This Is What Actually Works in 2026!

Back in 2023, I published one of the first prompt engineering courses here on YouTube. More than 1 million views and 40,000 likes. But it's 2026 now and a lot has changed. While preparing this video, I watched the latest prompting videos from top creators on YouTube. They are great, but I noticed there are five advanced tactics no one talked about and these can completely change your results. In the next few minutes, I break down each with a practical example and I'll give you my free playground so you can understand easily, test, and visualize. If you are ready, let's get started. Now, before we dive into our first advanced tactic, let's do a quick one minute refresher on prompt engineering in general. Regardless of the debates around the term itself, the idea is super simple. The better you prompt AI, the better results you get. And every good prompt has four pillars: roll, task, context, and rules. Here's a quick example. the role, who should the AI be, the task, what you want the AI to do, the context, the background info it needs, the rules, the guidelines, and the examples they I should follow when generating the output. Simple, but most people skip one or two, and this is why they don't get exactly what they want. That's enough. Let's now get into our advanced tactics. Number one is what we call recursive brain storm. The idea is simple. The implementation is a bit tricky but I'll make it simple too. So here I am include you can use charge, Gemini, Deepseek, whatever you prefer. I will start with this simple example, a very basic prompt so you can grasp the idea easily. What are the top five business ideas that will last in the AI era in the next 10 years? Return as a bullet list. So here I'm just prompting Clo to give me a bullet list of ideas. Let's test this. It says, "Hey Hassan and Ali because my child sometimes uses clothes." So here we have these five ideas. Now here where it gets interesting. We grab this first idea and we ask code now again brainstorm five child ideas under AI education upskilling. So we had the first layer with five ideas. We grab the first one and then we brainstorm five child ideas under this current topic. So now we have five sub ideas under the first one. We can repeat the same concept with all these five burnt ideas. But now I want to go more deep and ask it to brainstorm under this child one. Now brainstorm ideas under you see what's happening here. We are going recursively brainstorming going deep deep to discover ideas that you never thought about. And this can work on any topic on anything you want to brainstorm ideas around. Now the problem is if you want to do this manually suppose we have 10% ideas for each one you want to brainstorm and then go deep and deep and let's say you want to go 10 levels deep. Honestly it will be boring and will take a lot of time. So this is why we need a way to automate this prompting technique. So what I did is I implemented it inside simple LM and a few lines of code you can automate this process. But maybe you are not into coding. So what I did is I built this free playground. So you can test and prompt learn and visualize these techniques easily. So here in tools I will click on brainstorm and simply here select the provider and let's go with 4.0. It's faster now. And simply enter the prompt you want to brainstorm ideas around. Let's go with the same prompt. What are the top business ideas? Here in the parameters, you can select the max depth. How many layers you can you want to go down? Let's go now with two. And ideas per level. Let's go with three just to understand the concept now. And by the way, it will give you also the code example. So you can copy and test directly. Let's run now start brainstorm. And you will see now in the canvas here in this section the tree the brainstorm tree will start to appear. So this is the first level we have three ideas. Now it's going into the next layer which is level one and brainstorm three ideas for each parent topic. This is the first depth and in few seconds you'll get the second ideas and then the last one and the brainstorm completed. You can zoom in, you can drag these nodes as you like and more importantly you can click on any of these nodes and you can read the details here the data inscription and the score of the ideas. So each of these ideas will have detailed section and a score. Let's do this with three as max depth. Run again. And perfect. In 127 seconds, we get 21 ideas with with average score of 8.2. And here in the idea tree, you can see the full ideas. And by the way, you can click export results and export into JSON and you will see in detail how the brainstorm worked. Maybe you want to use this maybe in your application, another prompt and so on. Okay, let's now move on to tactic number two, the web JSON technique. The idea is simple. It's all about prompting AI to search the web and then return in a structured JSON format. Let me show you a direct example. So this is a prompt and the key terms here is search the web and return JSON in this format. So you see now after the research we get this JSON response. Now you may be wondering okay but what's the use case here? How is this helpful? Let me show you an example I build and I'm sure will shock you. If you copy this JSON here and I will take it into my carousel generator tool and if I run now the carousel generator this simple command here in a few seconds I will get this output. It converted each of the JSON records into a carousel slide. You see and this is how I can generate carousels easily. So it is the engine the webjson technique is the engine of micousal generator and we have many other use cases like building dashboards with live data creating comparison tables for different products and much more we have infinite possibility if you understand this technique it will be a game changer especially if you are building products with AI and by the way I'll be sharing the car generator fully open source for free in my next videos. So don't forget to subscribe and notifications so you don't miss any new update. And of course you can test this web JSON technique in the playground I built for you. We see here web search plus JSON. So I will paste the prompt here. Select the provider the model and now we can construct how you want the output to be or you can simply paste the code in Python. So I'll paste the structure I want. So, we have a list of stories. Generate. Wait a couple of seconds. Perfect. And now we have this structured out. You see the JSON here. Perfect. Let's now move on to tactic number three, the feedback loop. This one is interesting. The idea again is really super simple. So you prompt the language model but instead of accepting the first answer you go through a loop of feedback and improvement until we get better result. Let's see a practical example. Let's open a new chat and I will go with this simple basic prompt. Generate a sentence to explain quantum computing the best way for an 8-year-old kid. Imagine a magic code that can be both head and behaves at the same time while it's spinning. And quantum computers use tiny things like that magic coin to solve really hard puzzles much faster than regular computers can. Perfect. So this is the first answer. Now with the feedback loop, we get this answer and we tell the model like clo in our case or maybe other model to analyze and improve. So we have this feedback improvement loop. I will simulate this now. I will open a new chat. I will paste this result here and I will tell it here is a sentence explaining quantum computing to an 8-year-old. Let's add double quotes and then I will add a simple prompt like this one. Critique this explanation. Is it too complex? Are there simpler words we could use then write an improved version. This is the first iteration trying to improve the first answer with the same model. As I mentioned, we can grab this to JPT. So we can ask other model to do this and so on. Maybe Gemini. So let's run this. It gives me the critique what could be simpler and then it gives me the improved version and so on. You can go with three, four, five, 100 iterations maybe. But it will maybe overoptimize then. Now let's open our playground. And here we have the LM feedback. So I will paste the prompt and then you can even add initial answer to be evaluated based on but now we will leave it for the AI. We can select the architecture single dual or multi. So when you switch between those you see here we can select different providers in dual mode. You can select two providers one for generation one for critique. Let's go with single. So the same provider will critique itself and go into this feedback loop. I will go here with open AAI GPT40 and you can set the max iterations the quality threshold and so on and the evaluation criteria like clarity creativity and so on and then start improvement let's wait a little bit and you can see here the iterations with the answer with the critique for each iteration and then it outputs the final answer here you can see the scores and it stopped in three because the score didn't change. This is what I mentioned about over optimization. So the playground will handle this automatically. If the score stays the same, it will output and stop the feedback loop. Let's now move to tactic number four, the LM judge. And here we are moving from feedback to having a judge. So the idea instead of having the language model, the AI improving itself with iterations, with feedback and improvement, we generate multiple responses from different models, then grab one as a judge to evaluate them all. And here we have three judge modes. Number one, select best. So the judge will select the best answer from all the models. Number two, synthesize which combines all the elements from different models into one improved answer and the compare. It will return detail analysis breaking down the weakness and the strength of each provider. So in short, you open a code, you paste the prompt and get a result. You go to charge PT, you paste and get a result. Then you copy those results into the judge. for example here Gemini and ask it to generate the best or synthesize or compare. I will skip now the manual implementation to save some time. Let's go to the playground and see this in action. So we have the judge. You see the judge mode select best synthesize or compare. And you can add here the providers you want to compare like open AI GPT4 and open AAI for example GPT4 mini. And let's ask the same question on quantum computing in one sentence. And the judge LM will be for example OpenAI GPD 5 and then start evaluation. As I mentioned before, you can always grab this code snippet and try it in code if you want. Let's wait a little bit and in a few seconds you'll get the results from both models with the score with the winner and you can click to see the full breakdown the evaluation strengths weakness and the scores and you can see here the final answer from the judge. This is really super powerful when you can't afford just using the first answer from AI. I use this especially in some NLP and language tasks and things that require intensive reasoning from the language models and when building data sets. For example, in my case, I need to get results from different models then judge each response and see the weaknesses and this will help you spot a lot of details and points that you can improve your results based on. Let's now move on to tactic number five. LM retrieval. This one is a bit tricky. I'll explain directly with examples. So here I am in simple LM website. If you go in documentation, you see we have the full documentation for this library. So we have the parent sections like getting started, core features, providers. These are the parent sections. In each we have child sections, introduction, quick start. Here we have reliable LM structured output and so on. So you can see we have like a tree of structured content and in each of these sections that you will open we have a lot of content that can be split into multiple chunks too. So visualize with me we have like a tree of content that resembles this documentation. Now what I want is to use prompts and AI to navigate through this tree and grab and retrieve accurate data. And here we use something called semantic routing. So we will ask the AI based on the user query select the best parent nodes then select the best child nodes from the parents you selected and so on navigating through this structure to get the best answer based on the user query. Now a practical example I developed based on this technique is onto digest. If you open it now and grab any video, you will see we have summaries, quiz, flashcards, study notes, a lot of content around this video. The topic, the idea here is we have this feature to chat with this video. So instead of passing the full video transcript and chatting with it and this may increase hallucinations, I transform the video text or content into a structured format where the language model can navigate. If I ask it for example in three words summarize the video. What will happen now is using semantic routing it will decide to pick the full video or a specific chunk. In our case we will need the full transcript to summarize the video. But if we ask it summarize the 12 month plan the speaker outlined in five words. In this case, instead of grabbing the full transcript AI using semantic routing will go and grab the section in the video that is talking about the 12 month plan and get back the answer. This is very important to save on tokens to get accurate results especially in chat boats. I hope you get the idea with semantic routing. It has a lot of real use cases that transforms the way you prompt AI and you build applications with AI. Going back to our playground here, we have the LM retrieval and I created this demo. So you can build the index. Let's go with 40. build an index to test with like AI news articles or you can paste your own text here because with LM retrieval we don't use embeddings like in rag if you are familiar with rag we have vectors and embeddings and vector databases no we have just pure text and the language model will navigate through this structured index based on semantic meanings we got this simple chunks of data now we go here and query let's say what about AI safety for example search and a few seconds now the language model without using any databases or vectors and so on will get the best answer from these chunks and you can see this the chunk here so we have the cluster and AI with semantic routing will pick the best answers we are not done yet I still have four prompting tips I want to share with you quickly in like 1 minute number One, start with a criticism. Especially if you are brainstorming ideas because AI by default will make you feel like you're Einstein. It was a great idea and you go with it. We don't want this. We want AI to actually challenge your ideas and thinking and give you honest feedback. So always start with something like before you respond critically analyze my idea. Point out weaknesses, gaps, and potential failures first. Be brutally honest. I need real feedback, not encouragement. And this will transform totally the way AI will respond. The second tip, the creative mix formula. If you want truly unique ideas, you can mix completely two different topics together and ask AI to find intersection. This way, you force AI to make connections no one else is making. For example, combine the business model of Netflix with the problem of learning a new language. Give me five unique product ideas that merge these two worlds in unexpected ways. So, we mix different concepts together and brainstorm really interesting ideas. Try this. I'm sure you will love it and you read things you never thought about. Number three, context is everything or context is the king. Especially today when coding with AI, the idea is simple. For example, let's say you want to build the LM judge we talked about with simple LM. Instead of telling AI build an LLM judge application for me or whatever, what we do is we inject the documentation or the code snippets in the context. So AI can follow the same pattern. For example, here you can go here to my application, go to judge and go to the code example, copy the code and you can say here docs to use and paste the code. So we inject the way we want the AI to write the code so it knows exactly how this works and what you want exactly. This is what I call documentation injection when building and coding with AI. Tip number four is learn by building. Instead of asking AI to explain something, ask it to teach you by building a project. For example, include here we have also the learning mode. You can select it and you can prompt it like this. I want to learn how rest APIs work. You can change this topic whatever you want. Don't just explain the theory. Teach me by building a simple API project together. So you will go step by step learning by building and this is really a game changer and this by the way how I developed and built my solo builder course. So you can go from zero to building your dream project by building while learning. If you interested you can check the course in the description below. Now what about the playground? How to get it and use it? Simply the description below you'll find a link to download this application. It's totally for free. You can run it even without installations as a portable application. And when you open it, you'll go to settings and enter the API keys of providers you want to use. You can even use local models open source with for example here in this basic chat playground. You can select here and select the model you have it in after you install it and then you can chat with local models or use it in any tool you want. For example, if you go here to brainstorm, you can select here and select the model and use local models for testing. We have also this tool I didn't mention where you can compare models real time and chat with two models at the same time. For example, here you can select OpenAI GPT4 and OpenAI GPT4 or mini and ask it the same question. high and then you get answers from both models. You can compare them real time in the same chat window. If you learned something new today, if you enjoyed this video, don't forget to support my simple library with simple star on GitHub. And more importantly, smash the like button so the video can reach more people and benefit more people. Thank you for following and see you next week.