This NEW Claude Prompting Technique is blowing people's minds (gauntlet-loop) — Transcript
Full transcript
- 0:00There's a new prompting technique for
- 0:01Claude that's been blowing people's
- 0:02minds over the past week because in a
- 0:04single prompt it can build fully
- 0:06playable games in hyper custom 3D worlds
- 0:09like these that even Karpathy says might
- 0:11be the future of prompting LLMs. So
- 0:13today I'll share with you this technique
- 0:15called the gauntlet loop which might
- 0:17just be the quickest way for you to
- 0:18learn how to fan out sub-agents to do
- 0:20work for you. So that even if you're not
- 0:22into game development, you can add this
- 0:24tool to your arsenal and instantly get
- 0:26better at agentic AI. And by the end
- 0:28I'll share with you a skill that lets
- 0:29you fully take advantage [music] of this
- 0:30technique in the easiest way possible.
- 0:33And if you're new, my name is J. I spent
- 0:34over a decade working with brands you
- 0:36may know, have been in AI since my
- 0:38masters in data science. Now I'm running
- 0:39an AI business and one of the largest AI
- 0:41communities globally. Let's dive into
- 0:43it.
- 0:43>> [music]
- 0:45>> So I first saw this prompting technique
- 0:47from Matt Shumer who posted this insane
- 0:49demo over at X where it already garnered
- 0:51something like 4.8 million views. And he
- 0:53says here that Claude Opus 5 one-shotted
- 0:56this entire game with everything you see
- 0:58in the demo being custom code without
- 1:00any single external asset. And if Matt's
- 1:02name is familiar and if you're in the AI
- 1:03space for a while, that might be because
- 1:05he actually wrote this article called
- 1:07something big is happening which a lot
- 1:08of people read a few months ago now
- 1:10sitting at 87 million views. Point being
- 1:12that he has been working with AI for
- 1:14quite a while already and is actually a
- 1:15good source for on prompting techniques
- 1:17like these. And if you see a claim like
- 1:18this where an AI model supposedly
- 1:20one-shots a game that looks as good as
- 1:22this complete with sound by the way. I'm
- 1:24not sure if you can hear that if I just
- 1:25turn on the sound. Usually with this
- 1:27your first reaction would be a bit
- 1:29skeptical if it was even made by AI
- 1:31which is quite understandable because
- 1:33really the level of graphics here is
- 1:34already quite extraordinary. But a few
- 1:36days ago Matt actually shared this
- 1:38article where he went through how he
- 1:40created this game and he's calling it
- 1:42the gauntlet loop. And since then people
- 1:44have used that gauntlet loop prompting
- 1:45technique to recreate that same level of
- 1:48build quality. So to show a few
- 1:49examples, here's one where he recreated
- 1:51the starting area for Pokémon in perfect
- 1:533D. Here is an example for a car racing
- 1:56simulator game. And this is one where
- 1:57it's more of like a Mario Kart type of
- 1:59game. And this is just crazy how
- 2:01well-built this looks. Like you can see
- 2:03the different textures of this
- 2:04environment like with the road, the
- 2:05houses there, and there's just so much
- 2:07detail that the AI model was able to
- 2:09build out in this one game. And you
- 2:10might not be into game development in
- 2:12particular, and later on we'll show some
- 2:14use cases of how you can apply this
- 2:15outside of just video games. But
- 2:17personally, I still like to pay
- 2:18attention to these demos because it just
- 2:21points to how much raw capability these
- 2:23AI models now have. And Andrej Karpathy
- 2:26was able to probably articulate it
- 2:27better than I can where last weekend he
- 2:29made this post where he's saying that
- 2:31we're starting to leave the territory
- 2:33where you would test an LLM by creating
- 2:35an SVG of a pelican on a bicycle, which
- 2:37is this old test that AI models were
- 2:39used to be run on. And he mentions here
- 2:41that these kinds of examples are great
- 2:42because no one in their right mind would
- 2:44ever spend the time to write something
- 2:46this custom, but LLMs and AI models have
- 2:48all the stamina and patience in the
- 2:50world. So, these hyper-custom worlds and
- 2:523D environments are a really great
- 2:54example of new capabilities that you
- 2:56yourself as an AI user are now able to
- 2:58tap into that you couldn't really do
- 3:00before. So, what is the gauntlet loop
- 3:02exactly? Well, thankfully Matt also
- 3:04shared his exact prompt here. And
- 3:06surprisingly, it is quite simple. It is
- 3:08only a three-line prompt. And so, you
- 3:10can see I just pasted that whole prompt
- 3:12in here. And what's actually more
- 3:13interesting here versus the actual
- 3:15verbiage of this prompt is just the
- 3:17pattern and structure of it. Because if
- 3:19you really break this down into these
- 3:21three lines, essentially what you have
- 3:22is a prompt structure that you can copy
- 3:24yourself. Where first you give it a task
- 3:26of what you want to happen. In this
- 3:27case, the task that Matt was gunning for
- 3:29is to build a first-person shooter game.
- 3:32And then the second part here is
- 3:33essentially the build method. So, how
- 3:34that agent is going to achieve that
- 3:36task. Where he's asking the main agent
- 3:38to fan out sub-agents and have each of
- 3:41those tackle each task individually. And
- 3:43to have a separate sub-agent check it
- 3:45visually to ensure that it looks really,
- 3:47really good. And then finally, the third
- 3:48part to this is the the to hit, which is
- 3:50essentially the standard where the agent
- 3:52can decide when it can stop. And so he's
- 3:55saying here to not stop until each
- 3:56sub-agent is utterly wild with the
- 3:58quality when compared with the actual
- 4:00Call of Duty game. And what actually
- 4:02makes this gauntlet loop so effective
- 4:04are these two parts right here. Because
- 4:06if you haven't tried using sub-agents to
- 4:08orchestrate your work before, then this
- 4:10might just be one of the easiest and
- 4:11quickest way for you to try it out. But
- 4:13just to step back in case you don't know
- 4:14what we're referring to when we talk
- 4:16about sub-agent orchestration,
- 4:17essentially when you prompt an agent or
- 4:20talk to an AI model, there's three
- 4:21levels to it, at least in how I think
- 4:23about it. The first level, which is the
- 4:25most basic and probably the most common,
- 4:27is when you do work with an agent, you
- 4:28send a prompt, it provides an output
- 4:30back to you, you verify if that output
- 4:32already matches your standard, and then
- 4:33you send another prompt until you get to
- 4:35what you want. But it turns out this
- 4:37role of being the verifier can actually
- 4:39be offloaded to an agent as well. And so
- 4:41this concept of loops came about where
- 4:44if you take this to the next level, you
- 4:46can actually have an agent work for you,
- 4:48and the agent also does the
- 4:50verification. And so this agent right
- 4:51here will assume the role of a critic,
- 4:53and you'll just have these two AI agents
- 4:55talk to each other until it meets a
- 4:57certain standard, a bar that you set,
- 4:59and only then will this critic agent
- 5:01actually pass to you the final output.
- 5:03And by the way, this whole idea of
- 5:04having a verifier agent in order to
- 5:06increase quality output is not new at
- 5:08all. In fact, this is an article by
- 5:10Entropic called Building Effective
- 5:12Agents, and as a part of their study,
- 5:13they're mentioning here that same
- 5:15finding that they have where if you have
- 5:17an AI model generate the output, they
- 5:19actually find you generally get better
- 5:21outputs if you have another AI model
- 5:23assume the role of an evaluator. And
- 5:25this is probably not surprising because
- 5:27if you think about how AI models usually
- 5:29behave, they usually convince themselves
- 5:31that the output that they generate is
- 5:33already good enough. And so it turns out
- 5:34that having another model just validate
- 5:36that is actually good practice. And mind
- 5:38you, this was an article from way back
- 5:40in 2024. So the concept of looping isn't
- 5:43really new. But what is newer and what
- 5:45this gauntlet loop has pretty much taken
- 5:47to the extreme level is that in the
- 5:49build method of that prompt is actually
- 5:51instructing the main agent, the one that
- 5:53you're talking to, to orchestrate and
- 5:55fan out to a fleet of sub agents with
- 5:57each of them having a critic partner in
- 5:59order to just make sure that the parts
- 6:01that they are creating are up to spec to
- 6:03the standard that you set before the
- 6:05final output comes to you. And this is
- 6:07just a nice way to actually visualize
- 6:09what's really happening under the hood.
- 6:10But the great thing is about the tools
- 6:11that we have now like Claude code is
- 6:13that for you to do something like this,
- 6:15you don't actually need to learn any
- 6:17extra technical tooling. All you need to
- 6:18do is have a well-structured prompt like
- 6:21this where you're instructing the main
- 6:22agent to fan out sub agents, to have
- 6:24each of them tackle a task individually,
- 6:27and to have a separate sub agent check
- 6:28their work in order to meet this bar
- 6:31that you set. And so if you dissect this
- 6:33gauntlet loop prompt, then I think that
- 6:35pattern is the one that's most important
- 6:37to learn here because there's really no
- 6:38reason for you to not adopt the same
- 6:40pattern across any of your builds. And
- 6:42so obviously I needed to try out this
- 6:44gauntlet loop prompt structure as well.
- 6:46And I actually wanted to try it in use
- 6:47cases beyond just games. And by the way,
- 6:49if you want to learn how to build and
- 6:50sell AI systems that businesses actually
- 6:53pay for, then that's pretty much all we
- 6:54do over at the Robo Nuggets community.
- 6:56Where not only do you get access to the
- 6:58Claude living master class, which we
- 7:00update every week and takes you from
- 7:01zero to mastery with the latest on AI,
- 7:03but you also get access to our agents as
- 7:05a service course which walks you through
- 7:07how to actually get paid for all these
- 7:09AI skills that you are learning. You
- 7:10also get to be part of a genuinely great
- 7:12community of AI builders. In fact, you
- 7:14can see just some of the recent wins our
- 7:16members are getting from the program
- 7:18right here. So if you want to start
- 7:19earning from AI, then check that just in
- 7:20the pin comment below. Now back to the
- 7:22video. And I think if this prompt is
- 7:23really good at virtual 3D environments,
- 7:25then just a few months down the road as
- 7:27these models become even more capable,
- 7:29then this will probably have a huge
- 7:30impact on sectors like architecture or
- 7:32even real estate. And so the test that I
- 7:34put out for Opus 5 here is that I gave
- 7:36it this floor layout of a real real
- 7:38estate listing at Darling Point here in
- 7:40Sydney. And I also gave it some
- 7:41reference photos to match against. So,
- 7:44there's a living room, there's a
- 7:45bedroom, and so on. And then for the
- 7:47prompt itself, if you read through this,
- 7:48you can notice that it is the same
- 7:49structure as the gauntlet loop prompt,
- 7:51where we have a task here at the top.
- 7:53We're saying here that we want Claude to
- 7:54build an explorable 3D walk-through of
- 7:56this apartment. We're giving it the
- 7:57build method, where we want the goal to
- 7:59be broken down into the smallest pieces
- 8:01and to fan out sub-agents. And we're
- 8:03giving it that bar to hit, where we
- 8:05won't stop until each critic is utterly
- 8:07wowed. So, each sub-agent will need to
- 8:09verify that that bar has been met. And
- 8:11this whole prompt I didn't write it
- 8:12myself, by the way. Near the end, I'll
- 8:13share with you a skill so that whatever
- 8:15task that you need, you can just
- 8:16instantly build a gauntlet loop prompt
- 8:18similar to this. And when I send that
- 8:20prompt over, you can see that it created
- 8:21a plan here, where it has these room
- 8:23builders sub-agent and their
- 8:25corresponding partners, which are these
- 8:27blind critics. And at least with the
- 8:28Claude desktop app, what's great about
- 8:30it is that you can actually view these
- 8:31dynamic workflows now as well, where you
- 8:33can clearly see the phases that Claude
- 8:35has planned. Where right now, it's
- 8:37working on the lighting, and then the
- 8:38rooms, and then there is a phase where
- 8:40those sub-agents will evaluate those
- 8:41rooms, and it will continue to loop up
- 8:43until that original bar that we've set
- 8:45has been satisfied. All right, so it ran
- 8:47for around 2 hours now, and it's still
- 8:49working, but I think it's already at a
- 8:51point where we can just showcase the
- 8:52strength of this prompt. Because if you
- 8:54can see here, this whole report, this
- 8:55HTML page uh Claude just created for us
- 8:58in order to give us updates of what it's
- 9:00seeing versus its original peg. So, you
- 9:02can see this left one is the actual
- 9:04photo that we gave it, and this one on
- 9:06the right is the screenshots that it
- 9:07took of that 3D world. And it's already
- 9:09looking pretty close. Like the kitchen
- 9:11counter here, this is the original, and
- 9:13this is the one that it created for us.
- 9:15And even if it's already pretty close,
- 9:17it's actually still not satisfied. So,
- 9:18you can see that this particular round,
- 9:21it's still marking as failed. And and
- 9:22it's actually still iterating and
- 9:24improving the look of this visual. And
- 9:26so, you can see that's where the
- 9:27importance of setting a really high bar
- 9:29is. Because if you actually want this to
- 9:31be really perfect, and you want to run
- 9:33this for a couple of hours in order to
- 9:35get a showcase build, then that's
- 9:37something that you can just let Claude
- 9:38do for you. But since I don't want to
- 9:40sit around here waiting for a few hours
- 9:42more just to complete this 3D app, let's
- 9:44actually just view what it created for
- 9:45us here. And there you go, you can see
- 9:47we are in this living area. It even
- 9:50captured the painting for us. Obviously,
- 9:52the couches are not perfect yet, but I
- 9:54think if we go around here, we can see
- 9:56the kitchen counter. It has that marble
- 9:58finish. And remember this whole thing
- 10:00was one-shotted by Claude using that
- 10:01gauntlet loop prompt that we gave it.
- 10:03And just to show a sample view, so this
- 10:06is the kitchen counter and this was the
- 10:08original pic that we gave to Claude. So
- 10:10it's pretty close, right? Then if you go
- 10:12to the bedroom, obviously this texture
- 10:14probably can be improved in later
- 10:16passes, but I think it was able to
- 10:18capture at least the look and the size
- 10:20of the layout of the photo. And again,
- 10:22just for reference, these are the images
- 10:24that we fed to Opus 5. So that's pretty
- 10:27close at least in terms of the layout.
- 10:29And then this is the other bedroom,
- 10:31which for reference, this is the image
- 10:33that we fed it. And probably if we gave
- 10:35it a bit more time, it'll probably be
- 10:36able to improve the textures of these
- 10:38some more. But that's just a quick demo
- 10:40of how you can use the gauntlet loop.
- 10:42Now apart from 3D worlds and 3D
- 10:44environments, what I also did is to just
- 10:45test out this gauntlet loop prompting
- 10:47structure to create a front-end website
- 10:49design for this KetoneIQ product. And
- 10:52when we launch this workflow, you can
- 10:53see it ran for around an hour and 19
- 10:56minutes. And it's the same thing where
- 10:57it fanned out several sub-agents in
- 10:59order to create our website and also
- 11:02have this judging phase, which is
- 11:04essentially that evaluator agents to
- 11:06check the worker agents' builds. And
- 11:08what it created for us is this. So let
- 11:10me just shift that so you can see. So we
- 11:12have the product here, we have Brain
- 11:13Fuel as the headline, we have a dark
- 11:15mode and a light mode. And if we scroll
- 11:17down, we have these nice animations that
- 11:19just provide you some more details about
- 11:21this product. And I think what Opus did
- 11:23here is it actually fanned out some
- 11:25research agents in order to just make
- 11:27sure that these numbers are correct. Now
- 11:30this is pretty good if you're just
- 11:31looking at the visual flare of it
- 11:33because obviously this is quite far
- 11:34already from the normal AI vibe coded
- 11:37designs that you may be used to or see.
- 11:39However, even though this looks pretty
- 11:41good, remember that a visual flare is
- 11:43not really the only thing that brands or
- 11:45clients look for especially when it
- 11:46comes to these websites. Because if
- 11:48we're to look at Keto and IQ's actual
- 11:50website, their brand design system is
- 11:53actually quite different. So, I think
- 11:54the gauntlet loop can still help you out
- 11:56quite a lot, but if you don't start with
- 11:58a really good minimum viable design or
- 12:01product, then what the gauntlet loop
- 12:03will do is just optimize towards
- 12:05probably the wrong thing. And this is
- 12:07really important to consider especially
- 12:08with powerful prompt structures like
- 12:10these because if you notice those
- 12:11gauntlet loop prompt that we ran, in
- 12:13fact, any looping prompt that you run,
- 12:15they usually take a lot of time and
- 12:16tokens for them to finish. And so, the
- 12:18way that I would use them moving forward
- 12:20personally is probably not to start with
- 12:22them as your initial prompt because what
- 12:24can happen there is even though the
- 12:26final output that you would get looks
- 12:28good, they might not actually be on
- 12:29brief and might be really far from what
- 12:31you want because you just let the agent
- 12:33decide the direction for you. But if you
- 12:34start with a really strong minimum
- 12:36viable product or in this case a design
- 12:38system, which I've taught in previous
- 12:40other tutorials in the channel and in
- 12:41our community, then you can just
- 12:43introduce this gauntlet loop prompt as
- 12:45sort of a war drive in terms of just
- 12:47sharpening or polishing the quality of
- 12:49that MVP so that the version two of the
- 12:51build that you're making not only looks
- 12:53good, but it is also sitting on a good
- 12:55foundation and is on brief. And as
- 12:57mentioned, if you want to try out this
- 12:58gauntlet loop prompting for yourself,
- 13:00what I've done is a build out this skill
- 13:02called {slash} gauntlet loop. And what
- 13:04this skill does is whenever you use it,
- 13:06you can just give a particular task and
- 13:08it will create a gauntlet loop prompt
- 13:11for you. And you can just grab this down
- 13:12in the description below. But there, I
- 13:14hope that was useful and informative. As
- 13:15always, thanks for watching until the
- 13:16end if you made it this far. I'll see
- 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.