Anthropic just shared how they use Claude differently (Master it in 12 mins) — Transcript

Full transcript

1. 0:00In 1939, a student turned up late to a
2. 0:02statistics class, copied two problems
3. 0:04off the board thinking they were
4. 0:06homework, and handed in his answers the
5. 0:08next day. They were two unsolved
6. 0:10problems in statistics, and he solved
7. 0:13them because he didn't know that they
8. 0:15weren't solved. And according to the
9. 0:16lead engineer of the Claude app at
10. 0:18Anthropic, that is the mindset that the
11. 0:20top 1% of agentic AI users have. His
12. 0:24article on it called how to be more
13. 0:26ambitious, a tutorial for humans with
14. 0:28models, now has over 4.5 million views
15. 0:31on X. So today I'll share with you all
16. 0:33the insights he shared and go through
17. 0:35the three methods by which you can
18. 0:37become more ambitious with AI. And for
19. 0:40each one, I'll show you what most people
20. 0:41do right now and what the top 1% of AI
21. 0:44users' version looks like so that you
22. 0:46can apply it on your own to become even
23. 0:48more effective when you're using Claude
24. 0:50for your business. And if you're new, my
25. 0:52name is Jay. I spent over a decade
26. 0:53working with brands you probably know,
27. 0:55have been in AI since my master's in
28. 0:57data science, and now I'm leading our AI
29. 0:59business in one of the largest AI
30. 1:00communities globally. So Felix Riesberg
31. 1:02is a lead engineer at Anthropic, and
32. 1:04he's led the engineering for the Claude
33. 1:06app, co-work, and the Claude code
34. 1:08desktop app, which means he's worked on
35. 1:09a lot of the agents that we use every
36. 1:11day. And what he noticed is that, as he
37. 1:13puts it, some people seem to be flying
38. 1:15with AI, while many others are still
39. 1:17struggling to get their new wings to
40. 1:19work. Now, his fix in just three words
41. 1:22is to be more ambitious. But he admits
42. 1:25that's pretty tricky and general advice
43. 1:27to follow, so he broke it down into
44. 1:29types of ambition. And today we'll go
45. 1:31through three of them. So we'll go
46. 1:32through each of those three, and to keep
47. 1:34it concrete, I'll also give you examples
48. 1:36from different kinds of businesses for
49. 1:38each one. So let's start with the first
50. 1:39method, which is altitude. Felix says
51. 1:42people often flippantly call this
52. 1:44letting the model operate at a higher
53. 1:46abstraction, which basically means that
54. 1:48instead of asking for the task, you ask
55. 1:50for the outcome. Because when you give
56. 1:52Claude a task, it's only carrying out
57. 1:54your best guess at how to get what you
58. 1:56want. So, the question Felix says to ask
59. 1:58yourself is, "What do you actually want
60. 2:00to happen next? Or what's the goal
61. 2:02beyond the goal?" His own example is
62. 2:04that at the start of the last year, he
63. 2:05was still asking Claude to fix the
64. 2:07crashes in his apps one at a time. And
65. 2:10eventually, he just asked Claude to keep
66. 2:12the apps crash-free. And what that
67. 2:14resulted in is what Felix calls a
68. 2:16durable workflow, one that is more
69. 2:18effective at taking the human out of the
70. 2:20loop. So, he no longer has to hand
71. 2:22Claude task after task. So, in your own
72. 2:25business, what you can do is to look at
73. 2:26the task that you're about to hand
74. 2:28Claude and ask what it's actually for.
75. 2:30For example, let's say you're putting
76. 2:31together a pitch deck for a client. And
77. 2:33maybe you already have a slide template
78. 2:35for this. Now, what most people do is
79. 2:37ask Claude to just fill in the slides.
80. 2:39And to be fair, that works for most use
81. 2:40cases. But if the real goal is to
82. 2:42impress the client, then if you give
83. 2:44Claude that goal instead, it might come
84. 2:46back with a completely different
85. 2:48approach. If the pitch involves a venue,
86. 2:50for example, maybe it builds an
87. 2:51explorable 3D world of the space that
88. 2:53you can walk the client through. Or
89. 2:55instead of a boring slide pack made in
90. 2:56PowerPoint, it now builds you beautiful
91. 2:58HTML slides that are interactive, can
92. 3:00animate, and clearly show what you're
93. 3:02going to deliver. Now, Altitude works
94. 3:04best when you really are after some
95. 3:06ultimate definable outcome, like keeping
96. 3:08the apps crash-free, for example.
97. 3:10However, at least with the models right
98. 3:12now, it works less well for creative
99. 3:14work. So, for example, if you just ask
100. 3:16Claude to go make money and not make any
101. 3:18mistakes, that doesn't really work for
102. 3:20most businesses by itself in 2026.
103. 3:23Because that part still takes a lot of
104. 3:25creativity from you. Although Felix also
105. 3:27does say that models seem to be able to
106. 3:29handle higher and higher levels of
107. 3:31abstraction the more capable they get.
108. 3:33So, we'll see if that changes down the
109. 3:35line. And also, as we go deeper into the
110. 3:37next two types, just know that I've put
111. 3:39every resource, prompt, and learning
112. 3:41from this video into this free guide
113. 3:42that includes a skill, which you can
114. 3:44just send to your agent. So that if you
115. 3:46want your agent to be more ambitious
116. 3:47with a certain task, then it invokes all
117. 3:50the key points and best practices from
118. 3:52this video. That's just in the
119. 3:53description below if you need it. So,
120. 3:55now let's take it to a second level with
121. 3:57depth because sometimes you just want a
122. 3:59task done really, really well. And being
123. 4:01ambitious in depth means getting the
124. 4:03task done the way that a top expert in
125. 4:06that field would. And the trap that
126. 4:07Felix points out is that when you just
127. 4:09tell a model to act like a professional,
128. 4:11what it often does is just change his
129. 4:13vocabulary. It uses the jargon without
130. 4:15really changing its methods. So, how do
131. 4:17you get your agent to work like an
132. 4:19expert just sound like one? Well, Felix
133. 4:22actually uses three techniques. To show
134. 4:24them, we'll use his example, a company
135. 4:26called Acme that's losing customers
136. 4:27every month, which is called churn, and
137. 4:29it needs to get ready for its quarterly
138. 4:31review. The first technique is to insist
139. 4:33on the same tools that the best in the
140. 4:35industry use. Because to work on Acme's
141. 4:37churn properly, Claude would need the
142. 4:39tools that the experts themselves use.
143. 4:42So, at least for this problem, maybe
144. 4:43you're thinking of a customer success
145. 4:44manager or a pricing strategist. Or for
146. 4:47large-scale churn reduction work, you
147. 4:49would also need a data scientist. So, if
148. 4:51Claude just reaches for a simple table
149. 4:53in the chat and just goes ahead and does
150. 4:55the task, push back and ask it if that's
151. 4:57really what the best in the field would
152. 4:59use. And the good news is you don't even
153. 5:01need to know or understand what the
154. 5:02experts use because these models,
155. 5:05including Claude, usually does. Second
156. 5:07is that you need to pick goals that
157. 5:09sound and reasonable and just tell the
158. 5:10model to have a crack at it. For
159. 5:12example, if you need your website to
160. 5:14load in 5 seconds instead of 10, don't
161. 5:16just ask for 5 and see if Claude can get
162. 5:18it down to load in 1 second or even 200
163. 5:21milliseconds. And if Claude says that's
164. 5:22impossible, Felix says that you can just
165. 5:24tell it that it's very much solvable and
166. 5:26it just needs to keep going. Another
167. 5:28technique is to use an outside judge,
168. 5:30which means never asking the model to
169. 5:32judge his own work for your most
170. 5:34important tasks. And this is important
171. 5:36because if you notice, a lot of the AI
172. 5:37models that we have now are quite
173. 5:39agreeable and it actually doesn't take a
174. 5:41few prompts to convince them that what
175. 5:43you are saying is right. Now, Felix does
176. 5:45note that he expects this will get
177. 5:47solved as the models get smarter, but
178. 5:49for now he says to insist on outside
179. 5:51validation. Ideally, this is a tool that
180. 5:53the model can call, like let's say for
181. 5:55that website example, it can be a speed
182. 5:57test that it runs after every change. Or
183. 5:59if it's more creative work, spawning a
184. 6:01sub-agent which acts as a judge is
185. 6:03proven by multiple independent tests to
186. 6:05be quite effective in refining the final
187. 6:07output. And the good news is you can get
188. 6:09most of the way there for this method
189. 6:11with just a single prompt. And Felix
190. 6:13actually shares the one that he uses
191. 6:14himself to get a deeper version of any
192. 6:17task. It asks Claude four things before
193. 6:19it starts. What would a top professional
194. 6:21do here that you are not asking for?
195. 6:23What's the rawest material that they
196. 6:24would start from? Which methods and
197. 6:26tools would they use? And what, other
198. 6:28than you or Claude, would show that the
199. 6:31result is right? Then it rewrites your
200. 6:33prompt to ask for exactly that. And this
201. 6:35prompt you can just screenshot it, or
202. 6:37you can also access that skill that I
203. 6:38mentioned below, so that it's always one
204. 6:40short command away. And by the way, if
205. 6:42you want to learn how to build and sell
206. 6:44AI systems that businesses actually pay
207. 6:46for, then that's pretty much all we do
208. 6:48over at the Robo Nuggets community.
209. 6:49Where not only do you get access to the
210. 6:51Claude Living Masterclass, which we
211. 6:53update every week and takes you from
212. 6:54zero to mastery with the latest on AI,
213. 6:56but you also get access to our Agents as
214. 6:58a Service course, which walks you
215. 7:00through how to actually get paid for all
216. 7:02these AI skills that you are learning.
217. 7:03You also get to be part of a genuinely
218. 7:05great community of AI builders. In fact,
219. 7:07you can see just some of the recent wins
220. 7:09our members are getting from the program
221. 7:11right here. So if you want to start
222. 7:12earning from AI, then check that just in
223. 7:13the pinned comment below. Now back to
224. 7:15the video. And finally, the third
225. 7:16component of being more ambitious is
226. 7:19range. Because if altitude is looking at
227. 7:21things at a thousand feet view, and
228. 7:23depth is pushing each task to its
229. 7:25maximum possible output, range goes
230. 7:27sideways, which means looking at your
231. 7:29work from the point of view of other
232. 7:31fields. And probably more points of view
233. 7:33than would have ever made sense to do
234. 7:35versus the past. Now, this isn't a new
235. 7:37idea, because every single heist movie
236. 7:39that you to about would spend the first
237. 7:4130 minutes putting together a crew of
238. 7:43five different people and talents,
239. 7:44right? So, that's basically the same
240. 7:46principle. But, now it's really to have
241. 7:47your agents split a job across several
242. 7:50helpers at once. And for a very
243. 7:52important task, it helps if those
244. 7:54sub-agents have varying points of view
245. 7:56so that you can figure out what we call
246. 7:58unknown unknowns, which means the part
247. 8:01of your work that you didn't even know
248. 8:03is an option to begin with. Now, Felix
249. 8:05admits this one isn't easy because it
250. 8:07takes some creativity. For example, he
251. 8:10said that until he saw his colleague use
252. 8:12Blender, which is this free 3D software
253. 8:15to render interior designs, he never
254. 8:16thought of using Claude to build 3D
255. 8:18environments for his own website. So,
256. 8:20that's an example of an option that he
257. 8:22didn't even consider in the beginning,
258. 8:24an unknown unknown. So, given this
259. 8:26challenge, how do you actually do it?
260. 8:28Well, the first trick Felix uses is an
261. 8:30old one from a book called The Art of
262. 8:32innovation, which is to try replacing
263. 8:35the nouns with verbs. So, if we talk
264. 8:37about an example of a QBR, which is a
265. 8:40quarterly business review that a lot of
266. 8:41companies do, some people may describe
267. 8:43that output as a PowerPoint
268. 8:45presentation, which is the noun in this
269. 8:47case. The verb, and I'll read verbatim
270. 8:49from what Felix wrote here, is getting
271. 8:51busy people to believe a story about
272. 8:53cause and effect and to decide under
273. 8:55uncertainty in 30 minutes. And having
274. 8:58the verbs makes it much easier for not
275. 9:00just your agent, but also yourself to
276. 9:02see what the actual goal is and which
277. 9:04fields could bring in a unique and
278. 9:06useful point of view, which leads us to
279. 9:08a follow-up trick, which is to assemble
280. 9:10the crew. Meaning that you can insist
281. 9:12that Claude bring in several roles and
282. 9:15points of view and not just its own. So,
283. 9:17for the quarterly review example, Claude
284. 9:19would then bring an analyst, a sales
285. 9:21manager, and if it's being presented, it
286. 9:23can even bring in a graphic designer,
287. 9:25too. And a good practice, as we
288. 9:26mentioned in the earlier point, is to
289. 9:28bring in the point of view of a red
290. 9:30team, which if in case you don't know
291. 9:31that red teaming is basically this
292. 9:33practice where their job is to find
293. 9:35holes in the work, sort of like a
294. 9:37devil's advocate group. Now, obviously,
295. 9:38deploying helpers can be
296. 9:40token-intensive, so you can do this for
297. 9:41just your most important work, and you
298. 9:43can get started with a prompt like this
299. 9:45one, which you can just screenshot. Now,
300. 9:47just to provide an example before we
301. 9:48leave this point. Say you're a real
302. 9:50estate agent with a house to sell, and
303. 9:52selling this house is extremely
304. 9:53important to meet your revenue target
305. 9:55for the year. Now, what most people do
306. 9:57is just ask Claude to output the real
307. 9:59estate listing, which is the noun in
308. 10:01this case. But, the verb is, for
309. 10:03example, helping a family picture their
310. 10:05life in this house before they ever
311. 10:07visit. And once you say it like that,
312. 10:09then the mixture of experts that Claude
313. 10:11can bring in can change and can become
314. 10:13clearer. You can have it assume the role
315. 10:15of an interior designer to stage each
316. 10:17room, maybe a 3D artist to build a
317. 10:19walk-through, a bit like Felix's a
318. 10:20blender example. You can even assume the
319. 10:23role of a local who knows the best
320. 10:25restaurants nearby, so that the final
321. 10:27listing page that you get in the end is
322. 10:29not just some generic real estate page,
323. 10:31but one that is customized for that
324. 10:33project given its importance to you. So,
325. 10:35those are the three lovers of ambition.
326. 10:37But, what I really also like is how
327. 10:39Felix ends, because it all leads up to
328. 10:41one idea which he mentions, which is
329. 10:43this term called self-efficacy. And
330. 10:46basically, that means how much you
331. 10:47believe that you can pull something off.
332. 10:49And if you can increase it, especially
333. 10:50now that we have AI, you also increase
334. 10:52your chances of success in whatever
335. 10:54field or game you're playing in. And the
336. 10:56great thing is that self-efficacy can
337. 10:58actually be hacked in some way. There's
338. 11:00apparently more than 35 years of
339. 11:02goal-setting studies behind this, and
340. 11:04there's actually three hacks that are
341. 11:05pretty well documented. So, very
342. 11:07quickly, I'll just highlight each one
343. 11:09because it is quite interesting. The
344. 11:11first is believing it's possible, like
345. 11:13that student George Dantzig who solved
346. 11:15two never-before-solved problems in
347. 11:17statistics back in the 1930s because he
348. 11:20thought they were homework. The second
349. 11:21is watching somebody else do it. For
350. 11:23decades, for example, nobody could run a
351. 11:25mile in under 4 minutes. Then, Roger
352. 11:27Bannister did it in the 1950s, and only
353. 11:30after a few days, someone else did it
354. 11:32too. And the third is giving yourself
355. 11:34permission to fail, which is just
356. 11:36related to the idea that if you just
357. 11:38ship things and put your work out there
358. 11:40improving each time, then you will
359. 11:42eventually succeed. And as Felix Enz and
360. 11:44puts it, self-efficacy isn't really a
361. 11:46new concept, but it's a more powerful
362. 11:48one now. Because in a world where our
363. 11:50abilities make leaps every couple of
364. 11:53weeks, we can actually be more ambitious
365. 11:55in the things that we can bring the
366. 11:57world. But I think this also needs a
367. 11:58counterpoint because when we talk about
368. 12:00AI, we also need to talk about pacing
369. 12:02that ambition and even finding the why
370. 12:05behind that ambition. Or else you just
371. 12:06end up going faster and faster and
372. 12:08getting more and more ambitious without
373. 12:10a clear reason why. So really,
374. 12:12definitely taking these techniques and
375. 12:14methods to become more ambitious given
376. 12:16how powerful AI agents have become, but
377. 12:18at the end of the day, remember why
378. 12:20you're even doing it and feel free to
379. 12:22pause when you need to. As usual, thanks
380. 12:24for watching till the end. I usually do
381. 12:26sessions like these to help you use AI
382. 12:27better for your business. So if that's
383. 12:28your thing, feel free to do subscribe.
384. 12:30I'll see you next time. Cheers.
385. 12:32>> [music]