00:08:34.200 just laying the foundation. Basic stuff, really. 00:08:34.399 --> 00:08:37.080 Make sure you have tools like Git, Node .js installed, 00:08:37.759 --> 00:08:40.480 and you'll need an AI programming assistant integrated 00:08:40.480 --> 00:08:43.259 into your IDE. Something like the Quad Code extension 00:08:43.259 --> 00:08:45.360 for VS Code, for example. Standard developer 00:08:45.360 --> 00:08:48.620 setup. Got it. Step two. Step two, acquire the 00:08:48.620 --> 00:08:51.379 template. For this specific example, you'd clone 00:08:51.379 --> 00:08:55.659 a GitHub repo. https .getup .com forward slash 00:08:55.659 --> 00:09:00.960 duress 00000 contextengineering -intro .git. 00:09:01.299 --> 00:09:03.159 Then just open that clone folder in your IDE. 00:09:03.519 --> 00:09:05.360 Okay, clone the repo. Now what? This is where 00:09:05.360 --> 00:09:06.940 the context starts coming in, right? Yeah, okay. 00:09:07.019 --> 00:09:09.360 Step three, configuring global rules. This is 00:09:09.360 --> 00:09:11.500 where the magic starts. You open a file, often 00:09:11.500 --> 00:09:14.139 name something like Claude .md or systemrules 00:09:14.139 --> 00:09:17.039 .md. This file holds the core instructions for 00:09:17.039 --> 00:09:18.879 the AI for this project. Think of it as your 00:09:18.879 --> 00:09:21.279 project's constitution. The constitution. I like 00:09:21.279 --> 00:09:23.039 that. So what goes in there? All your high -level 00:09:23.039 --> 00:09:25.240 standards and preferences. Code structure stuff. 00:09:25.399 --> 00:09:27.720 Use hexagonal architecture. Or React components 00:09:27.720 --> 00:09:30.000 must be functional with hooks. Testing requirements. 00:09:30.360 --> 00:09:33.220 Use Jest and Supertest for API tests. Minimum 00:09:33.220 --> 00:09:36.100 85 % code coverage. Reliability standards, too. 00:09:36.220 --> 00:09:39.379 Like error handling. Yep. Always use Trycatch 00:09:39.379 --> 00:09:42.220 for I .O. log errors using our standard logger, 00:09:42.559 --> 00:09:45.279 even task completion rules, like only mark a 00:09:45.279 --> 00:09:47.240 task done when code and tests are generated, 00:09:47.879 --> 00:09:50.919 and conversation rules for the AI itself. Think 00:09:50.919 --> 00:09:54.139 step -by -step, ask clarifying questions if unsure. 00:09:54.399 --> 00:09:56.440 Wow, okay. So you're really setting the ground 00:09:56.440 --> 00:09:58.779 rules for how the AI should behave and build 00:09:58.779 --> 00:10:00.799 things within your specific project context. 00:10:00.899 --> 00:10:02.779 And this applies every time you interact with 00:10:02.779 --> 00:10:05.799 it for this project. That's the power. Consistency 00:10:05.799 --> 00:10:07.940 across interactions, even across different team 00:10:07.940 --> 00:10:10.159 members using it, it reduces that review friction. 00:10:10.440 --> 00:10:12.919 Makes sense. So Constitution is set. What's step 00:10:12.919 --> 00:10:15.940 four? the actual request. Step four is creating 00:10:15.940 --> 00:10:17.940 your feature request. This usually goes in a 00:10:17.940 --> 00:10:20.539 separate file, maybe initial .md or featurerequest 00:10:20.539 --> 00:10:23.759 .md, and detail is absolutely king here. No more 00:10:23.759 --> 00:10:26.159 make a login API. Definitely not. In a feature 00:10:26.159 --> 00:10:27.820 section, you'd write something specific like, 00:10:28.019 --> 00:10:31.399 build a post of PAV1 off login endpoint. It takes 00:10:31.399 --> 00:10:33.779 email and password. Authenticate against the 00:10:33.779 --> 00:10:37.460 user's table. On success, create a JWT 24 -hour 00:10:37.460 --> 00:10:40.519 expiry, include user and rule, return the token. 00:10:40.730 --> 00:10:43.230 super specific. Okay, that's clear. What else 00:10:43.230 --> 00:10:45.830 goes in this request file? Critically, an example 00:10:45.830 --> 00:10:49.309 section. This is huge. You provide snippets or 00:10:49.309 --> 00:10:52.309 paths to existing files in your code base. Ah, 00:10:52.350 --> 00:10:54.769 so it learns your style. Exactly. An existing 00:10:54.769 --> 00:10:57.950 controller .js, maybe a database model .ts. You 00:10:57.950 --> 00:11:00.029 put these in maybe an examples directory and 00:11:00.029 --> 00:11:02.370 reference them. The AI learns your patterns, 00:11:02.570 --> 00:11:04.970 your style, how you structure things. Okay, examples 00:11:04.970 --> 00:11:07.970 are key. A documentation section. Paste links 00:11:07.970 --> 00:11:11.230 to relevant docs, Stripe API page, a confluence 00:11:11.230 --> 00:11:13.129 article on your internal standards, whatever 00:11:13.129 --> 00:11:16.049 it needs. And finally, additional considerations. 00:11:16.269 --> 00:11:19.289 Edge cases. Security. Precisely. Endpoint must 00:11:19.289 --> 00:11:21.769 be rate limited. Hash passwords using bcrypt. 00:11:22.129 --> 00:11:24.190 Performance notes. Anything else it needs to 00:11:24.190 --> 00:11:26.970 know to do the job right. Leave no stone unturned. 00:11:27.129 --> 00:11:29.470 Okay, foundation laid. Constitution written. 00:11:29.690 --> 00:11:31.830 Super detailed request crafted with examples 00:11:31.830 --> 00:11:35.919 and docs. Now this next step. Step five, generating 00:11:35.919 --> 00:11:39.139 the product requirement plan or PRP. This sounds 00:11:39.139 --> 00:11:41.019 like where the real departure from just asking 00:11:41.019 --> 00:11:43.600 for code happens. Why plan first? This is the 00:11:43.600 --> 00:11:45.899 game changer step, honestly. Instead of jumping 00:11:45.899 --> 00:11:48.120 straight to code, you ask the AI to generate 00:11:48.120 --> 00:11:51.580 a detailed blueprint first. Using an agent or 00:11:51.580 --> 00:11:55.139 tool, you might run a command like agent .planrequest, 00:11:55.539 --> 00:11:58.519 requestfeaturelogin .md, output plansloginplan 00:11:58.519 --> 00:12:01.620 .md. And what does the AI do then? It goes to 00:12:01.620 --> 00:12:04.100 work. It reads your request, deeply analyzes 00:12:04.100 --> 00:12:06.600 the documentation links you provided, examines 00:12:06.600 --> 00:12:08.580 your code base, especially those example files, 00:12:08.879 --> 00:12:11.220 and then it constructs a detailed step -by -step 00:12:11.220 --> 00:12:13.159 plan. It might take a few minutes, but it's doing 00:12:13.159 --> 00:12:15.320 actual research and architectural thinking. Okay, 00:12:15.379 --> 00:12:18.799 so it generates this plan. Step six is reviewing 00:12:18.799 --> 00:12:21.740 that PRP. What am I looking for and why is this 00:12:21.740 --> 00:12:24.480 review so critical? I guess catching mistakes 00:12:24.480 --> 00:12:26.480 here is way cheaper. Way, way cheaper. Yeah. 00:12:26.600 --> 00:12:28.179 You're catching errors on the blueprint, not 00:12:28.179 --> 00:12:30.759 after the walls are up. A good PRP is detailed. 00:12:30.860 --> 00:12:33.480 It'll have a requirement summary. The AI basically 00:12:33.480 --> 00:12:35.120 repeats back what it thinks you want, confirming 00:12:35.120 --> 00:12:37.289 understanding. Good sanity check. Absolutely. 00:12:37.429 --> 00:12:39.710 It lists documentation references it plans to 00:12:39.710 --> 00:12:42.590 use. It shows a file structure analysis. Here's 00:12:42.590 --> 00:12:44.309 the current tree. Here's what I'll create or 00:12:44.309 --> 00:12:47.129 modify. And the core is the detailed implementation 00:12:47.129 --> 00:12:50.169 plan, literally step by step. One, create auth 00:12:50.169 --> 00:12:53.350 .controller .js. Two, add route to routes index 00:12:53.350 --> 00:12:56.490 .js. So you can see exactly what it intends to 00:12:56.490 --> 00:12:59.090 do, file by file. Exactly. And often, a risk 00:12:59.090 --> 00:13:00.990 assessment section two, potential problems it 00:13:00.990 --> 00:13:03.809 foresees, and how it plans to handle them. Reviewing 00:13:03.809 --> 00:13:06.210 this catches misunderstandings, logical flaws, 00:13:06.490 --> 00:13:08.389 potential hallucinations before code is written. 00:13:08.809 --> 00:13:10.850 It saves so much time and tokens downstream. 00:13:11.149 --> 00:13:12.950 That feels like a huge shift. It's not just coding 00:13:12.950 --> 00:13:15.950 faster, it's improving the design process itself. 00:13:16.490 --> 00:13:18.450 Like having a super diligent architect review 00:13:18.450 --> 00:13:20.769 your plans. Okay, so plan reviewed, looks good. 00:13:21.250 --> 00:13:24.070 Step seven, execute the PRP. Yep. Now you tell 00:13:24.070 --> 00:13:27.169 the AI to build it, a command like agent .execute 00:13:27.169 --> 00:13:32.429 plansloginplan .md workspace .src. And the AI 00:13:32.429 --> 00:13:34.210 just systematically works through the plan. What 00:13:34.210 --> 00:13:36.169 does that look like in practice? It'll create 00:13:36.169 --> 00:13:38.529 directories if needed, install dependencies via 00:13:38.529 --> 00:13:41.309 npm or pip, write the code for each specified 00:13:41.309 --> 00:13:43.570 file, run linters or formatters you've configured. 00:13:43.879 --> 00:13:46.539 It might even try to run build commands or simple 00:13:46.539 --> 00:13:49.600 validations. And if it hits small errors, it 00:13:49.600 --> 00:13:51.639 often tries to fix them automatically based on 00:13:51.639 --> 00:13:53.919 the error messages. That sounds pretty intensive. 00:13:54.279 --> 00:13:58.000 Takes a while. Uses tokens. It can, yes. Especially 00:13:58.000 --> 00:14:00.480 for complex features. It might take several minutes. 00:14:00.899 --> 00:14:02.840 And it will use tokens as it thinks, writes, 00:14:03.080 --> 00:14:05.740 and potentially self -corrects. But the output 00:14:05.740 --> 00:14:08.159 quality, the adherence to your plan and standards, 00:14:08.480 --> 00:14:10.659 it's usually far superior to just wean it with 00:14:10.659 --> 00:14:12.759 vibe coding. Right. The upfront investment pays 00:14:12.759 --> 00:14:15.419 off in quality and reduced rework. So execution 00:14:15.419 --> 00:14:17.940 finishes. What's the last step? Step eight is 00:14:17.940 --> 00:14:20.909 basically test and deploy. The AI usually finishes 00:14:20.909 --> 00:14:23.029 by giving you the final instructions. Things 00:14:23.029 --> 00:14:25.529 like environment variables you need to set, API 00:14:25.529 --> 00:14:27.950 keys you need to configure, the exact commands 00:14:27.950 --> 00:14:30.190 to install everything, npm install, run the app, 00:14:30.269 --> 00:14:33.529 npm start, and run the tests, npm test. So it 00:14:33.529 --> 00:14:36.549 hands off a working tested based on its generated 00:14:36.549 --> 00:14:39.929 tests feature with instructions. Ideally, yes. 00:14:40.370 --> 00:14:42.649 You follow its final steps, do your own thorough 00:14:42.649 --> 00:14:45.230 testing, of course, and then deploy. But it's 00:14:45.230 --> 00:14:47.990 built to your specifications defined in the context 00:14:47.990 --> 00:14:50.440 in the plan. Okay, that whole process makes sense. 00:14:50.679 --> 00:14:53.000 To really hammer home the power, maybe let's 00:14:53.000 --> 00:14:55.259 look at a more complex, real -world example, 00:14:55.460 --> 00:14:58.639 something beyond a simple login, like that automated 00:14:58.639 --> 00:15:01.860 financial analysis system idea. Great idea, because 00:15:01.860 --> 00:15:03.620 that's where you see the massive time savings. 00:15:04.320 --> 00:15:06.600 Imagine you want an AI to build a system that 00:15:06.600 --> 00:15:09.299 pulls quarterly financial reports and drafts 00:15:09.299 --> 00:15:11.600 an email summary for your investment team. Okay, 00:15:11.659 --> 00:15:14.600 that sounds complex. What would the initial .md 00:15:14.600 --> 00:15:16.759 look like for that? Under features, you'd be 00:15:16.759 --> 00:15:19.820 very specific. build an automated Python system, 00:15:20.080 --> 00:15:22.379 takes a list of stock tickers, uses the Braids 00:15:22.379 --> 00:15:24.519 search API to find the latest quarterly earnings 00:15:24.519 --> 00:15:27.340 report PDF or web page for each, extracts key 00:15:27.340 --> 00:15:30.120 metrics, revenue, net income, EPS, future guidance, 00:15:30.840 --> 00:15:33.899 then uses the Gmail API to draft an email summarizing 00:15:33.899 --> 00:15:36.279 findings comparing current quarter to previous. 00:15:36.700 --> 00:15:39.269 Wow, okay. Very specific feature set. What about 00:15:39.269 --> 00:15:42.389 examples and docs? Crucial here. Under examples, 00:15:42.470 --> 00:15:44.990 you might provide a sample Python script showing 00:15:44.990 --> 00:15:47.090 how you prefer to use libraries like requests 00:15:47.090 --> 00:15:50.009 for fetching data and maybe Beautiful Soup for 00:15:50.009 --> 00:15:52.950 basic HTML parsing, and definitely provide a 00:15:52.950 --> 00:15:55.129 markdown template showing the exact format you 00:15:55.129 --> 00:15:58.649 want for the final email draft. Style, tone, 00:15:59.070 --> 00:16:00.990 structure. So it matches your team's communication 00:16:00.990 --> 00:16:04.600 style. Smart docs. Direct links under documentation 00:16:04.600 --> 00:16:07.299 to the Brave Search API docs and the Gmail API 00:16:07.299 --> 00:16:10.120 docs make it easy for the AI. And additional 00:16:10.120 --> 00:16:11.860 considerations. What kind of things, Zach? Things 00:16:11.860 --> 00:16:14.200 like handle cases where an earnings report can't 00:16:14.200 --> 00:16:16.500 be found for a ticker. Ensure the email tone 00:16:16.500 --> 00:16:18.789 is professional and objective. maybe mention 00:16:18.789 --> 00:16:21.590 API rate limits to be aware of, error handling 00:16:21.590 --> 00:16:24.330 for parsing failures. Okay, and if you run this 00:16:24.330 --> 00:16:27.370 whole context engineering process plan, review, 00:16:27.710 --> 00:16:29.809 execute, what's the potential outcome? The outcome 00:16:29.809 --> 00:16:31.850 could be a complete Python project, maybe with 00:16:31.850 --> 00:16:34.529 a main script like financialanalyzer .pi, helper 00:16:34.529 --> 00:16:37.909 modules, requirement files, a system that automatically 00:16:37.909 --> 00:16:40.850 does that research, scrapes the data, performs 00:16:40.850 --> 00:16:43.669 the analysis, and drafts that detailed email. 00:16:43.750 --> 00:16:45.669 Which would normally take a human analyst star 00:16:45.669 --> 00:16:49.639 hours. Easily. hours of manual, tedious work, 00:16:50.000 --> 00:16:52.259 researching, downloading, reading reports, extracting 00:16:52.259 --> 00:16:54.860 numbers, drafting. This system could potentially 00:16:54.860 --> 00:16:57.519 do it for, what, maybe a few dollars in API fees 00:16:57.519 --> 00:17:00.659 and token costs? That's a jaw -dropping productivity 00:17:00.659 --> 00:17:04.359 game, moving from manual slog to automated insight. 00:17:05.000 --> 00:17:07.579 So if we zoom out again, connect this to the 00:17:07.579 --> 00:17:11.190 bigger picture. What are the real compounding 00:17:11.190 --> 00:17:14.210 benefits here? Adopting this context -first approach 00:17:14.210 --> 00:17:16.930 across a whole project or team. The benefits 00:17:16.930 --> 00:17:19.769 really stack up, I think. First, reduced hallucinations. 00:17:19.849 --> 00:17:23.309 Big one. By giving the AI a clear, bounded world, 00:17:23.769 --> 00:17:25.849 your context it operates on the ground truth 00:17:25.849 --> 00:17:28.549 you provide. less making stuff up. Makes sense. 00:17:28.710 --> 00:17:30.990 What else? Superior planning. Like we said, forcing 00:17:30.990 --> 00:17:33.210 the AI to plan before coding catches so many 00:17:33.210 --> 00:17:35.690 architectural or logical flaws early. Fixing 00:17:35.690 --> 00:17:37.309 the blueprint is way cheaper than renovating 00:17:37.309 --> 00:17:39.369 the building. Right. Cognitive offloading, too, 00:17:39.430 --> 00:17:41.829 maybe? Absolutely. Cognitive offloading. Developers 00:17:41.829 --> 00:17:44.309 can focus more on what the high -level requirements, 00:17:44.609 --> 00:17:46.670 the business value, and let the AI handle more 00:17:46.670 --> 00:17:49.089 of the detailed how. Freeze up mental bandwidth 00:17:49.089 --> 00:17:51.369 for harder problems. I can see that. And the 00:17:51.369 --> 00:17:53.390 context files themselves. They become a form 00:17:53.390 --> 00:17:56.380 of living code -based specification. your CLAWD 00:17:56.380 --> 00:17:59.640 .md rules, your detailed initial .md requests, 00:18:00.119 --> 00:18:03.240 the generated PRPs, they act as always up -to 00:18:03.240 --> 00:18:05.160 -date documentation explaining how the system 00:18:05.160 --> 00:18:07.519 is built and why certain decisions were made. 00:18:07.680 --> 00:18:09.920 That's incredibly valuable. Documentation that 00:18:09.920 --> 00:18:13.019 doesn't get stale and overall quality. Consistent 00:18:13.019 --> 00:18:16.019 quality. With good context, the AI consistently 00:18:16.019 --> 00:18:18.380 produces code that meets your standards, uses 00:18:18.380 --> 00:18:22.160 your patterns, includes tests, fewer bugs, easier 00:18:22.160 --> 00:18:24.849 maintenance, more reliable system overall. OK, 00:18:24.930 --> 00:18:27.269 it sounds almost perfect, which always makes 00:18:27.269 --> 00:18:29.609 me ask, what are the catches? Are there pitfalls 00:18:29.609 --> 00:18:31.450 to watch out for when you start doing context 00:18:31.450 --> 00:18:33.589 engineering? Where can things go wrong? Oh, definitely 00:18:33.589 --> 00:18:35.670 pitfalls. It's not magic. It requires skill. 00:18:36.230 --> 00:18:38.890 One big one is over -constraining. Too many rules. 00:18:39.009 --> 00:18:41.309 Yeah, being too rigid. If you lock down every 00:18:41.309 --> 00:18:44.210 single detail, you might stifle the AI's ability 00:18:44.210 --> 00:18:46.349 to find a clever or simpler solution you didn't 00:18:46.349 --> 00:18:49.109 think of. It can sometimes be counterproductive. 00:18:49.329 --> 00:18:51.750 OK, so don't stifle creativity entirely. Yeah. 00:18:51.869 --> 00:18:54.349 What else? Conflicting context. This is a subtle 00:18:54.349 --> 00:18:57.430 one. If your global rules in Claw .md say one 00:18:57.430 --> 00:19:00.410 thing, but your specific feature request in Initial 00:19:00.410 --> 00:19:03.529 .md accidentally implies something contradictory. 00:19:03.730 --> 00:19:06.789 The AI gets confused. Mixed signals. Exactly. 00:19:07.069 --> 00:19:09.809 It can lead to unpredictable or nonsensical output 00:19:09.809 --> 00:19:12.089 because it doesn't know which instruction takes 00:19:12.089 --> 00:19:15.069 precedence. Consistency across your context is 00:19:15.069 --> 00:19:17.190 key. Makes sense. What about the planning step? 00:19:17.670 --> 00:19:19.809 Skipping PRP review? It was tempting, right? 00:19:20.029 --> 00:19:22.839 You just want the code. But blindly trusting 00:19:22.839 --> 00:19:25.519 the AI's plan without carefully reading it is 00:19:25.519 --> 00:19:27.819 risky. It might have misunderstood something, 00:19:28.259 --> 00:19:30.279 planned a flawed approach, and if you just hit 00:19:30.279 --> 00:19:32.259 execute. You bake in the errors right at the 00:19:32.259 --> 00:19:34.839 start. Got it. And the opposite of over -constraining. 00:19:35.380 --> 00:19:37.380 Insufficiently detailed context. If you don't 00:19:37.380 --> 00:19:40.720 provide enough detail, good examples, clear requirements, 00:19:41.059 --> 00:19:43.619 relevant docs you just drift back towards. Vibe 00:19:43.619 --> 00:19:45.880 coding. Right back where you started. The AI 00:19:45.880 --> 00:19:48.799 is forced to guess again. So, yeah, finding that 00:19:48.799 --> 00:19:51.720 balance detailed enough, but not too rigid and 00:19:51.720 --> 00:19:54.380 internally consistent, that's the skill. It really 00:19:54.380 --> 00:19:56.819 feels like context engineering is more than just 00:19:56.819 --> 00:19:59.259 a technique. It's a fundamental shift, isn't 00:19:59.259 --> 00:20:01.660 it? in how we even think about interacting with 00:20:01.660 --> 00:20:06.059 these powerful AI tools. We're moving away from 00:20:06.059 --> 00:20:08.599 just being command -givers, shouting orders into 00:20:08.599 --> 00:20:11.220 the void, and becoming more like information 00:20:11.220 --> 00:20:14.579 architects, curators of context. The critical 00:20:14.579 --> 00:20:16.740 skill isn't just writing a clever prompt anymore, 00:20:16.740 --> 00:20:20.319 is it? It's about systematically gathering, organizing, 00:20:20.819 --> 00:20:23.200 structuring, and presenting the right information, 00:20:24.000 --> 00:20:26.710 designing the entire conversation. That's perfectly 00:20:26.710 --> 00:20:29.069 put. It's really about more than just training 00:20:29.069 --> 00:20:32.190 an AI to write code snippets. It's about training 00:20:32.190 --> 00:20:35.470 it in a way to become a capable, reliable member 00:20:35.470 --> 00:20:37.849 of your development team. one that understands 00:20:37.849 --> 00:20:40.009 your project's specific needs and standards. 00:20:40.130 --> 00:20:42.789 It's truly augmenting what we can do. So, maybe 00:20:42.789 --> 00:20:44.890 the final thought for everyone listening is this. 00:20:45.450 --> 00:20:47.950 The era of unreliable, frustrating vibe coding 00:20:47.950 --> 00:20:50.650 feels like it's coming to an end. The age of 00:20:50.650 --> 00:20:53.029 systematic, context -driven software development, 00:20:53.269 --> 00:20:55.549 that seems to be just beginning. The question 00:20:55.549 --> 00:20:57.930 is, are you ready to be a part of that shift 00:20:57.930 --> 00:21:00.380 and transform how you build software? It's an 00:21:00.380 --> 00:21:02.680 exciting time to be a developer. It really is. 00:21:02.980 --> 00:21:05.160 We definitely encourage you to explore these 00:21:05.160 --> 00:21:08.059 ideas more, check out the resources, maybe try 00:21:08.059 --> 00:21:10.140 building a simple context template for one of 00:21:10.140 --> 00:21:12.799 your own projects. The future of coding is unfolding 00:21:12.799 --> 00:21:15.000 and it looks like context is right at the heart 00:21:15.000 --> 00:21:15.279 of it.