WEBVTT 00:00:00.000 --> 00:00:03.419 Anthropic just shared a truly staggering software 00:00:03.419 --> 00:00:06.440 engineering project. Yeah, the sheer scale completely 00:00:06.440 --> 00:00:08.720 blew my mind today. They migrated the entire 00:00:08.720 --> 00:00:12.439 bun runtime from Zig to Rust. Mm -hmm. That is 00:00:12.439 --> 00:00:14.580 a massive framework translation. It produced 00:00:14.580 --> 00:00:17.859 about 750 ,000 lines of code. Functional code, 00:00:18.160 --> 00:00:21.480 too. Yeah. And in just 11 days. Right. It achieved 00:00:21.480 --> 00:00:25.910 a 99 .8 % test pass rate. Which is astounding 00:00:25.910 --> 00:00:28.329 for a code base that huge. But human engineers 00:00:28.329 --> 00:00:30.690 didn't actually write this code. No, they didn't. 00:00:30.870 --> 00:00:33.789 AI agents did it. And they worked entirely in 00:00:33.789 --> 00:00:37.030 strict parallel. Intro music. Welcome to the 00:00:37.030 --> 00:00:39.450 Deep Dive. It's great to be here. Today we are 00:00:39.450 --> 00:00:42.649 unpacking Claude Code Dynamic workflows. It's 00:00:42.649 --> 00:00:45.630 a fascinating look at the future of digital labor. 00:00:45.829 --> 00:00:47.549 We really need to look under the hood carefully. 00:00:47.590 --> 00:00:50.250 We have some incredibly dense source material 00:00:50.250 --> 00:00:52.850 to explore. Here is our roadmap for the discussion 00:00:52.850 --> 00:00:55.020 today. First, we're going to explore what these 00:00:55.020 --> 00:00:57.340 workflows actually are. Then we'll see how they 00:00:57.340 --> 00:00:59.600 differ from familiar Claude features. You might 00:00:59.600 --> 00:01:01.539 already be using some of those handy features. 00:01:01.979 --> 00:01:04.219 After that, we'll discuss the massive hidden 00:01:04.219 --> 00:01:06.700 token cost trap. You definitely need to avoid 00:01:06.700 --> 00:01:09.459 that expensive pitfall. Finally, we will show 00:01:09.459 --> 00:01:12.159 you how to launch one safely. Step by step, so 00:01:12.159 --> 00:01:14.920 you don't blow your entire budget. AI is moving 00:01:14.920 --> 00:01:17.579 far beyond simply answering our basic questions. 00:01:17.739 --> 00:01:21.560 Yeah, it's now managing massive complex software 00:01:21.560 --> 00:01:24.120 development projects automatically. So let's 00:01:24.120 --> 00:01:26.099 just jump right into it. Sounds good. We already 00:01:26.099 --> 00:01:28.519 know that AI can write code. Right, we've seen 00:01:28.519 --> 00:01:31.540 that for a while now. But writing 750 ,000 lines 00:01:31.540 --> 00:01:33.819 is totally different. It requires a completely 00:01:33.819 --> 00:01:35.959 different processing architecture. We have to 00:01:35.959 --> 00:01:39.120 look at how it orchestrates these tasks. Because... 00:01:38.989 --> 00:01:42.349 a normal chat session simply cannot handle that 00:01:42.349 --> 00:01:44.450 workload. The context window would just break 00:01:44.450 --> 00:01:47.829 immediately. Exactly. Dynamic workflows are designed 00:01:47.829 --> 00:01:51.069 specifically for these massive tasks. Things 00:01:51.069 --> 00:01:53.829 like code base -wide bug hunts. Or comprehensive 00:01:53.829 --> 00:01:57.090 security audits across a whole system architecture. 00:01:57.469 --> 00:01:59.810 Tasks that are just too large for a normal session. 00:01:59.989 --> 00:02:02.290 I like to use a specific analogy for this. Oh 00:02:02.290 --> 00:02:04.849 yeah, let's hear it. Think about running a massive 00:02:04.849 --> 00:02:07.170 construction site. Okay, a construction site? 00:02:07.329 --> 00:02:09.349 Claude acts like your general contractor in this 00:02:09.349 --> 00:02:11.830 scenario. Right, the person overseeing the entire 00:02:11.830 --> 00:02:14.629 build. Exactly. It doesn't try to build the huge 00:02:14.629 --> 00:02:17.310 house alone. That would take way too long. Instead, 00:02:17.409 --> 00:02:19.909 Claude looks at your initial project prompt very 00:02:19.909 --> 00:02:22.710 carefully. Uh, it reads your requirements. Yeah. 00:02:23.069 --> 00:02:25.569 Then it creates a detailed blueprint. The source 00:02:25.569 --> 00:02:27.810 material calls this the orchestration script. 00:02:27.949 --> 00:02:30.729 Right. Then it breaks the task into much smaller 00:02:30.729 --> 00:02:33.530 parts. Like framing a single room or running 00:02:33.530 --> 00:02:36.889 the plumbing. Precisely. It basically hires specialized 00:02:36.889 --> 00:02:40.330 task workers for the job. So, different AI sub 00:02:40.330 --> 00:02:42.050 -agents handle the framing and the plumbing. 00:02:42.270 --> 00:02:44.629 And they execute their assigned components in 00:02:44.629 --> 00:02:46.689 strict parallel. They don't wait for each other 00:02:46.689 --> 00:02:49.830 to finish. Then it brings in the strict city 00:02:49.830 --> 00:02:52.650 inspectors. Ah, I see. The verification workers. 00:02:52.810 --> 00:02:55.750 Yeah. They check the work, then it hands you 00:02:55.750 --> 00:02:58.439 the keys. The keys being the final answer you 00:02:58.439 --> 00:03:01.199 receive. Everything runs completely automatically 00:03:01.199 --> 00:03:03.240 once you approve it. Let me make sure I have 00:03:03.240 --> 00:03:05.580 this process flow correct. Go for it. You give 00:03:05.580 --> 00:03:08.860 it a prompt. That prompt generates a workflow 00:03:08.860 --> 00:03:12.139 script. Right. The blueprint. Then task workers 00:03:12.139 --> 00:03:14.900 execute their specific parts. Simultaneously, 00:03:15.139 --> 00:03:17.759 yes. Then verification workers review it. Exactly. 00:03:17.919 --> 00:03:20.560 And finally, it produces the final answer. You 00:03:20.560 --> 00:03:22.840 nailed it. What triggers the verification workers 00:03:22.840 --> 00:03:26.169 to step in? The workflow script defines specific 00:03:26.169 --> 00:03:29.370 checkpoints for mandatory review. So reviewers 00:03:29.370 --> 00:03:31.509 check the work before moving to the next stage. 00:03:31.610 --> 00:03:34.370 Right. And this ensures absolute quality across 00:03:34.370 --> 00:03:36.669 the massive code base. You can't have hallucinated 00:03:36.669 --> 00:03:39.250 code breaking the entire migration project. It 00:03:39.250 --> 00:03:42.289 acts as an automated quality gate. Two secs silence. 00:03:43.090 --> 00:03:45.689 Understanding this sheer scale raises a very 00:03:45.689 --> 00:03:48.620 important question. Which is? Why do we actually 00:03:48.620 --> 00:03:51.080 need a brand new feature today? That's a very 00:03:51.080 --> 00:03:53.219 logical question to ask right now. Couldn't we 00:03:53.219 --> 00:03:55.740 just use the AI tools we already have? The source 00:03:55.740 --> 00:03:58.340 material breaks down these specific differences 00:03:58.340 --> 00:04:01.580 carefully. Let's unpack that. What about standard 00:04:01.580 --> 00:04:05.610 AI skills? Well, skills provide a reliable, repeatable 00:04:05.610 --> 00:04:08.909 process for the AI. They use reusable instructions 00:04:08.909 --> 00:04:12.509 for very consistent daily tasks. Right. You might 00:04:12.509 --> 00:04:14.949 use a skill for standard code formatting. Or 00:04:14.949 --> 00:04:18.110 for generating weekly release notes. But workflows 00:04:18.110 --> 00:04:21.370 manage the execution of much larger processes. 00:04:22.129 --> 00:04:24.389 So skills are just a repeatable process. Yeah. 00:04:24.430 --> 00:04:27.050 For a single worker, workflows manage the whole 00:04:27.050 --> 00:04:30.480 crew. OK, what about subagents? Subagents are 00:04:30.480 --> 00:04:33.360 really good for a few separate side tasks. Like 00:04:33.360 --> 00:04:37.139 asking an agent to check an API key. Right. Workflows 00:04:37.139 --> 00:04:39.779 coordinate many related tasks together at once. 00:04:39.899 --> 00:04:42.079 They handle a much larger scope of work overall. 00:04:42.360 --> 00:04:44.339 Exactly. Wait, I want to push back on something 00:04:44.339 --> 00:04:47.060 here. Sure, go ahead. Isn't an agent team basically 00:04:47.060 --> 00:04:49.639 the exact same thing? How do you mean? Well, 00:04:49.819 --> 00:04:52.160 it's just multiple AI minds working together 00:04:52.160 --> 00:04:54.779 on something. It sounds similar, but the focus 00:04:54.779 --> 00:04:56.839 is totally different. OK, break that down for 00:04:56.839 --> 00:04:59.920 me. Agent teams focus heavily on collaboration 00:04:59.920 --> 00:05:03.819 and different viewpoints. So they debate the 00:05:03.819 --> 00:05:07.100 best path forward. Yeah. They share complex information 00:05:07.100 --> 00:05:09.720 to solve a problem collaboratively. But dynamic 00:05:09.720 --> 00:05:12.569 workflows don't do that. No. Workflows focus 00:05:12.569 --> 00:05:15.990 purely on strict coordination and rapid execution. 00:05:16.209 --> 00:05:18.930 They divide and conquer in strict parallel paths. 00:05:19.370 --> 00:05:21.850 Exactly. They aren't debating philosophy. They 00:05:21.850 --> 00:05:25.170 just execute the plan. So teams are for collaboration. 00:05:25.689 --> 00:05:28.370 Workflows are for coordination. You got it perfectly. 00:05:28.750 --> 00:05:31.170 What about the slash goal command feature? The 00:05:31.170 --> 00:05:34.569 slash goal command loops continuously until a 00:05:34.569 --> 00:05:37.230 condition is met. It focuses strictly on reaching 00:05:37.230 --> 00:05:40.509 one very specific outcome. Like fixing a difficult 00:05:40.509 --> 00:05:42.970 bug through trial and error. Workflows don't 00:05:42.970 --> 00:05:45.470 loop endlessly like that. No, they execute a 00:05:45.470 --> 00:05:48.410 specific parallel plan across independent parts. 00:05:48.670 --> 00:05:52.029 Can a workflow use a skill? Yes. Individual workers 00:05:52.029 --> 00:05:55.389 use them during their assigned tasks. Yes, individual 00:05:55.389 --> 00:05:58.029 workers can use skills for specific instructions. 00:05:58.250 --> 00:06:00.850 It creates a very powerful combination of different 00:06:00.850 --> 00:06:03.769 AI tools. It leverages everything Claude knows 00:06:03.769 --> 00:06:05.810 how to do simultaneously. That's why it's so 00:06:05.810 --> 00:06:08.009 incredibly effective. Now we've isolated the 00:06:08.009 --> 00:06:09.829 core differentiator of this new feature. It's 00:06:09.829 --> 00:06:12.310 all about parallel execution at a massive scale. 00:06:12.470 --> 00:06:14.250 We should probably define parallel execution 00:06:14.250 --> 00:06:16.649 quickly. Good idea. It means running multiple 00:06:16.649 --> 00:06:19.670 independent tasks at the exact same time. Right. 00:06:19.790 --> 00:06:22.149 It removes the bottleneck of sequential processing 00:06:22.149 --> 00:06:26.009 entirely. Traditional AI works step by step in 00:06:26.009 --> 00:06:28.970 a straight line. But workflows expand a task 00:06:28.970 --> 00:06:31.970 into multiple simultaneous paths. Let's look 00:06:31.970 --> 00:06:33.829 at what that actually looks like in practice. 00:06:34.060 --> 00:06:36.399 When is it actually worth your time to use this? 00:06:36.519 --> 00:06:38.920 Yeah, when should I spin up a dynamic workflow? 00:06:39.120 --> 00:06:42.100 They provide the most value for very large, divisible 00:06:42.100 --> 00:06:44.779 tasks. The project must be easily broken into 00:06:44.779 --> 00:06:47.740 standalone parts. Exactly. Large code -based 00:06:47.740 --> 00:06:50.480 reviews are a perfect use case here. You might 00:06:50.480 --> 00:06:53.079 need to review hundreds of files across modules. 00:06:53.360 --> 00:06:55.720 Claude can analyze multiple modules at the exact 00:06:55.720 --> 00:06:58.100 same time. Then it combines all those findings 00:06:58.100 --> 00:07:00.839 into a single report. It's incredibly efficient. 00:07:01.180 --> 00:07:02.779 Skill audits are another really great example 00:07:02.779 --> 00:07:04.879 from the source. Yeah, that's a very practical 00:07:04.879 --> 00:07:07.079 application for developers. You might need to 00:07:07.079 --> 00:07:10.139 evaluate many skills using identical criteria. 00:07:10.459 --> 00:07:13.019 Each skill gets reviewed completely independently 00:07:13.019 --> 00:07:15.220 by different workers. Then the system combines 00:07:15.220 --> 00:07:17.839 the results perfectly. It keeps your AI tools 00:07:17.839 --> 00:07:21.060 sharp and optimized. Multi -part research tasks 00:07:21.060 --> 00:07:23.360 fit this parallel model perfectly well, too. 00:07:23.480 --> 00:07:25.879 You can have agents scouring different databases 00:07:25.879 --> 00:07:29.629 simultaneously. One checks API docs? Well, another 00:07:29.629 --> 00:07:32.370 checks legacy code bases. Then they combine their 00:07:32.370 --> 00:07:35.209 findings into a cohesive report. Beat. Let's 00:07:35.209 --> 00:07:37.449 bring back that bun migration example from earlier. 00:07:37.569 --> 00:07:40.829 Right. Sick to rust. This involved huge framework 00:07:40.829 --> 00:07:44.069 and language syntax changes. It required hundreds 00:07:44.069 --> 00:07:46.709 of agents running entirely in strict parallel. 00:07:47.050 --> 00:07:50.699 Whoa. Imagine scaling to hundreds of agents working 00:07:50.699 --> 00:07:53.060 simultaneously. It completely changes how we 00:07:53.060 --> 00:07:55.860 think about manual tech labor. Framework migrations 00:07:55.860 --> 00:07:58.720 usually require massive amounts of human engineering 00:07:58.720 --> 00:08:02.300 effort. API upgrades normally take weeks of tedious 00:08:02.300 --> 00:08:05.560 manual review. Imagine manually rewriting every 00:08:05.560 --> 00:08:08.420 single API call in a project. Humans get tired, 00:08:08.480 --> 00:08:10.839 they make typos, they lose focus easily. But 00:08:10.839 --> 00:08:13.540 the AI agents maintain perfect consistency across 00:08:13.540 --> 00:08:16.060 hundreds of files. And they tackle the boring 00:08:16.060 --> 00:08:19.060 parts of software engineering in Why are file 00:08:19.060 --> 00:08:21.360 migrations a perfect fit for this? Because Claude 00:08:21.360 --> 00:08:23.379 can analyze different directories without getting 00:08:23.379 --> 00:08:26.379 confused. Because Claude can analyze multiple 00:08:26.379 --> 00:08:29.439 affected areas all at the same time. It saves 00:08:29.439 --> 00:08:32.320 countless hours of painful, repetitive manual 00:08:32.320 --> 00:08:36.039 code checking. It frees human engineers to focus 00:08:36.039 --> 00:08:39.500 on high -level system design. Mid -roll sponsor, 00:08:39.700 --> 00:08:43.179 Read. Seeing hundreds of agents spin up is undeniably 00:08:43.179 --> 00:08:45.389 thrilling today. It really is. But this brings 00:08:45.389 --> 00:08:47.649 us to the dark side of this feature. We definitely 00:08:47.649 --> 00:08:50.370 need to discuss the hidden token cost here. It 00:08:50.370 --> 00:08:53.090 is a very serious consideration for anyone using 00:08:53.090 --> 00:08:55.730 this. This is the biggest risk with dynamic workflows 00:08:55.730 --> 00:08:59.330 by far. They can consume substantially more tokens 00:08:59.330 --> 00:09:02.289 than a normal coding session. Anthropic specifically 00:09:02.289 --> 00:09:05.210 warns users about this massive potential burn 00:09:05.210 --> 00:09:07.490 rate. It can get incredibly expensive if you 00:09:07.490 --> 00:09:09.580 aren't paying attention. We should quickly define 00:09:09.580 --> 00:09:12.440 token usage for clarity here. It's the cost units 00:09:12.440 --> 00:09:15.360 AI uses to read and write text. Exactly. It's 00:09:15.360 --> 00:09:17.940 how you get billed. Why does token usage explode 00:09:17.940 --> 00:09:20.399 so rapidly with these workflows? Well, a normal 00:09:20.399 --> 00:09:23.379 chat session uses one main context window. You 00:09:23.379 --> 00:09:25.100 just chat back and forth on a single thread. 00:09:25.240 --> 00:09:27.779 Right. But workflows are very different structurally 00:09:27.779 --> 00:09:30.309 behind the scenes. I'm sure. Multiple workers 00:09:30.309 --> 00:09:33.169 need access to the exact same initial instructions. 00:09:33.190 --> 00:09:36.190 Oh, I see. They all need the same files and overarching 00:09:36.190 --> 00:09:39.210 project context. So the same context gets sent 00:09:39.210 --> 00:09:42.269 over and over again. Yes. Dozens of workers mean 00:09:42.269 --> 00:09:44.809 massive token burn across the board. Even if 00:09:44.809 --> 00:09:47.029 the final output report is completely short. 00:09:47.389 --> 00:09:51.269 Yes. Because each worker performs its own independent, 00:09:51.529 --> 00:09:54.509 isolated analysis. Each one needs full system 00:09:54.509 --> 00:09:57.590 context to do its job. If 50 agents need a 10 00:09:57.590 --> 00:10:00.759 ,000 token file, That multiplies fast. It's like 00:10:00.759 --> 00:10:02.980 paying 50 people to read the same massive book. 00:10:03.100 --> 00:10:06.279 And every single worker reads chapter one before 00:10:06.279 --> 00:10:08.960 starting work. Wow. If your initial context window 00:10:08.960 --> 00:10:13.620 is 50 ,000 tokens long and you spin up 20 different 00:10:13.620 --> 00:10:16.419 parallel task workers simultaneously. You just 00:10:16.419 --> 00:10:18.759 consume one million tokens just setting up the 00:10:18.759 --> 00:10:21.340 board. Exactly. The math gets incredibly scary 00:10:21.340 --> 00:10:24.240 very quickly. So how do we avoid this massive 00:10:24.240 --> 00:10:27.320 token waste? The solution is defining a very 00:10:27.320 --> 00:10:30.120 narrow specific scope. I still wrestle with prompt 00:10:30.120 --> 00:10:32.639 drift myself, honestly. It's a common struggle. 00:10:33.480 --> 00:10:36.039 Defining a hyper exact scope is surprisingly 00:10:36.039 --> 00:10:38.940 difficult to do well. It's difficult but it's 00:10:38.940 --> 00:10:41.580 absolutely essential for dynamic workflows today. 00:10:41.740 --> 00:10:44.759 You must define the exact objective clearly before 00:10:44.759 --> 00:10:47.480 starting anything. You must Isolate the exact 00:10:47.480 --> 00:10:50.340 files or folders involved initially. You must 00:10:50.340 --> 00:10:53.519 declare the expected output format very precisely, 00:10:53.779 --> 00:10:56.200 too. You can't just ask Claude to, you know, 00:10:56.340 --> 00:10:58.860 fix your app. A smaller scope helps Claude create 00:10:58.860 --> 00:11:01.259 a much more focused workflow. It prevents the 00:11:01.259 --> 00:11:04.320 AI from exploring completely irrelevant code 00:11:04.320 --> 00:11:06.620 paths. Which saves your valuable tokens in the 00:11:06.620 --> 00:11:09.179 end. A well -defined workflow uses far fewer 00:11:09.179 --> 00:11:11.620 tokens than broad requests. Always ask if you 00:11:11.620 --> 00:11:14.019 really need massive parallel execution first. 00:11:14.240 --> 00:11:16.559 Sometimes a simpler feature solves the problem 00:11:16.559 --> 00:11:19.700 much cheaper anyway. Does Claude warn you before 00:11:19.700 --> 00:11:22.240 spending all those tokens? Thankfully, yes. It 00:11:22.240 --> 00:11:24.700 shows a clear confirmation screen first. Yes, 00:11:24.720 --> 00:11:26.700 Claude always shows a confirmation screen before 00:11:26.700 --> 00:11:29.240 the workflow actually starts. It estimates the 00:11:29.240 --> 00:11:31.899 complexity and asks for your explicit permission. 00:11:32.600 --> 00:11:34.759 That's a very thoughtful design choice by the 00:11:34.759 --> 00:11:37.500 Anthropic team. It ensures you never spend tokens 00:11:37.500 --> 00:11:40.399 completely by accident. We know the massive token 00:11:40.399 --> 00:11:43.100 burn risks involved with this now. Right? How 00:11:43.100 --> 00:11:45.059 do we practically put our hands on the keyboard? 00:11:45.220 --> 00:11:47.340 You want to know how to trigger one safely. Yeah. 00:11:47.759 --> 00:11:49.919 How do we do this without blowing our budget? 00:11:50.480 --> 00:11:53.340 First, you don't write the workflow scripts yourself 00:11:53.340 --> 00:11:57.000 manually. Claude builds the entire workflow automatically 00:11:57.000 --> 00:12:00.159 for you. It bases everything on your clear initial 00:12:00.159 --> 00:12:02.519 project description prompt. You just tell it 00:12:02.519 --> 00:12:04.940 exactly what you want to achieve. Let's walk 00:12:04.940 --> 00:12:07.659 through a concrete example from the source material. 00:12:07.759 --> 00:12:10.139 Okay, what's the scenario? We want to audit a 00:12:10.139 --> 00:12:13.019 library of local Claude skills. Those are usually 00:12:13.019 --> 00:12:15.740 stored in the .Claude slash skills directory. 00:12:16.100 --> 00:12:19.659 Exactly. Step one is defining the strict project 00:12:19.659 --> 00:12:23.399 scope first. I can tell Claude to identify duplicated 00:12:23.399 --> 00:12:26.159 instructions across files. Or you might ask it 00:12:26.159 --> 00:12:28.559 to find outdated guidance there. And I must ask 00:12:28.559 --> 00:12:31.240 for a report with specific actionable recommendations. 00:12:31.580 --> 00:12:34.960 Right. Step two is letting Claude build the workflow 00:12:34.960 --> 00:12:38.179 automatically. So Claude handles all that complex 00:12:38.179 --> 00:12:40.519 orchestration behind the scenes. You don't design 00:12:40.519 --> 00:12:43.320 the execution process manually at all here. It 00:12:43.320 --> 00:12:45.740 breaks the task apart and assigns the virtual 00:12:45.740 --> 00:12:48.279 workers. It writes the orchestration script based 00:12:48.279 --> 00:12:51.480 on your initial prompt. Step 3 is reviewing the 00:12:51.480 --> 00:12:54.179 execution plan carefully before proceeding. You 00:12:54.179 --> 00:12:56.240 really need to check this before you click approve. 00:12:56.509 --> 00:12:58.809 Right, and this prevents a surprising amount 00:12:58.809 --> 00:13:01.850 of wasted token usage. You can see exactly what 00:13:01.850 --> 00:13:04.649 files the agents plan to read. Make sure it isn't 00:13:04.649 --> 00:13:06.769 analyzing more than strictly necessary. If it 00:13:06.769 --> 00:13:08.950 wants to read your node module's folder, stop 00:13:08.950 --> 00:13:11.809 it immediately. Step four is reviewing the final 00:13:11.809 --> 00:13:14.669 comprehensive report it generates. This is honestly 00:13:14.669 --> 00:13:17.190 the true beauty of the entire system. It shows 00:13:17.190 --> 00:13:20.309 affected files, supporting evidence, and specific 00:13:20.309 --> 00:13:23.370 recommended fixes. The workflow synthesizes all 00:13:23.370 --> 00:13:26.049 those parallel findings into one document. Without 00:13:26.049 --> 00:13:28.809 this, you'd be opening and reviewing every single 00:13:28.809 --> 00:13:31.309 skill file manually. It highlights exactly what 00:13:31.309 --> 00:13:33.909 needs to change in your code base. Step five 00:13:33.909 --> 00:13:36.730 is saving good workflows for future repetitive 00:13:36.730 --> 00:13:38.970 use. Right. You don't want to recreate the wheel 00:13:38.970 --> 00:13:41.409 every time. Can I reuse a workflow once it's 00:13:41.409 --> 00:13:43.610 built? Absolutely. You just save it for the next 00:13:43.610 --> 00:13:45.490 time. Absolutely. Just save it to your local 00:13:45.490 --> 00:13:47.929 Claude workflows folder for next time. You don't 00:13:47.929 --> 00:13:50.230 have to generate the orchestration script from 00:13:50.230 --> 00:13:53.110 scratch. So you basically build a personal library 00:13:53.110 --> 00:13:55.990 of reusable processes. It standardizes how you 00:13:55.990 --> 00:13:58.389 audit your code across the board. Two sec silence. 00:13:59.549 --> 00:14:01.610 Let's zoom out from these tactical steps for 00:14:01.610 --> 00:14:03.889 a minute. Good idea. We need to understand the 00:14:03.889 --> 00:14:05.909 underlying philosophy of the source material. 00:14:06.190 --> 00:14:08.289 The golden rule is choosing the right feature 00:14:08.289 --> 00:14:11.190 for the job. You don't always need the most advanced 00:14:11.190 --> 00:14:13.970 tool available today. That is the core message 00:14:13.970 --> 00:14:16.789 we want to convey today. Use standard Claude 00:14:16.789 --> 00:14:20.679 code for quick, simple, daily coding tasks, use 00:14:20.679 --> 00:14:23.480 standard skills when you have repeated consistent 00:14:23.480 --> 00:14:26.720 automated tasks, use subagents when you have 00:14:26.720 --> 00:14:29.740 minor supportive side tasks, use agent teams 00:14:29.740 --> 00:14:32.799 for collaborative multi -viewpoint problem -solving 00:14:32.799 --> 00:14:35.980 sessions, use the slash goal command for iterative 00:14:35.980 --> 00:14:39.139 condition -based loop tasks, and save dynamic 00:14:39.139 --> 00:14:42.000 workflows for large truly independent parallel 00:14:42.000 --> 00:14:45.320 tasks. Dynamic workflows only make sense with 00:14:45.320 --> 00:14:48.419 massive parallel execution architectures. If 00:14:48.419 --> 00:14:51.500 tasks must happen sequentially, do not use a 00:14:51.500 --> 00:14:55.000 dynamic workflow. If step B requires step A to 00:14:55.000 --> 00:14:58.700 finish, avoid workflows entirely. Match the AI 00:14:58.700 --> 00:15:01.019 tool directly to your specific problem shape. 00:15:01.200 --> 00:15:04.120 It saves time, saves money, and produces much 00:15:04.120 --> 00:15:06.500 better results. It's all about working smarter, 00:15:06.799 --> 00:15:10.539 not harder. If AI can orchestrate hundreds of 00:15:10.539 --> 00:15:13.750 parallel agents successfully today, if it can 00:15:13.750 --> 00:15:17.590 port entire massive code bases in 11 days. With 00:15:17.590 --> 00:15:21.190 99 .8 % accuracy, no less. What does the future 00:15:21.190 --> 00:15:24.090 of a one -person tech startup look like? It looks 00:15:24.090 --> 00:15:26.070 completely different than it did just last year. 00:15:26.409 --> 00:15:28.570 Perhaps your primary job is no longer writing 00:15:28.570 --> 00:15:31.480 code manually. You transition from a manual coder 00:15:31.480 --> 00:15:33.820 to a general contractor. Perhaps your job is 00:15:33.820 --> 00:15:36.679 simply defining tight scopes for your AI workforce. 00:15:36.899 --> 00:15:39.259 You manage the grand vision while the AI swings 00:15:39.259 --> 00:15:41.360 the hammers. I want you to audit your repetitive 00:15:41.360 --> 00:15:44.120 tasks this week. Look closely at your daily work. 00:15:44.240 --> 00:15:46.740 Ask yourself, could this task be broken into 00:15:46.740 --> 00:15:49.419 parallel, independent parts? Think about how 00:15:49.419 --> 00:15:51.899 you construct your daily blueprints going forward. 00:15:52.299 --> 00:15:54.120 Thank you for joining this deep dive together 00:15:54.120 --> 00:15:56.100 with us today. It's been a great conversation. 00:15:56.519 --> 00:15:58.740 Take care and keep exploring the possibilities 00:15:58.740 --> 00:15:59.700 safely out there.