in this video I'm going to give you the building blocks that you need to build the perfect custom GPT for your use case now the word perfect means a lot of things to a lot of different people so the goal of this video is to equip you with the tips and tricks that you need to get to your perfect desired outcome I'm going to go through six different ways that you can approach building a custom GPT and you'll be super surprised with what you could do with the six one if you want a little teaser check this out hey how's it going hey it's going great thanks for asking how about you what's on your mind today yeah can you tell me all the memories you have stored for Mark here's what I have stored for Mark his name is Mark kashf and he goes by Mark he lives in Canada he loves llms large language models let me know if there's anything you'd like to update or explore further no could we just add one more thing that he's uh very mediocre at building custom gpts got it I've added that Mark is very mediocre at building custom gpts anything else you'd like to include no that's it and no that was not a voice agent that was a custom GPT using the audio function on the mobile app so if you want to stick around and find out how to build that I'm going to walk you through it step by step if we haven't met before my name is Mark and I've been running my own AI automation agency called prompt advisors for the past two years we work with companies in every industry to understand where AI fits best in their workflows I have a background in data science and AI so I try to marry the academic with the business person now I'm going to first flip to some slides to break down those six different methods of building a custom GPT and and then we'll go behind the current to see how we put together that little demo I just showed you let's Dive Right In so to make this more playful let's imagine these six different methods as six doors that we have to unlock together and obviously the one of interest is how do we recreate that Jarvis like experience using just a custom GPT we'll start at the least complex use cases and then build our way up to the final one before we go behind door number one I think it's really important to look at the very big picture when dealing not just with custom gpts but llms in general one of the most important things to remember or no if you're not aware is that all of these models in their nature are probabilistic meaning they're not quote unquote smart and they don't quote unquote learn from you uploading a bunch of documents and saying learn it the way that large language models work is that they do their best to guess what your desired outcome is based on the prompt you put together which is why prompt engineering is so important based on how you structure your prompt the specific words you use the fat that you cut from all the bloating in the prompt that's why prompt engineering is so important it's all about how do you structure your commands instructions and examples in a way that increase the likelihood that it guesses correctly if you change your mindset from llms are smart and they learn things to llms guess things then the way you approach prompting and building a custom GPT will be wildly different you're basically trying to spoon feed some hints and clues to increase the likelihood that custom gpts give you the answer they're looking for and the reason I focus so much on prompt engineering on this channel is typically all kinds of applications whether it be a custom gbt or web application all have the same problem where there's either too much prompt or too little prompt or too vague of a prompt or too assumptive of a prompt meaning someone's assuming that this new model is so smart that it knows exactly what they want and can read their minds when it's actually your job to guide the llm to what your desired outcome is and with that in mind we can now go behind door number one now behind door number one we have a very common use case which is a style and cone assistant this is all about either distilling your brand voice your company's brand Voice or a colleague's brand voice so you can easily upload a file or give some form of instruction like hey generate me this email generate me this blog post generate me this new product announcement and it will write it in your quote unquote style of either writing and embed basically your voice inside the actual text the general structure you want to take with this is you have some form of role assignment and the reason why all these role assignments where you say you are this or you are that are very popular and common is if you go back to what we just talked about when you tell the llm or custom GPT what it should role play as it increases the likelihood or the probability that it will know what the desired outcome is so if you keep connecting probability to what the assignments are you'll understand why these pieces of the prompt actually exist so you can say something like you are a writing expert specializing in style and tone and your job is to help refine writing so it feels insert some form of variable and then ideally you give some examples for it to learn from so when it comes to examples to provide it's not about uploading 10 different documents or injecting 10 different examples in the prompt because just because you add more doesn't mean it gets better ideally pick somewhere between two to four different examples that you provide it that say hey here's example number one here's two and here's three and ideally they're different enough to capture the different elements of your voice or your brand voice when it comes to examples and doing some monkey see monkey do less is definitely more and one thing you'll see consistently in all of these examples is I like to always end with final instructions for the AI which could be formatting related which is remember to always output between one to two pages always avoid words like XYZ let's say you don't want to say standard words like delve or seamlessly and all the recurring words that llms love to print out this type of GPT typically takes some iteration where ideally you create what's called a meta promp and if you don't know what that is I have tons of videos about it where you have a new chat gbt window that says hey you are a prompt engineer here's the prompt that I gave the other GPT window that I have open here are my flaws with it or here are my issues with it it's not capturing my voice correctly it keeps using the wrong words so I typically have two tabs open one that is my custom GPT prompt that I'm actually writing and the other that is my tester and editor where I go and vent about why that first custom gbt prompt is not working well enough so in your your second tab if you say hey you are a prompt engineer I'm going to provide you with my prompt that I created for my custom GPT and here are all my Grieves or my complaints about how it's not working if you go back and forth and you tell hey output in a code block you'll be able to easily copy paste it until you're actually happy with that custom GPT this is a hack I use with myself and client gpts that we put together because it helps us get to that final mile where you want that Perfection so that was door number one let's go into door number two so this is similar to the first and the fact that there's not too much action going on in terms of API calls or anything fancy this is just a different frame of a custom gbt where you frame it more in hey help me prepare for this interview or help me respond to these customer inquiries that we're getting or help me with this sales script right and in this case you have a different role assignment which is you are a conversation coach or you are a customer service helper that kind of thing your job is to guide users through and then desired outcome the next one is examples to learn from again this is going to be consistent from Custom gbt to custom gbt one thing I really like to do in this use case is I like to actually give in quotes examples of Ideal responses to questions and if you want to go fancy I would give it in quotes questions received from customers and then answers or ideal answers to those questions now I used to be hesitant that maybe adding quotes into a custom gbt would force it to actually follow those quotes verbatim but typically if I just say hey here are some examples to be inspired by on how you should handle these conversations it's good enough to quote unquote generalize how it should deal with new scenarios it hasn't seen before as an expert tip one thing I love to do is integrate what are called conversation starters into my custom gbt now you might think it's because it's a better user experience it helps people understand what they can ask and that's one very small part of it the main thing I noticed is that if I created conversation starters and then integrated verbatim those conversation starters in my prompt there was a high likelihood of getting similar experiences every time if we take an example here like write me a thousand-word analysis on the future of GPT what I would do is I would put this in quotes put it in the prompt and say when the user says and then this statement then do the following and then in bullets here are different things that you should do and what I noticed is that when I linked the conversation starter and put it into the prompt itself it increased the likelihood that I would get the same experience when I click those four buttons so if you have a workflow that you think you can fit between one to four different conversation starters that is an ultimate hack to help you stay on track and keep the custom GPT on track as well all right and on to door number three so this one is meant to be slightly different where instead of just providing you some form of output which is a generated email a generated response this is meant to be more of an interactive experience where there's a bit of back and forth and in the prompt itself the biggest differentiator here and I won't go over these two you'll have access to these slides Es as well as a notion template that I put together in the bottom of the description below so I'm just going to focus on the key part here so when it comes to the instructions for the AI I like to tell it hey try and not just be playful when you go back and forth with the user but if the user seems confused about a certain topic or expresses confusion try and double down on that area until they're satisfied meaning ideally the custom gpts instructed to say hey are you happy now or do you think you got it enough or do you want me to go another route so creating that exper where the custom GPT organically offers up diving deeper versus having the user having to ask that themselves so the core differentiator with this one is that you're designing the experience to be more about reciprocity and being proactive with the user experience where the other ones are more so how do you dial in the response how do you increase the likelihood that you get the output they're looking for so behind door number four we have a document analysis system and this is something that a lot of folks ask for because the idea of uploading 15 files and saying learn this and then help me do some analysis is very tempting and appealing now the most common use cases where this is applicable is when you want some form of review think resume reviews think legal review think any document like a contract that has a typical process of looking for key let's say Clauses or pain points and addressing them or at least highlighting them the way you structure this one is you say something like you are a document reviewer your job is to evaluate and suggest Improv M ments for whatever you're talking about and then examples to learn from in this case instead of uploading 10 to 15 different files and saying hey learn from all the comments I wrote down on this document the ideal scenario is that you provide small Snippets of saying hey here are the types of things that I want you to focus on maybe in quotes Clause this Clause that and if you want to go again more fancy you can write a little rationale under it saying I would pull this out because XYZ of importance and then you provide Maybe three to five examples of that and that's the core part of your actual underlying prompt and then depending on the document here when it comes to instructions for AI you want to tell it the structure that it should expect so it kind of is kind of briefed on it and warmed up on it before it sees it and then ideally you can go into what the next thing is is it just that you're going to analyze it and pull out some Clauses after the Clauses are pulled out maybe is the user going to confirm that they're happy with it before the LM knows the next step these are the questions that you need to answer when you're designing and creating a mental model for how this will work one super helpful tip for this one is try and space out the different requests to the LM so instead of saying hey take this document and then turn it into all these social media posts and then write me some form of email and you have like tons of next steps in one shot try and space them out saying hey when we upload the file first thing I want you to do is extract all the Clauses that denote or are related to X after I confirm that I'm happy with those then I want you to put together a small brief saying all of the different things that I want you to outline that we want to highlight in the contract and then I would have it more peace meal and the reason why if we go back to the premise is if these LMS are probabilistic meaning they try to predict whether or not the answer they're outputting is correct if you try to shove way too many instructions or actions into one shot the likelihood that you get a jack of all trades that is deep enough for your satisfaction is very low so ideally maybe you do things that are related so output all the social media posts from this document then once I'm happy with it put together some form of content calendar to then organize them and then you keep going step by step and if you isolate it in this way every step is contained meaning the likelihood that you get the outcome or the output you're looking for is much higher now we're approaching door number five now this one's called hotkey based interactive assistant and this is not a real name I made it up but what it notes is a very different way of actually structuring your prompts if you remember on door number two we mentioned these conversation starters but I did say if you have one to four because that's kind of the limit you can have on the front end at least as of today what if you have seven different operations you want to execute that's where this hotkey method is super helpful because instead of saying hey refer to conversation starter you can say when the user writes w or writes I execute this and these are the things that you should execute when you see that hotkey now how do you inform the user of what those hotkeys mean what you would do is is create a main menu so it's easier to actually show you than tell you so if I say let's create a quick custom GPT here I'll go to explore gpts create and I'll open another window just for laziness if we say something like can you create a menu of hot keys where different keys they know different actions a write me a poem B write me an email c talk to me about meaning of life D teach me how to write prompts okay so if we wrote something like this and I we say add an emoji to each and just make it look pretty so here's a hotkey menu okay and then what we can do is we can take this and then in a custom GPT we can make our conversation starter main menu and then you can say when the user says main menu display the following okay and then what I'll do is I'll just take these out and put them at the bottom so let's just full screen this to make this easier so I'm going to put the what these do in a separate section just so they're segmented and just so I can tell it hey verbatim output these in the main menu okay and then I'll just remove the spaces here and I'll say when the user sends any of these letters execute the following okay uh and if we just close this as is without even prompt engineering much more let's do main menu and you'll see here you have a b c and d if we say d right now it's become a proxy or a shortcut for adding in a layer prompt and you can imagine you can make make this a lot more sophisticated imagine under a you can say you know when the user types a right and you can have tons of detail under this saying like you know talk about I don't know uh the color blue uh say it in one sentence and all these sub restrictions that you can do and now you're basically nesting and organizing your prompt in a way where one it's easy for the user to utilize it and it's easy for you to create these different distinct sections where you Nest all your additional logic and asks this is again another way to space out your requests and dial in to make sure that you have a very consistent experience from menu item to menu item now we're getting to the door that you've been waiting for for which is door number six which is multi-agent custom GPT and the reason why I call it multi-agent is I recorded a video that many of you seem to like a few weeks ago where I showed you how to create seven or eight manyi different agents that you can then call using the at sign in a normal chat GPT session now a few of you in the comments called out hey could you create one custom GPT that had multiple functions within it itself and the answer to that is a big yes which is how I created that demo at the beginning of the video so this prompt design is widely different you have a rooll assignment if you wish it's actually not important that you have one here the most important thing is that you call out all the different functions you have where you put them in these little asterisks like this so I can say um let's go to a custom GPT that I've actually put together to make it easy to understand so if we go to edit GPT here and we go to custom actions and I scroll to the very bottom you'll see that this action is called submit proposal email details okay so what I like to do in these prompts to make it easier is I call it out by name right so you can see when you have these inputs send them and Trigger trigger response proposal email so I call the name of the function in the single quotes and that helps it understand hey when you hear this thing or when you see this thing sent to you trigger this function that does XYZ so if we go back it's not about necessarily examples to learn from in terms of like knowledge but more so maybe provide it tidbits of instructions saying hey when the user says something like this this or this then invoke this function that we've created in our custom actions and if you're not familiar with what Customs actions are they're pretty much a way for you to access the backend of third party tools from a custom gbt which make it much more powerful than just using the mundane prompts and knowledge based file that you're typically used to using it basically allows your custom GPT to unlock new skills that aren't embedded within the chat GPT framework as it is today now the custom GPT I'm going to show you is called Jack of all trades GPT so in this case it is attached to four different tools One is using make. com an automation workflow tool to connect to a Google Calendar anytime I ask for a new schedule or a new meeting to be created then we have anthropic that allows us to speak to PDFs dynamically but just uploading any Google drive file and having it talk to very complex PDFs which I also recorded a whole video on and then we've hooked up to new tools which I haven't spoken about before which are fir crawl and mzero Fir crawl allows you to dynamically scrape websites so when you put a URL in it's very good at quickly scraping that and outputting everything that's on that web page and when it comes to mzero that's actually the memory thing in the demo that I showed you where you can log memories in a data database and you can segment it by user so imagine you have Mark you have Jim you have Sarah each one of us would be able to submit our own memories to this database and when we talk to that Jarvis we can say hey can you give me all my memories it's Sarah so it's a cool new tool I actually stumbled on this week so I thought I'd integrate it to show you how it works and at a high level this is how the custom gbt will work we'll have memory management where we can search memories add memories Etc we'll have the ability to scrape URLs and process that through fir craw for Google calendar events we'll be able to gather event details but more so create a brand new event that's the core goal of the actual GPT as it is and the last one is to be able to upload PDFs that are more complex and have it handle that all within the realm of one custom GPT so I think the time has come let's go into GPT and see how the Jack of all trades GPT was constructed so if we go behind the curtains here and go to my edit GPT and then pop this up I'll read most of the prompt with you just so you get a sense for how I structured it so you are a multi-functional assistant capable of managing memories scraping data creating caler events and analyzing PDF documents follow these instructions based on the user's input and basically for memory management I do have those little hot Keys If I want to use them but in general if you go down here I've segmented it by function so if I say hey here's all the stuff you need to know about the memory management system based on all the custom ACC I've created and then this is an example interaction so if I say this hotkey this is what you should ask me for the user ID is it marks memories Sarah's memories Robert's memories and then the actual query or question you have about those memories with web scraping the prompt is what URL would you like to scrape and then provide the whole URL and then I basically tell it for each one of these sections what do you need from me as the user to make sure that you can do your job with the Google Calendar event Creator basically I'm giving it a framework for what are the things that are required ideally you want a start and an end date for a meeting invite right cuz you can't just have it all day and for Google Calendar you can see here we have a few things that are optional like attendees or reminder settings but the one thing that I have or a couple things we have that are required are the name of the meeting when the meeting is and the duration which is the standard things You' need to actually put together a meeting on the make. com side I have this web hook that I'm sending the request to and I'll show you what that looks like right now so if we go into make and we go into this calendar update it is an unbelievably simple automation the hard part is basically just structuring the request from Custom GPT to make sure that we're getting the different aspects here which is the name the attendees Etc so we know where to plug it in and if we go into here you can see that we have the calendar connected and what you have to do and I'll show you basically how to structure this how to put it together all in a video in the description below just to save some air time on this video I'll walk you through how to set this up yourself but pretty much the crash course is you sign in using Google assuming you have a Google account and then you just sync your calendar ID which will be available if you have it linked up and then I just put in you know this is the event name this is the start name and I'm getting all of these from this web hook or ear that we've set up to listen for chat you request so we basically just like drag and drop start end uh the description of the event the location of the event uh any attendees I just put the attendees here so you can add multiple people to the event and then at the bottom you can configure it so this meeting has a Google meet invite already embedded within it which is a super cool feature so all we did is just hook this up to the Jack of all trades custom GPT and then in here here we go we just have a reminder at least I've put one out of paranoia that this is what you should be doing and it also has a little piece of feedback here which says like if it's a success say create it successfully otherwise saying invalid data check your inputs and this is an example interaction that I have here and the last one again I've gone through this specific Automation in detail in another video that you can watch it's the PDF analysis API from Claud which allows you to upload more complex documents let's say more on the visual first side versus having words where I can actually take screenshots in real time and analyze what's in them so I've basically enabled that through this prompt as well and then told it EXA an example interaction which is when I say Analyze This PDF ask me ideally for the Google Drve link that I want and then I love to end with General notes saying always guide users through required inputs validate data formats provide clear error messages combine Clarity with efficiency for a seamless user experience so these are more so for the gentrification of the back and forth now that's the prompt and that's the underlying brain what I've enabled to make that Jarvis experience possible is a few things so if we go back to my slides here my little Easter egg is that I figured out how to avoid having that constant confirm or deny pop up when you hook up a custom action so in this case there is one line of code here that you can add to a schema and a schema is basically the map of all the different functionalities of a custom GPT that are outlined here so if I go into one of these settings this is a schema this tells chat gbt or custom gbt here are all the different functionalities you have here are the keys you need to go access this third party to do whatever I want you to do if you go back if we add this x open AI is consequential false this helps us get a different popup instead of confirm and deny it'll say allow or always allow and if you're in the front end you can click on always allow and then it shouldn't bother you every single time you make a request which is something you must be used to and I've been used to even on this channel hitting constantly naturally this allows you to have that Jarvis experience so if we go back here we have these four custom actions and if you're looking at this saying oh my gosh I have no idea how I'd implement this this all is going to be available for you in the gumroad link in the description below you're going to have all the schemas the prompt as well as this bad boy here my custom gbt templates where you can go through all the templates I showed you but a bit more in depth with an actual prompt template that I've put for you so I'm trying to set you up for success to be able to create the custom GPT of your dreams now if we go back the main thing here is that these three this one this one and this one I've built on repet and the reason why I keep using repet is if you try to connect these Services individually and create a different schema for each one sometimes you'll run into some headaches where you don't know how to enter your API key which you'd enter here are you going to use a normal API key do you have to click on this thing called Bearer is it a custom header some of these really cool tools have very poor documentation on where you put the keys to be able to access the service so when I create services in repet and then I hook that up one that lets me have very consistent schemas meaning I can take one of these schemas put it into Chachi BT give it a new piece of code and say hey replicate this and then allows me to easily Bridge it and just put the URL of the actual repet app that I end up deploying and if you don't know what that repet is um in my original video that I did a few weeks ago on how to create these different custom agents I went into depth on how to actually set these up now the rest of these look very similar and if you don't know what our memories even is I'll show you what the front end of this tool is as well as fir crawl so you're more familiar with what we're even talking about um but these are all going to be similar except for this one this one is the make automation that connects to the Google Calendar and this is the web hook that we've created to be able to speak to and access make.
com and if you're completely confused how to even go about generating this let alone the rest of this I'll also have this covered in a video in the description below now these four components are what allow us to actually have this Jarvis experience and if we go into any one of them here you'll see that if I do a little search you'll see I've added this line that I just promised you in the slides called X open AI is consequential false based on my understanding it means that this is not a secure operation that you need to protect by constantly saying confirm or deny so when you enable this the first time you use your custom GPT it will ask you hey do you want to allow or disallow but once you click always allow you should be enable to talk to it back back and forth and not deal with that constant confirmation all right and now if we go into m0o this is what it looks like if you want to check out the tool you can use it for free up to a certain limit like all of these AI tools and if you see here if I go and click on a specific user let's refresh and let's click into Mark here you'll see we just added this at the beginning of this video Mark is very mediocre at building custom gbts and you have all of these different memories and what's cool about it is it Auto Tags those memories with different tags that you can also look up as well if you want to go down the rabbit hole you can go into the playground here and say things like remember that I love my husky named Moon and then once you save that if it doesn't exist in the memory then it will add it to your memory there you go and then it will remember that and then you can change users so if we go back here to the playground I believe or yeah users I could create a brand new user and just say hey you know I want to now store Sarah and for Sarah here are her memories and she can enter them herself through the custom GPT or the platform itself you can also imagine this as not just users but also possibly locations so imagine you have a franchise and you have five or six different locations you could theoretically add location number one you know South San Francisco and then North San Francisco and then each one of those would have have different memories that have different procedures or opening or closing hours any form of detail that you don't necessarily want the chat gbt memory to know or because in a custom gbt right now memories updating are not enabled this is a way to get around that if you want to be able to take advantage of memories within a custom gbt experience all right so that's mzero if we go into fir crawl it's unbelievably straightforward they're really good at scraping websites and then putting the actual website into a markdown format so if we just go to my website prompt advisors and we take that and we go into fir craw and then paste this and say start for free in my case I have a paid account but you'll be able to send I think up to 500 requests on a free account and you'll see here it scrapes the entire website and it puts it into markdown to make it large language model friendly so if you give it to chbt it will like this format a lot more so basically just accessing this service but through their API where I've created some code in repet using python where I'm connecting this API key and then hooking that up to our jack of all trades automation so that is the mem zero that is the fir crawl and Cloud PDF again just watch that video If you're not familiar with that specific function but the tldr if we want to go to Claude is this API came out a few weeks ago and it allows you to speak to much more beefy PDFs that are more maybe image heavy than text heavy and don't have a lot of explanations where it can take that analysis to the next level all right so with all that out of the way in terms of actually interacting with it after updating all of this if we go and ask for all memories I'll show you what we can do on the desktop so what you want to do as a shortcut is go to this drop down go to privacy settings and then you see how all of these are set to ask I can say always allow always allow always allow and then always allow and then when I close this and I say provide me with marks memories sometimes it'll still pop up one more time but in this case because I've enabled it from the actual top it didn't ask me to confirm or deny which is my Hallelujah moment that I've been looking for for months which is how figure this out and on mobile this is slightly different but if you want to be able to use that chat interface with voice you can't access that on the desktop app you can only do that on the actual mobile app so I'm just going to screen record from my actual phone and then I'll show you here and I'll make sure my editing team puts it on screen so what you want to do on a phone and I'm going to screen record right now and then I'll have my editing team put it on screen side by side is you want to click on Jack of all trades in this case and then click on C details and then you want to go to privacy and safety click on actions and then you can see here they've all been set to always from my desktop so basically the mobile has inherited the desktop and this is instrumental for me being able to have that seamless conversational experience because otherwise what happens is it'll say okay uh it seems like this app is trying to send requests to third party do you allow this and sometimes when you say it on audio it doesn't work as well so once you've enabled everything you'll be able to have conversations via the desktop app you'll be able to use the old version of voice on the mobile app if you choose to do so and now you'll be able to do things like create calendar invites so if I pull up my jack of all trades once again hey are you there yes I'm here what's on your mind yeah could we create a calendar invite for next Friday November 22nd uh for 10 a. m. EST and call it standup and make it with my colleague Taha sure can you share taha's email address so I can include him in the invite yeah it's Taha prompt advisor.
com the calendar invite for standup has been created successfully for next Friday November 22nd at 10: a. m.
