WEBVTT 00:00:00.000 --> 00:00:02.600 It almost feels like ancient history now, but 00:00:02.600 --> 00:00:05.000 do you remember that alt tab dance from, what, 00:00:05.040 --> 00:00:07.740 2023? Oh, yeah. You'd be writing code, you hit 00:00:07.740 --> 00:00:11.359 a bug, and then it was this whole ritual. You 00:00:11.359 --> 00:00:14.400 switch windows, copy the error. Paste it into 00:00:14.400 --> 00:00:16.899 a chatbot, wait for the little dots to stop blinking. 00:00:17.120 --> 00:00:20.359 Copy that text, tab back, paste it in, and hope 00:00:20.359 --> 00:00:22.719 for the best. We call that the Stone Age of AI. 00:00:22.940 --> 00:00:25.379 Yeah. You know, the chatbot era. And look, it 00:00:25.379 --> 00:00:27.940 was revolutionary at the time, but... Looking 00:00:27.940 --> 00:00:31.820 back from where we are now in early 2026, it 00:00:31.820 --> 00:00:34.240 just feels incredibly clunky. We aren't just 00:00:34.240 --> 00:00:37.079 chatting anymore. We've really entered what people 00:00:37.079 --> 00:00:39.520 are calling the agent native era. Right. We're 00:00:39.520 --> 00:00:41.240 talking about tools that actually live inside 00:00:41.240 --> 00:00:43.560 your terminal. Agents that can refactor massive 00:00:43.560 --> 00:00:45.799 code bases while you sleep. And systems that 00:00:45.799 --> 00:00:48.500 catch bugs before you even commit the code. It's 00:00:48.500 --> 00:00:50.280 the difference between having, I don't know, 00:00:50.299 --> 00:00:53.380 a consultant on the phone versus... A partner 00:00:53.380 --> 00:00:54.939 sitting right there at the desk with you. Or 00:00:54.939 --> 00:00:57.119 sometimes inside the keyboard itself. Welcome 00:00:57.119 --> 00:00:59.979 back to the Deep Dive. Today, we're unpacking 00:00:59.979 --> 00:01:03.579 the state of AI coding tools here in 2026. And 00:01:03.579 --> 00:01:05.900 we're basing this on a really comprehensive guide 00:01:05.900 --> 00:01:09.439 by Max Dan called Best AI Coding Tools 2026. 00:01:09.840 --> 00:01:11.480 Yeah, and it pits the heavyweights against each 00:01:11.480 --> 00:01:14.700 other. Cursor, Copilot, Cloud Code. But we're 00:01:14.700 --> 00:01:16.060 not just going to do a feature list. No, because 00:01:16.060 --> 00:01:18.439 lists are boring. They are. What we're really 00:01:18.439 --> 00:01:21.200 doing is mapping how the developer's role itself 00:01:21.200 --> 00:01:24.620 is changing. Max Anne's article poses this really 00:01:24.620 --> 00:01:27.500 great core question. It's not if you use AI anymore. 00:01:27.680 --> 00:01:30.019 That ship has sailed. Long gone. The question 00:01:30.019 --> 00:01:33.239 now is, which of these tools actually work? Yeah. 00:01:33.379 --> 00:01:35.299 You know, in a real production environment. Because 00:01:35.299 --> 00:01:38.159 the novelty has completely worn off. 100%. In 00:01:38.159 --> 00:01:40.319 2024, it was cool if the AI wrote you a poem. 00:01:40.420 --> 00:01:43.540 In 2026, the AI breaks the build. It's a liability. 00:01:43.840 --> 00:01:45.719 We need things that are effective. Efficacy. 00:01:45.819 --> 00:01:47.680 So we're going to break down the full stack from 00:01:47.680 --> 00:01:50.819 the IDEs where you write the code to the sidekicks 00:01:50.819 --> 00:01:53.180 you brainstorm with to the agents doing the heavy 00:01:53.180 --> 00:01:56.140 listing and then the safety net that stops us 00:01:56.140 --> 00:01:58.870 from. you know, accidentally destroying the database. 00:01:59.189 --> 00:02:01.670 And you should definitely stick around for the 00:02:01.670 --> 00:02:04.730 end because we're also going to cover three specific 00:02:04.730 --> 00:02:07.370 habits, Max Ann calls them, the human element, 00:02:07.549 --> 00:02:10.289 that are designed to stop these tools from hallucinating. 00:02:10.370 --> 00:02:12.469 And these are things you can start using literally 00:02:12.469 --> 00:02:15.650 today. They're the difference between an AI that 00:02:15.650 --> 00:02:18.310 feels like a genius and one that feels like, 00:02:18.490 --> 00:02:21.490 well, a heavy burden. Okay, let's get into it. 00:02:21.710 --> 00:02:25.689 Segment A, the new environment. The source makes 00:02:25.689 --> 00:02:27.710 a pretty bold claim right at the top. It says 00:02:27.710 --> 00:02:29.909 development doesn't really start with a text 00:02:29.909 --> 00:02:32.430 editor anymore. No. It starts with an environment 00:02:32.430 --> 00:02:35.729 that understands. This is such a huge shift away 00:02:35.729 --> 00:02:38.189 from the dumb editor. I mean, for decades, text 00:02:38.189 --> 00:02:40.789 editors were just fancy typewriters, right? They 00:02:40.789 --> 00:02:43.250 knew syntax highlighting, maybe some basic linting. 00:02:43.330 --> 00:02:44.770 Right. But they didn't know what you were actually 00:02:44.770 --> 00:02:46.969 trying to build. Right. And the source argues 00:02:46.969 --> 00:02:49.270 that the standout winner in this new category, 00:02:49.490 --> 00:02:52.550 the one everyone is chasing. is cursor i see 00:02:52.550 --> 00:02:55.590 cursor everywhere but i have to ask is it just 00:02:55.590 --> 00:02:58.129 vs code was a really good plug -in or is there 00:02:58.129 --> 00:03:00.469 something fundamentally different happening that's 00:03:00.469 --> 00:03:02.710 the common misconception and it's an easy one 00:03:02.710 --> 00:03:04.550 to make because it's a fork of vs code so it 00:03:04.550 --> 00:03:06.789 feels familiar it looks the same it looks the 00:03:06.789 --> 00:03:10.210 same but the source really emphasizes that cursor 00:03:10.210 --> 00:03:13.430 is a complete rebuild ai is at the center of 00:03:13.430 --> 00:03:16.090 how it indexes everything the key isn't the chat 00:03:16.090 --> 00:03:19.090 window it's the context cursor indexes your entire 00:03:19.090 --> 00:03:22.300 code base okay so Explain that in practical terms. 00:03:22.340 --> 00:03:24.659 If I'm working in one file, let's say it uses 00:03:24.659 --> 00:03:29.099 Roth .ts, and I change a variable name, what 00:03:29.099 --> 00:03:31.419 happens? So a standard plug -in might help you 00:03:31.419 --> 00:03:33.759 rename it in that one file. The cursor understands 00:03:33.759 --> 00:03:36.840 the dependency graph. It knows changing that 00:03:36.840 --> 00:03:39.860 variable name impacts a function in, like, a 00:03:39.860 --> 00:03:43.020 utility file three folders away. And it updates 00:03:43.020 --> 00:03:45.259 the context for the AI automatically. It's what 00:03:45.259 --> 00:03:47.400 we call global context awareness. So it reduces 00:03:47.400 --> 00:03:50.650 that friction of context stuffing. Where you're 00:03:50.650 --> 00:03:53.550 just copying and pasting 10 different files into 00:03:53.550 --> 00:03:55.189 the check. Just to get a good answer precisely. 00:03:55.389 --> 00:03:56.849 You don't have to explain the entire project 00:03:56.849 --> 00:03:59.050 to it every single time. It just, it knows. And 00:03:59.050 --> 00:04:01.530 on top of that, you have features that just streamline 00:04:01.530 --> 00:04:03.830 the actual typing. You have the tab feature, 00:04:04.090 --> 00:04:06.900 which is like a super powered autocomplete. And 00:04:06.900 --> 00:04:10.199 then CMD plus K for inline edits. The thing that 00:04:10.199 --> 00:04:12.439 really caught my eye in the source, though, was 00:04:12.439 --> 00:04:15.180 this idea of modes. It seems like they've finally 00:04:15.180 --> 00:04:17.600 acknowledged that coding isn't just one single 00:04:17.600 --> 00:04:20.000 activity. This is a huge breakthrough for user 00:04:20.000 --> 00:04:23.139 experience, yeah. They've split the AI's behavior 00:04:23.139 --> 00:04:25.360 into these different modes. You have agent mode, 00:04:25.560 --> 00:04:27.759 which is for those big autonomous multi -file 00:04:27.759 --> 00:04:29.800 edits. The heavy lifting. The heavy lifting. 00:04:30.000 --> 00:04:34.060 Then debug mode scans for errors. Ask mode is 00:04:34.060 --> 00:04:36.730 for just exploring ideas. But the most interesting 00:04:36.730 --> 00:04:39.730 one to me is plan mode. Plan mode. That sounds 00:04:39.730 --> 00:04:41.389 like something a project manager would love. 00:04:41.529 --> 00:04:45.370 It is. But for the developer, in plan mode, the 00:04:45.370 --> 00:04:48.329 AI doesn't write any code. It just writes a step 00:04:48.329 --> 00:04:50.910 -by -step implementation plan. It forces the 00:04:50.910 --> 00:04:53.550 system to think before it acts. That actually 00:04:53.550 --> 00:04:55.910 solves a huge problem I have, where an AI just 00:04:55.910 --> 00:04:58.050 rushes in and starts writing code that doesn't 00:04:58.050 --> 00:05:01.470 make any sense architecturally. Exactly. It prevents 00:05:01.470 --> 00:05:03.930 what you could call premature optimization by 00:05:03.930 --> 00:05:07.279 the AI. Now, Cursor is really for that back -end, 00:05:07.319 --> 00:05:10.180 full -stack work. But the source also highlights 00:05:10.180 --> 00:05:13.540 another environment, Builder .io, that solves 00:05:13.540 --> 00:05:16.800 a totally different problem. Let me guess. The 00:05:16.800 --> 00:05:19.060 war between designers and engineers? The endless 00:05:19.060 --> 00:05:22.040 war. The classic pixel -perfect battle. Move 00:05:22.040 --> 00:05:24.680 this button two pixels to the left. Oh, I know 00:05:24.680 --> 00:05:27.420 it well. Builder .io is what they call a visual 00:05:27.420 --> 00:05:31.329 AI IDE. It's designed to be the bridge. It can 00:05:31.329 --> 00:05:33.470 take design straight from Figma or even just 00:05:33.470 --> 00:05:36.189 read a description from a JIRA ticket and turn 00:05:36.189 --> 00:05:38.490 it directly into component code. So it's basically 00:05:38.490 --> 00:05:40.910 automating that translation layer between design 00:05:40.910 --> 00:05:43.050 and code. Yes. If you're a front -end team and 00:05:43.050 --> 00:05:44.730 you just want to kill the friction of the handoff, 00:05:44.850 --> 00:05:47.430 this is the tool. It's less about deep logic, 00:05:47.550 --> 00:05:50.519 more about visual fidelity. The source also mentioned 00:05:50.519 --> 00:05:52.480 a couple of ones to watch. I was really intrigued 00:05:52.480 --> 00:05:55.819 by Google Antigravity. The name alone. It sounds 00:05:55.819 --> 00:05:57.699 like a sci -fi movie, right? And the tech is 00:05:57.699 --> 00:06:00.439 genuinely futuristic. Unlike Cursor, which is 00:06:00.439 --> 00:06:02.939 on your desktop, Antigravity is a cloud environment 00:06:02.939 --> 00:06:05.779 that watches your browser. And it spins up multiple 00:06:05.779 --> 00:06:08.579 parallel agents. Wow. So you could have one agent 00:06:08.579 --> 00:06:11.639 fixing a bug, another writing a test, and a third 00:06:11.639 --> 00:06:14.259 updating documentation all at the same time. 00:06:14.420 --> 00:06:16.560 So it seems the editor isn't just a typewriter 00:06:16.560 --> 00:06:19.579 anymore. It's becoming a teammate. That's the 00:06:19.579 --> 00:06:21.519 perfect way to put it. You're shifting from writing 00:06:21.519 --> 00:06:24.779 code to directing a system. Sometimes you don't 00:06:24.779 --> 00:06:26.740 need a whole system. You just need to talk to 00:06:26.740 --> 00:06:28.779 someone, you know, a rubber duck that actually 00:06:28.779 --> 00:06:31.560 talks back. The sidekicks. Right. Segment B. 00:06:31.980 --> 00:06:35.540 The source calls this Category 2 Conversational 00:06:35.540 --> 00:06:39.139 Assistance. It says the chatbot isn't dead. It's 00:06:39.139 --> 00:06:41.439 just become a specialized tool. It's your second 00:06:41.439 --> 00:06:44.939 brain. And Max N. breaks them down into these 00:06:44.939 --> 00:06:47.740 three distinct personalities because they really 00:06:47.740 --> 00:06:52.100 aren't all the same. First up, the classic, ChatGPT. 00:06:52.259 --> 00:06:54.839 The Swiss Army Knife. Exactly. It's fast. It's 00:06:54.839 --> 00:06:56.720 flexible. It knows a little bit about everything 00:06:56.720 --> 00:07:01.720 from Python to poetry. It's your go -to for brainstorming 00:07:01.720 --> 00:07:04.160 or getting unstuck on a tricky bit of rejects. 00:07:04.220 --> 00:07:06.600 But then you have Claude, and the source specifically 00:07:06.600 --> 00:07:10.160 points to Opus 4 .5. Claude is the deliberate 00:07:10.160 --> 00:07:13.529 thinker. It's noticeably slower than chat GPT, 00:07:13.629 --> 00:07:16.689 but that's a feature, not a bug. It handles these 00:07:16.689 --> 00:07:20.009 massive context windows way better. So if you 00:07:20.009 --> 00:07:22.069 have a really complex architectural decision 00:07:22.069 --> 00:07:25.209 to make, you ask Claude. It thinks before it 00:07:25.209 --> 00:07:27.790 speaks. It reasons. It shows you the logic. And 00:07:27.790 --> 00:07:29.189 then you've got the third one, the dark horse, 00:07:29.389 --> 00:07:31.879 Gemini. I feel like Gemini is always the one 00:07:31.879 --> 00:07:33.839 people forget about. And they shouldn't. The 00:07:33.839 --> 00:07:36.620 source highlights it as the perfect tool for 00:07:36.620 --> 00:07:39.480 a second opinion. Because it's trained on Google's 00:07:39.480 --> 00:07:42.500 huge data set, it just, it reasons differently 00:07:42.500 --> 00:07:46.639 than the open AI or anthropic models. If ChatGPT 00:07:46.639 --> 00:07:48.279 and Claude are both giving you an answer that 00:07:48.279 --> 00:07:50.959 doesn't quite work, Gemini often comes at the 00:07:50.959 --> 00:07:52.879 problem from a totally different angle. I have 00:07:52.879 --> 00:07:54.819 to admit something here. I fall into this trap 00:07:54.819 --> 00:07:57.339 all the time. I call it model monogamy. Ah, yes. 00:07:57.600 --> 00:08:00.259 I get stuck in this loop with ChatGPT. I ask 00:08:00.259 --> 00:08:02.439 it a question, it gives me a bug, I paste the 00:08:02.439 --> 00:08:04.579 error back in, and it just gives me the same 00:08:04.579 --> 00:08:06.800 buggy code again, and I just keep hitting enter, 00:08:06.899 --> 00:08:09.000 hoping for a different result. We have all been 00:08:09.000 --> 00:08:12.220 there. It's model lock -in, and the source has 00:08:12.220 --> 00:08:16.160 a pro move. for this exact problem. Don't marry 00:08:16.160 --> 00:08:18.879 one tool. It sounds so simple, but we really 00:08:18.879 --> 00:08:21.660 don't do it enough. We don't use all three. Most 00:08:21.660 --> 00:08:24.620 have pretty generous free tiers. If one stalls, 00:08:24.879 --> 00:08:28.120 just switch tabs. Paste the same prompt into 00:08:28.120 --> 00:08:30.120 Claude. It stops you from getting locked into 00:08:30.120 --> 00:08:32.039 a single way of thinking. It's like getting a 00:08:32.039 --> 00:08:33.639 second medical opinion, right? You don't just 00:08:33.639 --> 00:08:35.480 sit there and ask the same doctor the same question 00:08:35.480 --> 00:08:37.360 again and again. That's a great mindset shift. 00:08:37.799 --> 00:08:40.539 Okay, let's move on. Segment C. Let's talk about 00:08:40.539 --> 00:08:44.230 speed. The launch pads. Category three. App builders. 00:08:44.809 --> 00:08:46.950 This is for when you have an idea at two in the 00:08:46.950 --> 00:08:48.830 morning and you just need to see if it has legs. 00:08:48.970 --> 00:08:51.049 Rapid prototyping. Exactly. The source calls 00:08:51.049 --> 00:08:53.590 these launch pads, not factories. You've got 00:08:53.590 --> 00:08:55.610 tools like Replit, which has grown from just 00:08:55.610 --> 00:08:58.549 a browser editor into a full stack AI workspace. 00:08:58.950 --> 00:09:01.509 You can literally write in plain language. I 00:09:01.509 --> 00:09:04.490 want a to -do app that gamifies tasks. And the 00:09:04.490 --> 00:09:06.990 code just appears. And what about Lovable? Lovable 00:09:06.990 --> 00:09:08.950 is interesting because it's what we call highly 00:09:08.950 --> 00:09:11.870 opinionated. It makes a hard commitment to a 00:09:11.870 --> 00:09:15.519 very specific stack. React, Tailwind, Invite. 00:09:15.620 --> 00:09:18.059 So if you love that stack, you're flying. You're 00:09:18.059 --> 00:09:20.639 building at light speed. But if you want to use, 00:09:20.639 --> 00:09:23.840 say, Angular or Python on the back end, you're 00:09:23.840 --> 00:09:26.779 out of luck. You're trading flexibility for pure 00:09:26.779 --> 00:09:28.860 velocity. And if you don't want to be locked 00:09:28.860 --> 00:09:31.000 in like that. Then you look at something like 00:09:31.000 --> 00:09:33.879 Bolt. The source calls it the Explorer. It supports 00:09:33.879 --> 00:09:38.059 Views, Felt, Expo. It's much more agnostic. But 00:09:38.059 --> 00:09:40.629 I want to push back on this a bit. These tools 00:09:40.629 --> 00:09:43.409 generate a lot of code very, very quickly. Isn't 00:09:43.409 --> 00:09:46.090 that just a recipe for massive technical debt? 00:09:46.289 --> 00:09:48.250 That is the catch. You wouldn't build the next 00:09:48.250 --> 00:09:50.409 Netflix on Lovable, would you? You wouldn't. 00:09:50.590 --> 00:09:53.450 They're great for starting fast, for validating 00:09:53.450 --> 00:09:56.389 an idea, or for building internal tools. But 00:09:56.389 --> 00:09:58.669 they're not designed to carry a complex, long 00:09:58.669 --> 00:10:01.850 -term code base. At some point, you have to eject 00:10:01.850 --> 00:10:03.929 and take over. And if you don't understand what 00:10:03.929 --> 00:10:06.019 the AI wrote for you... You're in real trouble. 00:10:06.179 --> 00:10:08.620 I think that leads us perfectly into the deepest 00:10:08.620 --> 00:10:11.039 and maybe slightly scarier part of the stack. 00:10:11.240 --> 00:10:13.500 Segment D, the automation layer. This is the 00:10:13.500 --> 00:10:17.519 biggest shift in 2026. Category four, agents. 00:10:17.879 --> 00:10:20.100 The source draws a really sharp line here. We've 00:10:20.100 --> 00:10:22.919 been talking about brains in a box, you know, 00:10:22.919 --> 00:10:24.580 chatbots. But now we're talking about brains 00:10:24.580 --> 00:10:27.340 with hands. It is a profound difference. Up until 00:10:27.340 --> 00:10:30.419 now, AI just gave you text. And you had to copy 00:10:30.419 --> 00:10:32.159 and paste it, run it. You were the gatekeeper. 00:10:32.779 --> 00:10:34.720 Agents have permissions. They can access your 00:10:34.720 --> 00:10:36.720 terminal. They can edit your files. They can 00:10:36.720 --> 00:10:40.159 run shell commands. Whoa. So theoretically, we're 00:10:40.159 --> 00:10:42.360 talking about a tool that can not only suggest 00:10:42.360 --> 00:10:45.679 a fix, but actually go into my directory, delete 00:10:45.679 --> 00:10:47.740 a file, and rewrite it while I'm making coffee. 00:10:47.960 --> 00:10:50.220 That is the reality of tools like Claude Code. 00:10:50.669 --> 00:10:53.049 It reads your entire repository. It can spin 00:10:53.049 --> 00:10:55.649 up subagents to work in parallel. The source 00:10:55.649 --> 00:10:58.429 compares it to a junior engineer that never sleeps. 00:10:58.730 --> 00:11:01.529 It is working in the system, not just talking 00:11:01.529 --> 00:11:03.610 about it. That sounds incredible for productivity, 00:11:03.889 --> 00:11:06.990 but brain with hands. My first thought is, what 00:11:06.990 --> 00:11:09.690 if it slips? And that is the trade -off. So the 00:11:09.690 --> 00:11:11.370 trade -off for this autonomy is higher risk. 00:11:11.590 --> 00:11:14.350 Much higher risk. If a chatbot gives you a wrong 00:11:14.350 --> 00:11:17.149 answer, you just ignore it. It's text on a screen. 00:11:17.389 --> 00:11:20.529 If an agent with terminal access is wrong, It 00:11:20.529 --> 00:11:22.730 breaks the build. Or worse, it could delete the 00:11:22.730 --> 00:11:24.769 database. Which is why you can't just let these 00:11:24.769 --> 00:11:27.669 things run wild. You need a referee. Exactly. 00:11:27.990 --> 00:11:30.029 And that sets up our next segment perfectly. 00:11:30.509 --> 00:11:33.669 If we have robots writing code at the speed of 00:11:33.669 --> 00:11:36.769 light, who is checking their homework? But first, 00:11:36.950 --> 00:11:40.490 we need to take a quick break. Okay, we are back. 00:11:40.570 --> 00:11:42.789 We just talked about agents, the brains with 00:11:42.789 --> 00:11:45.129 hands that can code while we sleep. But as you 00:11:45.129 --> 00:11:46.909 said, if they mess up, they can mess up big. 00:11:47.070 --> 00:11:49.429 So let's talk about segment E, the safety net. 00:11:49.529 --> 00:11:52.389 Right. This covers categories 5, 6, and 7 in 00:11:52.389 --> 00:11:54.350 the source. This is the review layer. Because 00:11:54.350 --> 00:11:56.730 the volume of code being generated is just exploding, 00:11:57.149 --> 00:11:59.809 human review can't keep up. You physically cannot 00:11:59.809 --> 00:12:02.610 read every line of code generated by an agent. 00:12:02.809 --> 00:12:05.830 So we need AI to check the AI's work. We do. 00:12:06.070 --> 00:12:07.970 And the source highlights a few key players. 00:12:08.519 --> 00:12:10.539 First, there's BugBot, which is actually built 00:12:10.539 --> 00:12:13.019 into Cursor. It scans your pull requests, your 00:12:13.019 --> 00:12:15.620 PRs before they even land. It's that first line 00:12:15.620 --> 00:12:17.820 of defense. And then there's CodeRabbit. I have 00:12:17.820 --> 00:12:19.919 to say, I love that name. It's a great name. 00:12:20.179 --> 00:12:22.480 But what's fascinating about CodeRabbit is that 00:12:22.480 --> 00:12:25.700 it's a learning reviewer. It's not just a static 00:12:25.700 --> 00:12:28.460 tool looking for syntax errors. You can actually 00:12:28.460 --> 00:12:31.779 give it commands in natural language. Like what? 00:12:31.879 --> 00:12:34.019 Like check that this logic handles edge cases 00:12:34.019 --> 00:12:37.399 for leap years or write unit tests for this function. 00:12:37.620 --> 00:12:39.720 So it adapts to your coding style over time. 00:12:40.059 --> 00:12:42.580 Exactly. It learns what you care about. It's 00:12:42.580 --> 00:12:45.120 like having a very patient senior engineer who 00:12:45.120 --> 00:12:48.019 reads every single line of code and gently suggests 00:12:48.019 --> 00:12:50.899 improvements. And for the really scary stuff, 00:12:51.120 --> 00:12:53.940 for security. That's where a tool like SNCC comes 00:12:53.940 --> 00:12:56.580 in. While CodeRabbit is looking for logic and 00:12:56.580 --> 00:12:59.940 style, Snyk is looking for vulnerabilities. Will 00:12:59.940 --> 00:13:02.620 this new code get us hacked? Are we using a dependency 00:13:02.620 --> 00:13:05.159 with a known exploit? It's the security guard 00:13:05.159 --> 00:13:07.379 at the door. Now, even with all of that, bugs 00:13:07.379 --> 00:13:10.059 still happen. Production still breaks. The source 00:13:10.059 --> 00:13:13.059 mentions a tool called Jam for debugging. How 00:13:13.059 --> 00:13:15.019 is that different from just staring at logs? 00:13:15.320 --> 00:13:18.000 Jam is all about solving the it works on my machine 00:13:18.000 --> 00:13:21.000 problem. Worst problem. The absolute worst. Usually 00:13:21.000 --> 00:13:23.279 a user sends you a screenshot and just says it's 00:13:23.279 --> 00:13:27.389 broken. Which is useless. Jam captures a video 00:13:27.389 --> 00:13:29.629 replay of their session, but, and this is the 00:13:29.629 --> 00:13:32.169 key part, it syncs that video with the network 00:13:32.169 --> 00:13:34.690 requests and console logs. So you can actually 00:13:34.690 --> 00:13:37.429 see the code failing at the exact moment the 00:13:37.429 --> 00:13:40.009 user clicks the button. Exactly. You are not 00:13:40.009 --> 00:13:42.350 guessing anymore. You are replaying the crime 00:13:42.350 --> 00:13:45.190 scene. And then finally, there's a tool called 00:13:45.190 --> 00:13:48.029 Mintlify for documentation. Because no one likes 00:13:48.029 --> 00:13:50.610 writing docs. No one. And no one keeps them updated. 00:13:51.090 --> 00:13:53.710 Metalify reads your code and keeps the documentation 00:13:53.710 --> 00:13:56.769 in sync automatically. You change the code, the 00:13:56.769 --> 00:13:59.250 docs update. Simple as that. It really feels 00:13:59.250 --> 00:14:01.750 like we're automating all the chores. You know, 00:14:01.850 --> 00:14:05.009 debugging, reviewing documents. Precisely. We're 00:14:05.009 --> 00:14:06.830 turning the parts of coding that nobody likes, 00:14:06.909 --> 00:14:09.990 the maintenance, the bureaucracy, into background 00:14:09.990 --> 00:14:12.029 infrastructure that's just handled by AI. And 00:14:12.029 --> 00:14:14.210 that frees up the human to focus on the actual 00:14:14.210 --> 00:14:16.330 problem solving. But that brings us to the final 00:14:16.330 --> 00:14:18.809 and I think maybe the most important point, segment 00:14:18.809 --> 00:14:21.690 F. The source ends with this warning about the 00:14:21.690 --> 00:14:24.610 human element. It says, quote, AI generates code, 00:14:24.769 --> 00:14:28.149 not judgment. And this is so critical. You cannot 00:14:28.149 --> 00:14:30.759 automate judgment. If you don't understand the 00:14:30.759 --> 00:14:33.220 fundamentals, you won't be able to spot the hallucinations. 00:14:33.379 --> 00:14:36.019 You won't know why the agent chose one architecture 00:14:36.019 --> 00:14:39.299 over another. Right. The source mentions Scrimba's 00:14:39.299 --> 00:14:42.179 full -stack AI engineer path as a good resource 00:14:42.179 --> 00:14:45.100 here. It teaches not just how to write prompts, 00:14:45.240 --> 00:14:48.100 but how the underlying code actually works. I 00:14:48.100 --> 00:14:49.500 think this is where a lot of people get tripped 00:14:49.500 --> 00:14:51.820 up. They think the tool does the job for them. 00:14:51.940 --> 00:14:54.820 But the source outlines three specific habits 00:14:54.820 --> 00:14:57.820 to make the AI better. And I think we should 00:14:57.820 --> 00:15:00.659 break these down. this is the real how -to part 00:15:00.659 --> 00:15:03.179 for anyone listening. Yeah, these are gold. Habit 00:15:03.179 --> 00:15:07.100 number one, plan before executing. So don't just 00:15:07.100 --> 00:15:10.139 say build this app. Right. Use that plan mode 00:15:10.139 --> 00:15:13.200 in Cursor or just ask the AI flat out. Outline 00:15:13.200 --> 00:15:15.639 the steps before you write any code. It forces 00:15:15.639 --> 00:15:17.840 the model to load the right context and think 00:15:17.840 --> 00:15:19.620 logically before it just starts spitting out 00:15:19.620 --> 00:15:21.860 syntax. It's like you're treating the AI as a 00:15:21.860 --> 00:15:24.259 junior developer. You wouldn't just tell a junior, 00:15:24.379 --> 00:15:26.360 build the checkout page and then walk away. You'd 00:15:26.360 --> 00:15:28.779 plan it with them first. Exactly. Yeah. Okay. 00:15:28.860 --> 00:15:31.779 Habit number two, teach AI your project. This 00:15:31.779 --> 00:15:33.700 is about those context files you mentioned. Yep. 00:15:33.980 --> 00:15:38.039 You can create a file that's like .plot .md for 00:15:38.039 --> 00:15:40.960 clod or .cursor rules for cursor, and you just 00:15:40.960 --> 00:15:43.320 dump your tech stack in there, your coding style, 00:15:43.580 --> 00:15:46.460 your preferred libraries. You do it once, and 00:15:46.460 --> 00:15:48.480 the AI reads it every time. So I don't have to 00:15:48.480 --> 00:15:50.559 keep repeating myself. I don't have to say, remember, 00:15:50.740 --> 00:15:52.980 I used Tailwind, I used Tailwind. It just knows. 00:15:53.340 --> 00:15:56.759 It aligns the AI's output with your project's 00:15:56.759 --> 00:15:58.779 preferences automatically. And the third one. 00:15:58.899 --> 00:16:01.139 Habit number three is the magic prompt. I love 00:16:01.139 --> 00:16:04.179 a good magic prompt. It's so simple. Before writing 00:16:04.179 --> 00:16:07.059 code, ask me at least three clarifying questions. 00:16:07.399 --> 00:16:09.379 Oh, that's really smart. Why does that work so 00:16:09.379 --> 00:16:12.139 well? Because it completely flips the dynamic. 00:16:12.620 --> 00:16:16.759 It forces the AI to stop and simulate a requirements 00:16:16.759 --> 00:16:19.700 -gathering phase. It has to ask, wait, what about 00:16:19.700 --> 00:16:22.240 mobile users? What about error handling? It turns 00:16:22.240 --> 00:16:24.659 a command into a collaboration. So the secret 00:16:24.659 --> 00:16:26.860 isn't necessarily a better model. It's just better 00:16:26.860 --> 00:16:30.759 inputs from the human. 100%. Planning, context, 00:16:30.940 --> 00:16:33.220 and clarification. That's what turns a random 00:16:33.220 --> 00:16:36.019 slot machine into a reliable teammate. I love 00:16:36.019 --> 00:16:39.100 that. Okay, let's zoom out. We've covered a ton 00:16:39.100 --> 00:16:41.639 of ground. What's the big idea here? The big 00:16:41.639 --> 00:16:43.700 idea, if you look at the whole 2026 landscape, 00:16:43.940 --> 00:16:46.860 is this. We've moved from simple chat windows 00:16:46.860 --> 00:16:49.940 to these full immersive environments like Cursor. 00:16:50.120 --> 00:16:53.279 We've gone from helpful chatbots to autonomous 00:16:53.279 --> 00:16:57.340 agents with hands like... Claude Code. And because 00:16:57.340 --> 00:17:00.019 that's inherently risky, we've wrapped it all 00:17:00.019 --> 00:17:03.139 in a safety net of AI reviewers like Code Rabbit. 00:17:03.340 --> 00:17:05.859 It feels like a complete ecosystem now. It's 00:17:05.859 --> 00:17:07.680 not just a bunch of scattered tools anymore. 00:17:07.900 --> 00:17:11.160 It is. But the big idea isn't the tools themselves. 00:17:11.759 --> 00:17:14.059 It's that the tools alone don't make you a developer. 00:17:14.619 --> 00:17:18.119 The real AI native mindset is about mastering 00:17:18.119 --> 00:17:20.359 the fundamentals so you can manage these agents. 00:17:20.519 --> 00:17:22.900 You're moving from being a bricklayer to being 00:17:22.900 --> 00:17:25.180 a construction manager. That's a powerful analogy. 00:17:25.599 --> 00:17:27.539 You still need to know if the wall is straight, 00:17:27.700 --> 00:17:29.720 even if you didn't lay every single brick yourself. 00:17:30.039 --> 00:17:32.200 Exactly. If you don't know what a straight wall 00:17:32.200 --> 00:17:34.200 is supposed to look like, the AI will build you 00:17:34.200 --> 00:17:35.900 a crooked house and you won't even know until 00:17:35.900 --> 00:17:38.759 it all comes crashing down. So for everyone listening, 00:17:38.940 --> 00:17:42.180 here is your challenge. Pick just one thing we 00:17:42.180 --> 00:17:44.690 talked about today. Maybe it's checking out Google 00:17:44.690 --> 00:17:47.230 anti -gravity just because it sounds wild. Or 00:17:47.230 --> 00:17:50.410 maybe it's as simple as adopting that magic prompt. 00:17:50.490 --> 00:17:53.029 Ask me three questions and just try it for a 00:17:53.029 --> 00:17:55.210 week. See how it changes your workflow. And, 00:17:55.289 --> 00:17:58.009 you know, just hope your new AI agent doesn't 00:17:58.009 --> 00:17:59.910 decide to delete the production database while 00:17:59.910 --> 00:18:01.730 you're taking a nap. Let's hope not. Check your 00:18:01.730 --> 00:18:03.690 permissions, people. Thanks for diving in with 00:18:03.690 --> 00:18:05.390 us today. Always a pleasure. We'll see you on 00:18:05.390 --> 00:18:06.230 the next Deep Dive.