Dynamic Firebase skills: Architecting agent ready codebases

Key Takeaways

Hey, it's Aja, and I'm here at Next, talking to folks in the GDE lounge about cool things they've built with AI and cool things they've built with AI dev tools. Today, I've got Rhody with me. Rhody, why don't you tell us a little bit about yourself, and then we can talk about some of the cool stuff you've built. Cool. Yeah, so my name's Rhody. I am am a dev rel on the anti-gravity dev rel team, and um working with dev tools, getting people excited about building with AI, whether it's non-coding tasks or even like building full applications and uh especially with existing code bases. Ooh, existing code bases. >> Yeah. We're going to come back to that. So, is there anything that you've built recently that you're like really proud of or you think people would find super cool? Yeah. One of the things uh so, I've been a huge like skills advocate. I think they're like the coolest way to package a bunch of information in a condensed space. So, I've been looking at putting skills on the client side. So, with uh Firebase, you can use like AI logic, and um by having Firestore ship your skills, you can basically create really awesome chatbots with a very small context windows. So, by loading in the skill For example, let's uh Tell me more. Tell me more. Yeah, so the classic support, you know, like I want to refund this order. Well, skill is a great way to say, here's how to call the right things, here's how you should respond, and what you should look for. But then you also have a skill for upselling, right? And it's going to have a completely different context. You don't want to have the AI be like um trying to upsell something when someone's having a bad experience. >> Ah, so you can Dynamically flavor. Yeah. Yeah, dynamically pick the skills and also kind of sandbox them so that they're not polluting each other's context windows. Exactly. Yeah. So, how'd you build that? What did you use to build that? Did you like write them by hand? So, I used anti-gravity to basically come up with a way to generate a bunch of scripts, and then uh created the Firestore functions that will um uh basically anytime something's uploaded to file storage, uh it'll automatically generate the YAML and the metadata. So, I didn't have to like write any of the Firestore logic itself. And then in the app, you know, I just uh all in anti-gravity, uh built it with Flutter and Firebase. Cool. So, what do you like about anti-gravity? So, one of the cool things about anti-gravity is I I like it helps me from any part of the journey, whether I'm just trying to update App Store screenshots and generate icons or non-coding tasks, all the way to like working with really deep uh code bases where I can have either a big mono repo or a bunch of different projects. I can open up each one of them in their own agent and have like an agent manager connect them all. So, you have multiple mono repo. What's been your experience working with anti-gravity on, you know, brownfield projects? I like to say that greenfield doesn't really exist because like you write the code and then it's brown it's brownfield immediately. What's your experience been using anti-gravity on projects that already exist, like add new features or to fix bugs? So, one of the important things to do when you're coming to working with anti-gravity in an existing code base is try to think about AI health, you know, like how do I make this code base ready to be agentive. And so, one of the first things you can do is just try to generate a markdown file for every directory in your project. And it should like basically have a summary of what it does and then all the files below it. And then from there, you're also working on like what are your common skills and things that I'm commonly doing in this code base. And then over time, you're just going to make it easier for you to to make these changes. Also, make sure you have tests and do other things because the agent will be able to make a lot of stuff pretty rapidly. So, what you're saying is tests, which I would have pushed for anyway. Yeah. Some sort of directory level read me that's designed for agents so that you can quickly get context and not have to read all the files. >> it the context MD. That's what I would call it. Yeah. >> And then on top of that, also just making sure that you've actually used reasonable engineering practices so that the agent has a good place to start from. >> Totally. Yeah, and also I feel like I've been noticing my architectures change over time. Like I'm going to a way more horizontal flat architecture where like the agents can put state in one layer and they can like, you know, easily refactor. Cuz you want to be able to tell the agent like break this large file into smaller functions and also try to find similar things that are, you know, in the same directory. And then that way like when you come back to it, you'll also have the better context as well. That's that's super cool. So, when you're building with anti-gravity, like have you learned any tips or tricks to make it work really well? I know that I watch you and I'm like, I didn't know you could do that. So, do you have any like tips or tricks to work effectively with a tool like anti-gravity? So, for me, I like to exploit the fact that humans are really good at pattern matching. So, uh for me, like looking at like a file system or um any type of streaming, you're wanting to find patterns and to see if the agent is following the pattern or falling out of the pattern. Because it's so easy to course correct and be like, "Oh, well, you're obviously in the wrong directory, so you must have not read this context." And so, skills and uh MCP servers and a bunch of other things can really help make that better. But really, it's about knowing where you want to head and be really prescriptive. Like, so I like to use like the Gemini app or uh NotebookLM to create prompts and context so that when I'm coming into anti-gravity, I have like a really clear plan. Uh so, it can just uh build what we need. That's definitely a thing that I've heard kind of this whole week is how important the planning phase is becoming. Yes. And how important it is to have a very clear idea, not necessarily how you want to build, but where you want to go and what success looks like going in, so that you can steer the agents in the right direction. Yeah, in fact, like I've seen a lot of people uh like even in the hackathons that have transitioned over time. Like, it used to be you're spending all the time just trying to get it working, get your database connected, and all that. But now, like 60% of the time should be spent in the design phase of like, "What do we actually want to do?" Because then it's just technology choices and how fast can we get to the MVP? Yeah, I know, I agree with that. And one of my most successful Viper project started with several hours in NotebookLM and in the Gemini app doing research, figuring out what was possible, what was in the scope of the problem I was trying to solve. It was great. So, uh folks, I've heard there's a lot of stuff out there. Uh every model provider's got multiple models. There's multiple agent tools. Yeah. There's a lot of different ways you can you have to use a live API, do you not use an live API, do Do use an LLM, or do you actually use an old-school uh machine learning model? Yeah. How do folks get started? Like lots of people have really cool ideas they want to build, but they don't know where to start. And like they know that it should be possible, but how would you get started? How would you tell someone to get started if they came to you? Yeah, totally. Uh I like to think about it like the universe of AI. Like there's going to be a lot of different parts of like the things that you're going to be exposed to. And honestly like try to find like the idea in everything that you want to do. And then sometimes it's going to be I'm going to like just want to build this prototype and I just want to share it with people. I'm going to go to AI Studio and I'm going to build that and I'm going to be able to easily just share with a link. But then other times I'm like I have this context and this problem I really want to break down. I'm going to go to Notebook LM or or some other type of app and then other times I'm like I have this idea I just want to take it further or I want to connect to real sources and stuff. So like I know people that are like reading their email or like creating a bunch of different types of files and you can do that inside of these editors with these all these MCPs that are like so powerful now. So like think about the workflows that you want to accomplish like any skills or tasks that can make that possible. And then also you just may find that like sometimes you're just in the dreaming phase. Like and that's where I really like Stitch. And so I could just go like oh like how would this look like before I could get started? >> I like I like using Stitch to figure out if my idea makes it out of my brain to something that another person might understand. Cuz if you can't describe it and make it visually you're like oh maybe I don't know the idea well. >> Yeah, maybe I haven't thought maybe I haven't thought this through yet. Yeah. I just want to thank you for your time and thank you for the ideas that you've shared with me and the audience. I really hope that this helps them get started with anti-gravity. Absolutely. Thanks.