Give Me 10 Minutes and I'll Make Your Claude 100X Better — Transcript

Full transcript

1. 0:00Boris Cherny just killed every prompt
2. 0:02template for Claude ever. You heard that
3. 0:04right. And Anthropic even took it one
4. 0:06step further and deleted more than 80%
5. 0:08of Claude's own instructions. So there's
6. 0:10a huge chance that the way that you're
7. 0:12using Claude right now is completely
8. 0:13wrong. [music] And so today I'm going to
9. 0:15show you each rule one by one so that
10. 0:17your Claude gets a hundred times better.
11. 0:19Let's go. On October 6th, Boris shared a
12. 0:21real prompt that he used that completely
13. 0:23broke the internet. [music]
14. 0:24The post got 1.4 million views because
15. 0:26people were completely shocked at how
16. 0:28simple this prompt was. And his response
17. 0:30was I'm surprised that people are
18. 0:32surprised about this. You see because up
19. 0:34until now the way that we taught
20. 0:35prompting was like this. First you give
21. 0:37it a role. You're a world-class
22. 0:38copywriter with 20 years of experience.
23. 0:41Great. Then you give it a goal. I would
24. 0:42like for you to create a viral post on
25. 0:44LinkedIn. And finally you would give it
26. 0:46a long list of instructions. Step one,
27. 0:49step two, step three. And a lot of
28. 0:50people started saving prompt libraries
29. 0:52like they were gold. And that actually
30. 0:54made sense. Boris himself said back in
31. 0:56Sonnet 3.5 days your prompt mattered a
32. 0:59lot. And that's because the models we
33. 1:01used to work with were a lot less
34. 1:02sophisticated. But now with Opus 5.5,
35. 1:05Fable, Sonnet, the new advice is not to
36. 1:07talk to Claude like it's a coworker.
37. 1:09Even mentioning that there is no secret
38. 1:11to prompting. In fact, Andrej Karpathy,
39. 1:13one of the leading people at Anthropic,
40. 1:15even said my prompts are getting shorter
41. 1:17and dumber. And so instead of telling
42. 1:19Claude you are an expert marketer with
43. 1:2110 years of experience, you could simply
44. 1:23tell Claude things like hey, could you
45. 1:24turn that webinar into 10 posts for next
46. 1:26week? And Karpathy even added it's
47. 1:28almost to the point where if you give it
48. 1:30too much you're constraining the model
49. 1:32with too many requirements when the
50. 1:34thing it would have done otherwise would
51. 1:36have been better. So these sophisticated
52. 1:37prompts could actually be holding back
53. 1:39your AI. But my friends, this doesn't
54. 1:41mean there isn't a new way to do it
55. 1:42which is a hundred times better. So step
56. 1:44one for this, you're going to ask in
57. 1:46plain English what you're asking for.
58. 1:48Then you can provide it a tiny bit of
59. 1:49context if needed. And finally attach
60. 1:51what good looks like. If you have any
61. 1:53examples of what you would like to
62. 1:54follow, then do that. So for the example
63. 1:56I gave you earlier, it would look
64. 1:58something like this. Turn this
65. 1:59transcript into 10 LinkedIn posts for my
66. 2:01client, a small accounting firm. They're
67. 2:03the business owners who hate doing their
68. 2:05books. I've attached three posts my
69. 2:07clients loved, so please match that
70. 2:09voice. But the next sentence is probably
71. 2:11more important because the new
72. 2:12instructions tell you to tell your AI
73. 2:14that if anything is unclear to ask you
74. 2:16specific questions. So you can even
75. 2:18start with a more simple prompt and then
76. 2:20it'll come back and ask questions like,
77. 2:21who's the client and what do they sell?
78. 2:23What do they send you and how often?
79. 2:25What would make them cancel? And so now
80. 2:27it's going to have all the context it
81. 2:28needs to go out and do that task. But
82. 2:30there's still one problem. Often times
83. 2:31people tell it, go and review your work,
84. 2:34but often times it doesn't really know
85. 2:35what to check for. So you need to be
86. 2:37able to give it a way to verify the work
87. 2:39that it's done so that it's able to do
88. 2:41it at the level that you want it to be
89. 2:42done. And if you follow those three new
90. 2:44rules, you're going to be prompting like
91. 2:45the top 1%. However, if you're anything
92. 2:48like me, you pretty much set your AI to
93. 2:50the same model and the same effort level
94. 2:53every time and then just ask all of your
95. 2:54questions, do all your work like that.
96. 2:56Well, it turns out that's actually not
97. 2:58the best way to do things. And that's
98. 2:59not just because you're blasting through
99. 3:01credits every time. It's also because
100. 3:03different effort levels are really best
101. 3:05suited for specific types of tasks. And
102. 3:08later in this video, I'll teach you how
103. 3:09to get three times more work out of your
104. 3:11existing subscription. But let me break
105. 3:12down the misunderstanding. People think
106. 3:14the effort level has to do with how
107. 3:15smart it is. And to a certain extent,
108. 3:17that is true. And a member of the
109. 3:19Anthropic team called Tariq actually
110. 3:21boils it down into one sentence. The
111. 3:23effort level depends a lot on how in the
112. 3:25loop I want to be. Meaning that for some
113. 3:27tasks, you want to basically be steering
114. 3:29at every step of the task and others,
115. 3:31you want them to be able to be done end
116. 3:33to end. So let's break down each level.
117. 3:35Low is for quick exchanges where you
118. 3:37review each result such as brainstorming
119. 3:40or maybe a first sketch. Give me 12
120. 3:42subject lines for this email. Medium is
121. 3:44the recommended default for most
122. 3:45day-to-day work. So for example, turn
123. 3:47this webinar into three LinkedIn posts
124. 3:49in the voice of these three. You know
125. 3:50exactly what you're looking for, medium
126. 3:52is the right choice. High is for work
127. 3:54where verification matters or edge cases
128. 3:57are likely. An edge case is pretty much
129. 3:58an exception to the rule. Build this
130. 4:00client's monthly report from these three
131. 4:02spreadsheets. Every number has to match
132. 4:04the source. And if the client's going to
133. 4:05see these, you want to be on high. Now,
134. 4:07extra high is for long, more complex
135. 4:09jobs. Go through all 12 months of these
136. 4:12invoices and flag every one that doesn't
137. 4:14match the bank statement. I'll check
138. 4:15back in the morning. And Max is pretty
139. 4:17much for the hard problems that you want
140. 4:19Claude to work through without you
141. 4:21involved at all. Find everything in this
142. 4:22client's contract that could cost me
143. 4:24money and check each one against these
144. 4:26three past disputes. And one thing to
145. 4:28note is that Tariq warns that if you go
146. 4:30for higher models, they will also make
147. 4:32more assumptions on your behalf. But
148. 4:34that brings up a huge question, which is
149. 4:36that if effort level is not directly
150. 4:37correlated to quality of work, then how
151. 4:40are we supposed to determine quality?
152. 4:42You see, for yours, the way we did this
153. 4:43was pretty much at the end of each
154. 4:44prompt say, make sure to check your work
155. 4:47or something like verify before you
156. 4:48respond. And that made sense because the
157. 4:50old models were simply less capable and
158. 4:52were not as good at doing things in the
159. 4:55first place. So you had to have them go
160. 4:56and review their work afterwards. But
161. 4:58Opus 5.5 actually reviews its own work
162. 5:00without you actually asking it, meaning
163. 5:02that adding that one sentence at the end
164. 5:03of your prompt is simply going to cost
165. 5:05you more with no actual improvement. So
166. 5:07here's the new way. You want to really
167. 5:08set the expectation before you start the
168. 5:10job. You basically tell it what done
169. 5:12looks like. And here are Anthropic's own
170. 5:14words. Give Claude something it can run
171. 5:16that comes back pass or fail. And I put
172. 5:18all of the instructions and examples
173. 5:20inside a card inside of my free
174. 5:22community. You can check that out for
175. 5:23free by clicking in the description. So
176. 5:25to have it check its own work, I would
177. 5:27do something like this for the webinar
178. 5:28to post example. Does every post use a
179. 5:30real quote from the transcript? Is every
180. 5:32number in the post also in the
181. 5:34transcript? Is every post under 150
182. 5:36words? You want to create small tests
183. 5:38that are really yes or no. If you ask it
184. 5:40something like, is it professional? It's
185. 5:42a vague answer. We need things that are
186. 5:43basically yes or no answers. And then
187. 5:45you can say, before you show me, go
188. 5:47through this list and give me pass or
189. 5:49fail on each line with the evidence. And
190. 5:51another way they recommend to check your
191. 5:53work is to create a new session, provide
192. 5:55it with the outcome, and ask it to
193. 5:57criticize it. So that's now what good
194. 5:58prompting looks like. But there's
195. 6:00another problem, instructions. Before
196. 6:02Claude does anything, it goes and reads
197. 6:04a called Claude.md, which provides it
198. 6:07with the context it needs to actually
199. 6:09start doing the job. And so all of us
200. 6:11started setting up these hyper complex
201. 6:13things with lots of rules. We put
202. 6:15important in capital letters before
203. 6:17certain things and capitalize never or
204. 6:19always and have certain instructions
205. 6:21under those. Well, it turns out that
206. 6:22that's no longer the right way to do
207. 6:24things. You see, Anthropic in July 2026
208. 6:27actually deleted more than 80% of its
209. 6:29own instructions and got similar or
210. 6:31better results. Their words exactly were
211. 6:33no measurable loss on our coding
212. 6:36evaluations, which further proves the
213. 6:37better models require less instruction.
214. 6:40And again, Anthropic says that every one
215. 6:41of these pieces of advice was great for
216. 6:43back in the day, but after tests, they
217. 6:45no longer hold up in 2027. So you want
218. 6:47to simplify these instructions. Also,
219. 6:49provide a little bit more context around
220. 6:51your goal. Give Claude more information
221. 6:53about your goals. So I would recommend
222. 6:55you ask Claude to actually review your
223. 6:57files and to recommend a simplified
224. 6:59version based on the latest instructions
225. 7:01that Anthropic is giving. And because a
226. 7:03couple months ago, memories became
227. 7:04automatic, what happens is that as
228. 7:06you're actually working with it, it
229. 7:08actually learns from each time that
230. 7:10you're working with it, meaning that it
231. 7:12starts creating rules around the things
232. 7:14that you ask for, focusing less on the
233. 7:16instructions you've given it at the
234. 7:17beginning and more at fine tuning the
235. 7:19way you like things done. So Claude is
236. 7:21getting way better, but the problem
237. 7:22still remains that you are blasting
238. 7:24through credits. So let me teach you
239. 7:26three clear techniques for you to be
240. 7:27able to save credits and get probably
241. 7:29three times the usage out of your
242. 7:31current subscription because I've
243. 7:32created three mods that you can
244. 7:34literally download and install right now
245. 7:36for free that I use in my own business
246. 7:38every day. Now if you don't know what a
247. 7:40mod is, it's pretty much a little card
248. 7:41that lives inside of Claude that appears
249. 7:44when you actually do certain things. Let
250. 7:46me give you an example. The first mod
251. 7:47that I've created is called Effort
252. 7:49Picker. This is a mod that appears like
253. 7:51a little card above my chat right as I
254. 7:54give it a prompt. And it basically
255. 7:55identifies the right model for that
256. 7:58specific task and recommends me to
257. 8:00choose that as the actual effort level.
258. 8:02Now I can choose to ignore it or I say
259. 8:04good to go, let's use that. That way you
260. 8:06never have to figure out on your own
261. 8:08which is the best effort level for the
262. 8:10task, meaning that you're saving so many
263. 8:12credits by just not blasting high effort
264. 8:15on every task. Now the second mod that
265. 8:16helps me save a bunch of credits has to
266. 8:18do with context windows. Basically, the
267. 8:20more you chat, the more that window is
268. 8:22filling up with context. And as that
269. 8:24chat gets filled up with more and more
270. 8:26context, it basically costs a lot more
271. 8:29to continue to work within that chat. So
272. 8:31instead of spending more credits by
273. 8:32using the same chat, I usually like to
274. 8:34hand it off to a new chat and provide
275. 8:37that new chat with all of the context it
276. 8:39needs. And so I installed a mod that
277. 8:41tells me how full that window is. And as
278. 8:43soon as it hits about 80 85% I hand it
279. 8:46off with a forward slash hand off. You
280. 8:48don't lose any of the context and it
281. 8:50saves a bunch of credits. Now the third
282. 8:51mod to save credits actually identifies
283. 8:53if a chat has gone cold. You see these
284. 8:56chat windows actually stay active for up
285. 8:58to 60 minutes without touching them. But
286. 9:01after 60 minutes they get cold. If you
287. 9:03come back from your lunch break and
288. 9:04start using that same chat, it will
289. 9:06basically have to reread the entire
290. 9:08context to start working again. And so
291. 9:10this mod simply tells me this has gone
292. 9:12cold, allowing me to again hand it off
293. 9:15to a new chat and keep going. [music]
294. 9:16And I've actually given you all of the
295. 9:17mods that I use on a daily basis to
296. 9:19download and install for yourself. And
297. 9:21you can check out this video if you'd
298. 9:23like to learn how these mods actually
299. 9:25work, what they do inside of my business
300. 9:27because they will just make your Claude
301. 9:28code 100 times better. Hope you enjoyed
302. 9:30this video. If you did, like and
303. 9:32subscribe to the channel. It helps us a
304. 9:33lot. I appreciate your time here. Let's
305. 9:36go.