КАК ЭКОНОМИТЬ ТОКЕНЫ в Claude Code и ChatGPT — контекстное окно и промпты с нуля (2026) — Transcript

Full transcript

1. 0:00Yeah. Hello, AI entrepreneurs. In
2. 0:04general, today I'd like to talk about a
3. 0:06very important topic. And, by the way,
4. 0:08surprisingly few people talk about this
5. 0:10, even though there are so many videos
6. 0:12on the internet. But nobody talks about
7. 0:15what tokens actually are. Tokens in
8. 0:18Claude, tokens in ChatGPT, and any
9. 0:21other artificial intelligence. So let's
10. 0:25finally figure out what they really are
11. 0:27. Because without understanding the
12. 0:31unit economics, it will be impossible
13. 0:32to get any effective responses to our
14. 0:34requests later. And to avoid beating
15. 0:37around the bush, let's start from the
16. 0:39very beginning. Let's imagine that one
17. 0:41token is one piece of LEGO. They can
18. 0:44even be different colors; it doesn't
19. 0:45matter. For us, the most important
20. 0:47thing is understanding the analogy of
21. 0:49how this happens. It turns out that
22. 0:51when we type a certain prompt into a
23. 0:53dialog box, we are filling in specific
24. 0:55characters that our artificial
25. 0:57intelligence will later have to break
26. 0:59down into those very small LEGO bricks.
27. 1:02This is called tokenization. To better
28. 1:05understand this process, I made this
29. 1:07special widget for this video. Let's
30. 1:10take a look at it. This widget clearly
31. 1:12shows what tokenization is. So, let's
32. 1:15type in the following text as an
33. 1:17example. Hello, world. My name is Vlad
34. 1:22Vugan. What did we just see? It turns
35. 1:27out that it broke our words down into
36. 1:30those small cubes, the very same LEGO
37. 1:32pieces. And we ended up with exactly 11
38. 1:35of them. An interesting point. It turns
39. 1:39out that since our Russian language has
40. 1:41very complex words for programming, and
41. 1:44since developers most often use English
42. 1:46, our language is the most expensive
43. 1:49and uses the most tokens. The same
44. 1:53thing probably happens for the French,
45. 1:55because they have some words where
46. 1:57letters repeat up to four times, which
47. 1:59is why these languages are not entirely
48. 2:01cost-effective for this program. After
49. 2:05our words have turned into blocks, now
50. 2:08they need to be turned into a language
51. 2:10that the Anthropic processor
52. 2:12understands. To do this, we click on
53. 2:15this button, which is called "in
54. 2:16numbers". This is that very language.
55. 2:19This is the second stage that the
56. 2:21processor understands. After the
57. 2:23processor has analyzed your entire
58. 2:26prompt, it provides an answer in these
59. 2:28exact same numbers. And now these
60. 2:30numbers just need to be decoded back
61. 2:33into our words. So now we click "decode
62. 2:36". And this is the third stage, where
63. 2:40the words are reassembled. And this is
64. 2:44what we get. Hello, world, exclamation
65. 2:46mark. My name is, etc., Vlad Vugan, and
66. 2:50exclamation mark. That is how
67. 2:54tokenization happens. Our dialog window
68. 2:57, where we enter our prompt, we will
69. 2:59represent it as an empty box. It turns
70. 3:02out that when we write some text, we
71. 3:04are putting words into this box. After
72. 3:06we have filled it, we now have to send
73. 3:08it to the artificial intelligence to be
74. 3:10understood. But, but, but there are
75. 3:14limits. Where do these limits come from
76. 3:16? To understand this, let's represent
77. 3:19our chat as an ordinary room that has a
78. 3:22door on your side and a door opposite
79. 3:24where the artificial intelligence
80. 3:27enters. Inside this room, uh, there is
81. 3:30a certain limit from floor to ceiling,
82. 3:33from one wall to the other. Exactly
83. 3:36200,000 tokens. That is, it is
84. 3:39impossible to load anything more than
85. 3:42200,000 tokens here. And so it turns
86. 3:45out that our chat is indeed that very
87. 3:48limiter, in effect, as a room in which
88. 3:51you cannot put more than 200,000 of
89. 3:53these tokens, these LEGO bricks. And
90. 3:57each time, the larger the box of
91. 3:59various tokens you send, the more space
92. 4:02you occupy. What happens now? After you
93. 4:07have written a prompt, come to the
94. 4:10middle of the room, and placed this box
95. 4:12with ready-made LEGO pieces, the
96. 4:15artificial intelligence comes in from
97. 4:17the other side, accepts the box, reads
98. 4:20it, and brings you a similar box in
99. 4:22response. And thereby, it also fills up
100. 4:27the space. What does this mean? That
101. 4:30for effective action, you need to
102. 4:33create such high-quality prompts that
103. 4:37the number of tokens doesn't end up
104. 4:40clogging all this space. And the
105. 4:44quality of the prompt should be such
106. 4:46that it immediately provides steps
107. 4:48ahead for what you wish to do, rather
108. 4:51than saying, "Oh, do this for me now,
109. 4:53and then when you've done that, I'll
110. 4:56order some more actions from you." For
111. 4:59this, it is better to consistently
112. 5:01describe, step by step, what you expect
113. 5:03from the artificial intelligence. In
114. 5:06doing so, in its response, it will
115. 5:07bring a box of the same size in which
116. 5:09you will receive almost everything at
117. 5:11once. Later on, by filling this space
118. 5:15further, you can detail any nuances
119. 5:18that you would like to improve in its
120. 5:22answers. I would like to pause here. In
121. 5:26this room, there is such a concept as
122. 5:28understanding context. Everyone talks
123. 5:31about it, and they somehow try to link
124. 5:34them—contexts, tokens, everything.
125. 5:36What is this context? Every time you
126. 5:39ask the next question, you have brought
127. 5:41the first box, brought the second box,
128. 5:43brought the third box. And now you are
129. 5:46bringing the fourth box. And before
130. 5:50giving any answer, the artificial
131. 5:52intelligence begins to analyze all
132. 5:54those previous three boxes. It
133. 5:57literally goes into every box, looking
134. 5:59through every piece of LEGO to
135. 6:00understand what is being discussed. And
136. 6:04then suddenly, after four boxes talking
137. 6:06about the topic of nature, for example,
138. 6:09you come along and say, "Bring in a box
139. 6:11.""And how do you pickle cucumbers?"
140. 6:14You bring in a completely different
141. 6:16type of box, for instance, a red one.
142. 6:19Well, now the artificial intelligence
143. 6:21starts to study all these points, but
144. 6:22it goes through all those four boxes.
145. 6:25We were talking about nature, and then
146. 6:26suddenly there's a question about
147. 6:28pickled or perhaps salted cucumbers.
148. 6:29What is it supposed to do now? And
149. 6:32because of this, the artificial
150. 6:33intelligence starts to hallucinate
151. 6:35because it doesn't understand what is
152. 6:37being discussed, since you were just
153. 6:39talking about nature, but it might give
154. 6:41you an answer in the context of what
155. 6:42you were talking about in those four
156. 6:44boxes. If it’s nature and cucumbers
157. 6:47are involved here, then it might link
158. 6:49something together like: "What if you
159. 6:51sprinkle salt on the cucumber beds so
160. 6:53they are pickled right away?" It can
161. 6:56come up with nonsense like that because
162. 6:57it might have read something similar
163. 6:59somewhere in one of the blocks. What
164. 7:01comes of this? That it is impossible to
165. 7:04work with 200,000 tokens while setting
166. 7:08certain different topics or different
167. 7:11contexts. Therefore, if you are looking
168. 7:15into, if you are trying to figure out a
169. 7:17specific point, try to set the exact
170. 7:20pace, rhythm, context, and topic that
171. 7:22you need to break down from start to
172. 7:25finish, all while tracking your tokens
173. 7:27so you don't exceed the 200,000 token
174. 7:29limit. But there are certain nuances,
175. 7:34such as, for example, with OpenAI, with
176. 7:36ChatGPT, this room has windows. If
177. 7:40Claude doesn't have these windows,
178. 7:42meaning it’s 200,000 tokens and
179. 7:43that's it, the limit is reached, it
180. 7:45just says: "That's it, nothing more is
181. 7:47possible, I won't answer you anything,
182. 7:49start a new chat." Then ChatGPT does
183. 7:52have windows. This means that every
184. 7:56time you load a new box, the topic that
185. 7:58was there before might fall out of the
186. 8:01windows. And the next time, thinking
187. 8:03that, well, I already told it about
188. 8:05this, you ask the question, but it’s
189. 8:07as if it doesn’t know it because it
190. 8:09just automatically deleted it. What
191. 8:11should you do in such cases now? And
192. 8:14for this, if you are using Cursor—
193. 8:17which is likely why you are watching
194. 8:20this video—you should enter the
195. 8:23following prompt, which I recommend
196. 8:26adding to your Cursor settings. This
197. 8:30prompt will help you track your tokens
198. 8:32in every window, in every chat. That is
199. 8:35, it will show you the number of tokens
200. 8:38that have already been used. So you can
201. 8:40always monitor these moments. But I
202. 8:43want to add some more interesting
203. 8:45nuances regarding Cursor. Cursor code.
204. 8:49While the web version gives us 200,000
205. 8:53tokens, Cursor has a context limit of
206. 8:56up to a million tokens. What does that
207. 9:00mean? It means a larger room, more
208. 9:02space where you can maintain a much
209. 9:04longer contextual dialogue. But there
210. 9:08is another interesting point, because
211. 9:11if you are programming, working on
212. 9:14certain web designs for webpages, or
213. 9:17other tasks related to repositories or
214. 9:20actions that might involve creating
215. 9:23applications, then you should pay
216. 9:26attention to the fact that Cursor has a
217. 9:29command called "compact." Once you
218. 9:34realize that your chat, let's say, is
219. 9:36getting too full and you need to free
220. 9:39up some space, you can enter the "
221. 9:41compact" command, and it will discard
222. 9:44half of those message boxes itself. It
223. 9:48goes through each box, takes the most
224. 9:51essential part of each dialogue box—
225. 9:54the main theme—and discards the rest.
226. 9:59By doing so, it frees up space so you
227. 10:01can continue loading more of your own
228. 10:03messages. By the way, this function
229. 10:06only exists in Cursor. That is why I
230. 10:08recommend using this function more
231. 10:10often than you might, because different
232. 10:12cases arise. It might seem like you
233. 10:15managed to create a webpage in two
234. 10:17prompts, but in the end, it turns out
235. 10:19that a client might say they are
236. 10:24unhappy with the text, the format, the
237. 10:26colors, or perhaps the infographics.
238. 10:29And as a result, you have to refine or
239. 10:32rework certain elements. And since you
240. 10:35wouldn't want to open a new chat and
241. 10:37explain other topics all over again, it
242. 10:39is advisable to use this "compact"
243. 10:41function more frequently. So, it works
244. 10:45out that every six or seven request
245. 10:47boxes, you just press "compact." You
246. 10:50already understand that you won't be
247. 10:52returning to certain specific actions,
248. 10:53for example. But you need this chat to
249. 10:55be able to understand you in the future
250. 10:58within the framework of this context.
251. 11:00There you go. And by using this
252. 11:02function, it simply takes a part,
253. 11:04removes these LEGO pieces, and frees up
254. 11:06space so that you can continue chatting
255. 11:08and keep going. working with code. Now
256. 11:11let's summarize. Understanding what
257. 11:15context is, what tokens are, and
258. 11:17understanding all the specific elements
259. 11:19regarding limits, we should now
260. 11:21consider the factors of prompt
261. 11:23engineering. For this, I will
262. 11:27definitely leave a link so that you can
263. 11:30use these prompts to approach these
264. 11:32limits more effectively. Since right
265. 11:35now, at this moment, I can show you one
266. 11:37of our accounts. With the Max
267. 11:40subscription, as you can see, we have
268. 11:41already reached those limits. And it
269. 11:44turns out that only next week, that is,
270. 11:47in 2 days, yes, we will be able to use
271. 11:50our subscription again. In general,
272. 11:54even with the Max subscription, you can
273. 11:57reach certain limits, so be careful.
274. 12:00Well, I hope it is now clear to you how
275. 12:03the economy of this artificial
276. 12:04intelligence, like Claude, works, so
277. 12:06that you understand how to use it, what
278. 12:09tokens are, and can navigate all these
279. 12:11limits, because if you don't know this
280. 12:13basic foundation, your limits will burn
281. 12:16out very quickly. Yes, I would also
282. 12:20like to add that every time you use any
283. 12:23additional third-party services, such
284. 12:25as MCP, or perhaps you have certain
285. 12:28other integrations where Claude tries
286. 12:30to analyze the work, everything comes
287. 12:33at a price. So take this into account
288. 12:36as well. And always keep it in mind,
289. 12:39every time you add it to certain
290. 12:41Telegram bots or some other cognitive
291. 12:43actions that it must perform. and watch
292. 12:47that ceiling. To avoid hitting those
293. 12:51ceilings, I recommend learning Prompt
294. 12:53Engineering, which you can find on our
295. 12:55Telegram channel; the link will be in
296. 12:57the description. These prompt examples
297. 13:01will help you save your tokens very
298. 13:04effectively and not waste them the way
299. 13:06we wasted ours in the beginning. So, I
300. 13:11wish you success in this new era. After
301. 13:15all, AI entrepreneurs are only just
302. 13:17beginning to emerge. This is a new
303. 13:19generation of people, new actions that
304. 13:22can be performed, let's say, by any
305. 13:24amateur who does not have any
306. 13:26programming education. Therefore, we
307. 13:28are trying to figure things out and
308. 13:30understand them in simple language. I
309. 13:32think this episode turned out to be
310. 13:33interesting. I will be glad to see you
311. 13:35in the next episode. Bye everyone. M.