The Fractal Fragmentation of AI Tools — Transcript

Full transcript

1. 0:00Hello and welcome back to another episode of AppStories.
2. 0:03Today's episode is brought to you by Vitally.
3. 0:06I'm John Voorhees and with me is Federico Viticci.
4. 0:09Good morning Federico, afternoon or evening.
5. 0:12It's almost evening I would say for me.
6. 0:14It's already afternoon here too.
7. 0:16Yeah.
8. 0:16Yeah.
9. 0:17How are you John?
10. 0:18I'm doing well.
11. 0:19I'm doing really well.
12. 0:20How about yourself?
13. 0:21I'm doing well.
14. 0:22I was just, I was messing around for MPC.
15. 0:26I was playing around with Game Hub with my Red Magic 11 Pro phone.
16. 0:31And I'm going to talk about this on the show later this week.
17. 0:34I started, I tried to play Resident Evil Requiem.
18. 0:37That's what I was wondering because you sent me that video.
19. 0:40It's not going to happen, man.
20. 0:41It's Resident Evil on that phone.
21. 0:42It's not going to happen.
22. 0:43What are people doing with these videos?
23. 0:45How do they do these things?
24. 0:46We'll have a conversation about this.
25. 0:48Okay.
26. 0:48I'm starting to not believe those videos on YouTube anymore.
27. 0:51I look, I've tried everything.
28. 0:53I don't know how these people are getting that image quality.
29. 0:57Look, I don't think I'm stupid with these things.
30. 1:01Right?
31. 1:01I think I know how to tweak, how to tinker around with things.
32. 1:04Sure.
33. 1:05I am just not getting that performance out of this gaming phone.
34. 1:08Don't they tell you the things to tweak in the videos too?
35. 1:12Like the parameters?
36. 1:12They do.
37. 1:13And I keep the same things and I do the same things and my game looks like crap.
38. 1:19Yeah.
39. 1:21I hear you.
40. 1:22Oh, well.
41. 1:22But anyway, we're not here to talk about gaming.
42. 1:25We have a separate show about that.
43. 1:27It's called MPC and Exportable Console on MacStories and on Patreon if you want to get extra content.
44. 1:31We're here to talk about something that is very frustrating for me and also I think for John.
45. 1:36Oh, yeah.
46. 1:37Which is making sense of all these new features and this new terminology for AI apps, especially on desktop.
47. 1:49And we are specifically going to focus on Claude and ChatGPT today.
48. 1:55Or I guess Claude and Codex.
49. 1:58Which is OpenAI's separate system slash application for programming and also kind of some agentic use cases.
50. 2:07Right.
51. 2:08This is very frustrating to us because not only are these things changing on a daily basis at this point, but these companies, Anthropic and OpenAI, they are adding feature on top of feature on top of feature and they're breaking things in the process and they are renaming things in the process.
52. 2:29If you thought that Apple renaming the efficiency course to performance course was frustrating.
53. 2:35That's nothing.
54. 2:37That's nothing.
55. 2:38It's really nothing.
56. 2:39No, I mean, obviously, I think the problem that we have here is that these companies are moving incredibly fast.
57. 2:45And they're not, you know, I think traditional software development, you would kind of expect things to be pretty smooth by the time they're launched.
58. 2:52There are always bugs.
59. 2:53But what's happening here is, you know, ChatGPT will come out with some feature in the chat bot.
60. 3:00And so the Anthropic people will jump all over code and they'll release whatever they were working on that might be similar right away, whether it's really ready for primetime or not.
61. 3:09And so you've had a lot of things like, you know, a lot of things issued, released as previews or betas.
62. 3:15Research previews.
63. 3:16Research previews, which is essentially an alpha, I would say, because when like Cowork first came out, and we'll talk about Cowork in a minute.
64. 3:23Cowork was super rough when it first came out, but it got better very fast.
65. 3:27I mean, it's only been in research preview for what, a month, three weeks, something like that.
66. 3:32It's not been very long, but it's gotten a lot better.
67. 3:34So, yeah, that's kind of what's causing it, I think.
68. 3:38But I feel like they need to like pause for a moment and rethink like all of these features, how they fit together, what they're called, what apps have them, what apps don't, and how the apps are designed even.
69. 3:51Because I've got some ideas for how the Claude app on the iPhone should look, for instance.
70. 3:56Yeah, let's start with that.
71. 3:57Let's start with Claude and Anthropic.
72. 4:00So I really spent some time over the past couple of days playing around with this.
73. 4:06The easiest way to describe this is that Anthropic has stuffed Claude for Mac, Claude for Mac, with the three Cs.
74. 4:15That's what I'm calling them.
75. 4:17Chat, Cowork, and Code.
76. 4:19So these are three separate modes that are now coexisting in the Claude app for Mac.
77. 4:26I'm not talking about the website.
78. 4:28I'm not talking about Claude coding the terminal.
79. 4:31I'm talking about Claude for Mac.
80. 4:33Chat, obviously, used to be the default chatbot mode with Claude, where you can have your projects and you can have your, used to be your MCP connectors.
81. 4:43That's a separate conversation now.
82. 4:45But now you have these three segments at the top of the screen.
83. 4:48You can actually move between them by doing command one, two, and three, I discovered.
84. 4:53That's nice.
85. 4:53You can move between chat, Cowork, and Code.
86. 4:56So chat, we don't need to talk about the chatbot.
87. 4:58You know what Claude the chatbot is.
88. 5:02Cowork and Code.
89. 5:03This is where I think most people get confused.
90. 5:06So Claude Code started as a terminal-only affair last year.
91. 5:12In fact, it launched about a year ago.
92. 5:14Then they added, if I'm not mistaken, Claude Code for Web.
93. 5:20It's horribly named, but Claude Code for Web is Claude Code.
94. 5:25But instead of working on your local computer in a folder, you point it to a GitHub repo and you work in that virtualized location.
95. 5:34It creates a little virtual computer in the cloud.
96. 5:36You point it to a repo and you say, work on this project for me.
97. 5:40Right.
98. 5:40Very different from Claude Code on your computer.
99. 5:43Claude Code on your computer can now be accessed.
100. 5:47I'm going to give it to you, John, because you use Claude Code much more than I do.
101. 5:51You can use Claude Code the old-fashioned way with a terminal window.
102. 5:56Yes.
103. 5:57And that's the OG Claude Code.
104. 6:00Or you can use desktop Claude Code inside Claude for Mac.
105. 6:05You go to the code section of the Claude app.
106. 6:09And there you can use Claude Code, which is kind of similar from a design perspective to Claude Code on the web.
107. 6:17But it behaves more like Claude Code in the terminal.
108. 6:21Yes.
109. 6:22In that you can point it to a folder and you can work in a local folder and you don't need to use GitHub at all.
110. 6:29Although you can point it at GitHub.
111. 6:30Although you can.
112. 6:31Yeah.
113. 6:32But you can also point it to a folder.
114. 6:33Which is why all this gets confusing.
115. 6:34It's like the focus is not on GitHub as much when you're in the app, although it's still possible to do it in the app.
116. 6:41And, of course, there's remote control, which allows you to work in the app, like one of the mobile apps I'm talking about specifically now, the iPhone or the iPad.
117. 6:51And you're connecting back to a terminal session, which is not the same as Claude for Mac.
118. 6:57It's something different.
119. 6:58So you wrote about remote control.
120. 7:00Yeah.
121. 7:01Remote control lets you remotely control a Claude Code session that is running in the terminal.
122. 7:09Yes.
123. 7:10Not inside the Claude app.
124. 7:12No.
125. 7:13You have Claude actually running in the terminal already.
126. 7:16And you type in a command.
127. 7:18This is already confusing.
128. 7:19You do remote-control.
129. 7:21It starts up a remote session.
130. 7:23Or you can make it the default so all your sessions are remote controllable, which is what I do.
131. 7:28Okay.
132. 7:29Because I never – that way you don't have to be thinking about it in advance.
133. 7:32Because otherwise you have to start your Claude session in remote control.
134. 7:36So you, like, you go to the terminal.
135. 7:38You're not in Claude yet.
136. 7:39You type Claude space remote-control.
137. 7:42And now you've got a remote session.
138. 7:43That's no good for my purposes because I never know in advance whether I'm going to be remotely controlling a session or not.
139. 7:50So I just have it as the default.
140. 7:52And then what you do is – I would say the best thing to do is to rename your session, which is just a slash rename and then give it a name.
141. 8:00You can, like, a lot of times all I'll do is I'll rename it for the computer that it's running on, like my Mac Studio.
142. 8:07And then just use it for whatever comes up when I'm away from my desk.
143. 8:11Or you can name it, you know, for a project if you're in the middle of a project and you have to leave your desk and you're not – you don't have the terminal available anymore.
144. 8:18And then you can use it from one of those apps.
145. 8:22And I wouldn't say I use it a lot, but I do use it.
146. 8:26I mean, for instance, there was one night I was working on a project and I went off to bed.
147. 8:32I had my iPad with me and something occurred to me.
148. 8:35It was like, oh, I should try to do this.
149. 8:37And I had it in a remote session on my desktop computer.
150. 8:40So the session was already running on the computer.
151. 8:45Yes, it was already running, right?
152. 8:46I had kind of quit for the day.
153. 8:48I didn't exit out of Claude or anything.
154. 8:49I just left the terminal window open for picking up where I left off for the next day.
155. 8:54I went to bed.
156. 8:55I thought, oh, you know, I should try this.
157. 8:56I had this idea.
158. 8:58I went into the iPad app.
159. 8:59I went to the code section.
160. 9:01I found my listed remote session, went into it.
161. 9:06And once you open it up, you actually see what you would see in the terminal.
162. 9:09I mean, it's formatted more like, you know, for an iPad or iPhone app, but it's the same words are there.
163. 9:16And you can just type commands straight into the remote session from the code tab.
164. 9:21So I did a couple of things.
165. 9:24And, you know, it was nice to be able to do that while it was still on my mind.
166. 9:28I mean, the downside is, of course, is that you're not, this is not like screen sharing because you don't have anything else available.
167. 9:35So if like you're working on an app, you could build the app, I guess, but you wouldn't be able to see the app or a web app or whatever you're working on.
168. 9:43You've only got that insight into what's in the terminal.
169. 9:47You don't, you know, you don't have actually the other things on your computer, which is why I use this kind of sparingly.
170. 9:54Because a lot of times if I'm on my laptop or something or even on the iPad, using screen sharing on the Mac or something like Jump Desktop gives me access to the other things, too.
171. 10:05So I can take screenshots and feed those into Claude, you know, other things that you might want to do beyond the terminal window itself.
172. 10:13So that's Claude and Claude code.
173. 10:16Yeah.
174. 10:17There's a third.
175. 10:19Can I, before we, before we get to that, or maybe this isn't the time to raise this, but I want to talk about the difference between Cowork and projects.
176. 10:25Because I think that's confusing to a degree.
177. 10:28Okay.
178. 10:29So projects are basically.
179. 10:31How do you read them?
180. 10:32So projects are folders for the chatbot.
181. 10:37Projects are chats in the chatbot that have a common system prompt.
182. 10:43Yep.
183. 10:43That can use remote connectors, but don't have access to the file system.
184. 10:48Right.
185. 10:48And they have their own memories.
186. 10:50They have their own memory that are separate from each project and separate from chat.
187. 10:54And you can attach documents for reference materials.
188. 10:58Sure.
189. 10:58And things like that.
190. 10:59Yeah.
191. 10:59And so the way I think about these things, and they're related but different.
192. 11:03I think of projects as long-term things.
193. 11:07Like, you know, the name suggests a project where you're going to be coming back to it over and over, whether it's just a thing that's going to.
194. 11:14And you don't want to type the same prompt every single time.
195. 11:17Yeah.
196. 11:17Yeah.
197. 11:17It's either it's something that's going to take you a long time to complete, or you're going to ask similar things and want a similar prompt over and over again to do that.
198. 11:28So I have a bunch of projects.
199. 11:29Whereas Cowork is more of like a one-off thing where you're.
200. 11:33Kind of.
201. 11:34See, that's where it gets confusing sometimes to decide which to use.
202. 11:39So I think Cowork is the most confusing of the bunch so far.
203. 11:42Also because it's the latest one.
204. 11:44And it's labeled as a research preview.
205. 11:47And it's desktop only.
206. 11:48So Cowork has no mobile presence at all.
207. 11:53Cowork, the technical description is that they took the agentic guts of Claude Code.
208. 12:00The way that Claude Code responds to you, calls external tools, comes back, compacts the conversation, continues.
209. 12:09They took that thing that John loves to call the agentic harness of Claude Code.
210. 12:17My favorite word.
211. 12:18And they turned it into something for more general productivity.
212. 12:23That's what Cowork is.
213. 12:25But there's a difference.
214. 12:26So Cowork, it doesn't organize your chats in projects.
215. 12:31You can choose to work.
216. 12:34You can choose to start a Cowork session in a folder on your Mac or not.
217. 12:41Which is a little bit like how Claude Code works, right?
218. 12:45Yes.
219. 12:45Because you're doing it inside some sort of project.
220. 12:48Yes.
221. 12:48Typically.
222. 12:48If you work in a folder and if that folder contains a CLAUDE.md instruction file, it will read that.
223. 13:00Yeah.
224. 13:01Even if your CLAUDE.md is sim-linked to an agents.md, because OpenAI and Anthropic cannot agree upon a naming standard for this instruction document, it will read the sim-linked document.
225. 13:17Right.
226. 13:17Okay.
227. 13:18So you can basically pre-inject some system instructions into Claude Cowork.
228. 13:25For example, in my markdown document that I have in my agent folder, there are some preset instructions that say, I'm Federico Viticci.
229. 13:33I use Todoist.
230. 13:34I use Notion.
231. 13:35This is the Notion database where I keep my daily notes, like basically preset information for Claude Cowork to get started.
232. 13:43So that it doesn't have to search for my daily notes every time, for example.
233. 13:47Right.
234. 13:48The thing about Cowork that I really like is that it behaves like Claude Code, but presents you information like regular Claude.
235. 14:00Right.
236. 14:01So it's got that sort of sequential tool calling nature without showing you that it's running terminal commands, without showing you that it's doing like programmary things.
237. 14:12And that's sort of the pitch for Cowork, right?
238. 14:15It's an agent.
239. 14:17It's an agent.
240. 14:17It's a general purpose agent.
241. 14:19And they have done this because people were using Claude Code, which is meant for developers and programmers for general productivity, which is not designed for.
242. 14:29So, you know, because there are bash commands to move files around your computer.
243. 14:33And, you know, there are scripts being written in the background to maybe export, marked down as a PDF using Pandoc.
244. 14:41And there's all kinds of things that can be done.
245. 14:44Yeah.
246. 14:44So I am going to, I think I mentioned all the things.
247. 14:50There's three more things I need to explain.
248. 14:52Okay.
249. 14:53First is a series of complaints.
250. 14:55Then there's a series of things that Anthropic are doing right.
251. 14:59And third is a final feature that I want to mention.
252. 15:02So the complaints is, I call this section in our document, John, the great fragmentation of tools.
253. 15:08Yeah.
254. 15:08There are too many words, Anthropic.
255. 15:11There are too many terms that you're asking people to learn and keep up to date with.
256. 15:16Yeah.
257. 15:17Just to go in succession here, John.
258. 15:19Skills, connectors, MCP, plugins, hooks, slash commands, agents.
259. 15:26I would add artifacts to that.
260. 15:29And artifacts.
261. 15:30Sure.
262. 15:31It's not technically a tool, but.
263. 15:32So let's try and simplify this for our listeners.
264. 15:36Yeah.
265. 15:37Skills are those skill.md.
266. 15:41They're markdown documents that basically teach Claude how to do something.
267. 15:45They're packaged instructions.
268. 15:48They're basically prompts as markdown files.
269. 15:51And they can be combined with a script, for instance.
270. 15:53They can contain scripts.
271. 15:54They can contain image assets.
272. 15:56They can contain whatever you want as long as you store them in a folder.
273. 16:00Connectors and MCP, they're basically the same thing now.
274. 16:04Anthropic is trying to rebrand the MCP standard as the English word connector.
275. 16:11They're the same thing.
276. 16:12You are connecting to Todoist.
277. 16:15It means you're using the Todoist MCP server.
278. 16:17And then this is like a step in the right direction, I think, in the sense that people don't care if it's technically built on the MCP standard or whether it's like some one-off tool that Anthropic built for Notion or Gmail or something.
279. 16:30Because that's the difference, right?
280. 16:31Is under the hood, some of these things may not be strictly MCP servers.
281. 16:37They may have some additional code that's working on the Anthropic side.
282. 16:42Slash commands, I never use this.
283. 16:45But these are basically ways to come up with your own text shortcuts, I think.
284. 16:49Like you can create a slash command where you type slash, I don't know, J-O.
285. 16:55And you have a slash command for John that sends an email to John or something.
286. 17:01Yeah, I use a few of these for certain things that I've built where I want to clean up a transcript or something.
287. 17:10I have one that calls some scripts that's called Transcribe Cleanup.
288. 17:16And it will call some things.
289. 17:19So they're basically abbreviations, essentially.
290. 17:21Yeah, they're just easy to remember ways to kick off some sort of thing that you don't want to have happen automatically.
291. 17:27It's like a manual process, basically.
292. 17:32This episode of AppStories is brought to you by our friends at Vitally,
293. 17:36the AI-powered workspace for customer success managers.
294. 17:40Vitally's purpose-built AI is for scaled customer success.
295. 17:44It's there to help you better understand customers to reduce churn, unlock growth, and always stay a step ahead.
296. 17:52Basically all the things that make you great at your job.
297. 17:55Vitally delivers the clarity, automation, and AI-driven insights your team needs.
298. 18:01So you can move faster, stay aligned, and drive customer outcomes at scale.
299. 18:08Vitally AI surfaces insights from every customer interaction to unlock another level of CS productivity, visibility, and collaboration.
300. 18:17It's why Vitally is trusted by more than 600 leading B2B SaaS customer success teams.
301. 18:24And how is this for a great offer?
302. 18:27Vitally is offering a free pair of AirPods Pro for every AppStories listener who takes a qualified demo call.
303. 18:34So if you're a customer success decision maker, schedule your call by visiting vitally.io slash appstories.
304. 18:43Again, that's vitally.io slash appstories for a free pair of AirPods Pro when you take a qualified demo.
305. 18:50Our thanks to Vitally for their support of AppStories.
306. 18:57Hooks.
307. 18:58Have you ever used hooks?
308. 19:00Maybe once or twice, but not really.
309. 19:02Hooks.
310. 19:02So if I'm not mistaken, hooks are busy.
311. 19:04So we're talking about Claude Code here.
312. 19:06It's hooks into the guts of Claude itself.
313. 19:09They're basically triggers and events for things that can happen automatically when something happens in Claude Code.
314. 19:17You know what I use them for, Ticci?
315. 19:19I use them for notifications.
316. 19:21I hooked up Simon Stovering's new beta app called Brrr, B-R-R-R, and I send myself notifications because you can do pretty rich notifications with that beta.
317. 19:32And that uses a hook to different events that happen within Claude, such as like when something's pushed to GitHub, I get a notification that something has been pushed to GitHub.
318. 19:44I don't have this beta, by the way.
319. 19:46I sent it to you over the weekend.
320. 19:48You did?
321. 19:49I did.
322. 19:50I sent it to you on Saturday.
323. 19:51How many weekends ago?
324. 19:52No, just like last Saturday, I think.
325. 19:55Maybe Friday afternoon.
326. 19:57But yeah, so I know when GitHub happens, when things are pushed to GitHub, I get notifications when Claude needs my attention on my remote server, that kind of stuff.
327. 20:08What's it called again?
328. 20:10Brrr, B-R-R-R.
329. 20:12Scroll up to Saturday.
330. 20:13There aren't that many messages.
331. 20:14We exchange too many links.
332. 20:16Yeah, we do.
333. 20:17I'll send you.
334. 20:18I'll resend it to you.
335. 20:18Thank you.
336. 20:19You should check it out, though.
337. 20:20It's very good.
338. 20:20And he's building a CLI for it, too, even in the time since I installed it.
339. 20:25Nice.
340. 20:25Which allows you to use it, I think, without necessarily using it.
341. 20:30There's advantages to the CLI versus hooks.
342. 20:33I've got it installed with hooks.
343. 20:34Okay.
344. 20:36The part that is most confusing to me is plugins, which I also think is, again, it's the latest thing that Anthropic has done, and it's the one that they have explained the least.
345. 20:48So I pulled some links from the Claude documentation and the Claude Code documentation.
346. 20:57Yeah.
347. 20:57In Claude Code documentation, they say plugins extend Claude Code with skills, agents, hooks, and MCP servers.
348. 21:04Plugin marketplaces are catalogs that help you discover and install these extensions without building them yourself.
349. 21:10It's a bundle is what it is.
350. 21:13It's a bundle of, so a plugin can contain an MCP server, some slash commands for that MCP server, maybe some skills for that, and it's all prepackaged as a plugin.
351. 21:23And it's on a marketplace where presumably you can trust it more than if you just pull something off of a random GitHub repo.
352. 21:30The problem is that these plugins in the Claude app for Mac, they're very new.
353. 21:35There is a gallery right now that only contains Anthropic plugins, no third-party plugins.
354. 21:42Superpower, I think, is in there now.
355. 21:44One for one.
356. 21:45That one I know of.
357. 21:47That's the one that I put in our notes because I was just reading about it last night, and it is in there.
358. 21:52I haven't installed it yet, but I read that it was in there.
359. 21:56I tried to install an Anthropic one.
360. 21:58Yeah.
361. 21:59I watched a few of them.
362. 22:01I clicked.
363. 22:02There was one in the Anthropic web directory called Claude Cowork Plugin Creator.
364. 22:09So it was like a plugin to create plugins, very meta.
365. 22:13I thought, this is interesting, and I clicked a link in my web browser, and he redirected me to Claude for Mac, and it said the plugin didn't exist.
366. 22:22You can get at them easier, I think, through Claude Code.
367. 22:25I haven't tried them on the Mac app.
368. 22:27I've only tried them in Claude Code, where you can do things like get a code review plugin, for instance, which I think is made by Anthropic.
369. 22:36They have a design one, I think.
370. 22:38There are too many words.
371. 22:39There are too many words.
372. 22:42But the one thing that Anthropic, I think, are doing right now is they are unifying a lot of things.
373. 22:52For example, something that I've noticed lately is that skills now sync all across your Claude apps.
374. 23:02Right.
375. 23:02If you upload a skill in the Claude website, it shows up in Claude for Mac on another computer.
376. 23:11I just booted up my Mac mini, and I opened Claude on the Mac mini, and it had all the same skills from Claude on the MacBook Pro.
377. 23:19So they're now syncing all your skills with your Claude account.
378. 23:22That's the right thing to do.
379. 23:23They're also now properly syncing connectors.
380. 23:25So anything that you connect to from claude.ai or Claude for Mac will show up in every other instance of Claude.
381. 23:33And not just that, but something that...
382. 23:36I'm sorry.
383. 23:36They seem to have stabilized, too.
384. 23:38For instance, like the Notion one and the Todoist one, which would get disconnected all the time, now seem to be...
385. 23:46And this is literally only over the last week or two that they've seemed to have stabilized.
386. 23:50Yes.
387. 23:50And something that I've noticed a few days ago, and I told you, is now when you fire up Claude Code in the terminal,
388. 24:00and you do the slash command slash MCP to see a list of your connectors,
389. 24:05it'll show you all the existing connectors that are tied to your account that are coming from claude.ai,
390. 24:13which means you previously connected from either the Claude website or Claude for Mac.
391. 24:18So you don't need to reset up all of those connectors in Claude Code for terminal again.
392. 24:24They all follow you around as long as you're using your account.
393. 24:28Right.
394. 24:30Finally, I want to talk about scheduled tasks.
395. 24:35Yeah, I want to hear about this because this is actually something I've been meaning to try,
396. 24:38and I haven't gotten around to it, and it's very, very new.
397. 24:40It's like, what, like a week old or something?
398. 24:42Yeah, this is very new.
399. 24:44Now, this lives in the sidebar of Claude for Mac, which I think you have some thoughts to share about the sidebar.
400. 24:54It's a new entry there, and it's basically a way for you to schedule tasks that Claude Cowork will run for you,
401. 25:05which means you can work in a folder, you can take all of your existing connectors,
402. 25:12and you can schedule a task with a time and a prompt.
403. 25:17So here's what I did, John.
404. 25:19I took a previous project from the Claude chatbot.
405. 25:24It was a very detailed prompt that told Claude how to process my email, look for important messages,
406. 25:33and turn those important messages into deep-linked Todoist tasks.
407. 25:40I used to have to run that prompt manually every couple of days or every morning.
408. 25:45I took that prompt, created a new scheduled task in Claude Cowork, running in the Claude app on my Mac Studio.
409. 25:55Not on my MacBook Pro, not on my Mac mini.
410. 25:58I chose the Mac Studio because my Mac Studio is always on,
411. 26:01and I can keep the Claude app open all the time because that's one of the limitations of scheduled tasks.
412. 26:09Right.
413. 26:10They need the Claude app to be always running.
414. 26:14Otherwise, they will not trigger.
415. 26:16That's why, like with the remote control, I'm doing the same thing on always on Macs
416. 26:21because the terminal window has to be open.
417. 26:23Let me ask you where this is because I don't see it.
418. 26:27Huh.
419. 26:29Have you updated Claude for Mac?
420. 26:31I just did while you were talking because there's another complaint right there,
421. 26:35is that sometimes you realize after a while that you're out of date because you have to restart the app.
422. 26:41Try and go to Cowork, and you should see schedule.
423. 26:44Oh, it's under Cowork.
424. 26:45I was under chat.
425. 26:46No, it's in Cowork.
426. 26:48I see.
427. 26:48I see.
428. 26:49Okay, no, I got it.
429. 26:50I got it.
430. 26:51Your scheduled tasks don't sync across devices.
431. 26:55So right now on my Mac mini, I do not see the scheduled task that I created on the Mac Studio.
432. 27:02Okay?
433. 27:03So more fragmentation.
434. 27:05Yeah.
435. 27:05However, I took that prompt, turned it into a scheduled task.
436. 27:10Yeah.
437. 27:11And it just worked.
438. 27:13Oh, that's good.
439. 27:13I did a first test.
440. 27:15It successfully found my Gmail and Todoist connectors.
441. 27:20Ran a first pass of the prompt.
442. 27:23And I can even tell you, scheduled tasks in Cowork can use your existing skills.
443. 27:34Because as part of my prompt, I added one more step that said,
444. 27:38every time you're finished processing my email and creating tasks in my Todoist,
445. 27:43send me a notification with pushover using your pushover skill, which I personally created.
446. 27:50And it did.
447. 27:52So your Cowork scheduled tasks can use connectors, can use your file system,
448. 27:58and can use your skills.
449. 28:00Oh, that's nice.
450. 28:01And they can be scheduled.
451. 28:02So now, every day at 3.30 p.m.,
452. 28:05Claude Cowork running on my Mac Studio should be able to process my email for me.
453. 28:11All right.
454. 28:12Well, so let me ask you about it.
455. 28:14I want to ask you, this is like an aside, what about your email processing?
456. 28:17What kind of, give me an example of the kind of things you tell it to,
457. 28:21float to the surface for you?
458. 28:23Is it like people who I've responded to before and developers who are pitching apps and that,
459. 28:31or somebody from, with an apple.com email, that kind of thing?
460. 28:36I mean, you are answering yourself.
461. 28:38Okay.
462. 28:40Those are just the things that I kind of already have set up in a way with like splits and superhuman, right?
463. 28:45But you're doing it all as one task and it becomes...
464. 28:48I do have my splits.
465. 28:49But the thing is, if I don't create a task for me to nag me, I'm not going to...
466. 28:53Yeah, no, I'm the same way.
467. 28:55I'm terrible about email these days and I want to set this up myself.
468. 28:57Yeah, Claude MacStories support requests.
469. 29:01Yeah, sure.
470. 29:02Emails from Apple PR.
471. 29:03It excludes things like newsletters, junk, you know.
472. 29:09Marketing emails of various kinds.
473. 29:10Marketing emails.
474. 29:11It tries to look if that's a frequent contact.
475. 29:15And obviously it only processes emails that I have not responded to.
476. 29:19Sure.
477. 29:20Emails were the last message.
478. 29:22Threads were the last message is not from me.
479. 29:25Yeah, that's the kind of thing that I do.
480. 29:26So overall, I will say Anthropic has obviously nailed the productivity space and the programmer space.
481. 29:38Yep, I agree.
482. 29:39They have too much going on in the Claude app.
483. 29:45Way too much, especially when you're on something like an iPhone.
484. 29:48I would at least spin off Claude Code into a separate app.
485. 29:55See, I don't think they need to have three apps, right?
486. 29:58I don't think there needs to be Claude Chat, Claude Code work, and Claude Code.
487. 30:02Code work and code together, maybe?
488. 30:05That's what I was going to say.
489. 30:07Maybe code work and code together and just keep the chatbot the chatbot.
490. 30:11Yeah.
491. 30:12I don't know.
492. 30:13But three products into one application is too much.
493. 30:17Three separate applications is also probably too much.
494. 30:20So maybe the middle ground is too, which, surprise, is what we're going to talk about with OpenAI.
495. 30:27Yeah.
496. 30:28No, because it's, I mean, the thing about Cowork and Claude Code is they're very similar in the way that they operate.
497. 30:34And so I think they make sense together.
498. 30:36And you could just do it as like a, you know, a tab bar or some other kind of UI element to separate the two.
499. 30:43The thing that drives me a little crazy about the Claude app is what you're saying is that the way this app works, especially on the iPhone, is that it's got a little menu button at the top that opens the sidebar.
500. 30:57And this sidebar just goes on forever because it's chats, projects, artifacts, code.
501. 31:03And then it's starred things, which are going to be both projects and chats, and recents.
502. 31:09And to me, it feels like this.
503. 31:11First of all, I think this should be a tab bar app because you have this.
504. 31:16Oh, my God.
505. 31:16Right?
506. 31:17Because it should be a tab bar that has chats.
507. 31:21I think artifacts should be combined with something called documents.
508. 31:24Just call them documents.
509. 31:25And your documents are either reports you get in like Word or Excel or Markdown and artifacts, which are just, they're just web apps.
510. 31:34They're like little mini web apps.
511. 31:35So it's just HTML or React or whatever usually.
512. 31:38So there should be a document section, a chat section, and a code section if you were to keep code in here.
513. 31:44And then you have the search button, which they could use the new tab bar thing where search shrinks down to the far left.
514. 31:53Right?
515. 31:54And then you would move the new chat thing up to the top, to the upper right.
516. 31:59I'm just doing, this is just like a really down and dirty design review here.
517. 32:04Instead of like opening it up and you've actually got this long list of chats, which, you know, I don't know.
518. 32:11Yeah.
519. 32:12It's too much.
520. 32:13It's overwhelming.
521. 32:14It's too much.
522. 32:15So, and I think we agree that one app is too confusing.
523. 32:19Three, we probably don't want three apps.
524. 32:21No.
525. 32:22So maybe two.
526. 32:22That said, I really like Claude.
527. 32:24I mean, I'm a big fan of Claude.
528. 32:26All the apps, actually.
529. 32:27I mean, I find what we've just been talking about confusing and hard to keep up with.
530. 32:32And I think it could be simplified a lot.
531. 32:34But I like the product in general and most of these different things, features we've talked about.
532. 32:39Yeah.
533. 32:39Now, OpenAI is doing things very differently.
534. 32:43And I think I use ChatGPT and OpenAI a lot more than John.
535. 32:49Yes.
536. 32:50Yeah.
537. 32:50So, OpenAI, they also have the confusing thing where they have a bunch of things called Codex.
538. 32:58Yes, they do.
539. 32:59If I'm not mistaken, Codex started as a separate model.
540. 33:03Then they did the Codex on the web, which was basically Claude Code, but by OpenAI.
541. 33:10Then they did Codex CLI, which is Codex in the terminal.
542. 33:14And isn't Codex both open source and proprietary?
543. 33:20The Codex command line utility is open source.
544. 33:23Okay.
545. 33:24The whole thing is open source.
546. 33:26I see.
547. 33:27Codex for Mac, which is N for Windows now.
548. 33:31The Codex app is not open source.
549. 33:34Right.
550. 33:34But it's based on...
551. 33:36But it's using the open source tool.
552. 33:38It's using the open source hardness.
553. 33:40And yeah, Codex is entirely open source.
554. 33:44Now, I really like Codex.
555. 33:46I really do.
556. 33:48It's very good.
557. 33:49It's very good.
558. 33:50I really like the new GPT 5.4 model, but it's a very different experience and it's a very different vibe from Claude.
559. 33:59So, first of all, I would urge everybody to use the Codex app instead of the terminal.
560. 34:06I think OpenAI has done a pretty good job with the Codex app in the sense that it takes all of your pre-existing instructions or MCP connectors.
561. 34:19If you were using Codex in the terminal before, and it gives you a UI for everything.
562. 34:27Right.
563. 34:27That happens because the settings for the Codex app are based on the same settings files for the Codex CLI.
564. 34:38I see.
565. 34:38Everything lives in a hidden folder on your computer, .codex, where inside you will find a file called config.toml.
566. 34:49That is a fancy JSON file that contains all of the settings for Codex.
567. 34:56That same file can be configured both with Codex in the terminal and the Codex app.
568. 35:03The Codex app is more intuitive because it gives you a UI to manage those settings.
569. 35:10So, that means that if you use Codex before, once again, this is totally separate from ChatGPT, right?
570. 35:17ChatGPT has a website.
571. 35:19ChatGPT has ChatGPT for Mac, which I think is a pretty good Mac app, actually.
572. 35:23Yeah, it is pretty good.
573. 35:24They have adopted Liquid Glass.
574. 35:26They have the feature called Work with Apps, which lets ChatGPT control some Mac apps like Notion and Apple Notes or TextEdit.
575. 35:39Codex is totally separate.
576. 35:41And Codex is, right now, primarily for developers and programmers for doing things with code.
577. 35:46In fact, when you open Codex and you point it to a folder, it says, let's build.
578. 35:53Right.
579. 35:55It's designed for coding.
580. 35:57And if you are primarily on Claude like I am, I think one good way to get a sense for Codex is to take whatever you've been doing, building in Claude and dump it into Codex and have it do a backstop review.
581. 36:11That's one of the things that I do all the time with some of the web apps and things that I've been building is I just point it at the same folder and say, hey, I'm working on this thing.
582. 36:19I want you to dispatch some agents and do a thorough review of the code and tell me what you think and come up with recommendations.
583. 36:26And then you can either fix those things if it finds bugs or whatever right there in Codex or you can take that output report, bring it back to Claude and say, hey, I had Codex do this.
584. 36:36What do you think of these recommendations?
585. 36:38And then you can play them off of each other and kind of come down.
586. 36:42Because one of the things I think in a lot of code reviews that these chatbots all do, that these LLMs all do, is that they will sometimes suggest things that are more theoretical problems than real problems.
587. 36:55And so it can be useful to kind of see, all right, here's a list of things and then ask, play one chatbot off the other and see which ones are the ones that are actually real meaningful bugs that need to be fixed versus things that are just either hallucinated or just minor issues.
588. 37:11That is an excellent point.
589. 37:13Also, because Codex, by nature, loves to analyze and read a code base up front.
590. 37:20It spends a lot of, even if you set the thinking level to low or medium, it spends, by nature, a lot of time reading and analyzing up front.
591. 37:29And then it executes and it's fast.
592. 37:31But it's a very different sort of harness from Claude Code.
593. 37:36Yeah.
594. 37:36Claude Code usually just springs into action, sometimes to a fault.
595. 37:41Yes.
596. 37:42Codex loves to pre-think and pre-analyze everything.
597. 37:46So very different behavior.
598. 37:47Anyway, the Codex app, I mentioned, it picks up all of your existing MCP servers.
599. 37:53If you were using Codex in the terminal before, if you were not, you can connect to MCP servers right in the Codex settings.
600. 38:02They have some recommended connectors, things like linear, notion, but you can also connect to some custom MCP servers with the UI.
601. 38:11And it's confusing.
602. 38:13The UI for connecting to a custom MCP in Codex is too confusing.
603. 38:18There are fields for headers and token values.
604. 38:26What did they try to reduce?
605. 38:31I don't know.
606. 38:32They tried to give a UI to the different types of MCP authentication that exist, like STDO and streamable HTTP.
607. 38:41But in the process, they made everything less intuitive than just doing what I did today, which is I Googled, how do I add the Todoist MCP to Codex?
608. 38:53And I copied a terminal command, and I opened my terminal, and I connected.
609. 38:59Yeah.
610. 39:00So that part needs to be rethought a little.
611. 39:05Yeah.
612. 39:05They did a good job, I think, in Codex with the skills directory.
613. 39:10They have a built-in visual gallery with some pre-baked examples of skills that you can add to Codex.
614. 39:20And not just that, but Codex is also able to automatically read skills that you have in other agents folders.
615. 39:29Meaning, it can read skills that you have in your Claude slash skills folder on your computer.
616. 39:35So they're doing that sort of like, oh, we see that you're using another agent.
617. 39:39We can also sort of import your other skills for you.
618. 39:45OpenAI, since Codex is open source, these curated skills that they make, they're also open source.
619. 39:52You can find them on GitHub.
620. 39:53They have an OpenAI Codex skills repo that you can browse.
621. 39:57Because funnily enough, John, they used to make a skill for ChatGPT Atlas, the browser that they have semi-forgotten about.
622. 40:08And they used to make a skill to control Atlas and to fetch your Atlas tabs.
623. 40:13Yeah.
624. 40:14In a recent commit to the repo, that skill is gone.
625. 40:18They have replaced the ChatGPT Atlas skill, which I still have on my computer because I didn't delete it.
626. 40:24But they replaced it with a generic Playwright skill.
627. 40:28Now, Playwright, if you're not familiar, is a virtualized Google Chrome MCP server thing that you can install to let a chatbot or an agent test web apps that you're building on its own.
628. 40:44Right.
629. 40:45Right.
630. 40:45So, I don't know what OpenAI is doing with Atlas, but the pace of art is...
631. 40:49I wonder if it just is because Atlas is a bit...
632. 40:51I mean, Playwright's been around so long, it has probably a lot more features than what the Atlas one had.
633. 40:57Probably.
634. 40:59You can, as of a few days ago, in Codex, enable fast mode.
635. 41:05This is another fragmentation problem.
636. 41:07Fast mode for Codex is not the same as, you may recall, Codex 5.3 Spark.
637. 41:17Oh, right.
638. 41:19They're not the same thing.
639. 41:20Codex 5.3 Spark is a special, smaller, less capable model that OpenAI released on Cerebras, the fast inference provider that they entered an agreement with.
640. 41:37Fast mode is a setting that is exclusive to the new GPT 5.4 model, parentheses here.
641. 41:46It's just called GPT 5.4.
642. 41:49It's not called Codex 5.4.
643. 41:52Right.
644. 41:52So, it seems that they are unifying the model names.
645. 41:55They no longer have a separate Codex model, at least for now.
646. 41:59I thought that the Codex models were optimized for Codex.
647. 42:04I also thought so.
648. 42:05I also thought so.
649. 42:07Okay.
650. 42:07But then, why did they launch everything and said, oh, GPT 5.4 is the new coding model also?
651. 42:13Yeah, I guess that's true.
652. 42:14They did say that.
653. 42:15I don't know.
654. 42:16Anyway, fast mode goes, if I'm not mistaken, speed is increased by 1.5x, but your credits are consumed at a 2x rate.
655. 42:31Great.
656. 42:32So, you're going 50% faster, but you're spending twice as much.
657. 42:38Yeah.
658. 42:39Right?
659. 42:39Right.
660. 42:40100% more.
661. 42:40Yeah.
662. 42:42So, it didn't seem that much faster to me when I tried today.
663. 42:48It's a little bit faster, but it's not Cerebras' faster, you know?
664. 42:52Yeah.
665. 42:53So, you're mostly just burning tokens for no reason then.
666. 42:56Unless you're in a real hurry.
667. 42:58Unless you're in...
668. 42:59I suppose the accuracy is less too, though, isn't it too?
669. 43:03Isn't fast?
670. 43:04No, no, no.
671. 43:04No?
672. 43:04No, no.
673. 43:05I don't think they are dumping down the model when you're...
674. 43:08They're just asking you for more money.
675. 43:10They're asking you for more usage.
676. 43:12All right.
677. 43:13But it's the same model.
678. 43:15It's...
679. 43:15Because Spark is a completely different thing.
680. 43:19Yeah.
681. 43:19I would be curious to see if eventually we get a proper GPT 5.5 on Cerebras' model.
682. 43:28Yeah.
683. 43:29I mean, it's weird to me that they would both...
684. 43:32They would combine a dumbed-down model, which presumably would be dumbed down to make it faster,
685. 43:37with Cerebrus, instead of putting 5.4 on Cerebrus, and seeing what that chip can actually do.
686. 43:44It must not have been enough for it to be...
687. 43:47I think since Cerebrus...
688. 43:49To market it, I guess.
689. 43:49Yeah.
690. 43:50I guess Cerebrus is a very small company.
691. 43:52I don't think they have the resources right now.
692. 43:54Because Cerebrus, they're not like renting NVIDIA GPUs.
693. 43:58They're actually making their own chips.
694. 44:00Yeah.
695. 44:00And I think the problem is that they don't have enough.
696. 44:02They couldn't handle the load.
697. 44:03Yeah.
698. 44:04Yeah.
699. 44:04Yeah.
700. 44:07Finally, I want to mention automations.
701. 44:11Automations in Codex are kind of similar to scheduled tasks in Claude,
702. 44:20but they are more geared toward programmers.
703. 44:26So they have a similar UI.
704. 44:28You can create an automation in Codex for Mac to say,
705. 44:32every morning at 10 a.m., do a review of all my open PRs in this GitHub repo
706. 44:39and give me a summary.
707. 44:40Yeah.
708. 44:42The problem is that Codex, you can use it if you want as a general agent,
709. 44:50but it's very slow.
710. 44:52And the agentic loop inside of it is clearly not being designed for general productivity.
711. 45:00Right.
712. 45:01I tried today to connect my Notion and my Todoist MCP to Codex,
713. 45:05and I tried asking,
714. 45:07hey, can you give me a briefing of what's going on in my Notion daily notes
715. 45:10and also what's going on in my task manager today?
716. 45:13It took forever.
717. 45:15Really?
718. 45:15Like, he spent five minutes thinking about it, even with a low thinking level.
719. 45:21Oh, wow.
720. 45:22And so that's because, like, Codex is designed for analyzing code bases,
721. 45:30reviewing code, writing code.
722. 45:33It's not designed to be a digital assistant, right?
723. 45:37So I am not going to use automations in Codex because I am not a developer.
724. 45:41I don't need to run code reviews every morning.
725. 45:44I don't need to, you know, check out my open PRs every morning or things like that.
726. 45:49But if you are a developer, you can do that in the Codex app.
727. 45:51And it's the only way, I think, to schedule something with Codex,
728. 45:56with the Codex app.
729. 45:59And, yeah, that's essentially Codex.
730. 46:01The story with Codex right now is a lot less complicated than Anthropic and Claude
731. 46:06because they are not doing personal productivity yet.
732. 46:11That, I think, is the big missing item from the OpenAI offerings.
733. 46:16They're doing a chatbot, which is insanely popular,
734. 46:20and they're doing Codex.
735. 46:22It's interesting to me, though, when you think about that, too,
736. 46:25because they're the one that's, like, angling to be the consumer model,
737. 46:28whereas Anthropic is a little more business enterprise-y in general,
738. 46:33productivity-oriented.
739. 46:34You would think that tools like Calendar and Tasks and things
740. 46:38would be a lot more important to OpenAI, that they'd be on top of that.
741. 46:41You know, and this goes back to an old episode of AppStories,
742. 46:44you know what's missing from all of this?
743. 46:46The prosumer angle.
744. 46:48Once again, once again, we have one company that's catering to the enterprise
745. 46:53on one side, and then we have OpenAI on the other that's just doing
746. 46:57friendly consumer things.
747. 46:58What is the product for somebody like me or you?
748. 47:02Right in the middle.
749. 47:03No, I've thought about this, too.
750. 47:04I've thought about this a lot.
751. 47:05Is it OpenClaw?
752. 47:06But it's too confusing, and it's not intuitive.
753. 47:09Yeah, it's certainly not OpenClaw right now.
754. 47:10No.
755. 47:11Is it perplexity computer?
756. 47:13Maybe.
757. 47:14Maybe.
758. 47:15I think it's closer to...
759. 47:17I think Claude's closer.
760. 47:19I think it's something in between perplexity computer and Cowork.
761. 47:25Yes.
762. 47:25But the big problem with these two, and maybe we'll talk about perplexity
763. 47:29computer next time.
764. 47:30The big problem for these two, there is no mobile version.
765. 47:35There is no mobile story.
766. 47:37If I'm out and about, and I want to do something, or if I just got in bed,
767. 47:41and I'm like, oh, I forgot to do something, I cannot do it from my phone.
768. 47:46Well, that's why I was so happy when remote control came out,
769. 47:49because it's not like I use it every day.
770. 47:51But on the other hand, I'm not tied to my desk if I'm working on something
771. 47:54in Claude code in the terminal.
772. 47:56So, I mean, I think as we wrap this up, Federico, I guess we should explain
773. 48:01maybe which of these tools are the ones we're using the most.
774. 48:04And I think for me, with Claude, it's the chat interface.
775. 48:08I use that quite a bit, mostly for research type stuff.
776. 48:12And the terminal using Claude code for building web apps and some native apps
777. 48:17and other things like that.
778. 48:19And then I'm using Codex to bounce some of my code off of Codex and then
779. 48:26reintegrate it back with Claude code when I'm finished.
780. 48:29I think the piece that I haven't used a lot that I want to try out and use more
781. 48:34is Cowork because I've really enjoyed using Claude code as a coding tool.
782. 48:39And I think, you know, there's a lot of potential for that to integrate well
783. 48:44with productivity tasks.
784. 48:45I just haven't spent a ton of time in it.
785. 48:47I tend to use projects more probably because I have a lot of recurring tasks
786. 48:53that I have to do over and over.
787. 48:55Or I have long running tasks where I'm going to work on something
788. 48:58for an hour to today and then I might not pick it up again for a couple of days.
789. 49:03So it's nice to have a project for that kind of thing.
790. 49:06How about you?
791. 49:07I mostly use ChatGPT and OpenClaw, which eventually we'll do an episode
792. 49:17about OpenClaw, I guess.
793. 49:20And I use Codex a lot.
794. 49:21I don't use Claude code.
795. 49:23I mostly use Codex.
796. 49:24I just, I really like the nerdy sort of like cold vibe of Codex,
797. 49:29which is very different.
798. 49:33Codex is the cold, pragmatic nerd.
799. 49:38Claude code is the friendly, funny, approachable nerd.
800. 49:42And I don't know, for some things I just, I just prefer, I really like Codex.
801. 49:47Here's my problem.
802. 49:48I mostly use ChatGPT and Codex,
803. 49:51but I know in my heart that the tool that I want is Cowork.
804. 49:56It just kills me that it's not on the iPhone.
805. 49:58It kills me that-
806. 49:59It'll get there.
807. 50:00It'll get there.
808. 50:01It'll get there.
809. 50:01I really, really like Claude Cowork.
810. 50:06And I hope that they can make it faster, easier to access, easier to use.
811. 50:13It's very early, but it's exactly the sort of thing where it's like,
812. 50:17it's an LLM and a bunch of connectors,
813. 50:19but it can also use your skills and schedule tasks.
814. 50:22Like that's the kind of tool that I want to use for work,
815. 50:26but it's Mac only.
816. 50:27And it, you know.
817. 50:28Yeah, well, I mean, look, they brought code over to the iOS and the iPadOS app,
818. 50:34and that is even less of a good fit, I think, with those platforms than Cowork is.
819. 50:40So I'm sure Cowork will get there eventually.
820. 50:42Like we said at the top, I mean, I think Cowork has been out for like a month at most.
821. 50:47It's not been, you know, it feels like forever ago in the AI world,
822. 50:51but it really hasn't been that long.
823. 50:52Yeah.
824. 50:54Boy, I'm sure there will be some new words next week that we haven't covered here.
825. 50:59Yeah, we'll talk about the overhang and the harness again.
826. 51:03Oh my God.
827. 51:04Yes, sure.
828. 51:06All right, everybody.
829. 51:07Thanks for joining us for another episode of AppStories.
830. 51:10And thank you to Vitalee for supporting our efforts.
831. 51:14And of course, you can find the two of us over at MacStories.net
832. 51:17and on social media where Federico is at Viticci.
833. 51:21That's V-I-T-I-C-C-I.
834. 51:24And I'm at John Voorhees.
835. 51:26J-O-H-N-V-O-O-R-H-E-S.
836. 51:30Talk to you next week, Federico.
837. 51:31Ciao, John.