YouTube2Text

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

by Jay E | RoboNuggets · 3,135 words · 434 segments · language en · Watch on YouTube

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. 0:04single prompt it can build fully
  5. 0:06playable games in hyper custom 3D worlds
  6. 0:09like these that even Karpathy says might
  7. 0:11be the future of prompting LLMs. So
  8. 0:13today I'll share with you this technique
  9. 0:15called the gauntlet loop which might
  10. 0:17just be the quickest way for you to
  11. 0:18learn how to fan out sub-agents to do
  12. 0:20work for you. So that even if you're not
  13. 0:22into game development, you can add this
  14. 0:24tool to your arsenal and instantly get
  15. 0:26better at agentic AI. And by the end
  16. 0:28I'll share with you a skill that lets
  17. 0:29you fully take advantage [music] of this
  18. 0:30technique in the easiest way possible.
  19. 0:33And if you're new, my name is J. I spent
  20. 0:34over a decade working with brands you
  21. 0:36may know, have been in AI since my
  22. 0:38masters in data science. Now I'm running
  23. 0:39an AI business and one of the largest AI
  24. 0:41communities globally. Let's dive into
  25. 0:43it.
  26. 0:43>> [music]
  27. 0:45>> So I first saw this prompting technique
  28. 0:47from Matt Shumer who posted this insane
  29. 0:49demo over at X where it already garnered
  30. 0:51something like 4.8 million views. And he
  31. 0:53says here that Claude Opus 5 one-shotted
  32. 0:56this entire game with everything you see
  33. 0:58in the demo being custom code without
  34. 1:00any single external asset. And if Matt's
  35. 1:02name is familiar and if you're in the AI
  36. 1:03space for a while, that might be because
  37. 1:05he actually wrote this article called
  38. 1:07something big is happening which a lot
  39. 1:08of people read a few months ago now
  40. 1:10sitting at 87 million views. Point being
  41. 1:12that he has been working with AI for
  42. 1:14quite a while already and is actually a
  43. 1:15good source for on prompting techniques
  44. 1:17like these. And if you see a claim like
  45. 1:18this where an AI model supposedly
  46. 1:20one-shots a game that looks as good as
  47. 1:22this complete with sound by the way. I'm
  48. 1:24not sure if you can hear that if I just
  49. 1:25turn on the sound. Usually with this
  50. 1:27your first reaction would be a bit
  51. 1:29skeptical if it was even made by AI
  52. 1:31which is quite understandable because
  53. 1:33really the level of graphics here is
  54. 1:34already quite extraordinary. But a few
  55. 1:36days ago Matt actually shared this
  56. 1:38article where he went through how he
  57. 1:40created this game and he's calling it
  58. 1:42the gauntlet loop. And since then people
  59. 1:44have used that gauntlet loop prompting
  60. 1:45technique to recreate that same level of
  61. 1:48build quality. So to show a few
  62. 1:49examples, here's one where he recreated
  63. 1:51the starting area for Pokémon in perfect
  64. 1:533D. Here is an example for a car racing
  65. 1:56simulator game. And this is one where
  66. 1:57it's more of like a Mario Kart type of
  67. 1:59game. And this is just crazy how
  68. 2:01well-built this looks. Like you can see
  69. 2:03the different textures of this
  70. 2:04environment like with the road, the
  71. 2:05houses there, and there's just so much
  72. 2:07detail that the AI model was able to
  73. 2:09build out in this one game. And you
  74. 2:10might not be into game development in
  75. 2:12particular, and later on we'll show some
  76. 2:14use cases of how you can apply this
  77. 2:15outside of just video games. But
  78. 2:17personally, I still like to pay
  79. 2:18attention to these demos because it just
  80. 2:21points to how much raw capability these
  81. 2:23AI models now have. And Andrej Karpathy
  82. 2:26was able to probably articulate it
  83. 2:27better than I can where last weekend he
  84. 2:29made this post where he's saying that
  85. 2:31we're starting to leave the territory
  86. 2:33where you would test an LLM by creating
  87. 2:35an SVG of a pelican on a bicycle, which
  88. 2:37is this old test that AI models were
  89. 2:39used to be run on. And he mentions here
  90. 2:41that these kinds of examples are great
  91. 2:42because no one in their right mind would
  92. 2:44ever spend the time to write something
  93. 2:46this custom, but LLMs and AI models have
  94. 2:48all the stamina and patience in the
  95. 2:50world. So, these hyper-custom worlds and
  96. 2:523D environments are a really great
  97. 2:54example of new capabilities that you
  98. 2:56yourself as an AI user are now able to
  99. 2:58tap into that you couldn't really do
  100. 3:00before. So, what is the gauntlet loop
  101. 3:02exactly? Well, thankfully Matt also
  102. 3:04shared his exact prompt here. And
  103. 3:06surprisingly, it is quite simple. It is
  104. 3:08only a three-line prompt. And so, you
  105. 3:10can see I just pasted that whole prompt
  106. 3:12in here. And what's actually more
  107. 3:13interesting here versus the actual
  108. 3:15verbiage of this prompt is just the
  109. 3:17pattern and structure of it. Because if
  110. 3:19you really break this down into these
  111. 3:21three lines, essentially what you have
  112. 3:22is a prompt structure that you can copy
  113. 3:24yourself. Where first you give it a task
  114. 3:26of what you want to happen. In this
  115. 3:27case, the task that Matt was gunning for
  116. 3:29is to build a first-person shooter game.
  117. 3:32And then the second part here is
  118. 3:33essentially the build method. So, how
  119. 3:34that agent is going to achieve that
  120. 3:36task. Where he's asking the main agent
  121. 3:38to fan out sub-agents and have each of
  122. 3:41those tackle each task individually. And
  123. 3:43to have a separate sub-agent check it
  124. 3:45visually to ensure that it looks really,
  125. 3:47really good. And then finally, the third
  126. 3:48part to this is the the to hit, which is
  127. 3:50essentially the standard where the agent
  128. 3:52can decide when it can stop. And so he's
  129. 3:55saying here to not stop until each
  130. 3:56sub-agent is utterly wild with the
  131. 3:58quality when compared with the actual
  132. 4:00Call of Duty game. And what actually
  133. 4:02makes this gauntlet loop so effective
  134. 4:04are these two parts right here. Because
  135. 4:06if you haven't tried using sub-agents to
  136. 4:08orchestrate your work before, then this
  137. 4:10might just be one of the easiest and
  138. 4:11quickest way for you to try it out. But
  139. 4:13just to step back in case you don't know
  140. 4:14what we're referring to when we talk
  141. 4:16about sub-agent orchestration,
  142. 4:17essentially when you prompt an agent or
  143. 4:20talk to an AI model, there's three
  144. 4:21levels to it, at least in how I think
  145. 4:23about it. The first level, which is the
  146. 4:25most basic and probably the most common,
  147. 4:27is when you do work with an agent, you
  148. 4:28send a prompt, it provides an output
  149. 4:30back to you, you verify if that output
  150. 4:32already matches your standard, and then
  151. 4:33you send another prompt until you get to
  152. 4:35what you want. But it turns out this
  153. 4:37role of being the verifier can actually
  154. 4:39be offloaded to an agent as well. And so
  155. 4:41this concept of loops came about where
  156. 4:44if you take this to the next level, you
  157. 4:46can actually have an agent work for you,
  158. 4:48and the agent also does the
  159. 4:50verification. And so this agent right
  160. 4:51here will assume the role of a critic,
  161. 4:53and you'll just have these two AI agents
  162. 4:55talk to each other until it meets a
  163. 4:57certain standard, a bar that you set,
  164. 4:59and only then will this critic agent
  165. 5:01actually pass to you the final output.
  166. 5:03And by the way, this whole idea of
  167. 5:04having a verifier agent in order to
  168. 5:06increase quality output is not new at
  169. 5:08all. In fact, this is an article by
  170. 5:10Entropic called Building Effective
  171. 5:12Agents, and as a part of their study,
  172. 5:13they're mentioning here that same
  173. 5:15finding that they have where if you have
  174. 5:17an AI model generate the output, they
  175. 5:19actually find you generally get better
  176. 5:21outputs if you have another AI model
  177. 5:23assume the role of an evaluator. And
  178. 5:25this is probably not surprising because
  179. 5:27if you think about how AI models usually
  180. 5:29behave, they usually convince themselves
  181. 5:31that the output that they generate is
  182. 5:33already good enough. And so it turns out
  183. 5:34that having another model just validate
  184. 5:36that is actually good practice. And mind
  185. 5:38you, this was an article from way back
  186. 5:40in 2024. So the concept of looping isn't
  187. 5:43really new. But what is newer and what
  188. 5:45this gauntlet loop has pretty much taken
  189. 5:47to the extreme level is that in the
  190. 5:49build method of that prompt is actually
  191. 5:51instructing the main agent, the one that
  192. 5:53you're talking to, to orchestrate and
  193. 5:55fan out to a fleet of sub agents with
  194. 5:57each of them having a critic partner in
  195. 5:59order to just make sure that the parts
  196. 6:01that they are creating are up to spec to
  197. 6:03the standard that you set before the
  198. 6:05final output comes to you. And this is
  199. 6:07just a nice way to actually visualize
  200. 6:09what's really happening under the hood.
  201. 6:10But the great thing is about the tools
  202. 6:11that we have now like Claude code is
  203. 6:13that for you to do something like this,
  204. 6:15you don't actually need to learn any
  205. 6:17extra technical tooling. All you need to
  206. 6:18do is have a well-structured prompt like
  207. 6:21this where you're instructing the main
  208. 6:22agent to fan out sub agents, to have
  209. 6:24each of them tackle a task individually,
  210. 6:27and to have a separate sub agent check
  211. 6:28their work in order to meet this bar
  212. 6:31that you set. And so if you dissect this
  213. 6:33gauntlet loop prompt, then I think that
  214. 6:35pattern is the one that's most important
  215. 6:37to learn here because there's really no
  216. 6:38reason for you to not adopt the same
  217. 6:40pattern across any of your builds. And
  218. 6:42so obviously I needed to try out this
  219. 6:44gauntlet loop prompt structure as well.
  220. 6:46And I actually wanted to try it in use
  221. 6:47cases beyond just games. And by the way,
  222. 6:49if you want to learn how to build and
  223. 6:50sell AI systems that businesses actually
  224. 6:53pay for, then that's pretty much all we
  225. 6:54do over at the Robo Nuggets community.
  226. 6:56Where not only do you get access to the
  227. 6:58Claude living master class, which we
  228. 7:00update every week and takes you from
  229. 7:01zero to mastery with the latest on AI,
  230. 7:03but you also get access to our agents as
  231. 7:05a service course which walks you through
  232. 7:07how to actually get paid for all these
  233. 7:09AI skills that you are learning. You
  234. 7:10also get to be part of a genuinely great
  235. 7:12community of AI builders. In fact, you
  236. 7:14can see just some of the recent wins our
  237. 7:16members are getting from the program
  238. 7:18right here. So if you want to start
  239. 7:19earning from AI, then check that just in
  240. 7:20the pin comment below. Now back to the
  241. 7:22video. And I think if this prompt is
  242. 7:23really good at virtual 3D environments,
  243. 7:25then just a few months down the road as
  244. 7:27these models become even more capable,
  245. 7:29then this will probably have a huge
  246. 7:30impact on sectors like architecture or
  247. 7:32even real estate. And so the test that I
  248. 7:34put out for Opus 5 here is that I gave
  249. 7:36it this floor layout of a real real
  250. 7:38estate listing at Darling Point here in
  251. 7:40Sydney. And I also gave it some
  252. 7:41reference photos to match against. So,
  253. 7:44there's a living room, there's a
  254. 7:45bedroom, and so on. And then for the
  255. 7:47prompt itself, if you read through this,
  256. 7:48you can notice that it is the same
  257. 7:49structure as the gauntlet loop prompt,
  258. 7:51where we have a task here at the top.
  259. 7:53We're saying here that we want Claude to
  260. 7:54build an explorable 3D walk-through of
  261. 7:56this apartment. We're giving it the
  262. 7:57build method, where we want the goal to
  263. 7:59be broken down into the smallest pieces
  264. 8:01and to fan out sub-agents. And we're
  265. 8:03giving it that bar to hit, where we
  266. 8:05won't stop until each critic is utterly
  267. 8:07wowed. So, each sub-agent will need to
  268. 8:09verify that that bar has been met. And
  269. 8:11this whole prompt I didn't write it
  270. 8:12myself, by the way. Near the end, I'll
  271. 8:13share with you a skill so that whatever
  272. 8:15task that you need, you can just
  273. 8:16instantly build a gauntlet loop prompt
  274. 8:18similar to this. And when I send that
  275. 8:20prompt over, you can see that it created
  276. 8:21a plan here, where it has these room
  277. 8:23builders sub-agent and their
  278. 8:25corresponding partners, which are these
  279. 8:27blind critics. And at least with the
  280. 8:28Claude desktop app, what's great about
  281. 8:30it is that you can actually view these
  282. 8:31dynamic workflows now as well, where you
  283. 8:33can clearly see the phases that Claude
  284. 8:35has planned. Where right now, it's
  285. 8:37working on the lighting, and then the
  286. 8:38rooms, and then there is a phase where
  287. 8:40those sub-agents will evaluate those
  288. 8:41rooms, and it will continue to loop up
  289. 8:43until that original bar that we've set
  290. 8:45has been satisfied. All right, so it ran
  291. 8:47for around 2 hours now, and it's still
  292. 8:49working, but I think it's already at a
  293. 8:51point where we can just showcase the
  294. 8:52strength of this prompt. Because if you
  295. 8:54can see here, this whole report, this
  296. 8:55HTML page uh Claude just created for us
  297. 8:58in order to give us updates of what it's
  298. 9:00seeing versus its original peg. So, you
  299. 9:02can see this left one is the actual
  300. 9:04photo that we gave it, and this one on
  301. 9:06the right is the screenshots that it
  302. 9:07took of that 3D world. And it's already
  303. 9:09looking pretty close. Like the kitchen
  304. 9:11counter here, this is the original, and
  305. 9:13this is the one that it created for us.
  306. 9:15And even if it's already pretty close,
  307. 9:17it's actually still not satisfied. So,
  308. 9:18you can see that this particular round,
  309. 9:21it's still marking as failed. And and
  310. 9:22it's actually still iterating and
  311. 9:24improving the look of this visual. And
  312. 9:26so, you can see that's where the
  313. 9:27importance of setting a really high bar
  314. 9:29is. Because if you actually want this to
  315. 9:31be really perfect, and you want to run
  316. 9:33this for a couple of hours in order to
  317. 9:35get a showcase build, then that's
  318. 9:37something that you can just let Claude
  319. 9:38do for you. But since I don't want to
  320. 9:40sit around here waiting for a few hours
  321. 9:42more just to complete this 3D app, let's
  322. 9:44actually just view what it created for
  323. 9:45us here. And there you go, you can see
  324. 9:47we are in this living area. It even
  325. 9:50captured the painting for us. Obviously,
  326. 9:52the couches are not perfect yet, but I
  327. 9:54think if we go around here, we can see
  328. 9:56the kitchen counter. It has that marble
  329. 9:58finish. And remember this whole thing
  330. 10:00was one-shotted by Claude using that
  331. 10:01gauntlet loop prompt that we gave it.
  332. 10:03And just to show a sample view, so this
  333. 10:06is the kitchen counter and this was the
  334. 10:08original pic that we gave to Claude. So
  335. 10:10it's pretty close, right? Then if you go
  336. 10:12to the bedroom, obviously this texture
  337. 10:14probably can be improved in later
  338. 10:16passes, but I think it was able to
  339. 10:18capture at least the look and the size
  340. 10:20of the layout of the photo. And again,
  341. 10:22just for reference, these are the images
  342. 10:24that we fed to Opus 5. So that's pretty
  343. 10:27close at least in terms of the layout.
  344. 10:29And then this is the other bedroom,
  345. 10:31which for reference, this is the image
  346. 10:33that we fed it. And probably if we gave
  347. 10:35it a bit more time, it'll probably be
  348. 10:36able to improve the textures of these
  349. 10:38some more. But that's just a quick demo
  350. 10:40of how you can use the gauntlet loop.
  351. 10:42Now apart from 3D worlds and 3D
  352. 10:44environments, what I also did is to just
  353. 10:45test out this gauntlet loop prompting
  354. 10:47structure to create a front-end website
  355. 10:49design for this KetoneIQ product. And
  356. 10:52when we launch this workflow, you can
  357. 10:53see it ran for around an hour and 19
  358. 10:56minutes. And it's the same thing where
  359. 10:57it fanned out several sub-agents in
  360. 10:59order to create our website and also
  361. 11:02have this judging phase, which is
  362. 11:04essentially that evaluator agents to
  363. 11:06check the worker agents' builds. And
  364. 11:08what it created for us is this. So let
  365. 11:10me just shift that so you can see. So we
  366. 11:12have the product here, we have Brain
  367. 11:13Fuel as the headline, we have a dark
  368. 11:15mode and a light mode. And if we scroll
  369. 11:17down, we have these nice animations that
  370. 11:19just provide you some more details about
  371. 11:21this product. And I think what Opus did
  372. 11:23here is it actually fanned out some
  373. 11:25research agents in order to just make
  374. 11:27sure that these numbers are correct. Now
  375. 11:30this is pretty good if you're just
  376. 11:31looking at the visual flare of it
  377. 11:33because obviously this is quite far
  378. 11:34already from the normal AI vibe coded
  379. 11:37designs that you may be used to or see.
  380. 11:39However, even though this looks pretty
  381. 11:41good, remember that a visual flare is
  382. 11:43not really the only thing that brands or
  383. 11:45clients look for especially when it
  384. 11:46comes to these websites. Because if
  385. 11:48we're to look at Keto and IQ's actual
  386. 11:50website, their brand design system is
  387. 11:53actually quite different. So, I think
  388. 11:54the gauntlet loop can still help you out
  389. 11:56quite a lot, but if you don't start with
  390. 11:58a really good minimum viable design or
  391. 12:01product, then what the gauntlet loop
  392. 12:03will do is just optimize towards
  393. 12:05probably the wrong thing. And this is
  394. 12:07really important to consider especially
  395. 12:08with powerful prompt structures like
  396. 12:10these because if you notice those
  397. 12:11gauntlet loop prompt that we ran, in
  398. 12:13fact, any looping prompt that you run,
  399. 12:15they usually take a lot of time and
  400. 12:16tokens for them to finish. And so, the
  401. 12:18way that I would use them moving forward
  402. 12:20personally is probably not to start with
  403. 12:22them as your initial prompt because what
  404. 12:24can happen there is even though the
  405. 12:26final output that you would get looks
  406. 12:28good, they might not actually be on
  407. 12:29brief and might be really far from what
  408. 12:31you want because you just let the agent
  409. 12:33decide the direction for you. But if you
  410. 12:34start with a really strong minimum
  411. 12:36viable product or in this case a design
  412. 12:38system, which I've taught in previous
  413. 12:40other tutorials in the channel and in
  414. 12:41our community, then you can just
  415. 12:43introduce this gauntlet loop prompt as
  416. 12:45sort of a war drive in terms of just
  417. 12:47sharpening or polishing the quality of
  418. 12:49that MVP so that the version two of the
  419. 12:51build that you're making not only looks
  420. 12:53good, but it is also sitting on a good
  421. 12:55foundation and is on brief. And as
  422. 12:57mentioned, if you want to try out this
  423. 12:58gauntlet loop prompting for yourself,
  424. 13:00what I've done is a build out this skill
  425. 13:02called {slash} gauntlet loop. And what
  426. 13:04this skill does is whenever you use it,
  427. 13:06you can just give a particular task and
  428. 13:08it will create a gauntlet loop prompt
  429. 13:11for you. And you can just grab this down
  430. 13:12in the description below. But there, I
  431. 13:14hope that was useful and informative. As
  432. 13:15always, thanks for watching until the
  433. 13:16end if you made it this far. I'll see
  434. 13:18you next time. Cheers. [music]

About this transcript

This page contains the full transcript of This NEW Claude Prompting Technique is blowing people's minds (gauntlet-loop) by Jay E | RoboNuggets, generated from the public captions YouTube serves with the video. The transcript has 3,135 words across 434 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.