Never Pay For Claude Upgrades Again — Transcript

Full transcript

1. 0:00Let's be honest with each other. Claude
2. 0:01usage limits absolutely suck. I'm sure
3. 0:04you know the feeling. You're deep in a
4. 0:05Claude prompting session and all of a
5. 0:07sudden, bang, you've hit your usage
6. 0:09limit. And it doesn't matter if you're
7. 0:11on the $20 plan, the $100 plan, or even
8. 0:13the $200 plan, which I'm on, it still
9. 0:16seems to be an issue. Up until 3 weeks
10. 0:18ago, I was still running into rate
11. 0:20limits on the $200 plan and this was
12. 0:22happening just after an hour session of
13. 0:24Claude code. So, I started questioning,
14. 0:26why is this actually happening? And I
15. 0:28started digging deep into how Claude
16. 0:29actually works and I worked out that I
17. 0:31was using Claude wrong. And since then,
18. 0:33for the last 3 weeks, I haven't hit a
19. 0:35single Claude usage limit. And I
20. 0:37literally use Claude 6 hours a day. So,
21. 0:39in today's video, I'm going to give you
22. 0:40my full framework to never hit your
23. 0:43Claude usage limits again. After
24. 0:44watching this video, you're going to be
25. 0:45able to extract the maximum value out of
26. 0:48your Claude subscription. And some of
27. 0:49these tricks, I haven't heard a single
28. 0:51other creator talk about yet. If you
29. 0:53don't know who I am, I'm Miles
30. 0:54Deutscher, I'm 25 years old and I've
31. 0:56been using Chat GPT since it first came
32. 0:58out 3 years ago. Since then, I've been
33. 1:00going deep down the AI rabbit hole and
34. 1:02I've used AI to help power my agency
35. 1:04business to over $20 million in revenue
36. 1:06and my digital product business to over
37. 1:07$5 million in profit. As an AI
38. 1:10enthusiast, I created this channel to
39. 1:11help share my learnings with you. So,
40. 1:13hopefully, you can also use AI to
41. 1:15achieve your goals as well. All right,
42. 1:16so the first step to never hitting your
43. 1:18Claude limits again is all in your
44. 1:20workflow. A really simple thing you can
45. 1:22fix is being very intentional with how
46. 1:24you prompt Claude. Most people figure
47. 1:26out what they want whilst talking to
48. 1:28Claude. And although it's a great
49. 1:29brainstorming partner, don't use the
50. 1:31most powerful model like Opus 4.7 for
51. 1:34brainstorming. Make sure you really
52. 1:35think about your prompt first before you
53. 1:37use Claude. Or if you have to brainstorm
54. 1:39with Claude, simply use a cheaper model,
55. 1:41which we'll go through later in this
56. 1:42video. Model selection is a huge trick
57. 1:45that most people are missing and if
58. 1:46you're just using the default model,
59. 1:48that's probably one of your biggest
60. 1:49issues. The next thing you need to
61. 1:50understand is how Claude actually works.
62. 1:53Certain tasks take up more tokens. For
63. 1:55example, if I just chat to Claude, "Hey
64. 1:57Claude, how are you?" and ask it
65. 1:59questions and go back and forth, this
66. 2:00actually isn't using up many tokens.
67. 2:03Where you start running into token
68. 2:04issues is whenever Claude has to build
69. 2:06something. So, if you're creating a
70. 2:08drawing, if you're coding, if you're
71. 2:10creating an artifact, a dashboard,
72. 2:12that's what burns through tokens. So,
73. 2:14you're actually better off before
74. 2:15building something, doing more planning.
75. 2:17So, going back and forth to make sure
76. 2:19the build is going to be right. Let's
77. 2:20say you're building a finance dashboard,
78. 2:22going back and forth to make sure the
79. 2:24reason and architecture is fully correct
80. 2:26before you physically build it. Because
81. 2:28having to rebuild the dashboard is going
82. 2:30to chew up much more tokens than simply
83. 2:33speaking about building the dashboard.
84. 2:35That's a major learning for me now. So,
85. 2:36whenever I go and I build something, I
86. 2:38always plan beforehand. If you're a
87. 2:40coder, you can actually go into Claude
88. 2:41Code specifically, and you can click
89. 2:43plan mode. This will by default do no
90. 2:46code and only focus on planning. So, if
91. 2:48you're building stuff on Claude Code,
92. 2:50going into plan mode first, doing all of
93. 2:52your planning before you actually code,
94. 2:55is going to save you a lot of tokens.
95. 2:57So, that simple mindset shift, just from
96. 2:59having the understanding of how tokens
97. 3:00are actually burnt, makes a big
98. 3:02difference. All right, now let's get
99. 3:03into one of the major needle movers,
100. 3:06chat length. Long chats are a silent
101. 3:08killer. If you use a single chat and you
102. 3:11keep talking to it, you have to
103. 3:13understand how Claude works. It keeps
104. 3:15having to go and scrape thousands of
105. 3:17prior messages to understand the
106. 3:19necessary context for giving you a
107. 3:22response. So, what I prefer to do is
108. 3:24actually set up projects. I have one for
109. 3:25this YouTube channel, AI Edge, for
110. 3:26example, and every time I start a new
111. 3:28task like a video analysis or a new
112. 3:31video script, I open up a brand new
113. 3:34chat. This way, Claude is using up less
114. 3:36tokens trying to understand the context.
115. 3:38And the other really powerful thing you
116. 3:40can do here is in your project folders,
117. 3:42put in your memory and instructions. So,
118. 3:45tell Claude exactly what the purpose of
119. 3:46these chats are. For example, for AI
120. 3:48Edge, it has a memory of exactly what AI
121. 3:50Edge is, my goals with the business, my
122. 3:52goals with the channel. Then in
123. 3:53instructions, I tell Claude exactly what
124. 3:56I want from it, how I want it to format
125. 3:57its answers. Here's a trick if you're
126. 3:59really trying to extract the maximum
127. 4:00value from each plan. In your
128. 4:02instructions, you can actually write in
129. 4:04something like be cognizant of response
130. 4:07length, try and speak in a concise
131. 4:09manner, try and act in an efficient
132. 4:11manner. If you just write that in to
133. 4:13your instructions, your responses by
134. 4:15default, whenever you're in that
135. 4:16project, will actually be shorter, thus
136. 4:18saving you tokens. Oh, and by the way,
137. 4:20so you remember all of this stuff going
138. 4:22forward, there'll be a free link in the
139. 4:23description below which contains all of
140. 4:25the tips from this video in a single PDF
141. 4:28that you can then actually put into
142. 4:30Claude to help it configure your chats.
143. 4:32So, you simply claim that by going to
144. 4:33link in the description, signing up for
145. 4:35the AI newsletter, and you'll unlock the
146. 4:36Instagram where the assets can be found
147. 4:38inside the Google Drive. So, to
148. 4:40summarize the chat setup, I'll say it
149. 4:41like this. Three highly focused chats
150. 4:44beats one long chat every day of the
151. 4:46week if you set up your memory
152. 4:48correctly. So, I've spoken about the
153. 4:49memory system for the project folders
154. 4:52themselves if you're just using the
155. 4:53chats. Something that you may want to
156. 4:55consider that I've done another video
157. 4:57on, my second brain video, is actually
158. 4:59setting up custom memory cuz you have to
159. 5:01understand, you don't have full control
160. 5:04if you're just using the chats over
161. 5:05which memory Claude saves. It's
162. 5:07essentially a black box. Yes, it is
163. 5:09technically updating the memory, but you
164. 5:11don't know exactly what and how it's
165. 5:14updating. So, the best workaround for
166. 5:16this is to actually create a local
167. 5:17folder on your computer, like your AI
168. 5:19brain folder, which stores all of your
169. 5:21local memory. So, in a folder, you want
170. 5:23to configure it something like this.
171. 5:25Have an instructions.md file that
172. 5:27basically has the instructions for
173. 5:29Claude to follow. You'll notice there's
174. 5:30a very important line here. Whenever new
175. 5:33information surfaces during a session,
176. 5:35the relevant MD file must be updated.
177. 5:37Don't let facts slip through. This is so
178. 5:40important because the instructions are
179. 5:42actually getting the memory to auto
180. 5:44update. Because without this, Claude is
181. 5:46very random in terms of what it actually
182. 5:48adds to the memory file. If you have a
183. 5:50line like this in instructions, it means
184. 5:52memory is constantly updating. Don't
185. 5:54worry because in the free guide below,
186. 5:55I'm going to give you the exact prompt
187. 5:57to set up your memory and folder system
188. 5:59in the correct way. And if you want to
189. 6:01watch the dedicated video on that after,
190. 6:03I've got a video on my second brain
191. 6:04guide, which also connects this to
192. 6:06Obsidian. But for now, all you need to
193. 6:07know is that you need an instructions
194. 6:09file and a memory MD file. Then, if you
195. 6:12go back into Claude and you utilize
196. 6:14co-work, the cool thing about co-work is
197. 6:16you can work inside a specific folder.
198. 6:18So, I can choose my AI brain folder, and
199. 6:21every single time I'm typing a prompt
200. 6:23that's connected to that folder on
201. 6:24co-work, it's going to be tapping in to
202. 6:26the instructions and all of the specific
203. 6:29memory to that folder. And the great
204. 6:30thing about this is that you actually
205. 6:32own your own memory folder. It's not in
206. 6:34a Claude black box, so if you do want to
207. 6:36ever switch models in the future, or if
208. 6:38you want to plug it into another tool,
209. 6:40you actually own your memory. I
210. 6:41personally think memory sovereignty, so
211. 6:43owning your own memory, is the most
212. 6:44important thing in AI because memory is
213. 6:47transportable. If you have your memory,
214. 6:49you're going to be a very effective AI
215. 6:51user. The people that I find that aren't
216. 6:52very effective with AI don't own their
217. 6:54own memory and don't have the systems in
218. 6:56place to update their memory. The good
219. 6:58thing is, it's super simple and I've
220. 6:59actually developed a prompt for it to
221. 7:01save you time. All right, that was layer
222. 7:03one, your workflow. Let's talk about
223. 7:05layer two, and this is just so important
224. 7:07and I don't see enough people doing it.
225. 7:09This is model stacking. If you are using
226. 7:11Opus 4.7, the most expensive model to
227. 7:14ask questions, do basic scraping, and
228. 7:16basic research, you're just burning
229. 7:18through tokens. There are other models
230. 7:20that are 90% as capable on these basic
231. 7:23tasks that you can default to, and then
232. 7:25only use Opus 4.7 for the work where you
233. 7:29really need a smart model to shine. I'll
234. 7:30show you what I mean. So, if you go into
235. 7:32create a new chat on Claude, you can
236. 7:34select your model. So, Opus 4.7 is
237. 7:36clearly the smartest. So, if I was
238. 7:37making a big business decision, if I was
239. 7:39coding a financial dashboard, and if I
240. 7:41was analyzing really important data, of
241. 7:43course, yes, I will use this. Still with
242. 7:46the framework that I showed you in step
243. 7:47one. However, the other models are still
244. 7:49pretty good. Specifically, Sonnet.
245. 7:51Sonnet for most real work is great. If
246. 7:52you're analyzing a PDF, if you want to
247. 7:55update some text, if you want to do a
248. 7:56bit of writing, Sonnet in my opinion is
249. 7:58fantastic. And Haiku is still pretty
250. 8:00capable for quick tasks like extracting
251. 8:02data, scraping. There are probably
252. 8:04people that use Open Clo that are
253. 8:05watching this video. I do too. For
254. 8:07certain workflows, Haiku is a really
255. 8:09quick and easy way to scrape data from
256. 8:10the internet. And the way I do it is I
257. 8:12use the cheaper models for the scraping,
258. 8:14for the repetitive tasks, and then I'll
259. 8:16use Opus 4.7 as my curator. It's the
260. 8:19model that curates all of the research
261. 8:21from these other models. If you're not
262. 8:22using Open Clo, you can still implement
263. 8:24the strategy on the desktop application
264. 8:27of Claude itself through simply being
265. 8:29cognizant of what model you're using,
266. 8:31using Sonnet 4.6 if you think that the
267. 8:33task is less important or slightly
268. 8:35easier, and then just manually switching
269. 8:37to Opus 4.7 when you think the task is
270. 8:39slightly more advanced. On Claude Code
271. 8:41as well, if you're a coder, this is
272. 8:43something to be very cognizant of.
273. 8:45There's actually an effort level that
274. 8:46you can select. If you're on extra high
275. 8:48or max, you're going to chew through
276. 8:50tokens. Yes, you're going to get a
277. 8:51better result, but let's say you're only
278. 8:53using Claude Code to update local
279. 8:55folders, you're not going to need max
280. 8:56effort. You can get away with medium or
281. 8:58even low in some cases. I would only use
282. 9:00extra high or max effort if you're doing
283. 9:02a task where you really need precision
284. 9:04and output. And Claude Code, just like
285. 9:06the chats, allows you to select the
286. 9:07model. Now, I'm not going to lie, Opus
287. 9:094.7 is way better than the other models,
288. 9:11but if you need something quickly vibe
289. 9:13coded, like let's say, you know, a
290. 9:14little dashboard or just a interface or
291. 9:17a web portal or something relatively
292. 9:18easy, you don't need to use it. I would
293. 9:20just prioritize using it for the more
294. 9:21advanced use cases. And here's the
295. 9:23reality as well. You don't just need to
296. 9:25use Claude. I use Quen and Kimi on my
297. 9:28Open Clo. I use Grok for real-time news,
298. 9:31which is built into my X subscription. I
299. 9:33use Gemini as well, which is built into
300. 9:34my my subscription. You don't just need
301. 9:36to use Claude. So, just be cognizant of
302. 9:38when you're using the model. If you're
303. 9:39really running into issues and let's say
304. 9:41you're on one of the lower plans, just
305. 9:43make sure you use Claude for the work
306. 9:44where you need it. Understand where
307. 9:45Claude shines. It really shines when it
308. 9:47comes to interactive dashboards, when it
309. 9:48comes to coding. But if you just want a
310. 9:50little bit of news, you can use Grok.
311. 9:52You can have another cheaper model. You
312. 9:53can even use GPT. For voice prompting
313. 9:55and brainstorming, I do that all the
314. 9:56time and then I use Claude as my curator
315. 9:58because I know what it's good at. So
316. 10:00just understand the strengths of each
317. 10:01model and combine multiple models in
318. 10:03your workflow. I actually did a tools
319. 10:04video on this where I break down my use
320. 10:06case for my top seven AI models and in
321. 10:08that video, I break down the strengths
322. 10:10of each model. And I feel like that's
323. 10:12also one of the reasons I was able to
324. 10:13reduce my token expenditure cuz I've
325. 10:15started leaning on other models for
326. 10:17things where those models genuinely
327. 10:18shine. Because although Claude is great,
328. 10:20other models actually do some things
329. 10:21better. So, keep that in mind next time
330. 10:23you're doing a task. All right, now you
331. 10:25understand the second step, which is
332. 10:27selecting the right model for your use
333. 10:29case, let's talk about something
334. 10:31underrated, tool splitting. The
335. 10:33important thing to understand is that
336. 10:34although Claude is one application, it's
337. 10:36not just one product. You have the chat,
338. 10:39you have co-work, and you have Claude
339. 10:40Code. And you need to treat each
340. 10:42individually. And you even have extra
341. 10:43products now like Claude Design with its
342. 10:45own separate usage limit. Now, Claude
343. 10:47treats the token limits for Claude Code
344. 10:50and the Claude Chat the same with design
345. 10:52being separate. But this is where you
346. 10:54need to understand who you are as a user
347. 10:55because if you are a heavy Claude Code
348. 10:57user and you're chewing through your
349. 10:59overall plan with Claude Code, then what
350. 11:01you can actually do is hook up the API
351. 11:03to Claude Code. It will run separately
352. 11:06through the API. You need to go onto
353. 11:07Anthropic's website to the API section
354. 11:10to create one. And then this will
355. 11:12actually now be separate from your chat.
356. 11:14If you're a normal user who only uses
357. 11:15Claude Code sometimes, you can get away
358. 11:17with having it under your main plan. The
359. 11:19only thing I'd urge is that you actually
360. 11:21go into claude.ai/settings/usage
361. 11:26and actually look at where your usage
362. 11:27is. So just being aware of where your
363. 11:29limits are. For example, right now, I've
364. 11:31used up 12% of all models and it resets
365. 11:33on Friday at 9:00 a.m. so I'm all good.
366. 11:35But I've actually run out of Claude
367. 11:37design cuz I was going crazy on that
368. 11:38yesterday and that's a completely
369. 11:40separate topic because Claude design is
370. 11:42so usage-intensive. And now if I want
371. 11:44more tokens for Claude design, I
372. 11:46actually need to go in and I need to
373. 11:48purchase pay-as-you-go usage credits,
374. 11:50they call it, for Claude design. This is
375. 11:52also a strategy where instead of you
376. 11:54having to upgrade your entire plan, if
377. 11:56you notice that you're about to hit your
378. 11:57limits in let's say 24 hours and you
379. 12:00just need to do a few more prompts, you
380. 12:02can actually just buy more usage credits
381. 12:04instead of upgrading and spending the
382. 12:05extra $100, jumping from 100 to $200,
383. 12:08you might only need, you know, 5 to 10
384. 12:10extra dollars worth of credits to get
385. 12:12you over that hump. So
386. 12:13claude.ai/upgrade,
387. 12:14this will allow you to manage your plan.
388. 12:16So just being aware of it, I think is a
389. 12:17big thing and actually checking in on
390. 12:19your limits so, you know, you're not
391. 12:20blindsided when something happens. And
392. 12:22this will can also change your behavior.
393. 12:23If you notice you're at 75% and you need
394. 12:25the model for work tomorrow, then maybe,
395. 12:28you know, you won't go crazy vibe coding
396. 12:30application the night before. I think
397. 12:31that awareness is really key and
398. 12:32although it's annoying, it's something I
399. 12:33think a lot of people are missing. Now
400. 12:35unfortunately, Claude is actually
401. 12:36getting stingier with their limits. This
402. 12:38is a conversation for another day but
403. 12:40energy and compute is a real problem for
404. 12:42them and that's why we've actually seen
405. 12:44degraded performance across Claude over
406. 12:46the last month or so. You've probably
407. 12:47noticed Claude in some cases has become
408. 12:49dumber and although the new models are
409. 12:50great, like 4.7, it's also really
410. 12:53usage-intensive. And they've actually
411. 12:54taken it a step further and now as you
412. 12:56can see in the pro plan, you no longer
413. 12:58have access to Claude code. You can only
414. 13:00get the Claude code plan from the $100 a
415. 13:02month plan or the $200 a month plan. So
416. 13:04$17 subscription users no longer have
417. 13:07access to Claude code. I kind of get why
418. 13:09they're doing this. They are obviously
419. 13:11making a bet that serious users who want
420. 13:13to use Claude code are willing to pay
421. 13:14$100 and more and that they need to
422. 13:16conserve energy somewhere so they would
423. 13:18rather, you know, ice out the pro cohort
424. 13:21but it still does suck and this is just
425. 13:23something we need to accept about the
426. 13:24current state of AI. Prices are not
427. 13:27going to get better. They are only going
428. 13:28to get worse. So, I do feel like a video
429. 13:30like this actually might be one of the
430. 13:31most important ones of the year, so you
431. 13:33can save money longer term, because this
432. 13:35problem is only going to get worse. You
433. 13:36need to tweak your usage habits now,
434. 13:39unless you've just got money to burn,
435. 13:40because these features are only going to
436. 13:42become more restrictive. I mean, we're
437. 13:43already seeing features like Claude 2
438. 13:45design, which are amazing, the average
439. 13:46person can't afford. I blew through my
440. 13:48entire plan in a 12-hour session, and if
441. 13:51I want to keep using it now, I'm
442. 13:52probably going to spend $50 to $100 a
443. 13:54day if I want to use it seriously. So,
444. 13:56it is a problem, and if you're on a
445. 13:57budget, you might want to reconsider, as
446. 13:59I mentioned in the last step, which
447. 14:01tools you're actually using. And if you
448. 14:02do use an interface like Hermes Agent or
449. 14:04Open Claude, you might want to consider
450. 14:06even running local source models if you
451. 14:08want to front-load an investment into
452. 14:09hardware, like a solid Mac Studio that
453. 14:11can actually run local models, then
454. 14:13you're literally running an open-source
455. 14:14model on your computer, and then you're
456. 14:15not actually paying a subscription, but
457. 14:17you have to bear the upfront hardware
458. 14:18costs. So, your architecture will depend
459. 14:20on what you want to get out of AI, but I
460. 14:22think now we actually really need to be
461. 14:23cognizant of where our subscription
462. 14:25money is going. I still think Claude is
463. 14:27the best subscription to have overall,
464. 14:29but I think no longer is it a viable
465. 14:31option for absolutely everything just
466. 14:33due to the sheer cost issue that we are
467. 14:35experiencing. I think it's good to have
468. 14:37other subscriptions or even utilize
469. 14:39other free tools to spread the load up.
470. 14:41Now, I want to give you one of my
471. 14:42biggest tips in the video. If you really
472. 14:44want to save usage cost, if you have
473. 14:46tasks that you do on a repetitive basis,
474. 14:48if there are things that you do in your
475. 14:50day-to-day or in your weekly life,
476. 14:51simply create Claude skills for them.
477. 14:53It's a bit of work up front, but what
478. 14:55this will then generate is the ability
479. 14:57for you to hit {forward slash} and then
480. 14:59select a skill, for example, creative
481. 15:01session, which now loads up my morning
482. 15:03brief skill that I follow every day. So,
483. 15:06I've front-loaded the work, I've trained
484. 15:08Claude to do something, be my morning
485. 15:09assistant, and now whenever I want to do
486. 15:11my morning session, I simply enter the
487. 15:13skill. You can do the same for script
488. 15:15writing. You can train a Claude skill to
489. 15:16write like you. You can do the same for
490. 15:18portfolio analysis. You can train Claude
491. 15:21to analyze the portfolio in the way that
492. 15:23you want with the exact parameters that
493. 15:24you want. And the reason why this
494. 15:25actually saves tokens over time is
495. 15:27because Claude knows exactly what to do.
496. 15:29You're not constantly correcting it and
497. 15:31re-prompting, which is what ends up
498. 15:33burning tokens. So, I've actually done a
499. 15:34full guide on Claude skills, how to
500. 15:36develop skills from scratch. I'll leave
501. 15:37that in the top corner if you want to
502. 15:39check it out. And that's another feature
503. 15:40of Claude that you can utilize to
504. 15:42significantly reduce your token
505. 15:43expenditure. Now, let's say you're in a
506. 15:44situation where you've done everything
507. 15:46I've told you about in today's video,
508. 15:48and you are still running into usage
509. 15:50limits. Let's just be frank with each
510. 15:52other. Maybe it means you need to
511. 15:54upgrade your subscription. Now, I don't
512. 15:55think you need to jump to the $200 plan
513. 15:57if you haven't implemented anything I've
514. 15:59spoken about today. But if you have and
515. 16:01you're still running into issues, you
516. 16:02might just be a power user of AI, and
517. 16:05that is just the cost of using AI right
518. 16:06now. So, if you really want to use
519. 16:08Claude, if you love Claude code, if you
520. 16:09love Claude co-work, and you're still
521. 16:11running into issues on the $17 plan, you
522. 16:13may have to upgrade to 100. And if
523. 16:14you're still running into issues like I
524. 16:16was on the $100 plan, you may have to
525. 16:18upgrade to 200. That is just the name of
526. 16:20the game right now, unfortunately. Of
527. 16:22course, as I said, you can use other
528. 16:23models in order to alleviate some of
529. 16:25that burden, but if you're a Claude
530. 16:27power user, this is just the cost of
531. 16:29using the best frontier model in the
532. 16:30world right now. What I will say,
533. 16:31though, is don't just jump into the max
534. 16:33plan if you haven't optimized. Optimize
535. 16:35first and then make subscription
536. 16:37upgrades later. That's obviously the
537. 16:38best way to go about it. If you want all
538. 16:40of my tips condensed into one framework
539. 16:42from today's video, make sure to use the
540. 16:44link in the description below to claim
541. 16:46your free PDF. It's going to contain all
542. 16:48the tips and tricks, including the set
543. 16:50up prompt for an efficient instructions
544. 16:52and memory system. Thanks everyone for
545. 16:53your time today. Hopefully, I was able
546. 16:55to help you out, save you some money,
547. 16:57and help you extract [music] more out of
548. 16:58your Claude. I will see you in the next
549. 17:00video. Have a lovely rest of your day.
550. 17:02Peace out.