WEBVTT 00:00:00.000 --> 00:00:03.140 You type a command into an AI, nothing happens. 00:00:03.259 --> 00:00:05.919 Beat. It's not broken. It's just waiting for 00:00:05.919 --> 00:00:08.339 you to teach it how to think. We really do treat 00:00:08.339 --> 00:00:10.679 it like static software. We assume we're just 00:00:10.679 --> 00:00:12.900 passive consumers. But you can actually rewrite. 00:00:13.000 --> 00:00:15.960 It's cognitive pathways. Welcome to the deep 00:00:15.960 --> 00:00:19.500 dive. Today we're exploring five custom slash 00:00:19.500 --> 00:00:21.640 commands. Yeah, these are built specifically 00:00:21.640 --> 00:00:24.859 for the Claude Code app. But the catch is you 00:00:24.859 --> 00:00:27.019 have to build them yourself. Right. We are your 00:00:27.019 --> 00:00:30.379 hosts, and we're looking at Plan, Clear, Goal, 00:00:30.640 --> 00:00:34.420 Rewind, and BW. These commands bridge a really 00:00:34.420 --> 00:00:37.560 vital gap. They bridge the gap between AI just 00:00:37.560 --> 00:00:40.280 doing things and AI doing things your way. But 00:00:40.280 --> 00:00:42.640 before we build them, we need to understand a 00:00:42.640 --> 00:00:44.740 common frustration. You watched a developer demo 00:00:44.740 --> 00:00:47.119 on YouTube? Oh, yeah. They attack a slash command. 00:00:47.159 --> 00:00:49.820 It runs perfectly. Right. You try it on your 00:00:49.820 --> 00:00:52.799 machine, and nothing happens. People assume their 00:00:52.799 --> 00:00:55.460 software is just outdated, or they missed a secret 00:00:55.460 --> 00:00:57.799 plugin. But let's look at what a slash command 00:00:57.799 --> 00:01:00.060 actually is. It starts with a forward slash. 00:01:00.179 --> 00:01:03.159 It triggers an action instantly. Exactly. But 00:01:03.159 --> 00:01:06.459 the secret here is it's not native software. 00:01:07.099 --> 00:01:10.099 It's literally just a plain markdown file. We 00:01:10.099 --> 00:01:12.260 call it a skill. And it lives locally on your 00:01:12.260 --> 00:01:15.719 machine. Yeah. Claude reads that exact file when 00:01:15.719 --> 00:01:18.359 you type the command. That is why lists vary 00:01:18.359 --> 00:01:20.459 between users. I mean, does this mean we are 00:01:20.459 --> 00:01:23.159 basically programming the AI's internal habits 00:01:23.159 --> 00:01:26.420 from scratch? Yes, absolutely. You are dictating 00:01:26.420 --> 00:01:28.260 its operational boundaries. You write them in 00:01:28.260 --> 00:01:31.000 plain English. So they were just text files directing 00:01:31.000 --> 00:01:33.219 the AI's behavior behind the scenes. Exactly, 00:01:33.400 --> 00:01:36.280 like stacking Lego blocks of data. So let's look 00:01:36.280 --> 00:01:38.340 at the first command, the plan command. Because 00:01:38.340 --> 00:01:40.439 Claude has a tendency to rush, right? Yeah, the 00:01:40.439 --> 00:01:43.939 problem is Claude starts coding instantly. On 00:01:43.939 --> 00:01:46.340 big tasks, it picks the wrong approach. It wastes 00:01:46.340 --> 00:01:48.500 so much time. It's like a driver accelerating 00:01:48.500 --> 00:01:50.200 before checking the map. That's a great way to 00:01:50.200 --> 00:01:53.109 put it. The planned skill fixes this. It forces 00:01:53.109 --> 00:01:56.109 Claude to stop entirely. It forces it to read 00:01:56.109 --> 00:01:58.590 the task first. Right. It breaks the task into 00:01:58.590 --> 00:02:01.370 steps, it lists the affected files, it flags 00:02:01.370 --> 00:02:04.090 risks, and then it waits for your approval. So 00:02:04.090 --> 00:02:06.689 like building an event signup page. Exactly. 00:02:06.849 --> 00:02:10.629 Using HTML and CSS, Claude stops and asks question. 00:02:10.849 --> 00:02:13.050 It asks if you need email validation. And it 00:02:13.050 --> 00:02:15.430 does this before writing a single line of code. 00:02:15.610 --> 00:02:19.669 Yeah. But exception here. You skip this for one 00:02:19.669 --> 00:02:21.990 line, one file fixes. Right. Why shouldn't we 00:02:21.990 --> 00:02:24.050 just type out these instructions manually every 00:02:24.050 --> 00:02:26.250 time? Well, because human discipline fades, you'll 00:02:26.250 --> 00:02:29.110 forget, and then the AI messes up. Typing it 00:02:29.110 --> 00:02:31.870 repeatedly wastes your time. This automation 00:02:31.870 --> 00:02:35.009 saves your sanity. Perfectly said. Once you finish 00:02:35.009 --> 00:02:38.250 that task, you have a new problem. The AI memory 00:02:38.250 --> 00:02:41.169 becomes a liability. We should quickly define 00:02:41.169 --> 00:02:43.909 context window here. The memory space where the 00:02:43.909 --> 00:02:47.180 AI reads your current... Chat history. Spot on. 00:02:47.219 --> 00:02:50.020 The problem is old context slows down responses. 00:02:50.360 --> 00:02:52.740 Claude tries to connect totally unrelated tasks. 00:02:52.879 --> 00:02:55.360 Like debugging a login form today, then switching 00:02:55.360 --> 00:02:57.599 to writing a course product description. Exactly. 00:02:57.620 --> 00:02:59.439 You do not want code in your marketing copy. 00:02:59.680 --> 00:03:02.319 The clear skill fixes this. It wipes the slate 00:03:02.319 --> 00:03:04.360 clean. It tells Claude to treat the prompt as 00:03:04.360 --> 00:03:06.900 brand new. Yeah. It ignores past files completely. 00:03:06.979 --> 00:03:09.580 But there's a crucial detail here. New commands 00:03:09.580 --> 00:03:11.979 only load when a session starts. So you have 00:03:11.979 --> 00:03:14.639 to start fresh. Right. You must open a new chat. 00:03:14.830 --> 00:03:17.569 Then the clear command will appear. Does the 00:03:17.569 --> 00:03:21.150 AI naturally realize when a topic has completely 00:03:21.150 --> 00:03:24.969 changed? Uh, no, not at all. It lacks that awareness. 00:03:25.069 --> 00:03:27.330 It just tries to blend everything. No, it assumes 00:03:27.330 --> 00:03:29.330 everything connects together unless you explicitly 00:03:29.330 --> 00:03:34.189 tell it otherwise. Sponsor. Beep. We're back. 00:03:35.150 --> 00:03:38.460 So we've started a clean task. But how do we 00:03:38.460 --> 00:03:41.300 stop the AI from getting lazy? Yeah, this is 00:03:41.300 --> 00:03:43.680 the huge issue. Instructions usually say what 00:03:43.680 --> 00:03:46.300 to do. They don't specify when it counts as done. 00:03:46.599 --> 00:03:48.740 So Claude just stops halfway through. Exactly. 00:03:48.919 --> 00:03:51.219 And it requires you to constantly reprompt it. 00:03:51.580 --> 00:03:54.000 The goal skill changes this. It establishes a 00:03:54.000 --> 00:03:56.039 clear finish line. It gives it a strict condition 00:03:56.039 --> 00:03:58.500 to meet. Right. Claude works through the steps. 00:03:58.840 --> 00:04:00.240 It checks the result against your condition. 00:04:00.539 --> 00:04:02.860 It does this after every single step. And it 00:04:02.860 --> 00:04:05.759 only stops when fully met. Yeah. Beat. Whoa. 00:04:06.360 --> 00:04:08.340 I mean, imagine it iterating through 50 steps 00:04:08.340 --> 00:04:10.180 overnight, perfectly checking its own work. Like 00:04:10.180 --> 00:04:13.000 building a landing page. Exactly. Where it done 00:04:13.000 --> 00:04:16.639 strictly means a hero section, one CTA, and mobile 00:04:16.639 --> 00:04:19.100 responsive. But wait, how does it actually know 00:04:19.100 --> 00:04:21.100 it hit the goal without our human eyes on it? 00:04:21.180 --> 00:04:24.560 It runs internal verification loops against your 00:04:24.560 --> 00:04:28.920 text. It reads its own output like a human reviewer. 00:04:29.079 --> 00:04:31.360 It constantly checks its own output against your 00:04:31.360 --> 00:04:34.300 strict definition of done. Exactly. But even 00:04:34.300 --> 00:04:37.240 with clear goals, things go wrong. Sometimes 00:04:37.240 --> 00:04:40.480 the AI takes a completely wrong turn. We need 00:04:40.480 --> 00:04:43.560 an undo button that works. Because Claude can 00:04:43.560 --> 00:04:46.939 make a really bad mid -task edit, and rolling 00:04:46.939 --> 00:04:49.939 back code by hand takes forever. I still wrestle 00:04:49.939 --> 00:04:52.639 with prompt drift myself. Sometimes I let it 00:04:52.639 --> 00:04:55.019 go too far before fixing it. We all do, man. 00:04:55.259 --> 00:04:57.399 That's why the rewind skill is amazing. It lists 00:04:57.399 --> 00:04:59.740 recent session checkpoints. With short little 00:04:59.740 --> 00:05:02.019 descriptive labels, right? Yeah. You pick one, 00:05:02.319 --> 00:05:04.540 and it rolls back both the conversation and the 00:05:04.540 --> 00:05:06.720 code. It takes everything back to that exact 00:05:06.720 --> 00:05:09.170 state. So if Claude restructures files in a way 00:05:09.170 --> 00:05:11.569 you didn't ask for? You just rewind to the checkpoint 00:05:11.569 --> 00:05:13.470 right before it happened. Does rewinding mean 00:05:13.470 --> 00:05:16.410 we also lose any genuinely good code it wrote 00:05:16.410 --> 00:05:19.189 after that checkpoint? Uh, yes, unfortunately. 00:05:19.589 --> 00:05:21.769 Rolling back wipes the slate clean moving forward. 00:05:21.829 --> 00:05:24.310 You have to act fast. Yes. Rolling back wipes 00:05:24.310 --> 00:05:26.209 everything after that point, so you must act 00:05:26.209 --> 00:05:29.430 quickly. Right. So say the task is flowing nicely 00:05:29.430 --> 00:05:32.149 now, but suddenly a random thought pops into 00:05:32.149 --> 00:05:35.170 your head. How do you ask? without derailing 00:05:35.170 --> 00:05:38.089 the AI. That's the problem. Asking a side question 00:05:38.089 --> 00:05:41.230 breaks the flow entirely. Claude mixes the answer 00:05:41.230 --> 00:05:43.850 into the main task's code. It pollutes the context 00:05:43.850 --> 00:05:47.829 window. Exactly. The B2 skill is the fix. It 00:05:47.829 --> 00:05:49.990 answers the question directly. It doesn't affect 00:05:49.990 --> 00:05:52.810 ongoing files at all. And then it returns focus 00:05:52.810 --> 00:05:55.490 to the main task. Yeah. Like, say you are coding 00:05:55.490 --> 00:05:58.649 a landing page. You ask for SEO image naming 00:05:58.649 --> 00:06:01.740 patterns. It isolates that query. Right. It answers 00:06:01.740 --> 00:06:03.540 it and acts like the interruption never even 00:06:03.540 --> 00:06:05.500 happened. What happens if you actually want the 00:06:05.500 --> 00:06:07.579 side question to influence the code? Oh, well, 00:06:07.620 --> 00:06:09.839 then just ask it as a normal prompt. Only use 00:06:09.839 --> 00:06:12.040 this for isolated thoughts. Then just ask it 00:06:12.040 --> 00:06:14.819 normally. Only use this command to keep tangents 00:06:14.819 --> 00:06:18.480 completely isolated. Exactly. So to recap our 00:06:18.480 --> 00:06:23.220 big idea today, these five tools, Plan, Clear, 00:06:23.620 --> 00:06:26.740 Goal, Rewind, and B2. They don't exist until 00:06:26.740 --> 00:06:28.500 you build them. Right. And you definitely don't 00:06:28.500 --> 00:06:30.779 need all five today. Just pick the one your workflow 00:06:30.779 --> 00:06:33.259 misses most. Save the markdown file and try it. 00:06:33.379 --> 00:06:35.899 Yeah, exactly. See how it changes your workflow. 00:06:36.439 --> 00:06:38.740 Thank you for listening today. We spend so much 00:06:38.740 --> 00:06:41.720 time marveling at what AI can do. We really do. 00:06:42.060 --> 00:06:44.120 But perhaps the real leap forward is realizing 00:06:44.120 --> 00:06:47.259 we have the power to shape exactly how it thinks. 00:06:47.959 --> 00:06:50.620 You are no longer just a user. You are the architect 00:06:50.620 --> 00:06:53.079 of your own AI. Outtio Music.