I Quit Claude Code. Here's What I Use Now — Transcript

Full transcript

1. 0:00Everyone's asking the same question since I quit Claude Code.
2. 0:02What do you use now?
3. 0:03There's two parts to the answer of that question,
4. 0:05which I'm going to cover in this video.
5. 0:07For the first half of this video, we'll talk the what and the why,
6. 0:10and then the second half, we'll build.
7. 0:11So you can see the general idea for the scaffold of workspace
8. 0:14that will work as well or almost as well as Claude Code
9. 0:17using much cheaper intelligence.
10. 0:18And every prompt you see me use,
11. 0:20you'll be able to download the template for free.
12. 0:22By the end of this video,
13. 0:22you should have a roadmap of how to approach
14. 0:25leaving the Claude Code trap if it's right for you.
15. 0:28We're also going to explore how to decide if it's right for you, because it's not going to be for
16. 0:33everyone. A big part of the conclusions I've drawn are because of, I believe, a narrative being put
17. 0:39out by the world, by AI, that we need to keep up with token maxing the best frontier models in
18. 0:44order to stay on top of things. I think there is some propaganda in there that is not actual reality,
19. 0:50because when you have all these tokens to just build, build, build, how much are you actually
20. 0:53shipping. How much that your shipping is actually bringing value back in the form of revenue,
21. 0:59for example. So there's many ways to approach this whole dialogue. I want to share that all
22. 1:03this is from my perspective. I've been doing the agentic AI stuff for about two years now,
23. 1:08really hardcore in it. I'm not a developer. I'm not an engineer. I'm just a builder. So everything
24. 1:12I share on this channel is really about what I've learned in my experience and my honest take and
25. 1:16recommendations. This is what I've been learning in my 30 days now without Claude Code. And this
26. 1:21leads us to part one of the answer, what am I using instead? Well, first we need a brain,
27. 1:25intelligence. I'm personally using a Venice API key, and that leads to another question you guys
28. 1:30were asking in the comments. Do I work for Venice because that video felt like an ad? Yes, I work
29. 1:34for Venice, but no, Venice was not associated with that video at all. I'm just a huge believer
30. 1:39in the product. I was making videos about Venice before they ever decided to hire me. Before that
31. 1:44ever even came up, you can check my channel for the proof. You can also use OpenRouter. You can
32. 1:49use Requesty. You can use PPQ, NanoGPT. You can go straight through Kimi, straight through DeepSeek.
33. 1:54Whatever provider you want to use, you need intelligence. And that's part one. Where are you
34. 1:57getting your intelligence from? How much does it cost? Does it offer you what you need for your
35. 2:01system? That's part one. You need intelligence. Part two is pretty much everything else. Cloud
36. 2:05code isn't just intelligence. You can chat with intelligence like a chat box. Claude Code is
37. 2:09agentic. So you need a harness, a runtime environment for the AI to do things, a sandbox,
38. 2:14if you will. And that was my favorite part of the last couple of months was the agents getting out
39. 2:19of the sandbox memes. So in this video, we're going to play with OpenCode, but you can also use Agent Zero,
40. 2:24or you could use DeepSeek V4 Flash. There's plenty of open-source alternatives out there to Claude Code.
41. 2:29I happen to like OpenCode, so that's what I'm going to show to you. But just like the provider,
42. 2:33you can choose whatever you want. The point is how do we get closer to achieving the results that we
43. 2:37get with Claude Code without using Claude Code. The difference here is you're going to be paying for
44. 2:41the inference instead of having it subsidized by that subscription cost, which, as I mentioned in
45. 2:47that last video is not going to be here forever. We already saw Claude start changing the limits on
46. 2:51us. Limits are going down. The writing is on the wall. These companies have to figure out how to
47. 2:56be profitable eventually. So we need to get ahead of the game. So you have your intelligence. You
48. 3:00need your open-source harness. And I'm going to show you how to do all this in just a little bit.
49. 3:04But what's the last crucial part of this conversation? Is canceling Claude Code the
50. 3:08right move for you? Okay, let me split myself up into two people right now. Person A and person B.
51. 3:13Person A is a solopreneur, entrepreneur, small business owner that, yes, I do build apps sometimes,
52. 3:20websites, things, but most of my AI usage is actually just automated workflows.
53. 3:25So I need good intelligence to run through my AI software, if you will, my harness workspace,
54. 3:31which has skills, knowledge base, all the important stuff to do a good job on my specific tasks in my
55. 3:37business, in my projects, in my YouTube channel, whatever. I just need to run them. I don't need
56. 3:42to be building them. But guess how I built all of these? With Claude Code, with subsidized inference.
57. 3:47So I use the frontier models to build my software, and then I just use the regular open-weight models
58. 3:54to run them. So I'm blessed to be in a position where I already token maxed all that time with
59. 3:59Claude to build the things I use. And now as person A with my business, I can just run them
60. 4:05theoretically cheaper than the $200 a month, as long as I'm not building new things all the time.
61. 4:10But I can run complex repurposing workflows, script writing, work on the new membership
62. 4:14software to replace the Skool academy.
63. 4:16I can work on all this stuff with open-weight models, pay for what I consume, and I'm being
64. 4:21sustainable with the resources.
65. 4:23It's not reality to get a free lunch, like all the tokens, every single day.
66. 4:28And that goes back to the psychological behavioral aspect of all of this.
67. 4:32Training your behavior to think, oh, I could just use 500 million tokens in a month.
68. 4:37Oh, I can just use 500 million tokens in a day because it's within my five-hour limit.
69. 4:41Twice.
70. 4:42That's not reality.
71. 4:43It's a dreamland.
72. 4:44It's fantasy land.
73. 4:49And why is this important?
74. 4:50This goes back to the propaganda I was talking about earlier.
75. 4:52Oh, if you don't use the AI, you're not going to keep up.
76. 4:55If you don't build your business, you're not going to keep up.
77. 4:58If you're not generating more revenue, you're not going to keep up.
78. 5:02You need to be doing something and focused on the prize.
79. 5:04Alex Hormozi talked about this recently.
80. 5:07just because you're doing more with AI,
81. 5:09does that mean you're actually moving your business forward?
82. 5:11That's what I learned as person A.
83. 5:13I was doing more, shipping less, getting less done,
84. 5:16just distracted by all the shiny objects I could create
85. 5:20with all these essentially free tokens.
86. 5:22This is the psychological aspect of all of this.
87. 5:25It's very American consumer-like, if you know what I mean.
88. 5:29More, more, more, bigger, bigger, bigger,
89. 5:31faster, faster, faster, that's obviously better.
90. 5:33But the truth is, as I've learned as person A,
91. 5:35that that's not the case.
92. 5:36going deeper into a project or a task and giving it my creative genius that only I have as a human.
93. 5:44AI doesn't have that. Not yet anyway. They don't innovate like we do. But that is the value I bring
94. 5:49to these projects. So I don't need all the extra tokens. I need me. Then I need AI to automate the
95. 5:54labor. AI is that leverage for labor now. Okay, so that's person A. Meanwhile, there's person B.
96. 5:59Say you do have a job where you're building all the time. Look, back here, I'm building an AI
97. 6:03editor. Now this is for my job, okay? This is not for person A running my small business, my side
98. 6:09hustle, if you will. This is for the job where I need the tokens all the time. I'm using frontier
99. 6:13models because I need the best of the best to create software that's like kind of crazy and
100. 6:17cutting edge, like next level stuff. So that's when I need to build. And then eventually when
101. 6:22the open-weight models catch up, I'll use open-weight models to run the AI editing software that
102. 6:27I make. But we're not there yet. So person B does need these tokens. It does need to build. But that
103. 6:32can also just be a phase. So at times I can be person B and then I become person A again. We don't
104. 6:39always need to be building things non-stop even if it's fun. I have so many little side projects that
105. 6:43I've just been left out to try because I had the tokens to try it. So I thought why not? But really
106. 6:48it was just a distraction from the goals. So person A, person B, what am I really trying to say here?
107. 6:53They're cycles. They're not identities to be attached to all the time. Sometimes you're person A,
108. 6:57sometimes you're person B. I think maybe for the rare person, yes. If you're a software developer
109. 7:01or whatever, probably shouldn't be watching this.
110. 7:03But if you're making things all the time,
111. 7:05you have to be making new things,
112. 7:06then yeah, a Claude Code or Codex subscription makes sense.
113. 7:09But for people like person A,
114. 7:11the small to medium-sized business owners
115. 7:13that aren't building things all the time,
116. 7:14you really don't need it
117. 7:16and you can get away with it financially as well.
118. 7:18Person B, on the other hand, is in a building phase.
119. 7:22It's making things.
120. 7:23It's creating the tools necessary to run the business.
121. 7:26It's doing the thing for the client.
122. 7:27You might even have multiple projects,
123. 7:29a job where your person B,
124. 7:31and side hustles where your person A.
125. 7:34In my opinion, this is not only perfectly valid,
126. 7:36but important to segment those differences
127. 7:39in the different types of work you do.
128. 7:41If your employer or client is paying for your inference,
129. 7:43you're in a much different boat
130. 7:45than you as a small to medium-sized business owner
131. 7:48who has to pay for every credit.
132. 7:50So you need to weigh this accordingly.
133. 7:51I'm just here to open up the discussion
134. 7:54and hopefully inspire you to think about this a little bit.
135. 7:56So now let's check out how to build these workspaces
136. 7:59that can replace Claude Code.
137. 8:01Now what I will say before I begin,
138. 8:02while we're still on the person A, person B thing,
139. 8:04Build with frontier, execute with open-weight.
140. 8:07It's not a bad plan, whether you're on Claude Code,
141. 8:10or whether you're already paying per request.
142. 8:13Use the best possible models to build your workspaces.
143. 8:16So that way the open-weight models
144. 8:17that aren't as intelligent but are far more economical
145. 8:20can have state of the art instructions
146. 8:22for how to do the tasks that you need
147. 8:23to run your business, your project, whatever that is.
148. 8:26Okay, so that is an intro to all of this.
149. 8:28Now we're going to dive into creating your own harness workspace.
150. 8:31Get your API key from whatever provider you choose and we'll jump right in.
151. 8:35Okay.
152. 8:35I've shown this before on the channel.
153. 8:37This is what my workspace looks like.
154. 8:39Obviously it looks a lot like Claude Code, except we have OpenCode over here.
155. 8:43It's got different colors and I'm using a different provider.
156. 8:46Let me just show you what happens in this workspace because it is important to understand
157. 8:51just how it works in case you didn't get too deep into Claude Code.
158. 8:54You were just using it because it's pretty awesome.
159. 8:56Even if you don't know how to use it.
160. 8:58we still have a Claude folder because that Claude folder was made with Claude Code and inside we
161. 9:03have sub-agent profiles markdown files and we have skills very important and then meanwhile we have
162. 9:10the AGENTS.md file which is called CLAUDE.md with Claude but we call it AGENTS.md and if we format it
163. 9:17here it basically has all the wiring for the OpenCode agent when it starts fresh in this workspace
164. 9:23for how to do things, how to find things, how it should behave, what it is. So for example, here,
165. 9:30it has a whole deploy map. Okay. So if we're working on jordanurbs.com, this is the command
166. 9:35to deploy it to Netlify, right? The Create Own Scale, these are the different projects, right?
167. 9:40This is what folder they're in. This is their Netlify ID for rapid deployment. So as a creator,
168. 9:45solopreneur, I'm often making new landing pages, the lead magnets, email sequences, things like
169. 9:51that. So this has everything there for a personal brand. And so the AGENTS.md file explains all
170. 9:57that, explains the agents in the workspace, the skills in this workspace, blah, blah, blah. And
171. 10:00then you can see the style, the colors, everything that's new has its place, has its own color scheme,
172. 10:07for example, based on the project. And everything will render a certain way. Everything will write
173. 10:12a certain way. For example, here, this is working in the background to make B-roll for the first
174. 10:16half of the video that you just watched. So that's working while I'm recording. And that is going to
175. 10:21do it in such a way that it matches the brand. The same thing is going to happen with writing.
176. 10:26When it writes the first draft of a Substack post, for example, it'll take a YouTube video,
177. 10:30then write the first draft post. But Substack posts are started from the repurposing pipeline
178. 10:36because it can follow my voice and it can write like me. This is my voice standard,
179. 10:41how to write like Jordan Urbs' voice. And then we have an auditor. So the agent sends the draft,
180. 10:49then I edit the draft. And then the final draft, I say, hey, run the voice tune skill to refine and
181. 10:56fine tune the writing skill even more. Okay. Anyway, so if you got halfway deep in a Claude
182. 11:00code, you know, you could do this with Claude Code. None of this is special to OpenCode or agent zero
183. 11:05or anything. You can do this with any agent harness. But the point I like to make on this channel
184. 11:11is that if you want Fable-level results, you need quality agentic workspaces such as this
185. 11:18so your open-weight models like Kimi K3 Fast, which is less than half the price of Fable,
186. 11:23can do the work the right way. It also saves a ton of tokens. So if you're working on a project
187. 11:29all the time and you don't have your knowledge infrastructure wired in properly, Claude Code is
188. 11:34going to consume way more tokens, but you don't notice it because those tokens are subsidized and
189. 11:39you're not paying for it. You will notice it a lot faster when you're paying for every inference call
190. 11:43you make. You want to save those tokens however you can. Okay, so let's get into our demo. I'm using
191. 11:48VSCodium here. It's a fork of VS Code with less telemetry. You can obviously use Cursor. You can
192. 11:53use whatever you want. You could even just use a terminal. And I type in OpenCode. By the way,
193. 11:59OpenCode. Open source AI coding agent. If you want to install it, just copy and paste this into
194. 12:03a terminal, boom. There's also a desktop app you can use now, and it can accomplish a lot of the
195. 12:08same stuff that we're going to do here in the workspace when you're in the right folder, but
196. 12:12we're going to stick to IDE classic style. So now we have our agent over here. We have our whole
197. 12:18folder over here, and we've got to get started building our workspace. Now, a lot of the times,
198. 12:22Claude will do some of this on its own. Hermes agent will do an even better job, but I don't like
199. 12:26using Hermes for building things. So I've got my little presentation here. Here are the prompts
200. 12:30that you can find below. I'll put a link. We're going to say, set up this workspace,
201. 12:35create a file called AGENTS.md, which is your job description. Put in it who you are, how we talk,
202. 12:40what you never do, and keep it under 40 lines. So I'll copy that. And actually, before I paste it,
203. 12:46I'm going to actually switch to an even cheaper model because the truth is I don't need
204. 12:50Kimi K3 Fast for this. This is very simple stuff. So I'll use GLM 5.2. I'll paste that prompt there,
205. 12:56set up this workspace. And we're probably going to get questions asked to us. Yeah, there it is.
206. 13:04What is your business? All right. So what should we say? Yeah, let's say we're a fitness coach.
207. 13:10Why not? What voice should I use when talking or writing for you? So this is going to be the very
208. 13:16baseline, right? Obviously, you're going to need to feed it samples of your writing eventually to
209. 13:21get there, but this is how you get started. Direct encouraging, zero hype. Perfect. What are your
210. 13:28hard rules? Things I should never do. No medical claims, no discount spam, max one emoji. Let's do,
211. 13:36yeah, I like one emoji. I mean, I guess that's fine. So we're just running through a demo here.
212. 13:44When you do this for yourself, I recommend doing it with a real purpose. Have a result you want
213. 13:51What is your actual business? Run it for real. And there we go. Here in AGENTS.md now,
214. 13:57we'll open up the preview. Who I am. I'm the operations assistant for a one person fitness
215. 14:03coaching business. How we talk, what I never do, et cetera, et cetera. That's our starting point.
216. 14:09Next, we're going to make the checklist and the specialist. So we're going to create a skill,
217. 14:14which is a workflow that shows the agent how to do something that's going to have to do a lot.
218. 14:19Pop that right there. Create your first skill. One job, we repeat. So skills are the first easy
219. 14:26step you can take to mimicking a Claude Code workspace. So a model that's less intelligent
220. 14:31can accomplish a task that Cloud doesn't need a skill for, for example. If you have the most
221. 14:36intelligent model ever, it might not need a skill to do something, especially if it's not a very
222. 14:41customized skill. But once you have the skill, it can do it faster, cheaper, and probably just as
223. 14:45Good. All right. So here we'll do weekly client check-in email, monthly progress summary.
224. 14:52Yeah, monthly progress summary. Then we'll send that to the clients.
225. 14:57There's our skill. Let's check it out. In our skills folder, SKILL.md. There we go. Skill name,
226. 15:05monthly progress summary, when to use. A client has completed a month of training and needs a
227. 15:09written recap of their progress notes and program adjustments. Structure, voice rules,
228. 15:14What done looks like.
229. 15:16So this is how when our agent, when I say, hey, watch, we'll start a new session.
230. 15:22Say, okay, test client Jill Jacobs needs her monthly progress report.
231. 15:31Make a sample.
232. 15:33Watch what it's going to do right out the gate.
233. 15:35Skills.
234. 15:36Monthly progress summary.
235. 15:39So now it's reading the skill file.
236. 15:42It's doing a little bit of thinking.
237. 15:44We're in high thinking mode.
238. 15:47And it's going to create a sample for us here.
239. 15:49Okay, the sample report's done.
240. 15:50It follows the skill structure.
241. 15:52So let's open that up.
242. 15:54Preview.
243. 15:55Client, Jill Jacobs, month.
244. 15:57What you did.
245. 15:58What changed.
246. 15:59Your back squat max increased.
247. 16:02What comes next.
248. 16:04And one line of encouragement.
249. 16:05So remember the structure here.
250. 16:08The header, what you did, what changed, and what comes next.
251. 16:10One line of encouragement.
252. 16:12And this is exactly what we got.
253. 16:14in the Jill Jacobs summary here. What you did, what changed, what comes next, one line of
254. 16:18encouragement. Now we can simply say, now add another skill to send this as an email.
255. 16:25Let's connect my Gmail account. And it'll help you do that. Now, how do you connect your Gmail
256. 16:32account? Well, that's not really the scope of this video, but it's not hard. You can use Composio.
257. 16:38Install Composio via CLI, add the MCP server, then you're going to need to authenticate
258. 16:42with Google. You could use OneCLI. You could use the official Google Workspace MCP. Plenty of ways
259. 16:47to do that. The point is, is that now, I'll even say here, Composio. We want this to be reusable
260. 16:57for any report. So now our agent is creating the skill to be reusable for any report. So we can send
261. 17:04that email. Now, obviously, you could use this for anything. Okay, this is just a super basic example.
262. 17:10Skills can be created for very complex and complicated workflows.
263. 17:13And it's not a bad idea to really get it figured out.
264. 17:17It's just as if you were going to train an employee.
265. 17:20Train your AI to do something really well, and then it can just keep getting better.
266. 17:24I'm a year into this workspace I just showed you, and it's really impressive.
267. 17:29Every quarter or so, I go in and I revamp it based on the new things I've learned and the things that have shifted.
268. 17:35And then it just gets better and better.
269. 17:37So you want to take this kind of thing seriously instead of just relying on Claude Code to be able
270. 17:41to do it because Fable is so damn intelligent. Relying on Astra to do it because it's just so
271. 17:46intelligent. Yes, they can do all those things, but keep those for the new things when you're
272. 17:51building them for the first time. Even right now, I'm building this with GLM 5.2. If this was a more
273. 17:56complex task, I would easily be choosing Fable or Opus or Astra over GLM 5.2, but this is a pretty
274. 18:03simple thing right now. So I'll save the inference. So that's building the next skill. Let's start
275. 18:08multitasking. New tab. That one's running in the background. And next prompt. Let's create a
276. 18:15specialist. This specialist is going to be a sub-agent. And now here we have the reviewer agent.
277. 18:21Here in the OpenCode folder, we have agents, reviewer.md. Reviews a draft against the style
278. 18:27checklist and AGENTS.md voice rules, returns pass or a numbered list of fixes. So now you have the first
279. 18:35agent that writes the monthly progress report then another agent that will proofread it and make sure
280. 18:41it's good. It'll do an audit. Claude Code is notorious for wasting credits in reviewing things
281. 18:47and double checking its work even if you're just trying to make a button blue. But just have your
282. 18:51workspace do this by default. Save tokens. Make it faster, make it more specific. Save tokens.
283. 18:59Meanwhile, here we go. It's ready for our Composio API key and to connect our Gmail. I'm not going to
284. 19:04do that for this demo. New session. Next prompt. Okay, so we built our sample workspace and I promise
285. 19:11you if you work on this for whatever project you're working on, give it time. Over a few months, it's
286. 19:17going to get really good and really helpful for you. And as I mentioned earlier, we'll help you
287. 19:21stay focused on your tasks instead of just constantly building new things because it's fun.
288. 19:25Once you have something really locked in, then the fun part comes. So I just built all this with GLM
289. 19:315.2. And as I mentioned earlier, when you're building complex stuff, I wouldn't advise not
290. 19:36using a frontier model if you can afford it, if you're using Claude Code, subscription, whatever,
291. 19:41use it. And you can see here, I could easily just switch to Fable 5.1 right now via Venice,
292. 19:47and I could use it. But what I want to show you now is how now that this workspace is all done,
293. 19:52I could use any provider. I don't have to be connected to Venice. So if I do forward slash
294. 19:59connect, look at all these providers. There's so many providers you can use. Use anything you want.
295. 20:05And so you have your provider for whatever reason, you pop in that API key and all of the models
296. 20:11will be able to follow the workflow in this workspace.
297. 20:14No problem.
298. 20:15I repeat this whole workflow a lot on my channel
299. 20:18because it consistently surprises me
300. 20:20how few people take their agentic workspace seriously.
301. 20:24This is not 100% of the difference between Claude Code,
302. 20:27between using Claude Code and not using Claude Code,
303. 20:29but it does make a big difference
304. 20:31having proper knowledge infrastructure.
305. 20:34Okay, so now let's come back into building the workspace.
306. 20:36I'll switch back to GLM.
307. 20:38I'll paste the prompt.
308. 20:40And what we're going to do is add a routing section to AGENTS.md, which is basically going to say, hey, when we're doing complicated stuff, use the better model.
309. 20:49We're making a plan right here.
310. 20:52When we're building, use a medium model.
311. 20:54When we're doing small mechanical jobs, use a cheap model.
312. 20:57It's called model routing.
313. 20:58And this is something that many people don't ever think about.
314. 21:00I never did.
315. 21:01Using subsidized inference with Claude Code.
316. 21:04But when you start to go off on your own, you have to figure out how to save on costs while still optimizing your performance.
317. 21:10So in my example here, I like Kimi K3 for planning. It's a powerful model. It's a Fable
318. 21:17of open-source models as far as I'm concerned. So I use that for making the plan, complex reasoning,
319. 21:22figuring things out, even creative stuff. But then I use GLM for drafting and building. Once
320. 21:27we have a good plan, okay, GLM, you build it. And I'm talking about more complicated stuff than this.
321. 21:31We're just copy and paste to make a workspace. I'm saying if I'm making an app, a website,
322. 21:36replacing a SaaS so I could save the subscription cost, right? You're going to want a good plan,
323. 21:41but you're going to want a good builder. And then use a super cheap model for mechanical tasks. Like,
324. 21:46hey, summarize this. Hey, I forgot, what do we have on the website right there? Oh, hey, fix that
325. 21:52tiny thing on the website. Like fix the grammar here or whatever. All of that can be done with a
326. 21:57super cheap model. Pennies. So we're looking at pennies, dollars, tens of dollars per million
327. 22:04tokens. So that's model routing. If we go back to my person A, person B thing, person A is going to
328. 22:11be running the systems with the models. Person B is going to be building them with maybe Fable or
329. 22:18Astra or whatever. So you want to think about that when you're routing. Maybe you want to use Fable
330. 22:23for the plan. Once again, it all depends on you, your budget, the privacy you want, etc. Okay, so we
331. 22:29made our specialist. Now we're going to run a prompt that's basically creating a loop, a review
332. 22:34audit loop. So we're going to say, hey, let's do the job and then hand your draft to the reviewer.
333. 22:40If it fails, do it again. So draft a new monthly progress report for Jordan and have the reviewer
334. 22:53check it, rewrite it if it needs work. Show me the final draft and the reviewer's verdict.
335. 23:01Let's clean this up.
336. 23:02And now we see it makes a list here,
337. 23:04and it's going to do all the things.
338. 23:05So, hey, I've said this before,
339. 23:07but if you've only used Claude Code,
340. 23:10I hope you're thinking by now, like,
341. 23:11wow, OpenCode is pretty neat.
342. 23:13You can do a lot of stuff, even on a pretty cheap model.
343. 23:15Okay, here we go.
344. 23:16The loop is complete.
345. 23:18Reviewer verdict: pass.
346. 23:20Great.
347. 23:21So here it is.
348. 23:23This was written using the monthly progress summary skill
349. 23:27and the voice rules in AGENTS.md,
350. 23:30which, by the way, was this.
351. 23:33Be direct, say what you mean in the fewest words possible,
352. 23:36be encouraging, use zero hype, write in plain language,
353. 23:39active voice, short sentences, what I never do,
354. 23:42make medical claims, blah, blah, blah.
355. 23:44So it followed the rules for how I talked to my clients,
356. 23:48and then to this client, Jordan,
357. 23:50we made this progress summary report.
358. 23:52Here's what you did, here's what changed,
359. 23:55here's what comes next, one line of encouragement.
360. 23:58So sometimes AI doesn't follow its rules if it's out of the sandbox.
361. 24:02So that's why you bring another agent in to review it.
362. 24:05So fresh context.
363. 24:06It doesn't know what's going on.
364. 24:07You say, hey, does this follow this rules?
365. 24:09Yes or no.
366. 24:10If it does, give it a pass.
367. 24:11And this passed.
368. 24:12So the next thing we would do is give it a prompt like this.
369. 24:15I just edited the draft.
370. 24:16Compare my version to yours.
371. 24:18Find what I changed.
372. 24:20And if any of my edits look like a rule, I probably always want.
373. 24:23Propose one line to add of the skill and wait for my approval before changing anything.
374. 24:28Now, I'm not going to do that in the demo, but you can see how this is the direction we go.
375. 24:33This is how we create a voice tune skill.
376. 24:36We keep the loop going where we only come in to enhance the work that the agent's doing.
377. 24:41So, for example, I'll say remove all em dashes.
378. 24:46You know, like here's one.
379. 24:48Never write with them again.
380. 24:50Propose to me a change in our workspace so that doesn't happen.
381. 24:57Here we go.
382. 24:58propose change, add one rule to AGENTS.md under how we talk, never use em dashes, and then also
383. 25:04tell the reviewer agent to add it to the checklist. Where's the reviewer agent? Your checklist.
384. 25:12No em dashes. Want me to make both edits? Yes. And we're going to see that it adds that rule to
385. 25:18the reviewer checklist here. And so we should never have to deal with em dashes again, though I do
386. 25:23think that we need to be more forgiving of non, it's not X, it's Y, em dashes, because it is a
387. 25:30great punctuation mark that I've always been a fan of. And it's a bummer not using it anymore,
388. 25:35because I don't want people to think that I wrote with AI just because there's a freaking em dash.
389. 25:39And there we go. We have a very basic workspace. There's a link below for you to get all those
390. 25:44prompts to start scaffolding your own thing, experiment with this. I do recommend having an
391. 25:48actual goal for something you want to accomplish as you play around and learn the thing, whatever
392. 25:53it is, especially in AI, have an actual goal, have an intention for where you want to arrive
393. 25:58in your learning process, and then something will come of it. Easier and better than just saying,
394. 26:01hey, I'm just going to try it just to try it. And then you just end up with something
395. 26:05kind of nonsensical like this. Like, okay, I'm a fitness coach now. Okay.
396. 26:09Thanks for watching, guys. Hit subscribe if you're interested in learning about agentic AI
397. 26:13with me and following my opinions. Real quick, I just want to tell you about the AI Captains
398. 26:17Academy. This is where you can find an actual curriculum for learning like everything about how
399. 26:24all this stuff works. So this is a curriculum that's meant to help you build the foundations
400. 26:28that will not expire. There's a seven-day free trial. We do two calls a week, one tactical call
401. 26:32where we do like a workshop, and then one call where it's more like housekeeping, come ask your
402. 26:37questions, and we just talk about what's going on in AI. We come share our solutions, the things
403. 26:41we're learning. So if that sounds interesting to you, hope to see you in there.