Transcription
Most people don't realize that there's a pattern hiding in their AI conversations. Every time you perfect an image after multiple attempts, nail an AI video after tweaking it around six times, or finally get that vibecoded app working just right after 30 iterations.
You might not know this, but you've created something very valuable that most of you are throwing away while everyone else opens a brand new chat in whatever platform it is and starts from scratch every single time. You can be busy building an entire database of super prompts that you can use for everything. Images, videos, voice agents, and AI coded apps. The principle and the way you apply this is exactly the same. So, in this video, I'm going to walk you through how to extract the DNA of the perfect outputs that you finally get and use them to make sure that prompt engineering is never an obstacle ever again. Let's dive in.
So, just in case you don't know what this technique is, it's called reverse prompting. And many times you'll find it called reverse metarrompting. And what metarrompting means is that you get the AI to write the prompt for you. But instead of writing it from the beginning, you write it in reverse. So to give some more color here, just for an example to start us out, we have write me a brief to make an AI app that helps people getting started with AI create a learning path for themselves. So in this case, Claude comes up with an output and maybe I don't like it and I say something like this where I say make it a bit more detailed and vary the sentence structure. Some long sentences, some short sentences, some bullets weaved in, mix it up overall, and then we get this output that has a mix of different syntaxes of text and is a lot more detailed.
Instead of me having to start a brand new chat from scratch and remind Claude every single time, I can just say something like, "Act engineer and write a prompt and output it in a code block." That would have saved me the effort of getting to your last output, the last thing I just did right now in one shot without me having to provide that feedback to begin with. And the result is something that you can easily copy paste that has everything I've given it. And you can see right here, critical style requirements, vary sentence structure dramatically, mix short and punchy sentences, medium descriptive ones, and longer complex sentences. And that's one small glimpse of how you can use this strategy. Like I said at the start though, you can apply this to all kinds of modalities like image, video, and what have you. So let's dive straight into those examples.
All right, so for the very first example involving image, I used Chatbt. In this specific example, I say go and research how to design beautiful infographics that look like a worldclass designer put them together. And this is sometimes what I like to do in general, Chatbt or not Chatbt, to prime the conversation to help it come up with a solid image prompt internally so that when I ask it for an infographic, there's a higher chance that it will output something actually proper. So in this case, it comes up with the research and then I go in right for the request right after. So with that in mind, design an infographic that uses these best practices to explain very elegantly how to learn AI if you're a business owner with no time. Now by priming it, we initially get something that looks pretty decent where it says how to learn AI. The spelling is good. The infographics pretty minimalistic. It's not too AI slop. But we have a problem here. We have AI essential that cuts off, right? So in this case I say it cuts off and then it outputs it again with the cutff and then I say still cutting off at the very bottom. Let's remove the part entirely and then finally it understands that because basically there's a length limitation of the actual image itself.
So, at the very bottom here, what we could do is, hey, so how could I have prevented me going back and forth with you on this image and made sure that it would fit perfectly the first time so I wouldn't have to repeat myself over and over again? Can you output a prompt, an image prompt based on what you know about yourself that would have helped me skip that step that's very concise, to the point, and would have made sure this never didn't happen? So, I'm doing this very off the cuff to show you how powerful of a technique it could be where now it's going to output the nuance. And this is an example of the optimize prompt. So I won't read it in full, but you can see it says design a complete vertical infographic called this. So this is one thing I didn't have at the beginning, which is I didn't say vertical. It I didn't actually specify exactly what I wanted. I didn't specify the colors. In this case, it did specify the colors. And now if I repeated this process, would it be 100% foolproof? No. But maybe if I set it to GPT5 thinking and I pasted this in at the very least, maybe I wouldn't have anything that cuts off. Now whether or not everything else would be perfect is another story. At the end of the day, we're still dealing with language models and images and videos are still very finicky. But having gone back and forth, starting your next chat would be that much more fruitful. And imagine you went back and forth in these kind of chats and stored these optimized prompts in what have you, a notes file, a Google sheet, and you were able to now go to your repository of prompts and start your next conversation on a much more stronger foundation.
And in this next example, let's try out video. So, I happened to be sitting in a bubble tea place when I was coming up with this prompt. So, it's very diabolical. Please excuse me, but it says, "A young adult sitting comfortably in a modern cafe, happily sipping a vibrant dill pickle bubble tea with neon pink tapioca pearls. The camera focuses on the evervescent bubbles and glowing pearls swirling inside the drink." And then we go through a little bit more style here. Okay, so we get a V1 which is actually pretty impressive. You won't be able to hear the audio. If I just press this right here, looks dickly for sure, but I wanted even more dill pickle. So I asked it create one with an actual dill pickle sticking out of the drink itself. And you get something like that. And then I said make the bubbles neon pink and slant the pickle so it doesn't look straight. So in this case, we get one final version of this where it's very clear. There's neon in the background and in the cup. And I really like this as an actual output.
So same technique, different application. And I say, I love the last video. What kind of prompt could I have written to get the same result from the start? It then thinks and comes up with this prompt where it says, "A young adult sits comfortably in a modern cafe sipping a vibrant dill pickle bubble tea. The drink is filled with glowing neon pink pearls, etc., etc. Then it says here, "A single dill pickle slants jaunty. I don't even know what that word means. I'm not even that good at English." Uh, close-up shot. This is probably going to do a much better job than the three back and forths I had that actually tapped out my credits for the day. And one more thing you'll notice, especially when you go from one modality to the other. And when I say modality, by the way, it just means from image to video to music to text, and you'll be able to start learning brand new vocabulary that will help you become a natural, better prompt engineer. And that word might not be jauntly in this case, but it could be the way it describes the setting, the scene, the tone, in this case, the vibrant hues. Again, words that we might know internally, but we don't know to actually surface it or when to surface it.
In the next example, we'll show you an application of how to use reverse metaparprompting in a slightly different way. We have a feedback loop. So, this voice agent is something we're putting together for my early AI adopters community. And basically what it will do is is on the about page, you'll be able to call a number and ask questions about what we offer, right? So, I gave it a brain dump prompt of exactly every single thing that's in there. And then when I tested it, I noticed some very alarming mistakes it was making, especially in terms of promises or lack of promises. So if I go into something like the call logs right here, and I pull the last chat right here, and then we export, let's say, as a JSON, the entire transcript, you'll see right here, this call log. If I open it up, there'll be a lot of metadata here that we don't really care for, but something like a Claude or a Chatbt, we'll be able to look through. If I just scroll through just a little bit and you'll see right here it says, "Hey, how can I help?" So, if I take this conversation and I go into our Amigo Claude, let's take uh this one right here. And let's take the original prompt for this actual agent. So, let's go into here and zoom in. Let's go into the early adopters one. Copy it. Paste it. There we go. Oh, looks like we're context limits issue. Let's reroute this to Chatbt. So in Chatbt it works just fine.
So what I'll do is the following. So I noticed going back and forth with the voice agent and actually go and read the transcript from my voice assistant. It made a few mistakes. It said that any of the private coaching for niten or make.com is an additional fee for my community, but it's actually included in the membership. How can we edit the prompt to make sure it never makes a mistake like that? And you'll notice that in the conversation, it forgets that the person said they're a beginner or starting from zero to one and it keeps asking that same question. Take the conversation, read through how cringe-worthy it was in terms of the actual flow of the conversation or lack thereof, understand the way that the person speaks, the user to the actual voice agent, and come up with an amended version of this prompt in a markdown code block so I can paste it back in and test it again. So in this case now we are using the power of AI to audit the AI. So now we have AI and in this case we're just giving it an artifact. And a lot of people get stuck on how to edit things let's say like a voice agent prompt but this is one of the most effective things to do. And at some point around last year my agency was going really heavy into voice assistance cuz it was brand new. We had a system where it would take all the logs of the conversation and then it would update based on a score whether or not the prompt should be updated. If yes, it would go via API and update the system prompt as it was getting new information. So in a way there was some form of feedback loop based on the conversations. You can only accomplish that using a version of reverse metaprompting. And you'll see right here, it acknowledges that it was incorrect by assessing the actual transcript and it logged in things like an additional paid service as well as a few other things that are hardcoded in the prompt. And then it comes up with a V2 of the prompt that we can just copy paste and continue this feedback loop.
Like I said though, you could have an automation to do the heavy lifting of this reverse metaprompting process to do it a lot more at scale. But just to show you the general concept because again the same thing could apply to a chatbot with a rag knowledge base where you take the knowledge base and you can compare it against a conversation or a series of conversations and say what can we do to make the system prompt better based on all the flaws in these interactions.
And for my next magic trick I'll show you how you can apply this in the world of vibe coding. So, this is an app I actually spent a while on in Replit that basically helps you create a prompt through a prompt wizard. And then when you get to this section, simulation, it will actually simulate a back and forth against an AI user that you can configure. So, you can basically theoretically create a chatbot assistant for, I don't know, real estate or insurance, and then test out the prompt that you use right here or you ask AI to really create for you. And then you simulate how the conversation would flow. And then it goes back and forth until it's good enough based on your instructions and what you're looking for. And then you get your prompt that you can use. So because this was harder to build, what you can do is in most of these vibe coding apps, you can go to the bottom and say ask mode. And then you could say based on all the back and forths, all the failures we had, all the bugs we had, what kind of prompt actually could you write the prompt and take the persona of a prompt engineer that would have helped me skipped a lot of steps here and avoid all the unnecessary mistakes in our architecture and the app itself. Can you create a prompt that I could use in one shot to output this whole application? So if I send that over in plan mode, it'll take just a bit of time and it will output after looking at all the code in the codebase, looking through the conversation in context and give us something that we can store to make our next attempt at building something like this a lot better. And you'll see here it comes up with a very very detailed prompt along with even telling it the database schema basically how to design the tables in the back end.
One thing to be wary of this, more of an expert tip, is if you do reverse metaprompting with a vibe coded app, sometimes language models hallucinate a lot if you dictate exactly how they should build the thing. So sometimes I will actually say, can you not advise from a very um code technical standpoint? Like don't give me code in the prompt. Just tell me conceptually what the agent should have done or what it should have done better or what specific stack without telling me too much. So in this case it says React 18 in TypeScript which is fine. But if you go to the very bottom prompt here like core features this is the kind of thing you want it to output. So prompt management simulation engine analysis and scoring user personas improvement tracking etc. And the rest of this are details that aren't actually not bad. But this is a way you can now learn from all the back and forths because sometimes, especially with vibe coded apps, even though they're getting better, you could spend a week finally getting exactly where you're looking for after tons of iterations. So, if you can store tens of these prompts over time, you will by osmosis become a better prompt engineer.
Last but not least, I wanted to show you how you can apply this to a lot more heavily AI coded apps. So, this is an application I put together using cloud code and it's super helpful. I've basically used it internally to replace things like Tableau or PowerBI where all I have to do just to show you a demo example is upload some data. So let me just upload a CSV file here. So let's do I think it's this one. Yes. And then automatically takes that CSV, breaks it apart into different parts and then creates analytics and then a full table here that I can keep going until the end of the data set. And I can do all kinds of things like units sold, sort by revenue, etc. And you can see all the summary level statistics here. And but wait, there's more. You can filter by region, by sales channel, by date. And I made it to be like a complete substitute to all of these programs that you'd have to otherwise learn.
So once again, same technique, same application, different modality. In this case, what we can do is I can go into Claude and I can go on to shift tab tab tab until we get to plan mode. So it doesn't actually do anything. And I say, can you go through this entire codebase here and come up with a very comprehensive prompt? Try to avoid dictating code in the prompt, but more so dictate infrastructure, how to approach, basically like a neural pathway of how an AI agent could have built this in entirely one shot. And I want you to create a MD file, but do nothing else. And basically after you create that MD file, I want it to be written in a way that I could copy it and paste it in a brand new session and there would be a very high chance that we could build it in one shot. And then bonus beneath that prompt, come up with examples of agents or sub agents I could have used in cloud code along with associated prompts for those agents that could have helped me also make this a lot more efficient by working in parallel. So I put it in plan mode, but I'm just realizing that if I wanted to make a file, it should be in non-plan mode. So I'm going to go on yolo mode and pray that it doesn't actually execute and just makes this file. Let's do one more comment here. Again, don't do anything except for create this MD file. Don't build, don't write code, don't do anything, just in case. And let's see what it comes up with.
All right. And it came back after 5 minutes with a mega prompt here with a project blueprint, core infrastructure requirements, the project structure. So, this is what I love about Cloud Code, the Flask API. So, I won't get technical there, but let's just keep scrolling to the very bottom. I want to see there we go. So, I came up with an idea for a back-end infrastructure agent, the actual purpose, and a prompt for exactly what that agent's job would be, which is awesome. Huge timesaver. And agents is something I explore a lot with cloud code. Once you go through back and forth with agents, you can also create sub conversations with those agents and learn how you could have made the instructions for that agent that much better. But anyway, you can see agent two, agent three, and now you got all this and this is something that you now own. You could just download this to your computer or keep it in that folder and you have this to be able to recycle and reuse a lot of the concepts and more importantly the learnings to become that much better the next time.
So, this is one of the most underrated best prompting techniques you'll ever see. It'll help you learn, it'll help you build, and most importantly, it'll help you have much higher leverage. If this helped you, then let me know down in the comments below. It helps the video, helps the algo, helps the channel. And if you want even more prompting sauce from the prompting hefe himself, then check out the first link in the description below. So much prompting sauce, it'll make you tear up. I'll see you the next one.