“Anyone can build an AI startup, this is how - AI Jason

this is Jason zo you probably know him as AI Jason he's a professional AI agent Builder and product designer in this podcast we talk about how to build an AI startup how owan is making people thousands of dollars and how to build anything with ker if you want to turn your AI idea into a startup this video is perfect for you so watch until the end so Jason what advice do you have for people who want to build their own AI startup I think probably there are few things there are a lot of the potential

ideas about what can you actually build with AI or with AI workflows first they just try to find a problem that you personally experience a lot that high frequent and also painful enough that is worth your time cuz the worst part is that you might solve a problem that no one actually end up care about but that's very general I guess to building any type of startups but what's really I guess interesting about the building AI startup is you want to figure out which part the AI can actually Empower and enable so that's the problem

like after you identify problem just try to build a very quick prototype I would say a lot of people either underestimate or overestimate what AI agents can do right almost nobody has a balanced view I think and only the people who really get their hands dirty and actually build with it they understand it the most there's like people who think you know AI can do everything there's people who think it's completely useless obviously as with most situations the truth is some in the middle so yeah I I agree like building some prototype and getting it

to the market as soon as possible is one of the best things people can do because what I see with a lot of people they stay in this like idea phase forever like they research preparation and they they think they have the world's greatest idea but then as you said nobody wants to use it yeah exactly cuz I often talk to uh people like when I say oh my idea is let's say F in model specific for my use case uh which required to have like huge amount of training data to begin with uh but

instead of like try to spend amounts to prepare training data maybe just try out with like simple strategy uh see what default outcome from the large model dat this will speed up the uh kind of idea to prototype a lot faster as well yeah there's like solutions that would be perfect but would take like one or two years to create right or solutions that would be like 80% and you can just use you know CLA 3.5 Sonet or gbd 40 and you can prove it that way so I think again people are always aiming for

that perfect thing but maybe we need the next generation of models to make that happen right I think there's a lot of ideas that are possible just not now right so like it people can waste maybe not I it depends if you you know take it waste but they can spend a lot of time developing an idea that isn't ready yet the classic example is Google Glass right it's like the AR glasses from I think 2012 or whatever was like way ahead of its time and now we see that meta AI is obviously with a

much better product is actually creating that possible like AR glasses are going to be a thing but just you know Google was super early with that and I think a lot of people are going to be the same with AI agents where they think like they can automate something really complicated especially in like content creation where the you know creativity part is really missing I see a lot of possible startup there that I think like you know I I have like thousands of hours in ch GB and cl I know like I don't think the

models can do it so yeah I think people either underestimate or overestimate and also what's funny is that a lot of beginners they want to build something that like you know is AI powered but could probably solved with just raw code that's also a funny phenomenon I noticed yeah yeah exactly like one of the when you think about it a lot of star star out ideas literally can be achieved with sprad sheets in in the old days right so uh instead of building your whole platform you could just build like a species automation um and

see that this actually works so there are a lot of like hacky things you can just quickly prove uh as well with that real quick if you want me to help you grow your business then click the first link in the description I will be choosing five people who make at least $2,000 a month and I'll help them scale to $40,000 a month and Beyond I've already done it twice myself so I'm certain I can help you do it too however this project is limited to five people only to make sure to fill out the

form today it's the first link in the description I also think with AI agents what people don't realize is like reliability is by far most important like the you know the obvious example that everybody uses is the plane tickets right like people think it sounds good to have 99% reliability right oh it's 99% but if you know every 100 times your AI agent booked a ticket to the wrong country or booked a hotel in a different state or you know different country you would not be happy right so in reality we need like 99 .9%

reliability if not more in AI agents and I think for most use cases that's simply not the case right now and people need to like people need to realize with their idea like how how quickly do you think you can get to that 99.9% because if if it's going to take you like two years in the AI space that's like you know two decades or more so um yeah reliability is for sure key yeah I agree I actually uh I think kind of learned it Hardway for past few months uh cuz we were trying to

build a sales and agent uh which is okay help let's say think about if a b2bs uh you want to outreach to customers uh and uh instead of hiring a person to do that Outreach uh which takes a long time uh and uh training uh why not have ai agents who can like you give a list of customer in research Outreach things like that and and also book meeting one thing we could realize is that use case like that exactly what you mentioned it require really high level of accuracy uh cuz it's customer facing you

don't want to mess things up uh so those type of use case can still work but you basically need to find the a specific segment of customer who are really high risk tolerance they okay with fully align with you it's not going to 100% it's going to 99% but we accept that uh so that's kind of for this type of use Cas you probably would need to find a very specific Sig customer there who can go live with that otherwise you guys just keep stocking at that uh iterating loop uh that's kind of for forever

going on so set up expectation correctly with this those uh those use cases is very important uh on the other hand uh what's probably more important is that you probably want to think about what a use case is actually it's okay to mess up for most of customers like it didn't require 99% of accuracy for example internal if you look at internal automation process that's probably a bit easier if you look at a inste of talking at the uh instead of automate whole process of let's say marketing uh writing content publish content but if you

just do the content generation but still have human in the loop as part of workflow then the risk tolerance became much higher sorry yeah the the bar for the going life let's say go life standard bar is much lower so that means you can go life easier as well so this kind of s I definitely agree like the accuracy really depends on so you really need to use accuracy almost at a benk try to figure out what a use case is most suitable for your uh target audience if you want lifetime access to new Society

then make sure to follow me on Instagram all you have to do is like and comment on my fre latest post and one of you is going to win free lifetime access my Instagram is linked below so I guess now is a good time to kind of introduce what your work at relevance AI is like what's your vision for the startup for the future and yeah what basically what you are what you guys are building we are building the home of AI Workforce uh so imagine we basically realize that there are a lot of opportunity

for the business to automate their work and when you look at the traditional automation platform there are platform like uh D here or there are platform like m.com or RPA platform uh those platform normally really focus on the integration and try to do a kind of Point too automation which means uh if you have some kind of the trigger or workflow uh oh sorry or some data from system a and you want to trigger some workflow to system B uh and a very kind of linear workflow but with with large L model and AI uh

with large L model and AI agents what's really cool about it is that AK can star make a lot of fuzzy decisions uh and it is able to some sort of more agentic automation that wasn't possible before and this kind of not a good platform there yet uh so we are try to build a platform enable anyone to build uh kind of agentic workflow easily so people can build uh AI agents and then embed this AI agents as uh can almost go to random AI to build any sorts of AI agents backend and then connect

to any sort of f you need uh but on the other hand we also introduce uh we also have a kind of chat UI for people to use and embed on the website directly yeah I think the like especially after CH GPT people can understand the power of a chat UI right like it it's much easier to tell the agent what to do in text than know you know all the settings or keyboard shortcuts or again and how to connect a different integration so I definitely think that's uh good like I'm considering adding a chat

interface to my app even though productivity tabs usually don't have it right just because people know how to interact with it it kind of invites more people to use that AI agent so basically you're trying to like disrupt the zapier word of you know automations where you still have to know what to do by making it more like easy for I guess businesses to automate right yeah exactly like when you build workflow a traditional automation you kind of need to know what is the process there and you normally pick up those I say high frequent

super high frequent use case and scenario because those are the ones that make sense justify the let's say cost of deploy but what this left out is a huge amount of like longale use case that uh traditionally very hard to uh automate so think about a use case as a meeting schedule scenario if you want to build a AI system or automation system that can B meeting for you uh tradition is actually very hard because people can responding any sorts of like millions of different ways when you ask them hey what time you want to

they will say hey I'm in sosit today but UK next week how about like Thursday right so that's really complicated uh to autom kind of longtail use case but with large L model this type of more kind of fuz decision is very easy to make uh once you train the agent properly so this is a kind of you use case and workflow there where kind of more focus on and what's your role in the company specifically uh so my my background is a product designer so I actually uh uh so I'm doing a lot of

like product design and also was doing a lot of like solution engineer at work for uh because what we realize that at beginning we are building this platform enable anyone to build agents and uh um but very quickly we realize that it's so hard to actually build a AI agents in production for Enterprise uh it require so much up front training and uh uh and and and uh and building so what do we de side like Also let's build a kind of Flagship use case to showcase the word that this is actually going to work

so that's what we start doing so internally we have a team to deploy build and deploy AI agents for Enterprise use case as well yeah I think like UI is one of the biggest bottle legs for AI agents right like we all talk about like reliability but also what's preventing most businesses and people from having their agents is that it takes some skill right some technical knowledge to build them so I think we need the same what have happened what basically CH GPD did for llms we need the same for happened with AI agents where

it's just so easy to build an a agent that it doesn't make sense having one same same way it happened with email as well right at the start only like the real programmers real ogis of the internet had their own email even before the internet right but after when it become so easy everybody had it so I think UI is probably another portal Nega it's kind of preventing AI Agents from becoming as popular as jgpt for example yeah totally so we actually learned a few things as well so at the beginning we kind of everyone

is think about chat UI so you just like chat to AI agents but when we look into automation use case quite often what they do is not like kind of chat when I think about chat is more like co-pilot use case so you build AI that can maybe retrieve information you talk to it but when we look at a kind of Enterprise automation a lot of them is actually autopilot use case which means that just you have workflow that you want you currently have a human to do uh but you want to build AI to

let it autonomously doing it and the automate whole process and for this type of use case the chat is probably not the main kind of entry point for you you if you have employee you don't want to like go and check every single task he was working on right uh so it's more like uh so we start introduce different type of experience for people to let's say how do I uh review the agents task at different level uh so for example you probably we need to design very clear uh let's say uh human in the

loop experience so when agent needs how can you easily build a workflow inside your agentic work system so that when it's unsure it can escalate human easily uh and after escalated how can human easily give instruction to the agent and uh so that it know what to do next and after the agent uh was given instruction and do the task how can it actually learn learn about this interaction from Human so that next time it doesn't need to uh make the same mistake again so there are kind of like a lot of U around human

in the loop experience um and the second part is that as we mentioned before we realize when you have a fully autonomic agent existent rounding in the production you actually don't care about chat UI too much at the point at least from the ones we observe what do you care about is like okay how can I know my agents how many tasks that my agent actually running for the past 24 hours so logs basically logs and the review system so we actually introduced quite a smart way for people to uh like let agent let's say

label tasks by itself so if I'm Outreach to a uh to a ppat who after research I realize that high uh High Fit versus low fit the agent will be able to autonomously like tagging and labeling this task so that later you can come to system say Okay among all the people we out register how many of them like high fit how many of them are low fit uh and what's that kind of email open rate uh reply rate things like that so we start building out this kind of a an analytics uh system as

well which is quite interesting because when you think about is actually large Range model system is doing task and then labeling General reporting for the for the review by autonomously which is quite different from the traditional kind of dashboard experience um yeah so there kind of a few things we explore as well like both in human in the loop experience and also how can we enable uh kind of human to review the work that agent has been done yeah actually you kind of give me an idea to implement myself because I didn't have anything to

kind of take care of when so for context for people my AI agent is kind of sorting tasks and I didn't have anything in place where if the confidence score was low the human would review that so I I just got an idea like what if I just pop up like a you know small warning icon or something like that let's say the if the confidence score is below five out of 10 or whatever then the human can see like the a agent wasn't certain where to put this task probably review it so yeah I'm

definitely going to add that so thank you for that U my next question is just for people right A lot of people watching have their own business right they're small mediumsized business owners and they're wondering okay how should I think about implementing AI Asians what do you think are some of the most obvious use cases that as like small and B mediumsized companies should focus on do you mean like the use case that they build for themselves as a internal use case or more micr basically what do you think is like some of these things

where they can get like nearly 100% reliability with a agents right uh yeah so one thing that large Lang model is really really good at uh is that extract insights uh extract structured data from unstructured information so there are few use case that we uh we saw is working really well really reliably I would say uh one is that there are people companies have huge amount of let say uh customer feedback data uh or meeting transcript um and uh it's probably a bit harder if you just give the agent agent if your expectation is like

give agent let's say those information and generate bunch of reports that expectation I would say is probably a bit higher and you will need to a lot fun to get it right 100% but if you just give agent uh a task let's say this is a meeting transcript help me extract what's action item from this meeting transcript and uh what are the pinpoint this customer talk about that one is really really good at uh and the accuracy is is is is to a level that you probably care you you probably don't need to worry too

much that it mess up things uh or extract things wrong because just extraction is really good so from there you can probably think about you like how do we enable uh the sales process after the meeting is finished can you just extract information out from meeting and then push those information to CRM so that all those information so the salesperson need to focus on moving things or uh if you have a uh a customer support ticketing system like intercon that here uh you can use this to actually automate process whenever a ticket come in automatically

extract the let's say categorizes is this a buget report a new feature request and then um trear and routering to the right people so there are a lot of huge amount of use case you can think about uh just about extracting information is structur information from unstructured data messy data um so that's kind of one big I would say big ability that's really reliably live mod is can really reliably doing um and the second part I think is that you can probably observe how people around you are using chpt at the moment uh because there

are huge among people uh just hacking things with TR today uh and uh when you look at that a lot of workflow is quite messy uh so my wife is uh uh is working Furniture business like has uh and uh not technical at all but should use chpt every single week to just build uh blog posts basically so but when you look at his her like um workflow is like she will pump all the internal data out from their own CMS or CRM uh go to the chat PT and then prompts three or four times

to ask it to like to get get to read out you want and also maybe go to the SEO website like HF to get keyword information keyword ranking information uh into the chbt so that you can like start generating content based on all those information so a lot of repetitive actions basically yeah exactly a lot of actions uh and so if you just observe how people are using that you can probably build a micro sess I would say fair easily um might might not be a billion dollar idea beginning but I think it should be

straightforward if you try to make let's say uh 10 or 50 uh K kind of Rue per monthly Rue I think that's probably doable if you're just looking those use cases yeah I think there's going to be a surprising amount of companies that are in that range right like maybe like low five figures mid five figures and a lot of people watching would you know give everything to have a 50k per month AI startup and I think I I I don't know who said this maybe it was naal ravikant who kind of predicted that there

would be like 1 billion companies in the future where you know thanks to Technologies like a agents people will be able to start their own company with zero employees right or maybe you have ai employees like AI agents but yeah I think it it will it is absolutely going to be the thing that let's say a year from now two years so now we're going to be paying even more for these AI subscriptions because we're going to be using you know this one thing that saves us 30 minutes a week this another thing that kind

optimizes our desktop a bit organizes tabs whatever right there's going to be so many different like as you said like 10 to 50K companies that are just like super Niche you know they're not in a multi-billion dollar market but they solve very Niche so problem that a certain Avatar is going to be more than glad to pay for yeah exactly and uh uh especially with like the service uh ai ai coding Trend we saw recently um the entry point became even you are not building AI solution you can just use a solution to build a

new kind of a platform pretty easily uh and this is one one kind of thing I I think I was experimenting the other day uh so uh probably kind of similar to the ability to extract structured information from a unstructured data uh when you look at the uh with this type ability the agent is actually very good at use case let's say web scripting uh it can just look at HTML and figure out okay what are data I care about uh and when you go to I remember I was going to platform like uh upor

or something uh if you search today there's still huge amount of hundreds of web scraping jobs there literally and uh it used to take a quite a bit amount of time to build every single website scraping because every single website is different right but a lot of people didn't realize that you can't actually use large L model and build a web agents to literally script any website and get information you out um so there are public opportunity like that as well you you don't even need to provide a software service you can computer agent that

actually deliver the job and sell the job and the uh work unit as well I think this is one of the biggest like right now this is not going to exist forever I think for the next 6 to 12 months this is one of the most untapped opportunities where people can make like $1,000 there $2,000 there up work there's so many things you can just solve with either CLA 3.5 Sonet or 01 like there are so many projects where people just offer like $700 $1,500 that if you just put into 01 in one to free

promps you would have it built it's kind of crazy yeah exactly okay so also the extraction part you mentioned I think that could be very useful with emails right A lot of people get especially if you have like a company or something you get a lot of emails that are low quality right so there there's going to be for sure AI agents that can like you can set custom criteria right that this is what I'm looking for let's say you get inbound leads and only certain amount of clients they have to have this type of

Revenue they have to live in this country that is going to be for sure a thing what do you think is like in the emails Department because to me that's kind of obvious why is hasn't that been like revolutionized it it feels like emails are a bit behind still in terms ofation with a agents yeah I actually think email email is a one use case I saw at least from the the use use Enterprise use case we're looking at it is actually something that has been touched a lot uh maybe not consumer side but on

business side it was done quite a bit like case we saw uh one is that like uh we saw a lot of VC company they were building AI agents to automate their internal process because they got a lot of inbound inquiry uh and uh when you think about it it's like the volume it's basically there are Ts that with so much huge volume before that didn't really make sense or couldn't justify the human cost before uh because you just there such mix of qu quality of those um but with agent start making sense for you

to for agent to look at millions of those kind of messages and then future out the ones that um and ones that need to be handled uh so we start seeing some actually workflow actually deployed into the company uh looking at the incoming emails categorize them based on different uh based on rules and then even trigger workflows uh so if it's let's say a startup company then it will get agent to like research about this company uh and then extract information like recent founding team size is a product already live or not what type of

product it is what's the revenue model so you can actually get all those information from the website uh of this company and then push those information somewhere so there are a bunch of things happening like that already uh I think part of reason or at least what we saw at beginning was stopping the adoption is that email if you if you want agent to take over p someone's personal actual email address that they have been using for years there's a lot of push back in lency they don't want to do that uh and especially if

you actually want agent to not just do the data processing but also taking actions that's a part is really kind of they they have bit afraid to do so uh so there but there are some mechanism you can actually Implement uh for example you can if you are building outbound agent you can build a w that a can only handle email they s out but if someone else send the email to this email inbox ignore that or for those type of action just to kind of more low risk uh lowrisk handling instead of sending an

email inad directly just draft that email in your email inbox so human Ste in the loop for every single action um so this kind of like the thing that we learn that kind of stop people for adopting it but I definitely think email is something that has been uh has been one use case that very much on the tip to actually drive a lot of adoption you mentioned also the recent kind of Boom in AI coding right like the last 3 four months have been insane with the new tools coming out and also the current

tools just being massively improved I think we're legit like 6 to 12 months away from people just typing one sentence and in 30 seconds they have a fully deployed startup that can accept payments on the web I think it's going to become so easy that literally all of us will be building custom software for the smallest of things like right now you can obviously still build custom software for internally for stuff but it's it has to has a certain bar right otherwise it's a distraction it's going to be everything I think in in the year

we're just going to have like oh it would be cool if we had like a Google meet internal clone that has a you know AI agent inside with GPT voice that can answer questions boom in 5 minutes you have it built I think it's like do you agree with this that we're all going to be building it's going to be so easy to build an entire AI app look any app with AI in about a year or two yeah I definitely think so because I think both unb has been uh like diving into AI coding

and we start seeing this what's really seeing this new market because what's really cool about rent AI coding trend is that uh let's say first generation of AI coding like GitHub ciler that's cool those kind of Auto compilation but the target market is still the programmers it didn't really unlock new things but with the cursor and especially prob both I don't know if you know both yeah both on you they can unlock this new market uh about the people who never know codes um and I actually talked to the founder of b. new the other

day uh and uh what they are trying to focus on is actually really talking at those people who are non coder before interesting uh so and this actually introduced a lot of like uh uh like kind of new challenge I would say that cursor because cursor still to be honest even though it's a lot of people using that to build uh um build those production ready application it still a bit require like previous knowledge about the coding but I think both our new or maybe some other new breid of company like rep agent they're focusing

on those kind of non-technical people that introduced a huge amount of newx problem uh what he told me it's like people just have wrong different expectation about what to tell AI uh they would give very vague answ so they introduced a bunch of a whole bunch of kind of U act like automatically improve the prompt uh for the for the user even though they give that vague ones uh and uh uh and you know like I think actually in one of my previous video I was for cursor wflow uh I I was trying to figure

out can we have a standard prompt for standard feature than like log in payment that I just need it but I don't care how it implemented uh so there are things that on one hand probably I think Community will come up with those kind of standard AI coding prompt to make those development easier but on the other hand I think platform like b. new they will probably do something integrated with super base directly and programmatically like set up the uh data table in your back end or uh automatically provide native integration with stripe so you

just like enable click on enable payment everything will be set up for you so I definitely imagine this would come uh in in in one year time uh we already saw people we already have kind of like early adopter start learning all this AI coding workflow do a whole bunch of things uh very very time in a very time consuming way uh but I expect that we're getting much better what you describe with the prompts is basically what vzer is doing with shed CN UI like they're taking like pre-built Snippets that are proven to work

and implementing them based on the what user wants right so maybe that is going to be real the answer the unlock for bolt and replate agent is where they try to understand understand what the user wants roughly and then take like expert prompts or expert Snippets of code and then Implement that because you know otherwise the LM is going to be confused because it's very easy to if you're a beginner to like point the llm in the wrong direction because the llm just generates the next token so it will gladly follow your bad instructions and

lead you down a path and then like shaking your fist like oh you know AI model suck when in fact your prompt was like completely terrible and you confused the AI and it didn't even know what to do so yeah I think that's probably among beginners that's the biggest challenge is that they don't even know how bad their proms are yeah exactly uh and actually by the way I think that's probably one good startup idea there as well uh because currently we have platform like cursor. directory um to be honest I but all the content

it has just the do cursor rules uh and from my experience that part the do cursor rules wasn't really that helpful what really helpful I think is those kind of the modular prompt that you can just take and then uh put into your current code base with some prompt that so the agent actually understand sorry the the AI actually understand your project structure so that it can Implement those feature into your platform so that's I think a pretty pretty good idea if someone someone doing it I would personally use it a lot I think so

that's kind of one part and uh so you mean like inserting it as a comment in the code instead of cursor rules uh so yeah so cursor rules I I found that is that the rules you're putting into cursor rules cursor actually didn't follow that strictly yeah like remind it yeah I need to remind this stuff yeah yeah exactly I found to actually get a work properly uh I my best practice is for every single feature I actually have a very specific markdown file so if I'm set implementing a let's say a user authentication I

will have a prompt that I use quite often inside this problem I would explain okay this is survey you should use uh this is the like the the file structure that exp and this is the current project structure so the second third section is actually very important because quite often the cursor don't really understand the the project structure at moment then it just create a file in the wrong place wrong dependencies so I to actually get cursor following things I will normally create those specific feature marked on file uh and then give it to C

and the newag file exactly uh that's probably the best way I found when to actually tame it uh and if that workf got to continue then what is needed is like what are the all also markdown file for specific feature uh because I observe myself actually keep a uh repository of those prompts for user authentication for payments for back half setup for backend setup so I I imagine probably mine is not the best as see since I'm not a professional front end rapper but there are all the other different use case and scenario that could

be pretty standardized so I do something slightly different I put still I use the cursor rules but I have like these different processes so I have a process for implementing new features process for fixing errors and then I have to still remind it in the actual prompt that follow the error fixing processor follow the building process but when I remind it it does it right so like I have different processes for different type work I do within Cur but still when I remind it it does the process so for me this this is what I

found to be like the most time efficient I guess yeah yeah so that's that's good I guess those are the general rules uh that is should is should photo so I think Cur rules actually good for those kind of General principle you should do whatever uh for different scenario but on top of that you probably you can combine that with a feature specific prompt so that it can do the feature with the best practice workflow uh so that it has less arrows okay so for people who are watching this most of them are not like

senior developers right so how would you approach that would you like consult with oan about like how you build a good login feature or good like payments integration and then take that put that into markdown and upload it yeah I actually have a uh a very kind of very specific wow uh that I don't see many other people are doing so basically what I do is that a few a few things I was there one is that cursor at default cursor at default quite often uh it didn't really do the planning that well um and

second one is that uh especially if you're using some new package it's almost guarantee it will have Arrow because it didn't have the knowledge baconing yeah um so if you just do go to cursor and give it instruction about like this is what I want you to do then it will often fail uh but instead of what I will do is maybe I can quickly share share my screen to show you as well um of course so this is normally what I do when I start a project uh so I actually want to generate very

oh sorry not this one give me one second okay here you go so this is kind of the normally the markdown file I would generated uh like pretty much like if you're part manager part designer you're pretty familiar with con project requirement doc uh if basically include the whole table of content about what's the overview of product what a core functionality required uh what a file structure uh so I will here break down okay this is what to try to build and then break down every single feature uh what are the requirement for every single

feature uh in this part it also have very detailed uh instruction about like which package you should use uh what are the kind of data point you should get from the API end point and then also list out like this is a final file structure you should work towards so it understand the dependencies uh and then some very specific notes about here I call additional requirements but basically are the things that I often observe that uh cursor might at default ignore or doing things wrong so and in the end I will include the very specific

documentation and code example to implement uh some let's say new uh some feature that using the new package so this is kind of like the final docs that I will try to get ready before I implement it um but I guess the question workflow I think will be interesting like how do we actually get to the doc because this might feel pretty stunning and a lot of work um to actually write this doc uh so what I normally do is that I will try to uh start with a kind of a draft doc instead of

having this hosting I normally just have like a a few sections I would have a section called I say uh let me check U I think I have template I can copy paste over it now looks something like this so I have a project overview Co core functionality do uh current file structure and additional requirements uh so I normally just write this part uh but not very detailed to be honest uh it's it's some kind of like uh simple doc so let me paste in a example I had just a side note when you mentioned

that you know you know people like when the model doesn't have access to the internet and uses outdated versions it's so surprising to me how many people still don't understand this like it it's almost every day I see somebody you know asking CH gbt like without telling it to browse the web to do something or clot which doesn't have internet access at all people like need to realize that you if the llm doesn't have internet access and you're asking for something recent it's going to hallucinate an answer that seems legit but it's not real so

this is like a super common mistake I see with people yeah totally you it's it's uh it's actually quite hard even though sometimes you give the URL and Doc of it still get things right it's crazy sometimes it's just like the training data is like so strong it it values it so highly yeah exactly uh that's why I actually always follow this process of some doing some pre-work so I were uh if I paste in like normally I will try to write down this uh basically I I found because with where it is at the

moment with large Range model it has all sorts of potential pass to get Arrow so what I do is like basically pel off those un risk uh pel off those risk and uncertainties by doing those planning uh early on so I might just write a things like project overview uh and then I will write down a uh kind of like core functionalities instructions uh so this might feel a bit stunny but if when you look deep into that it's really just a a list a requirement about what are the core functionalities needed uh but this

two I think are pretty easy to do but the third one is a is part I guess I will start spending a bit more time like uh basically I want to include the working code example for the agent or for for the cursor uh about how to use certain packages so it didn't hallucinate and it didn't like make things make Arrow later and what I normally do is like let's say in this case I want to use snow wrap as kind of package our uh sorry um oh sorry this is uh not cursor um or

open cursor instead you was about to say like the codium with of cursor that seems illegal yeah yeah say um okay so normally what I would do is that I will go there and then add a doc which is connect to the doc that uh I want to use mhm then I will just ask it to let's say so do you found it add dogs feature reliable CU from experience it it can be hit or miss as well yeah not really uh this is not reliable at all that's why I actually do this planning and

testing up front cuz normally what I do is that I would give it promp like this dog help me build this let say simple type script script uh to do this job uh and then I will run it uh it will try the best uh so it will try it best to read the doc and then get information uh but as you mentioned often it fail and when I observe that fail I would personally write read a doc and just copy paste the most important part in uh but through this process you you normally start

creating some kind of the testing script uh then I can run this and debug this very modular feature uh until I got result I want and after I got result I want then I can copy this thing into the doc and uh here I can say like code example of how to use uh snow uh snow wrap to fetch RIT data so I we pasting this code example that I know is working for my testing so that um cursor don't need to uh cursor were less likely make a mistake later so I basically repeat this

process just a side note for people watching this this a really good practice to kind of test out every small feature or every change you do because if you like if you go you know multiple hours without doing any testing and then something doesn't work it's like a nightmare to figure out what is causing that issue so like the more testing you can do in general the better yeah exactly so basically I will repeat this process a few times until I get the very modular Cod snip working and when I'm doing this the benefits is

that I actually don't need to worry about okay how whether how they can interact with other files we didn't mess up the things like you mentioned uh which will very very likely happen if you just go code directly um so I'll do this and then I will also kind of putting the uh file structure as well so normally this is like after I set up the project I were because I found often C of since it didn't have an understanding about the file structure which very surprising because I think they should be able to know

that pretty easily um I think that's a feature you should definitely at um but it also seems like it knows right because when you type the pr and you like control enter it it looks like the graphic kind of tells you like it knows the structure right but sometimes it feels like it doesn't yeah exactly it definitely feel it doesn't um that's why at moment I normally what I do is that I will use a this is this one package called tree so tree will be able to generate the let's say the file structure um

for you so if you are in node as project so I would do tree uh- L2 which means it will go to label deep mhm and then Dash I to ignore certain like folders file I just don't care yeah uh and then this will kind of show you the file structure this is a bit like Wick but normally what it would generate is like a actually proper project structure um for you so I can copy paste it over uh as part of the kind of file structure here yeah I mean it just becomes more valuable

your project grows right it's like with small projects if you have like five seven files it's not another a big deal but once you have like backend front end folders and like folders within that it it can become confusing yeah exactly exactly uh so I would normally like let's say uh here this one probably bad but uh let me see if I can find a better one um so normally for project structure you it will look something like uh something like this after you run it give you like very s detailed like project structure so

that it understand the dependency much better in this way um and the in the end I will include some kind of additional requirements so additional requirements is like based on the type of project you're creating I found it often can make similar mistakes uh and uh this is one that I so I have a collection of list of prompt I adding here for different P projects like if it's IOS app it has like list of things I want to add if it's web app then this is also list of things I want to add as

well so if this web app I would just paste in the things that I have here like things like where do you put the uh components uh where which which nextjs version we should use uh where should create new page things like that um so I were so this normally like the first draft uh but I often found for my experience if I just give it directly to the uh cursor it's not that great as well so let's that's where the 01 comes in in my workflow so I will copy paste this uh let's say

the basic one and then I will go to like 01 version um let me try to find the ones that I used last time we still need to see how good the actual full o1 is because o1 preview is not even the full version yeah exactly um but normally will go here uh pasting that in and then ask a few things like firstly I will ask you to figure out the file structure because I I basically say like this is the this is the feature I needed and this is a current file structure help me

like design architect how this uh how different files should be uh exist in my in my purchase based on the feature request then it will kind of generate the uh the file structure and after it get file structure then I ask you to okay now great now I'll try to generate very detailed product requirement doc that any engineer can just take look and then start picking up with without any ambiguities uh so with that I found 01 is really really good um so sorry yeah should probably switch to 01 uh preview and later the the

401 so with that one this is normally how I get like the the final instruction like this which is very very detailed and very clear but don't you find like sometimes o1 inser just unnecessary details and that complicate the project a lot uh yes so sometime it would do uh that's why I were uh when I asked it to generate file structure I would put in some specific instruction like uh try to create as less Pages as possible um so that it has uh it has a bit more kind of the um it it's less

likely to make arrrow on those file structures yeah that's good PRT yeah uh but yeah this this is kind of my my kind of current wlow you can see this huge amount up running up font planning uh but that's kind of a current hack of getting getting through those different arrows and limitations of lar model yeah just one caveat to that I would put for beginners watching this you actually want to make sure you don't you know get stuck in the planning phase cuz like you know you might spend like you Jason are compared to

average viewer you're like a genius programmer right so you might spend a few hours building that but you going to get to the building phase a lot of people are like okay I still need to do more research I still need to watch more videos I still need to ask perplexity and then they're like oh perplexity gave me a different answer than CH GPT like you know which framework should I use so one B caveat I had for people watching this don't get stuck in the preparation phase forever right don't skip it for sure do

some prep you know you need some clarity about what you're building but get to the building part because if you don't you're never going to make it yeah totally like you should build like this person looks long but part of that is actually I want to get build as fast as possible because I were ask to build small building blocks so I know especially P part that is more complicated you want to test to understand does this part actually work cuz other otherwise the rest actually didn't even matter um so yeah totally agree so I

want to ask a selfish question would you recommend me building fully the back end like the AI functionality first and then try to connect the front end I mean I already have like a basic design with vzer but I don't have it connected with the back end or would you recommend me build both at the same time yeah so my my personal worklow is actually I try to use cursor to build build a whole functionality first okay and then I were I basically so you know when you go to cursor it will actually help you

build a whole functionality P whole page um even the page looks not as good but I just don't care at the beginning I just don't want to build whole functionalities and later because if you're building web app it basically will create a bunch of page for you right this time actually I start using v0 so I will just copy paste Page by Page and ask okay now try to make this page look up a lot better uh and keep the style the same so that's kind of my workflow because I I basic just want to

separate cursor just focus on functionalities because it is really good at it and then v0 do the front end um but if I do the other way WR which I know actually is pretty popular kind of uh workflow uh my challenge is that is for building more complicated purest I found is actually make it a bit difficult because uh in VZ I believe it only create like one single page um yeah but often you would you would need the whole project yeah yeah that's what I found like I started with v0 then it was like

600 lines of code including some of the functionality I was like had to slowly remove it put it into the back end and like you know update stuff it was like yeah maybe it's good practice to stick to like less than 200 lines and then probably probably your your approach is definitely better just have cursor build everything forget about like looking nice just focus on the actual functionalities that it works and once you're happy with how the app works then you can start polishing the UI and you know making it easier to use so I'm

definitely going to do that because right now I have like a solid UI and a solid back end and a complete disconnect I'm wondering like should I because I want to keep building the agent right I want to keep building the python back end that's what I like to do but then still I need to connect the front end otherwise it's not an app so I'm like wondering whether I should do it and like I think you just answered it so I appreciate that yeah nice so any final advice for people who are new to

AI or new to AI agents I think my probably specific for agents one is that I found is sometimes it's easy to get into the feeling of like say shiny to uh kind of experience you you saw a new framework launch every day and you want to like okay which one I should use uh but what I found is really useful is actually just just build a agent don't use any framework from scratch yes and then you actually understand it's actually not that scary all those framework they pack things together in all sorts kind of

uh planning that they think is useful but maybe not so just build a extremely basic function calling agent by yourself by just calling open AI API D you will get so much better understanding about all those new Frameworks and which one will actually work for use case so that's probably my my kind of one suggestion um for people who are building agents man I couldn't agree more like I started with Frameworks and now I I I never use them I just do direct API calls to open out anthropic it's it just gives you complete control

and you know I think you can never go back but anyways Jason I appreciate taking you the time I think it was super great episode I've learned a lot and I'm sure the viewers have as well I'm going to link your channel and your community below the video so people can check it out and yeah thank you for taking the time man awesome thanks a lot have a great day man bye have a good day byebye