AI News: Software Development is Changing | EP01

[Music] with AI comes great power great responsibility but also a ton of bad content and today I want to talk about how I think that all product teams everything that you've ever known about shipping software is fundamentally changing I'm a practitioner of this I've been building software as a product designer a product manager and an engineer for about a decade and got really into all of the llm stuff when GPT 35 came out and April two years ago and I want to walk through all of my thoughts on how AI will build the first single person billion dollar company and we'll go backwards a little bit so for those folks that don't have all the context of what this means this will be really helpful and then for those Advanced Engineers Advanced developers that are always for the next you know Clyde or Claud Dev or AER or new model this will be good for you too just stick around so as I mentioned I think there's just so much information out there and these newsjacking videos where a new thing comes out you got to test it it's actually not that helpful because if you test every single one of them you're just wasting time so having a strong filter is the best way possible I like to let a little bit of signal come through through the noise and see some people actually using it and then build a little bit of conviction before I jump in and start using a new tool my current stack is I'm using cursor so this is the goto and I'm using the composer and I'm using a bunch of different prompts and different keyboard expanders and those are actually built into the settings on Mac you can get these on Windows as well but it's just using cursor and cla's newest model and that's for actual code execution and then I'm using a Chain of Thought model so o1 preview for planning I find that to be pretty helpful it will go through and think about the things and I think that that's really awesome too on some of the bits and pieces of the actual text expander so what you do in in Mac is you just type in keyboard and you go in and under keyboard you will see text Replacements and then you can see I have a bunch in here I have analyze architect ask bug comments copy critique don't Eli 5 explain L comp 5 and this is something that I was trying out where you can have a start and an end prompt for reloading the context basically destroying context of the previous window because as you know these models they have Windows and it's similar to the days of bandwidth when you didn't have enough bandwidth getting to your iPhone or your Android so you could stream a video context is having that same sort of bottleneck right now now and that's what this end workflow does explain feature graph next this one's language specific plan tree so I also have some bash things in here some some shell commands where it'll generate a tree types YouTube this one's actually just for ripping YouTube audio so that's really helpful uh think step by step and then shrug so all of these if you want I will put them in the description below let me actually write that down cuz sometimes I I just say these things and I don't actually do them so just for expanders and that's the current flow for the expanders and I also find myself using this toggle floating notes a lot so this is through raycast which is a launcher and this just having this like little notepad that floats around is really helpful for my own context and for reloading my cash because you sit down you get deep into work and then you come back and you want to actually write where you left off so when I left off at 2: in the morning last night I thought of this feature where there was a dynamic Tab and so okay then I have these different directories so a has a bunch of different things in it and it has a bunch of different team members in it so a number of years ago when I was working at a startup that my friend started we had a full staff we had QA engineering team that was offshore actually I'll start from the top we had a product manager which was the role that I filled I worked with directly with the CEO and we would figure out what is it that we need to build for the business then I would spend time with the customers we were a food delivery company so I'd go spend time with the drivers I would go and spend time with the customers and I would go and spend time with the restaurants and that was the role the product managers to figure out what is this problem that people are facing are they willing to pay money for it is there a theme to this problem that we're trying to solve is this a revenue generating product is this a life Improvement for a quality of life Improvement uh theme product is this a DX Improvement for developers there's different themes you could have but that was the role and then after that you know it kind of goes in sequence where every team kind of works like this where the product manager is the one that's responsible for getting the team to row in the same direction you know Steve Jobs famously was amazing at getting the team first at Pixar and then later Apple to get developers and designers to work harmoniously sometimes there's a battle there so that was the role of the product manager and as you can see I'm trying to emulate parts of that into an actual team member so what were the things that I did for seven years that was really helpful and can I distill That Into You know a 94 line prompt probably not but it can do some of it and I think that this is where it's going this is kind of what I'm getting at is that I don't think that these side windows which while they are helpful and even the voice assistance where you're unloading and re loing context and you should check out Indie Dev Dan on YouTube he is one of the biggest innovators in the space in terms of pushing forward where agents are going but I think that this is where it's going where you'll have each one of these roles so you know as I had explained first You' have that product manager and then typically I would go and I would I would sit with our engineer who Architects the system we're not writing any code he's not writing any code I'm not designing anything talking through this one pager that I've written which is the pitch which I would go through what is that problem that we're trying to solve I would figure out what our appetite for the problem was how much are we willing to spend on time because these are bets any sort of product work is a bet and we'd figure out from a high level what are the technical decisions we need to make and that's what you see with this the software architect and I've taken pieces from from The Prompt that you've seen with ader if you've seen my videos on that where there's different types of modes where you can prompt the coding agent an architect is one that does not write complete code it's not just spitting out a bunch of uh generative code it's thinking through how is this going to work in the system what are the tradeoffs do we need to reach for a library it's an examination of the current code base and all these try to emulate that and so once that was done in real life I would then go and we'd sit with the uh the whole team and we draw a diagram we take that you know first step was the pitch the one pager that I worked with the CEO on the second p uh the Second Step was the technical architecture of it how is this going to work then we get the whole team together make sure hey does this look good Draw Something on the board make sure we're not missing anything bring in the mobile team and so that would be the software architect mermaid expert and this is going and it's building out let me see if I have any in here looks like I deleted them but it would actually build out a mermaid diagram and mermaid is just what you know it's a diagram so builds out that flowchart after that someone mainly me because I was product manager and product designer would go in and take all those artifacts all those pieces that you've worked so hard and thought about about this feature you know you've have the customer stories you have the one pager you have this this uh conviction from the architect and the the diagram and then you're actually breaking those down into tickets and that's important because if you don't have those it's hard for the team to know where they're at if they're tracking so that's what this one does and then we get into step five which is that actual doer so this one would be language specific in my case you know you a fullsack engineer task with impl a new feature I'm you know teaching it hey we need to use react 19 now where this goes off the rails and where I think this will get better is this doesn't have context of documentation that I want it can like I can take this and put it into composer and say hey we have these tickets that are generated we've already generated the task take a look at those or sorry uh we already have the plan we already have the mermaid we have you know we have these three different things right we've already made the one pager we've already made the technical architecture decisions we've already made the diagram now we need you to take examples of what good tasks look like and make those and it can do a pretty good job it can put them in here but it's not as specific because it doesn't have as much context so I think this is where it gets better where if this was not in this medium of just a markdown file that you can get a lot more out of it I think this is where these tools will go uh and then after that full stack engineer then you need someone that is specific to the Polish because in real life you know you have people that are really good at certain areas they can go full stack but because they're really good at full stack that usually means that they're not fantastic at one area it means that they might not be as good at knowing the libraries or the CSS tricks and how use key frames and all these things so then once you have the thing working it's okay let's make this really cool let's get the design engineer involved and that's in real life what would happen is I would sit as a designer i' put the designer hat on and I'd go sit with the engineer and we'd take the non-real um you know figma designs and we turn them into crude prototype you know we call it steel threading where it's just get the core usage down that solves for that problem but then we need to make it better so then that would happen then we had a QA team sometimes that was internally sometimes that was externally they'd go through they'd test things then I test engineer and then I have these these last two that I've recently added which really this is about just making it so not only you as the single person on the team or the small team can quickly Traverse files and get that context going like I have with this floating note for the things that I want to do next it's adding TS docs to the different files so you would then say hey I want you to go through this is what we're doing here's the files go and Mark these things up and so it adds in the pams if you're familiar with JS docs it's that but for typescript and it seems redundant but it it makes it so it's easier to just get up to speed on what's going on for any new developer for any llm to be able to read that it's worth the tokens um something worth maybe mentioning is also at the bottom here you can see I have token count that's helpful because I can see across these 10 different team members it's about 5,000 tokens and so if you take that and then you already provideed examples provide documentation you know that documentation that gets scraped that goes into the context window and even as get larger there's the needle and the hyack problem where some things will get lost and they say that there's tricks where you put the most important parts at the beginning or at the end and it's model uh specific but that's sort of where I think this is is headed and now I want to talk a little bit about just this idea of the billion dooll single person company and I put some notes together but it's you know you have a list of these these billion dooll companies with small teams famously Instagram was one of the ones that was really small they think in terms of Revenue per employee but I think that that's going to get totally changed and I think that yes you know people do need to act now I think that if you're using a computer chances are that every single thing you're doing throughout your day can be improved I think of that all the time you know as I started writing those prompts it made sense where this is where it's going um if you go to Claude this is another tip if you go to Claud and you go to the developer console you can actually have them generate prompts so that's where all these came from where I have different ones that are not part of the team member they're not part of the text expander but they're to solve for a specific task so if I want to analyze a directory and fix lenting errors there's one for that and if I want to explain a file there's one for that if I want to generate Pixel Perfect clones there's one for that if I want to generate the TS docs you saw that in the team member if I want to extract reusable components there's one for that uh if I want an architect to generate the tickets there's one for that but it's really just about iterating on them and taking this and then putting it into CLA and letting it plus it up so I might take this one use it for a few days realize where it's falling short and then put in the prompt generator hey this is the current prompt here's the problems that we're facing with it plus it up and it does a great job with that and so if I go back to this article oh and I guess you know since I opened this up these are the different agents that I'm thinking about the different team members and we have ones that are very specific to languages I have a lot of uh postgress functions where you know it took me six months to get one of them right that sounds wild but it's because I hadn't written any of them before and and I am constantly adding things to what it needs to do so had I taken the time to build out the agent give it the context of what goal is trying to be accomplished give it examples because I could search GitHub and find versions where you know in this case this one would go out and it would capture all it would ingest every single calendar every single calendar event all of your Access Control list mechanisms and settings all of your free busy settings all of your uh I already mentioned events and then settings for Google Calendar it would go through all six of the methods that are provided or resource groups provided by Google and then Loop through them iteratively batch them paginate them make sure you have the tokens and bring all those into map my product in about 5 seconds so again had I had this specific one that would have been really great um and let's see now back to this right up if I can just find it so when I look across the day and I see things that I'm doing a lot of different times I try to figure out you know how am I spending my time and how can I make it better with AI and if you're looking at your schedule you basically have two things you have the inputs or the actions that you take to accomplish a goal in your day and then you have the outputs or the things that you'd like to get done you can see this in examples of some of the best performers where they set big goals and then they audit I know Keith R boy he's uh famous investor and he's made a bunch of uh unicorn companies happen with his help he always audits people's calendars and I think that that's also where AI is going and that's something that I'm focused on map is figuring out how people are spending their time and then generating those tasks not just development specific ones but actually generating the tasks that are actionable items you could do today to make progress towards your goals I think that even goal setting will be part of map and that'll be a big part of it and so in one example you know if you want to be a great writer it's really similar to logging hours when you want to become a pilot you have to log a lot of those hours and the common Trope it's been said for years is that 10,000 hours is the number to get really good or get competent as a at a skill but I think that that gets changed I think if you did 10,000 hours but each hour is more levered meaning you more effective per hour per unit of work then you're going to be twice as good three times as good four times as good and so you know dhh from base camp says it's fun to be competent it totally is and spamming an AI to give it what you want is not the way to get there it's double-edged sword it can be both the best thing that helps you and the thing that atrophies the muscles that you're trying to get so this is a chart that I had an AI pull together that kind of explains you know how do you get to 10,000 hours how do you get good at something quickly in this case development well if you do 100 hour weeks it's going to take about 2 years if you do 80 hour weeks 2.4 years but I think that this gets flipped on its head I think it challenges the parable you know I pulled together a bunch of examples of folks that didn't traditionally have the skill set they didn't go to college for certain things like Elon Musk is you know famously had never really thought about launching Rockets now he has them Landing in Chopsticks and it's because he just outworked everybody in the room and I think that had he had AI he could have taken the books that he was reading and thrown them into notebook LM and listen to it while he was on a walk or he could have trained a bunch of agents to teach him on the specifics of a manual of some NASA guide so then you know we have Edison who held over a thousand patents you can read about his work ethic you have Fineman Howard Schultz Serena the tennis player all these examples you know you just it in order to do stuff that is great and impact a lot of people you have to throw your hours at it and so I think that this this Mentor student relationship also changes a lot I think that you know as AI models reach a PhD level on many different subjects they can serve as the mentors and I get into that a little bit more I'll post this full article on my website but I think that by using the best an models and knowing how to apply them L effectively we create a pathway that resembles the mentor student Dynamic that we're in a point in history we knowing how to use the tools effectively gives a distinct advantage over those who do not this learning method is like having experienced guide who shows us where pitfalls and opportunities lie and benchmarking has always been about identifying excellence and striving towards it which is really similar to that relationship so you have examples you know this is not a new thing like ancient Greek philosophers from from Socrates mentoring Plato who taught Aristotle who then in turn tutor Alexander the Great this chain of mentorship laid the groundwork of for Western philosophy ethics and scientific thought and showed how knowledge can grow and shape entire cultures so forget making A1 billion doll company out of a photo sharing app think about the ability and the effectiveness of each person that's out there when they properly leverage these tools you know another example of this is the medieval Guild appren apprenticeships so you have the knight in the page you know the guild system passed down Knowledge from Master to Apprentice if you ever played Guild Wars I think I logged like 10,000 hours on that it was you know you're part of a guild and when you join the game if you got to level up in that game with someone that had already done all the missions before and knew where the bosses were and the enemies and what spells to cast you just get there faster and that's because that person had had the experience they read the manual so to speak they were trained they had the context and so that's the power you know I can get into more examples you know you have the Renaissance art Masters Samurai training the Scientific Revolution with Newton both the yoga and the Buddhist traditions and then and startups you know you probably care about startups because you're watching a tech video but Steve Jobs mentor Zuck a lot of people don't know that but he accredits a lot of his product thinking and wisdom to that and so those are some examples of the mentor student relationship and that's what I'm saying here is you've never been a product manager before well guess what and not so long even today with some of the markdown files that I'll share below you can get up to speed on what is a good product manager what are the outputs that they do what are the inputs that they do if you you know you can't totally substitute the work right but you can get up to speed much quicker you can make a podcast out of the top 10 product management books you can make an assistant on the assistance platform that asks you questions that come from cracking the product interview if you want to get a job there or with that and so and so I think that there's a lot of hate on the insurance of AI but it's actually been around for a really long time it's just that GPT is now able to generate texts now we're generating images we're also generating formulas all those things this is not totally new just so you guys know ai's been around since the 50s you know famously the touring test is can you tell if it's a human or a computer but here's a brief timeline of it and a little bit just on the end I want to give some more of the tactics as well as some predictions on where I think this is going which you should probably know by the the tone of the beginning of the video but little things like warp so warp is a terminal a lot of people don't like it because it requires you to log in but who cares it's helpful so if I just start typing and I say generate a bash script that you'll notice that there's this little icon that's going to then take what I'm about to type and then turn it into the command that I need to run I think that's really helpful you can also attach context so if there's a previous block that you need to use in the context of this chat like let's say I made a bash script on my own or made some script on my own and it didn't work then I could hit up and then type in and it would fix it that's really helpful I think the workflows are really helpful too oops so if I hit command P when I'm in here and I click on workflows and you can see these different text expanders where it's you know okay for a YouTube video that I want to pull the video it will CD into the proper thing the proper directory it will pull it it will pass these arguments where then I'm just pasting the YouTube link this is less AI related but still helpful if I want the audio version of it I can do that if I want to actually run a project I can have it CD in there start the database and then do Bund Dev uh The Voice Assistant python demo just check out indd Dev Dan in terms of CLI tools this is another one that I find pretty helpful this one's called repo pack I made a dedicated video on this but I've since changed the way that I use it so I have this config and what it's doing is it's saying hey here's the file path of where we want to take every single file or directory that you include in the include array and put it in this place put it in there and I use XML because Cloud's the best at coding hands down and they prefer XML I have an instruction file path that then can be uh concatenated on top of the output so it has the conventions so I say here are the conventions here's the structure here's how it works here's the core user flow I haven't filled that out here's where the types are here's how we do Au here's where the database is this is basically that context that it can know about the project it's kind of like a read me and then I can put in different texts I can have it remove comments if I want to remove empty lines this only works in certain languages but if you're using any popular language it works you have the include array like I mentioned the ignore and security check looks environment variables so that's repo pack and then all I would do is I just type in repo pack uh the right click to add this is another one that's pretty helpful if you're using dedicated chat meaning like it's in a browser and you're not using one of these then you want to add this extension so you do command shift X and then you type in there's the token count one but there's like a copy paste thing or copy file copy text of selected files it just is what it sounds like so if I right click something and I say copy content of selected files it will grab this one I can continue stacking those so then I can have you know six files 10 files 20 files in there and it has the tick marks around it with the file names on top that one's really helpful mermaid extension this is another one for when you're generating outputs that have mermaid uh diagrams in it you're going to want to have that extension so you can view them in markdown and then I mentioned these different commands so the predictions on is the future's bright I think that we're in the place where the models are just bottleneck by context but that's going to go away and I think that if you are constantly staying on the edge of where these things are going and you're not just blindly following what people are putting out there like videos like this I think if you're actually seeing the video and then testing it and then seeing what you could do to make it better that you'll then be the edge and that's where I think it's going if you found this helpful make sure you like the video these take a lot of effort and that's your way of doing a good deed for the day also subscribe to the channel I'll have more videos every single week thanks for watching byebye [Music]