Beyond Vibe Coding: How to Scale AI-Assisted Development Without Architectural Chaos

Again my name is Mushek Yorgan and um today we are going to talk hopefully um very interesting topic which is VIP coding or uh I more prefer to say AI coding. Uh so for me the difference w coding and when you don't really understand what's what it really generates and AI coding when you are an engineer and AI is basically generating the code for you but you still understand what it does right and um obviously um it's not a secret that with current tools that are available in the market like cloud code cursor windsurf codex from openai They make us engineers like 10 times more productive. You basically ask what you want. AI creates the code. It works. It feels like a magic. But what if it's not only you that are using the AI to generate the code, but your whole team is using it. So this is when a lot of problems are arising and basically this is u my speech is about how to solve this problem but what let's start from the problem first what's the problem so imagine there are like multiple engineers in the team and everyone is using AI obviously they are using some kind of inter isolation uh there's engineer and AI and they're chatting together and it might be possible that one engineer might create a table another one might create another table which duplicates some data The third engineer might create some service while one of the existing services might work. So basically each of these chains separately makes sense but together that actually can create a mess right but engineers make mistakes developers I mean we can make mistakes we can write bad code that's normal and that's why we had code reviews right so when you write some code you send it for review and your peers review your code give your feedback and this is how we have been preventing ing let's say bad code or we we have been preventing messy code when different engineers creating the same stuff in a different way but now in this new world when AI generates in a single PR sometimes 500,000 or 2,000 lines of code obviously like the old way of doing code review doesn't work anymore like as a human you can't really review the AI generated code I mean you can but probably you will miss a lot of things there you will be you will not you will not be able to review all the code and also I think when you review AI generated code to be honest in my opinion it's already too late because based on my experience especially when I'm reviewing junior or mid developers AI generated code what usually my feedback is like hey can you just rewrite this all please because it doesn't make sense at all because this is what AI it's true not the engineer sometimes So and also if if you if you catch these issues, I think it's already too late because now uh the engineer should go and start everything from scratch. So basically the old style of doing code reviews doesn't work anymore in this new world. And now let's multiply this this problems over five engineers across multiple screens. And let's say you don't have a proper solution for this problem. uh what you'll end up with is uh with codebase that is not maintainable and I was in this situation to be honest I was managing a team five engineers were shi like it was a startup and obviously we're shipping very fast like very very fast but after few weeks we realized that code is unmaintainable we are getting bugs and fixing the bugs becoming like kind of refactoring take taking sometimes days sometimes weeks of rework. So we noticed that we are getting duplicate tables in our database schema. We basically were not using any patterns. So basically code was garbage. Maybe for startups it's okay in the beginning because eventually yeah if you gain customers then you will be able somehow to um refactor or sometimes even start from scratch. But in general, I think it's very dangerous, right? You don't want to base your business on a codebase that is messy. It's it's not maintainable and you don't know what might happen in the production. So that's why basically why I came up with the idea of spec. It's an open source project. Uh and I credit it first for us and we have been using it um so far. Then the idea of spec is very basic, nothing complicated to be honest. So the it's it's a three-step process. So first is analyze analyze your codebase, then design a feature, then implement a feature. We'll actually review each of these steps today and I will also try to do real demo with real AI coding and hopefully it will go good. Um but yeah, let's let's review each step separately. So first step is analyze. So let's say you want to you want to start doing AI coding on an existing project not a new project and this is usually what happens in big companies or big teams or not new startups right you already have some code base that's there and you have a team and now you want to accelerate your development with AI so if you start using specind the first thing you would need to do is to run the analyze command and by the way this um um spec works in all existing uh coding assistance. So you don't need really to switch or you know uh you don't need to change your processes. You can still keep using cloud code cursor codex windsurf github copilot whatever. So it lives inside of this assistance. We'll review it later in more details. But in a higher level what you need to do first you need to run analyze command. So what analyze command will do it will basically analyze all your current existing codebase. it will parse it with three and whoever um doesn't know the tree is basically understand your code structure and not just text uh and and then spec also understands your service layers your patterns your databases your dependencies frameworks or you use and eventually based on all of this it creates a comprehensive documentation it's called systems which basically tells what your project is about and how it's implemented and covering all of this. Uh this is the first step and let's try run analyze command on one of um projects I actually vcoded by myself. Uh yeah so first thing you need to do basically you need to install um specind into your environment and as you can see we support different um coding assistance. I'm going to use cloud code for this demo purposes and I'm going to run spec setup command on my project and hopefully you can see it. Uh if not I will try to kind of uh go through the generated code. So I run this and what what this uh yeah first I set up the code. What it does basically it copies the commands which are basically the prompts into your codebase and also it installs a CLI that can be later be used. So now if you go to our clut folder here we'll notice we got three commands analyze design and implement review let's review analyze for now. So what analyze basically is it's a prompt for cloud code which says you know what whenever developer runs this command first go and run this MP spec mind analyze C cla which will generate for you you know what let's run it and see what it does so I'm going to run analyze command yeah because I just installed I guess I need to restart my cloud code. Let's try again. Yeah, here we go. Now it's available. So I'm going to run analyze command. And here hopefully you can see that first thing it will do, it will run the CLI, the analyze. So now this CLI will go over your codebase and will start generating JSON files. What you see it automatically detected services in my codebase. I didn't tell anything right but it went to my codebase and actually yes I have like a couple of services like I have admin service agent crawler. By the way this project is another AI agent uh for car buyers. Whenever you want to buy or lease a car, you you chat with this AI agent and it's connected to different dealers and it automatically makes suggestions for you and um schedules appointments. That's what AI agent does. But it doesn't matter in this context. So it automatically detected these services for example admin agent crawler DB is like module more not a service and web and then under each um under each uh service it also detected the layers. So it always tries to find any API layer any data layer external layer and service layer and under each layer it will generate chunk files. So this is basically uh kind of this is the high level explanation of your service. What methods does it have? What classes you know like everything whatever it could uh uh kind of get from your codebase and then the coding assistant which is in this case cloud code as you see it's it's running five different agents in the background. So now they are reading these chunk files. So instead of going and reading the all the uh all your codebase which sometimes might be so big that it will not fit in the context of AI right even now we have like 1 million uh tokens context but still it will not be able to fit in one um in one uh kind of context. So that's why first what we do we just like because we don't care about like implementation details yet we we um now more care about the architecture. So basically with the this chunk files we get a high level architecture of our project and then we give this to AI and say hey this is our project please create a comprehensive documentation. So I think now it's working on the documentation as uh itself. Let's go back to presentation and we'll come here uh again. So again let's summarize what analyze does right. So first it analyze your codebase. It generates descriptive JSON chunks that will fit in your AI context. By the way, even if your codebase is very big, this will still work. Um then it's basically runs a prompts against your chunk files and generate the system SM file. And what you will have in this system SM file? Basically you will have this high level system architecture for your project. Then you will have per service architecture if you have multiple services. By the way, now we are talking about monorreper project, right? Imagining that you have multiple services in one repo. Also, it will automatically generate sequence diagrams showing the cross service flow between your services. And event and last but not least, it will generate the entity relationship diagram. Basically, your database schema or your database schemas if you have multiple ones. And now this system SM file will uh kind of play and be a source of truth for us. Whenever we want to understand what our project is, we need to uh check the system SM file and also the AI by itself can go and read this SM file when it needs. Yeah, let's see if if it's done. It's still working. Sometimes you know like it it takes like few minutes. Sometimes it uh might take more but uh you know what let's continue presentation and we'll be back I guess in a minute or so. I see like it already generated 16,000 tokens. So probably it's almost done. It might take a little bit more. Let's go back to presentation. We will come here. So the next command let's say we have system smile right it means okay we have the base now we are ready to start implementing new features. So if you remember in the presentation I mentioned like before AI how we how we were handling the implementation right we're writing a code and then we are sending the code by itself for review what I propose I propose that wait before starting implementing the code let's design the feature but and here by saying design I mean design in terms of architecture and technical details let's create a specific ification file for our feature and only then we'll start implementing it. So basically to do it you just run /design command you tell your um your feature description feature description it can be as short as like I don't know like realtime notifications or it can be very comprehensive description about the feature and then it will start generating the um specification just for your feature. I want to show you this. So, okay. Cool. I think we got the uh systems SM file. Let's review it. Cool, cool, cool. Yes. So, now uh so as you see it generates spec folder in your project which obviously you need to commit and push to g. And here you will have the system folder which is basically the JSON representation of your codebase. You basically don't need to come here. It's it's just like for first analyze step. And then we got this system SM file. And by the way uh as part of spec I also built an u extension for VS code for vinser for cursor for every IDE. Uh so you can preview this file because basically for AI obviously this is enough it's it's markdown but for human if you want to read and understand what it generated you can basically preview it. And this is what what AI basically generated. So as you can see there's a high level overview what our project is about. Then there is high level architecture of our system. Sometimes this is more complicated. In our case it's pretty simple. So here what we see we see that okay here we have um couple of services. Then we have our data stores which is posgress and radius. Then it automatically detected all the external services we use in our project. For example we use openai for LLM. We use oz for authentication. Lagsmith for tracing our agent to understand what it generates. Uh this link is basically for uh iMessage communication. Google maps API and me zero for agent memory. Yeah. So it automatically detected what uh kind of services we use. Then yeah here's also some um explanation right what's uh which uh about each service like admin service, agent service and so on. the data stores, external integrations, some communication patterns and then oh sorry okay and then uh it generated the sequence diagrams. So basically by sequence diagrams it shows each flow in my application and here we mean like flow from user perspective. For example, this one is crawl job flow. In in my agent, I have a crawl job when when I click and it automatically goes and crawls and scraps the data from dealers, right? And here you can see that I hope you can see that it basically with a sequence diagram describes that. Okay, this is how the flow works. So, first it insert a job in posgress, then it cues a job with radius, and then it it uh triggers the crawl service. basically just explaining how our um agent is implemented. So basically if you onboard a new developer you know what you can even give them this system SM file and say hey go read this system SM file you will learn everything about our project you don't need to read even the code so there are more uh sequence diagrams obviously I hope it also generated the entity relationship diagram here we go so it automatically found our um entities and it automatically detects multiple uh OMS so it automatically detected our entities uh the types the relations and everything and it it just like wrote down and then it generated some more diagrams for each service and some let's say summary right so we have our system sn file and now as you remember we uh wanted to generate a new future now I'm going to ask AI to generate uh for example I want to add a new feature so the the car buyers can add for example I don't know favorite cars like and I'm going to call it like buyer favorite cars. I mean again I can provide very long description about my feature. Let's say I can even provide a PRD or my I can copy from my Jura ticket or whatever we have our task but for demo purpose I think this is enough. So what design uh command will do now? It will go and read your system SM file every time to understand okay what what's our what's our current state right and then if needed it will also go and review your codebase this is this is basically depends on your system SM file and your uh codebase uh by itself and then it will take your um description and it will start generate specification file just for that feature. Okay. And basically this specification file is um this specification file let's go back to our presentation. So now like instead of like sending your um code for this feature you you will send this specification file to your peers and now instead of like reviewing few thousand code that AI generated they are going to review just specification file to make sure that whatever is written there makes sense right for example yes we need a new field or yes we might need a new service for this feature or yes we might need to add some new relationship in our database and so on and and now human uh can read the specification and review it and just approve which means like okay now we are ready to implement this feature so that's that's the new process and everything still happens in your same code review process so you don't need to like change the code review process in general if you are using for example GitHub for your code reviews and you are using uh PRs pull requests you can still use them right I can generate the spec file. I can open a PR. I can send my to my team to review and then when when uh I got approval, I can start implementing it. So let's go back to our um code and see where we are. Okay, I see that it's actively now generating the feature specification. Okay, now I have everything. I let me create the feature spec specification. Okay, uh let's go back to presentation and we will come back here. So let's say we send the specification. Well, which by the way is very similar to the S system SM. It just um it just u only about the the feature specification, not the whole project. And by the way, I forgot to mention one thing to be honest. I'm very sorry. U if you have any questions during the presentation, I should mention this in the beginning. You can scan this QR code uh and submit your questions. We'll re Oh, sorry. We will review at the end or if you want to ask at the end. Yeah, you can do that as well. By the way, I vcoded my presentation as well. Yeah, I even I I even vive coded something more and I forgot to do that. I vipcoded a speaker notes and also I someone from my colleagues asked to record this and I forgot to start recording but I even implemented recorder. So, I'm going to turn it on now. At least we'll have some part of the presentation. Cool. And it will record and automatically put to cloud. Yeah. Uh I forgot to do that in the beginning, but that's okay. Okay. Now, let's go back where we were. Yeah. Implement. So, um hopefully our uh feature specification is ready. Yeah, it is. It's just writing now to the disk. And now for each feature you'll have first you will have this features folder here and under each you'll have your like feature slack which automatically obviously generated by AI and then you'll have your each feature specification here in the another SM file and we can preview it as well. I'm not going into much details but I what I want to highlight it's first it generated requirements right okay so what are the requirements functional and technical requirements for my feature so it just like just thinks in this way what what I need to do in order to build this feature and what is very important it highlights the services it's going to touch it says you know what I need to change the agent service and I need to change the database and that's right because it doesn't need to change crawler or the web I didn't ask any UI change or I didn't ask any admin change. So that's why it automatically highlighted that I'm going to change this service and me as a reviewer I'm going to review this and say yeah that makes sense that's cool and if if I want to get more details okay what are you going to modify I can read for example it's saying you know what I'm going to create a new um inm value and I'm going to create a new tool which is correct I'm going to create a new manage favorites tool for LLM I'm going to create a new server service for API endpoint and I'm going oh look and it says no new source existing buyer vehicles interest table used with new source value so it turned out that I have some table and I think that oh you know what instead of like creating new one I can reuse it and probably it's also highlight in the ER database uh ER diagram yeah here it's saying you know what I'm going just to change this table and I'm going to add this new field and me as a reviewer I'm looking if if I for example notice that it's creating a completely new table I must say hey wait but why you are doing that right let's say it's a new engineer join our team why you are creating a new table actually we had a we have a table that you can use so in this way I can review the specification and just give my kind of uh approval and then we can say okay me as engineer who who wanted to implement this feature I'm ready to implement it and that's the last step which is implement now I'm just going and telling AI hey this feature specification is okay by the way uh if I want to change anything here, I can do it manually. I mean it's not only AI. I can go and change here or even even I can chat to AI and say you know what uh no I want to create a new table for example let's say for some reason which is in this context is not correct but it's not that we are fully kind of uh automating the process it's I mean or delegating to AI we still can inter we can change whatever uh AI thinks but in general the idea is that I'm working on my spec file, not code. I'm talking to AI. I'm changing a lot of stuff, but I'm not changing the code yet. I'm changing the feature specification. That's it. And then when I'm ready, I'm uh running implement. And yeah, what implement does basically it detects your feature specification file. It just basically starts implementing based on it, right? Automatic and then automatically updates the system SM file. So your architecture is always up to date which is very important right because sometimes we create documents but we as human are lazy or engineers are late and we never go and uh update them and and if you're moving like super fast with AI I mean probably you will not be able to keep it up to date. So now Specman will do it for you and also it creates a change log a separate file when it just like writes every change you did. So if you want to understand how you ended up with this, you can read the change log. Now what is very important when you have a feature specification file and as you know with AI the cost of implementation now is low right basically you just ask AI to do it and they do it that's it now you can give the your feature specification to five different AI assistants you can ask cloud code to implement it then you can ask codeex to implement it and you can like look and choose any version that's the good thing about uh specification and by the way this specification thing is not something I came out with it's called like specdriven development. It's it's pretty new thing which means like let's first work on our spec and then think about specification and even if I implement it and I don't like my implementation I can just trash it because it's AI generated. I didn't spec spend hours on it, right? I can just like delete it completely and and start again because I have all my um engineering results in my specification file. And that's why we engineers are being paid, right? We are not being paid just for coding. We are being paid for engineering things, designing things, thinking and you know like designing a proper solution that can work. So with with uh future specification, you can just like give it to multiple assistants which sometimes I do. And they take the version you like most. Uh and finally let's just Okay, it's asking to change uh changes to my specification file. I will say no. You know what? Because I know that my request was not a good one. I will just now ask implement implement buyer favorite car. And I mean I can do typos here. it will it will probably detect uh which feature I'm referring to and it will start okay look what it's saying let me read the feature spec and understand what needs to be implemented and this will take more than other steps so let's go back to our presentation yeah so uh basically with specime we create this loop right so first we analyze which is one time operation we don't need to analyze our codebase every time then we design a feature, we send it to team team review, we implement it, the documentation is up to date and then we go to design, review, implement, update. That's the loop actually. Yeah, I mean test is is part of implementation. Yeah, I I I missed that. That's a good collab. Probably I need to add. Yeah, but basically um uh everything which is related to um testing, deploying and you know like running your unit test, running your integration test uh all of that part in this context is part of implementation. Uh yeah so as I mentioned in the beginning so spec is in already integrated with cloud code winds cursor uh I'm working on codex um soon um because I'm using three sitter uh basically adding a new language is not very very hard thing but it still requires some time so if you are interested to collaborate this is a good area you can collaborate Uh basically I only have Typescript, JavaScript, Python and C# um uh integration. So yeah, let's kind of understand what we get with Specmind. So before Specmind, the problem was that each developer was working isolated with with with their AI and which basically creates architecture drifts and usually you catch problems very late and which is costly which I faced. Um but with with spec mine basically you have this uh the AI has the full context. Uh you basically design before coding and uh architecture stays kind of in sync and consistent and docs are automatically up to date. So you kind of keep the speed but you don't um end up with chaos. So obviously as I mentioned it's a open source project. If you are interested um please scan this QR code. Uh if you like it please give me a star or fork it or whatever or you can even if uh you can try if you see issues you can open issues or or even open PRs as I said there are areas for contribution um and yeah now let's go to questions but before that so let's see what it's doing here. Yeah. So basically now it's implementing uh the feature based on our specification. Let me close this. Okay. So now AI has a clear picture what it needs to do. Cool. It started implementing it. So we can review this at very very end. Uh but if you don't mind I would like to do a selfie with you guys if you don't mind. Uh I'm so excited that so many people joined my presentation. So I want to fix this moment. Thank you. And now let's go to questions if you have any. Oh, we have questions here. Cool. I was not expecting this. So let's read them. Uh how does this compare to openspec? Um yeah to be honest uh I think it's one just um um I think it's just another implementation of specdriven development. Uh and there's another cool project by the way from GitHub. I think it's called spec kit. So you can check that as well. I'm thinking to maybe integrate also to to uh to this um kind of projects. Uh but uh I probably don't have deep understanding of the openspec so maybe I can't give you a good answer. Um now after spec generation does it become part of the system or does it live separately? Could you explain the long-term integration? after specking the generation sorry does it become part of the system file oh okay so yeah basically when you generate the system SM file it's now in your codebase so basically you commit as you do any other file right and you have also g history on top of it it generates change log and I think at the end I will show what the change log is but also you can commit it same about the features so features are in your codebase and they stay there and you commit them and obviously uh why you need to do this because whenever uh new developer joins or even during the work like when developers fetch and pull from repo obviously all of them need to have all of these feature specification files and system file right uh and sometimes you might even get merge conflicts on systems SM files as as with any other file obviously um is this thing on okay if it's about spec mind yes during analyze it creates a bunch of chunk files to form the system SM file I understand that we need system SM for any feature development but can we discard it actually that's a very good question uh yes now these chunks are used once so basically you can discard them I was thinking that maybe I need to you know like periodically or again after each implementation update this chunk file as well because you know what sometimes happens sometimes AI might hallucinate and I don't know forget to update the system SM file it it's happening very rarely but let's say if it happens I always want to have the source of truth which is my chunk file which is generated deterministically right which means like it's it's true there is no hallucination I want still keep it up to date as well bless you uh I was thinking about that but for now at this moment yeah you can basically discard it because you just used once. Um, curious YSM extension over MD. Um, yeah. So, obviously YSM because it's spec mine. Why not use MD? uh because actually what I did in my extension as you noticed I'm using mermaid diagrams and um I think uh the diagrams were not automatically being rendered in VS code but I think they are rendered in GitHub so I just wanted to render them and I couldn't with MD file and I didn't want to change the uh kind of the previewer for MD file so So I just came up with new extension which is SM. That's it. I don't know maybe I overthink uh but uh that's that's uh why I did SM just to have this extension so I can have full control over it to be honest which is also open source which is also in the project. Can the system SM file get so large to be problematic to break into the context or is there some logic? Um yeah I I think I have a issue in my GitHub project. So it's not happening but I think in theory it can happen. Uh it it depends uh on your project size but because spec now works only for monorreas. So if you have very very big projects most probably you are you are not on monorreo you have multiple repos. So this is how I control the size of system smile. But I was thinking that to have some um chunking logic for system SM file as well or some you know like u split into separate uh logical uh parts you know instead of having just one file. Will PHP be added to your road map anytime soon? Um so yeah again to be honest uh so far I was doing this totally alone and I was getting only feedback from teams I was working on. If I start getting feedback hey we need PHP yeah probably I will do it. Um and if you want you can even try to add that uh by yourself because you can use AI to add this integration to be honest right. So uh but yeah I mean uh if I start getting some feedback I will start adding more languages and uh PHP might be one of them. What do you think about having the test creation as part of the plan step rather than implementation step this will or against? Yes that's a very good question. So I have some more ideas. Uh to be honest I want to now um split implement step into more steps and one of them is test. Another one is a plan you know like I want to take a feature I then I want to create a test plan. uh and and then I also want to create a implementation plan you know like just get more understanding how AI is going to do and give less let's say judgment to AI how this should be implemented you know like I want to make sure that the specification file is so well documented that whatever AI does it just like just kind of different style of coding but it will not make bad or wrong decisions so but I I think creating a test plan separately is a very good colot and by the way uh now whenever I'm vibe coding or AI coding any anything uh I usually try to start from test so obviously unit test integration test but what uh I do more now with AI is simulation test so whatever AI I'm creating I create simulation test for them basically creating scenarios so I can give to to my agent get the output and assert on the output to see like okay how it will behave I don't know like is this part of integration or what but usually I do simulation test and um yeah and and I think probably this will be good another good step to add to generate you know this simulation but simulation tests are only for AI agents when you are building nondeterministic system if you are building deterministic one basically unit test and integration test will cover your scope language agnostic. What amount of token credit usage was used for your example? Uh good call out. Let me let me check. Um yeah, I I think it will be hard to uh >> sorry. >> Oh, right. But it's still doing that's the problem. I think it's still implementing. But let's see what it will give because I think the most of the tokens will go on the implementation by itself. Uh to be honest, I'm now using cloud code max and I just stopped thinking about tokens because uh before that I was using windsurf um that that was coming from my team and yes I was always checking my tokens. I was always buying extra tokens. Um I'm not promoting any of them but my personal preference is now cloud code uh because basically I just don't care about tokens anymore. There's a limit which I never reach and I and I AI code a lot nowadays. So yeah but uh um so basically if you generate the code without specind most probably you will spend the same tokens as it will spend when implementing. So you are spending extra tokens on uh specification generation but in this way you prevent the future ray implementations. So in a long term probably you just uh save tokens. So um but I don't have numbers to prove this to be honest. Uh do we have more questions? Language agnostic uh with question mark. I mean it's not language agnostic actually. uh in terms of that I need to support each language separately. So I support if that's the question to be honest I support these languages Typescript, JavaScript, Python and C#. Why? Because this is the languages I use. Uh but yeah pro probably like Go I know is pretty popular. Rust, Java, I mean C++ I don't know just popular language I wanted to have on my slides to be cool. Um what is the workflow of uh implementing the new language support? So basically what I do um so first uh I just explain what I want to do right with with u I use spec mine for my project as well now uh and then what I do usually I go and vibe got another project with that language some you know like mock project for example to do usually I do like task management system implemented by Java for example right uh and then I just run the spec mind on that project to to do human test and see what it what it generates. But I don't have to be honest um good designed automated testing yet. I think yeah I I have I have I have unit tests for of course uh but during this time I was just asking my friends to use spec and give me feedback. That's how to be honest I was improving it. Yeah, but that that's the process basically and again I'm using three-seater. So uh to understanding the a abstract syntax tree u that's part is delegated to three-seater. So the only part specbind um handles is understanding the patterns the layers some um let's say dependencies OMS to because it it's not looking to your database it's looking to your entities to build your entity relationship diagram. So you need to understand the syntaxes. So are if you're using drizzle or entity framework or whatever it automatically understand oh these are your entities. Oh then these are the relationship. So it pulls from your uh entities not the database. Um >> and we have a couple questions from the audience. >> Yes please. >> So uh notice one back here first. >> Great presentation. Um, could you just go over the basics of like uh say you're a beginner vibe coder, how you work with git and get github so that you when you've created something you don't lose it because sometimes I just go in circles for half a day and nothing's accomplished. Uh, yeah, let me rephrase your question. I don't think I understood it like how how I handled not to lose the context or can you repeat the question? Sorry. Well, I'm a beginner vibe coder and I tend to vibe code, but I don't really have a good workflow with GitHub yet. So that when I do work, >> okay, >> going back to it, I could spend all day and then lose what I did. >> Oh, I see. Okay. Uh, okay. So, yeah, if you're vibe coding, uh, so are you an engineer or I guess no, right? Okay. Yeah. So, if you're vibe coding, um, the good thing that u you can even ask AI to do the GitHub stuff for you. That's what I do nowadays. I I stopped manually running g or g status or g commit or whatever. So what you do so first thing you need to create a project in GitHub right so that's that's the part that you need to do it manually I guess and just copy the URL from the GitHub give it to AI and say hey this is my g project please pull it or do whatever so I need to have my code there and whenever you are creating a feature any feature and you see that it works just tell AI that hey please push this to GitHub that's it and it will push that chunk of code to the GitHub automatically what it does under the hood commits. It creates a commit and it push to main or you can say I mean if if if you work in a team this is not something you need to do you need to open PRs but if you are alone and if you're vive coding just say push to uh GitHub after every meaningful iteration okay and then you will have everything in your GitHub. Yeah. Any more questions? Yes, please. >> Hold on. No. Hey, so um I'm wondering though because my codebase is split across a an absolute boatload of services in completely different repos, but they all speak the same language and they all of course depend on uh you know libraries and similar to talk to each other. What's the recommended way in spec mine to handle a lot of repos given you only talked about the monor repo use case? Yeah. So I have like just for information I have that in my uh list to support uh multiple repos let's say multi-repo solutions. But for now what you can do basically you can um open uh you you can have like per repo kind of assistant right with it spec setup and then you can have another agent which you open in your root folder where all your repos are and you just ask and you just dis explain them hey I have this repos and there are feature specification files here here there and then if you want to come up with any let's say I don't like integration or whatever you can ask the agent but to be honest the uh ultimate solution will be just uh adding that support to spec in my opinion uh or any specdriven development uh tool so that's I have in my list but that's the probably the best advice I can give you yeah >> cool uh could we could you talk a little bit more about the disposable chunk JSON files right I believe those are created from trees sitter right um but could you explain why they're disposable or one-time use or did I get something confused there? >> Yeah, sure. So, um basically why what are these JSON files and why I need them? So, what I do so basically to in order to generate the system SM file the AI needs to understand your codebase, right? Basically, so uh the one way of doing this will be asking AI, hey, go and check my codebase and generate uh documentation. I tried that it I tried that it it it's not always working well because it's using different search algorithms and most probably it will pull some of the files and when it feels that oh I have high level understanding it will just start generating documentation on top of it. what I'm doing I'm generating JSON files from my codebase basically uh to ignore the implementation of my methods let's say if you go if our unit is one method right I just ignore the implementation of the method by itself but I want to understand all my relationship between methods classes and services uh which is less information which can be included in the context so that's why I run the CLI to generate these JSONs which basically are kind of some kind of a graph on top of my code and if for example if you look to this uh API layer I have this JSON files and I have a summary which um explaining my chunks and then each each chunk it's it's kind of it has some representation in terms of for example uh this one is because this is a API layer it has all the APIs for example it says call caller name post I don't know arguments location receiver you know for example if I go to my data layer I will have another representation probably about yeah uh about modules and entities and everything so then these chunks now are small enough to just inject into the context that's the reason only and because it's used only once for generating the system SM file because I hope that uh moving forward You will always use spec mind to generate these feature files. So you'll have commumulative updates in terms of feature specs. So you don't need this anymore. So that's why you can basically delete them. >> It's just around bootstrapping the project. >> Yes. Exactly. Okay. >> Yeah. >> Thank you. >> History is like set of versions of system files and and the log files that have the changes from one version to the other. That's it. >> Uh that's a good question. I think implementation is done. So I can show you. So >> the chunks you just throw away. You say you discard. So they're just like temporary. >> Yeah. Yeah. Yeah. So basically you just >> repeat the question for the uh >> sorry repeat the question. >> Yeah. For the audience. >> Yeah. So the question basically u let's say you you implemented multiple features. Uh what's the kind of version history? What what you have right? So I have only one system file. I'm not um it's only one. it just being updated after each implementation. But I have a change log and as you can see it generated one change log right. If I want to understand how this uh system SM file uh evaluated I can always come to this u change log and here it tells what it did and why it did with notes and also I get this u feature specification file. So basically you'll have bunch of feature specification files here. you will have just one system file and you have all your changes in this change log. That's it. >> Okay, thank you. Um, yeah, so this is very interesting. I I actually use a uh Amazon's hero uh coding agent which has a similar spec feature but it's not as comprehensive as this is basically start with green field create requirements design and then tasks. Uh so this kind of ends on top of that by uh examining existing code bases which is great but uh uh how do you handle uh having code get out of sync with the spec right? It seems like if somebody makes just changes to the code directly that's going to pretty much break the whole thing and then it's going to be hard to get back. >> Yes, that's a very good question. So the question is what if someone just u change directly in the code? So now the the system SM and and code are out of sync. uh that's why I was thinking to uh implement this feature so I can regenerate the SM file you know iteratively and that's why I was thinking that I still need to keep these chunk files which I don't do now so the use case you mentioned yes it's a problem so if if I start using spec mine and at one point I just stopped uh so my system SM file will be kind of uh out of date but uh let's say if if at one point I want to um resume. I mean I don't recommend this but uh let's say if there's you want to do in some way you can just ask AI to go and kind of try to update the system asm file but it will not work well. So basically that's a missing feature probably that uh which I have I think I even have in my GitHub issues. I I addit issues for myself just to remember. I think I have one of them to be able to you know like if you just temporarily stop using it and you want to resume be able to regenerate the system SM file not generate from scratch but just like you know like add whatever was added after last point. The good news that I have a change log so I know when last time you edited it or even I can look to the g history but yeah I think uh I think that's a missing point in spec now if you if you use it you need to keep using it uh always let's say >> any more question okay looks like that's it let's give our speaker some love >> thank you >> thank you for coming out Uh there by the way sorry there was a question where is the presentation uh I'm going to submit this to uh scale and I think there should be way of getting all the presentations so >> u uh or what you can do you can connect me on LinkedIn and I can uh you can DM me and I can send it to you as well. >> So this is my LinkedIn URL and my QR for my LinkedIn. There's a live stream right now on YouTube and the videos should be edited out uh very very soon. Can't say. >> Oh, cool. >> Thank you. >> Thank you everybody for coming for coming. Have a great rest of your day at scale.