MASTER the Prompt: TOP 5 Elements for Reusable Prompts, AI Agents, Agentic Workflows

have you ever wondered which parts of your prompts are providing the most value and which parts of your prompt you can delete there's an ocean of prompt engineering tips and tricks out there but do you know which ones actually matter in this video I want to share five essential elements of the prompt to help you get 80% of the results with 20% of the effort I don't know if this is cool or embarrassing to say but I've written thousands and maybe tens of thousands of prompts and I think I've cracked The Prompt equation on effort in versus value out so what are these five important items that make up high quality prompts that are very unlikely to change as the llm ecosystem evolves the five most important elements of the prompt are model purpose variables examples and the output if you're writing prompts and you want to get the most bang for your buck some of this will be recapped but I promise this Elemental breakdown will help simplify and speed up your work work with your prompts this video is all about the 8020 of incredible prompts let's get started so let's look at five examples of these prompts showcasing each element through a prompt and then we'll look at two full concrete prompts A N View component and the Omni complete we've covered on the channel That utilize these five elements of great prompt so let's get started first things first let's say the obvious dumb and most important thing the model you use dictates the performance more than any other one of these elements if you've run a prompt on gbt 40 and then you run that exact same prompt on a smaller lower class model llama 7B 53 you know that the quality of the output goes down drastically there's nothing more important when building your prompts than to pick a great model the model you choose will have the biggest impact on the performance of your prompt hands down full stop let's go ahead and move on the purpose is the entire reason you're writing the prompt it's your goal it's the target it's the task if you write a great clear goal for your prompt you will start to get those clear simple results after selecting the llm the purpose of your prompt becomes the next most important piece here we have a super simple prompt with an explicit purpose of generating blog posts about llms simple enough right this is the purpose now things get interesting let's talk about variables this prompt has three of the five elements right we have the model we have the purpose and we have a variable so not all variables are the same there are two types of variables static variables and dynamic variables we'll discuss static variables in a moment but this here is a dynamic variable Dynamic variables are variables that you're going to want to update when you're reusing your prompts in an ad hoc way so for instance here we can replace this with llms and run this and we'll get another result out we can update it again we'll say AI agents rerun and you know whatever typescript versus JavaScript right we can go on and on here and reuse this prompt the key value that variables offer you is that they allow you to reuse your prompts and you can use code to update your prompts in production environments the big kicker here is if you're building AI agents or agentic workflows you know that you need variables in order to chain together the outputs of prompts into the inputs of your next prompts we'll talk about that more when we get to Output let's go ahead and move on to our next element examples so the fourth most important part of your prompt is to include concrete example let's go ahead and walk through this and clarify what everything is so we have the purpose at the top we have our Dynamic variable here and then at the bottom here we have um a bunch of examples right so you can see here this is a prompt that helps us generate new python functions and let's just go ahead and generate something we'll say place string start end and I want to pass in a string and then we'll have start match and match and that's going to give back a string right so I'm doing a little bit of AI code prompting there by writing out the uh function definition right so go ahead and kick that off while we discuss this right so examples are really really important you can see here that with examples we're able to specify how exactly we want our output to look in the response of our prompt examples tie really closely to Output which we'll look at in a second but but you know if we go ahead and remove our examples and we'll remove the reference to our examples and we rerun this you'll see we get a very different output right even with the best model if you don't tell the model exactly what you're looking for it can't give it to you right so in this case it's giving us example usage it's giving us an entire explanation and you know this is great but when we're writing prompts that go right into the model in a non chat conversation interface we want to get our exact output structure and the best way to get that is to use examples to tell the model exactly what you're looking for right so if we throw those back in there we rerun this you can see once again we get that exact result examples are really important because we can come in here and update this right we can add a a function common and just specify exactly what we want right so detailed about function right we can come in here copy this make sure all the examples have the exact same format and this off colorize that and then we can rerun our prompt right and of course we're going to get exactly what we're looking for so we're going to get this function now with a really clean comment um and it actually improved the code a little bit because we forced it to explain itself a little bit above so really interesting stuff that's the power of examples and like I mentioned examples tie really really closely to Output let's talk about the output so let me go ahead and just label this we have the purpose we have a dynamic variable here and then we have an example output here right we have the example structure prompt output typically comes in two forms you can output text or you can output Json like we are here right so let me just go ahead and kick that off and you can see here we specified the exact structure and our awesome gp40 model created the output in the exact form we were looking for why is this important text output is default it's straightforward it's useful for simpler tasks but when you're starting to chain together your outputs and you want more reliable consistent outputs from your prompts Json is the way to go Json outputs enable structured and reliable outputs which is crucial for building the higher order systems like prompt chains AI agents and agentic workflows also if it's not clear function calling is essentially Json output with a couple additional wrappers on top when you make a request with a function tool you get back the function name and you get back the parameters you can fully replace function calling with Json outputs because well that's exactly what it is so these are the five most important elements of the prompt so we can come in here and update the topic let's go ahead and say you know we had that typescript uh versus uh JavaScript let's go ahe and just rerun that and you can see here it's going to generate a new blog post with all this information and it's got the SEO keywords in there we have the hook with all these five elements with variables with some example structure here by specifying the Json output right by using every one of these elements we can create concise high output prompts that are very very reusable this is why variables are number three right it's really really important to be able to swap in and out values of your prompt then it's also important to have examples to specify the exact format and then last but not least element five we have output right it's really important to have the option to use a raw text output versus Json output right which will give you really really structured outputs that you can use to chain together your prompts and we've talked about this on the channel before but if it's not super clear this is how your prompts start working together to build prompt chains your outputs of your previous prompts become the inputs to the variables of your new prompt and if you've been with the channel for a bit if you remember our agent OS video we looked at this very exact example you can see here that we can create composability between agent one and agent two by chaining together these agents right by taking the outputs from agent one inside the lpu and then using that as the input for agent two right so everything comes full circle here this is what we're building up to being able to control your prompt and build valuable prompts quickly with these five elements can help you get to that composability step faster and it helps you have a really dialed in way to create this composability between your prompts right it all starts with The Prompt which is why I created this video which is why I want to share this with you these are the five most important elements if you use these five elements you will get 80% of the value of writing a really really high quality prompt with 20% of the effort you select a model you clearly Define the purpose make sure it has a single purpose then you can set up variables to be reused in your prompt in an ADH way or with code your variables also serve as placeholders for your previous prompts to insert new values we then have examples which allow us to build out specifications for what we want the prompt output to look like and then we have the specific output type right so this is our text versus Json it ties really really closely to examples because the examples help guide both your text and your Json outputs so these are the five most important elements this is how we create composability and chain together our prompts let's look at a couple concrete examples so we talked about the Omni complete on the channel before a link that video basically it's a prompt that takes you know somewhere between two and 500 lines of code and put puts it all into a single prompt so what I'll do here is just go ahead and label these sections so it's super clear at the top we have our purpose we then have a static variable here and then we have Dynamic variables down here so we talked about Dynamic variables before these are variables that will update in an ad hoc way like we're doing right now or via code and then we have static variables static variables are simple they're basically things that you'll update as you're perfecting your prompt as you're improving your prompt right you'll just come in here and you'll add additional rules they don't really change outside of the development phase right so these are static variables right once you set them once your prompt is good you just kind of leave them the dynamic variables will update over time these are things that will be populated by your rag system these are things that would be populated by um you know again you coming in here ad hoc or by you know replacing it in production systems based on whatever user is logged in or whatever information is going to swap these variables out right let's just go ahead and walk through a simple example here the topic is going to be typescript I'm going to leave off previous completions and domain knowledge since we can fall back on the llm here and then I'm just going to set something up to autocomplete right so I'll just say typescript is and then we'll ask for an autoc completion there and then you can see our system is giving us a bunch of different Auto completes that we could use in our hypothetical drop down field or our intellisense let's do a couple more R interfaces and and we're looking for and types of course interchangeable so you can see here we're getting some really nice clean autocompletes let's go ah and do one more um let's change the topic this is why this is an omni complete right we come in here we just change the topic we'll F this to SEO content writing and then we'll say how can I and just you know look for auto completions here so how can I improve my website search engine ranking so imagine you have like an FAQ input field now your users can come in here use the obni complete prompt and you know just by replacing two variables you have an entire system here that can help you drive business results here via autocompletion that's a full example with the OM complete you can see we have a clear defined purpose at the top we have static variables here and then we have Dynamic variables down here we are using the Json output and of course we're using the top-of-the-line gbg 40 model for the best results and that gives us a clear concise consistent prompt so let's look at N View JS components as a final example let's go ahead and label these so we have our purpose at the top we have guidelines you can see I'm using this kind of guideline pattern again I find this to be a super useful kind of meta tip we then have our actual Dynamic variable that will replace as we reuse this prompt the component request and then at the bottom here we have a explicit example of what we want our VJs components to look like I'll come in here I'll make these we'll do great so you can see I'm just defining exactly what I want the component to look like and that's that so now all we need to do is make a request let's say we want to create a Circle progress bar and we'll pass in of course progress and I'll make that an INT let's do like gradient right so we'll have two color gradients I'll say blue red so I'm doing a little bit of AI coding prompting here by just specifying the component structure that I want to see completed and so we'll fire this prompt off and the output is going to be a vuejs component in the structure with the request completed so really cool stuff this prompt again is completely reusable thanks to the structure of The Prompt we have the clear guidelines we have the purpose we have an example of what we want the output to be in this case we're using just clean text output since we don't need this as Json and then we are using a dynamic variable via the component request and then we're clearly defining exactly what we're looking for here right and that's how we get this prompt generating consistent high value outputs okay so to recap these are the five essential elements of high quality prompts we have the model purpose variables examples and output by focusing on these elements you can achieve 80% of the results with just 20% of the effort so drop a comment down below if you agree with my top five list if you have another element that you think is higher priority than you know the variable or the examples or the output or whatever I'm curious definitely let me know what you think like I mentioned I've written thousands of prompts and these are the things that keep coming up that give me the most value and the least amount of time so I wanted to hop on here and share this with you when thinking about development in the age of AI and the age of generative AI I think it's always important to focus on the underlying technology that's giving everything its value and when it comes to the prompt the model and then you know these next five elements really serve as the groundwork for the prompt and all of its higher level compositions right so once you master The Prompt once you understand the prompt you can then build great prompt chains great AI agents all the way up right but it all starts here with the prompt on the base level whenever I'm creating new prompts now this is how I like to structure it right first I choose my model then I write a clean clear single purpose for the prompt I then Define my variables both static and dynamic I'll then add examples to kind of you know specify the exact output I'm looking for and then finally I'll determine if text output is okay or if I want some really explicit Json output if you're building chainable prompts and AI agents and agentic workflows you'll likely want Json output for the consistent and the ease of parsing right so that's it for this one drop a comment drop a like if you enjoyed this video definitely hit the sub if you're not subb we are on the journey to building a gentic software software that works for us while we sleep keep prompting keep building and I'll see you in the next one

Channel: IndyDevDan

Views: 4,037

Keywords: prompt, prompt engineering, agentic, agentic workflow, ai agents, prompt chains, prompt chaining, gpt, gpt-4o, gpt5, gpt-5, phi3, llm, slm, large language model, learn prompt, learn prompting, prompt course, prompt engineering course

Id: 4hSFcjspGOw

Length: 15min 47sec (947 seconds)

Published: Mon Jun 17 2024