This NEW Claude Prompting Technique is blowing people's minds (gauntlet-loop) — Transcript

Full transcript

1. 0:00There's a new prompting technique for
2. 0:01Claude that's been blowing people's
3. 0:02minds over the past week. Because in a
4. 0:05single prompt, it can build fully
5. 0:06playable games and hyper custom 3D
6. 0:08worlds like these that even Karpati says
7. 0:11might be the future of prompting LLMs.
8. 0:13So today, I'll share with you this
9. 0:15technique called the gauntlet loop,
10. 0:16which might just be the quickest way for
11. 0:18you to learn how to fan out sub agents
12. 0:20to do work for you, so that even if
13. 0:22you're not into game development, you
14. 0:23can add this tool to your arsenal and
15. 0:25instantly get better at agentic AI. And
16. 0:27by the end, I'll share with you a skill
17. 0:29that lets you fully take advantage
18. 0:30[music] of this technique in the easiest
19. 0:32way possible. And if you're new, my name
20. 0:34is Jay. I spent over a decade working
21. 0:35with brands you may know, have been in
22. 0:37AI since my masters in data science. Now
23. 0:39I'm running an AI business and one of
24. 0:40the largest AI communities globally.
25. 0:42Let's dive into it. [music]
26. 0:45So, I first saw this prompting technique
27. 0:47from Matt Schumer who posted this insane
28. 0:49demo over at X where it already garnered
29. 0:51something like 4.8 million views. And he
30. 0:54says here that Claude Opus 5 oneshotted
31. 0:56this entire game with everything you see
32. 0:58in the demo being custom code without
33. 1:00any single external asset. And if Matt's
34. 1:02name is familiar and if you're in the AI
35. 1:04space for a while, that might be because
36. 1:05he actually wrote this article called
37. 1:07something big is happening which a lot
38. 1:08of people read a few months ago now
39. 1:10sitting at 87 million views. Point being
40. 1:12that he has been working with AI for
41. 1:14quite a while already and is actually a
42. 1:16good source from prompting techniques
43. 1:17like these. And if you see a claim like
44. 1:19this where an AI model supposedly
45. 1:20oneshots a game that looks as good as
46. 1:22this, complete with sound. By the way,
47. 1:24I'm not sure if you can hear that if I
48. 1:25just turn on the sound. Usually with
49. 1:27this, your first reaction would be a bit
50. 1:29skeptical if it was even made by AI,
51. 1:31which is quite understandable because
52. 1:33really the level of graphics here is
53. 1:34already quite extraordinary. But a few
54. 1:36days ago, Matt actually shared this
55. 1:38article where he went through how he
56. 1:40created this game and he's calling it
57. 1:42the gauntlet loop. And since then,
58. 1:43people have used that gauntlet loop
59. 1:45prompting technique to recreate that
60. 1:47same level of build quality. So, to show
61. 1:49a few examples, here's one where he
62. 1:51recreated the starting area for Pokémon
63. 1:53in Perfect 3D. Here is an example for a
64. 1:55car racing simulator game. And this is
65. 1:57one where it's more of like a Mario Kart
66. 1:59type of game. And this is just crazy how
67. 2:01wellbuilt this looks. Like, you can see
68. 2:03the different textures of this
69. 2:04environment, like with the road, the
70. 2:05houses there. And there's just so much
71. 2:07detail that the AI model was able to
72. 2:09build out in this one game. And you
73. 2:10might not be into game development in
74. 2:12particular. And later on, we'll show
75. 2:14some use cases of how you can apply this
76. 2:16outside of just video games. But
77. 2:18personally, I still like to pay
78. 2:19attention to these demos because it just
79. 2:21points to how much raw capability these
80. 2:23AI models now have. And Andre Karpati
81. 2:26was able to probably articulate it
82. 2:27better than I can where last weekend he
83. 2:30made this post where he's saying that
84. 2:31we're starting to leave the territory
85. 2:33where you would test an LLM by creating
86. 2:35an SVG of a pelican on a bicycle, which
87. 2:37is this old test that AI models were
88. 2:39used to be run on. And he mentions here
89. 2:41that these kinds of examples are great
90. 2:43because no one in their right mind would
91. 2:44ever spend the time to write something
92. 2:46this custom. But LMS and AI models have
93. 2:49all the stamina and patience in the
94. 2:50world. So these hyper custom worlds and
95. 2:523D environments are a really great
96. 2:54example of new capabilities that you
97. 2:56yourself as an AI user are now able to
98. 2:58tap into that you couldn't really do
99. 3:00before. So what is the gauntlet loop
100. 3:02exactly? Well, thankfully Matt also
101. 3:04shared his exact prompt here. And
102. 3:06surprisingly it is quite simple. It is
103. 3:08only a threeline prompt. And so you can
104. 3:11see I just pasted that whole prompt in
105. 3:12here. And what's actually more
106. 3:13interesting here versus the actual
107. 3:15verbiage of this prompt is just the
108. 3:17pattern and structure of it. Because if
109. 3:19you really break this down into these
110. 3:21three lines, essentially what you have
111. 3:22is a prompt structure that you can copy
112. 3:24yourself where first you give it a task
113. 3:26of what you want to happen. In this
114. 3:28case, the task that Matt was going for
115. 3:29is to build a firsterson shooter game.
116. 3:32And then the second part here is
117. 3:33essentially the build method. to how
118. 3:35that agent is going to achieve that task
119. 3:37where he's asking the main agent to fan
120. 3:39out sub agents and have each of those
121. 3:41tackle each task individually and to
122. 3:43have a separate sub agent check it
123. 3:45visually to ensure that it looks really
124. 3:47really good. And then finally, the third
125. 3:49part to this is the bar to hit, which is
126. 3:51essentially the standard where the agent
127. 3:52can decide when it can stop. And so he's
128. 3:55saying here to not stop until each sub
129. 3:57agent is utterly wowed with the quality
130. 3:59when compared with the actual Call of
131. 4:01Duty game. And what actually makes this
132. 4:02gauntlet loop so effective are these two
133. 4:05parts right here. Because if you haven't
134. 4:06tried using sub agents to orchestrate
135. 4:08your work before, then this might just
136. 4:10be one of the easiest and quickest way
137. 4:12for you to try it out. But just to step
138. 4:13back in case you don't know what we're
139. 4:15referring to when we talk about sub
140. 4:16agent orchestration. Essentially, when
141. 4:18you prompt an agent or talk to an AI
142. 4:20model, there's three levels to it. At
143. 4:22least in how I think about it. The first
144. 4:24level, which is the most basic and
145. 4:26probably the most common, is when you do
146. 4:27work with an agent, you send a prompt.
147. 4:29It provides an output back to you. You
148. 4:31verify if that output already matches
149. 4:33your standard and then you send another
150. 4:34prompt until you get to what you want.
151. 4:36But it turns out this role of being the
152. 4:38verifier can actually be offloaded to an
153. 4:41agent as well. And so this concept of
154. 4:43loops came about where if you take this
155. 4:45to the next level, you can actually have
156. 4:46an agent work for you and the agent also
157. 4:49does the verification. And so this agent
158. 4:51right here will assume the role of a
159. 4:53critic and you'll just have these two AI
160. 4:55agents talk to each other until it meets
161. 4:57a certain standard, a bar that you set.
162. 4:59And only then will this critic agent
163. 5:01actually pass to you the final output.
164. 5:03And by the way, this whole idea of
165. 5:04having a verifier agent in order to
166. 5:06increase quality output is not new at
167. 5:09all. In fact, this is an article by
168. 5:10Entropic called building effective
169. 5:12agents. And as a part of their study,
170. 5:13they're mentioning here that same
171. 5:15finding that they have where if you have
172. 5:17an AI model generate the output. They
173. 5:19actually find that you generally get
174. 5:20better outputs if you have another AI
175. 5:22model assume the role of an evaluator.
176. 5:25And this is probably not surprising
177. 5:26because if you think about how AI models
178. 5:29usually behave, they usually convince
179. 5:30themselves that the output that they
180. 5:32generate is already good enough. And so
181. 5:34it turns out that having another model
182. 5:35just validate that is actually good
183. 5:37practice. And mind you, this was an
184. 5:39article from way back in 2024. So the
185. 5:42concept of looping isn't really new. But
186. 5:44what is newer and what this gauntlet
187. 5:46loop has pretty much taken to the
188. 5:48extreme level is that in the build
189. 5:49method of that prompt is actually
190. 5:51instructing the main agent, the one that
191. 5:53you are talking to, to orchestrate and
192. 5:55fan out to a fleet of sub agents with
193. 5:58each of them having a critic partner in
194. 5:59order to just make sure that the parts
195. 6:01that they are creating are up to spec to
196. 6:03the standard that you set before the
197. 6:05final output comes to you. And this is
198. 6:07just a nice way to actually visualize
199. 6:09what's really happening under the hood.
200. 6:10But the great news about the tools that
201. 6:12we have now like claude code is that for
202. 6:14you to do something like this, you don't
203. 6:15actually need to learn any extra
204. 6:17technical tooling. All you need to do is
205. 6:19to have a well ststructured prompt like
206. 6:21this where you're instructing the main
207. 6:22agent to fan out sub agents to have each
208. 6:25of them tackle a task individually and
209. 6:27to have a separate sub agent check their
210. 6:29work in order to meet this bar that you
211. 6:32set. And so if you dissect this gauntlet
212. 6:33loop prompt, then I think that pattern
213. 6:35is the one that's most important to
214. 6:37learn here because there's really no
215. 6:38reason for you to not adopt the same
216. 6:40pattern across any of your builds. And
217. 6:42so obviously I needed to try out this
218. 6:44gauntlet loop prompt structure as well.
219. 6:46And I actually wanted to try it in use
220. 6:47cases beyond just games. And by the way,
221. 6:49if you want to learn how to build and
222. 6:51sell AI systems that businesses actually
223. 6:53pay for, then that's pretty much all we
224. 6:55do over at the Robbernuggets community,
225. 6:56where not only do you get access to the
226. 6:58Claude Living Master Class, which we
227. 7:00update every week and takes you from
228. 7:01zero to mastery with the latest on AI,
229. 7:03but you also get access to our agents as
230. 7:05a service course, which walks you
231. 7:07through how to actually get paid for all
232. 7:09these AI skills that you are learning.
233. 7:10You also get to be part of a genuinely
234. 7:12great community of AI builders. In fact,
235. 7:14you can see just some of the recent wins
236. 7:16our members are getting from the program
237. 7:18right here. So, if you want to start
238. 7:19earning from AI, then check that just in
239. 7:20the pin comment below. Now, back to the
240. 7:22video. And I think if this prompt is
241. 7:24really good at virtual 3D environments,
242. 7:25then just a few months down the road as
243. 7:27these models become even more capable,
244. 7:29then this will probably have a huge
245. 7:30impact on sectors like architecture or
246. 7:32even real estate. And so, the test that
247. 7:34I put out for Opus 5 here is that I gave
248. 7:36it this floor layout of a real real
249. 7:38estate listing at Darling Point here in
250. 7:40Sydney. And I also gave it some
251. 7:42reference photos to match against. So,
252. 7:44there's the living room, there's the
253. 7:45bedroom, and so on. And then for the
254. 7:47prompt itself, if you read through this,
255. 7:48you can notice that it is the same
256. 7:50structure as the gauntlet loop prompt
257. 7:51where we have a task here at the top.
258. 7:53We're saying here that we want Claude to
259. 7:54build an explorable 3D walkthrough of
260. 7:56this apartment. We're giving it the
261. 7:58build method where we want the goal to
262. 7:59be broken down into the smallest pieces
263. 8:01and to fan out sub agents. And we're
264. 8:03giving it that bar to hit where we won't
265. 8:05stop until each critic is utterly wowed.
266. 8:07So each sub aent will need to verify
267. 8:09that that bar has been met. And this
268. 8:11whole prompt, I didn't write it myself,
269. 8:13by the way. Near the end, I'll share
270. 8:14with you a skill so that whatever task
271. 8:15that you need, you can just instantly
272. 8:17build a gaunt to the loop prompt similar
273. 8:18to this. And when I send that prompt
274. 8:20over, you can see that it created a plan
275. 8:22here where it has these room builder sub
276. 8:24agent and their corresponding partners,
277. 8:26which are these blind critics. And at
278. 8:28least with a Claude desktop app, what's
279. 8:29great about it is that you can actually
280. 8:31view these dynamic workflows now as well
281. 8:33where you can clearly see the phases
282. 8:35that Claude has planned where right now
283. 8:37it's working on the lighting and then
284. 8:38the rooms and then there is a phase
285. 8:40where those sub agents will evaluate
286. 8:41those rooms and it will continue to loop
287. 8:43up until that original bar that we've
288. 8:45set has been satisfied. All right, so it
289. 8:47ran for around 2 hours now and it's
290. 8:49still working. But I think it's already
291. 8:51at a point where we can just showcase
292. 8:52the strength of this prompt because if
293. 8:54you can see here, this whole report,
294. 8:55this HTML page uh Claw just created for
295. 8:58us in order to give us updates of what
296. 9:00it's seeing versus its original peg. So
297. 9:02you can see this left one is the actual
298. 9:04photo that we gave it. And this one on
299. 9:06the right is the screenshots that it
300. 9:07took of that 3D world. And it's already
301. 9:09looking pretty close. Like the kitchen
302. 9:11counter here, this is the original and
303. 9:13this is the one that it created for us.
304. 9:15And even if it's already pretty close,
305. 9:17it's actually still not satisfied. So
306. 9:19you can see that this particular round,
307. 9:21it's still marking as failed and and
308. 9:23it's actually still iterating and
309. 9:24improving the look of this visual. And
310. 9:26so you can see that's where the
311. 9:27importance of setting a really high bar
312. 9:30is because if you actually want this to
313. 9:32be really perfect and you want to run
314. 9:33this for a couple of hours in order to
315. 9:35get a showcase build, then that's
316. 9:37something that you can just let Claude
317. 9:38do for you. But since I don't want to
318. 9:40sit around here waiting for a few hours
319. 9:42more just to complete this 3D app, let's
320. 9:44actually just view what it created for
321. 9:45us here. And there you go. You can see
322. 9:47we are in this living area. It even
323. 9:50captured the painting for us. Obviously,
324. 9:52the couches are not perfect yet, but I
325. 9:54think if we go around here, we can see
326. 9:56the kitchen counter. It has that marble
327. 9:58finish. And remember, this whole thing
328. 10:00was oneshotted by Claude using that
329. 10:01gauntlet loop prompt that we gave it.
330. 10:03And just to show a sample view. So, this
331. 10:06is the kitchen counter. And this was the
332. 10:08original peg that we gave to Claude. So,
333. 10:10it's pretty close, right? Then if we go
334. 10:12to the bedroom, obviously this texture
335. 10:14probably can be improved in later
336. 10:16passes, but I think it was able to
337. 10:18capture at least the look and the size
338. 10:20of the layout of the photo. And again,
339. 10:22just for reference, these are the images
340. 10:24that we fed to Opus 5. So that's pretty
341. 10:27close, at least in terms of the layout.
342. 10:29And then this is the other bedroom,
343. 10:31which for reference, this is the image
344. 10:33that we fed it. And probably if we gave
345. 10:35it a bit more time, it'll probably be
346. 10:36able to improve the textures of these
347. 10:38some more. But that's just a quick demo
348. 10:40of how you can use the gauntlet loop.
349. 10:42Now, apart from 3D worlds and 3D
350. 10:44environments, what I also did is to just
351. 10:46test out this gauntlet loop prompting
352. 10:47structure to create a front-end website
353. 10:49designed for this ketone IQ product. And
354. 10:52when we launched this workflow, you can
355. 10:54see it ran for around an hour and 19
356. 10:56minutes. And it's the same thing where
357. 10:57it fanned out several sub agents in
358. 10:59order to create our website and also
359. 11:02have this judging phase which is
360. 11:04essentially that evaluator agents to
361. 11:06check the worker agents builds. And what
362. 11:08it created for us is this. So let me
363. 11:10just shift that so you can see. So we
364. 11:12have the product here. We have brain
365. 11:13fuel as the headline. We have a dark
366. 11:15mode and a light mode. And if we scroll
367. 11:17down, we have these nice animations that
368. 11:19just provide you some more details about
369. 11:21this product. And I think what Opus did
370. 11:23here is it actually fanned out some
371. 11:25research agents in order to just make
372. 11:27sure that these numbers are correct.
373. 11:30Now, this is pretty good if you're just
374. 11:31looking at the visual flare of it
375. 11:33because obviously this is quite far
376. 11:34already from the normal AI vibecoded
377. 11:37designs that you may be used to or see.
378. 11:39However, even though this looks pretty
379. 11:41good, remember that visual flare is not
380. 11:43really the only thing that brands or
381. 11:45clients look for, especially when it
382. 11:47comes to these websites. Because if we
383. 11:48to look at Ketone IQ's actual website,
384. 11:51their brand design system is actually
385. 11:53quite different. So, I think the
386. 11:54gauntlet loop can still help you out
387. 11:56quite a lot, but if you don't start with
388. 11:58a really good minimum viable design or
389. 12:01product, then what the gauntlet loop
390. 12:03will do is just optimize towards
391. 12:05probably the wrong thing. And this is
392. 12:07really important to consider, especially
393. 12:08with powerful prompt structures like
394. 12:10these. Because if you notice those
395. 12:11gauntlet loop prompt that we ran, in
396. 12:13fact, any looping prompt that you run,
397. 12:15they usually take a lot of time and
398. 12:16tokens for them to finish. And so the
399. 12:18way that I would use them moving forward
400. 12:20personally is probably not to start with
401. 12:22them as your initial prompt. Because
402. 12:24what can happen there is even though the
403. 12:26final output that you would get looks
404. 12:28good, they might not actually be on
405. 12:29brief and might be really far from what
406. 12:31you want because you just let the agent
407. 12:33decide the direction for you. But if you
408. 12:34start with a really strong minimum
409. 12:36viable product or in this case a design
410. 12:38system which I've taught in previous
411. 12:40other tutorials in this channel and in
412. 12:42our community then you can just
413. 12:43introduce this gauntlet loop prompt as
414. 12:45sort of a warp drive in terms of just
415. 12:47sharpening or polishing the quality of
416. 12:49that MVP so that the version two of the
417. 12:51build that you're making not only looks
418. 12:53good but it is also sitting on a good
419. 12:55foundation and is on brief. And as
420. 12:57mentioned, if you want to try out this
421. 12:58gauntlet loop prompting for yourself,
422. 13:00what I've done is to build out this
423. 13:02skill called slashgauntlet loop. And
424. 13:04what this skill does is whenever you use
425. 13:06it, you can just give a particular task
426. 13:08and it will create a gauntlet loop
427. 13:10prompt for you. And you can just grab
428. 13:12this down in the description below. But
429. 13:13there, I hope that was useful and
430. 13:15informative. And as always, thanks for
431. 13:16watching until the end if you made it
432. 13:17this far. I'll see you next time.
433. 13:19Cheers. [music]