Rubber Duck Thursday!

Black. Nah. You love me. Hey, back up. Heat. Heat. Heat. Heat. N. Hello, hello, hello friends. Hey y'all. Welcome to the stream. It's time for rubber duck Thursday. Hey everybody, I'm Cadaca, developer advocate here at GitHub. Thank you so much for joining the rubber duck Thursday stream. I have a really really fun and interesting stream for you today. So I hope you stick around and thank you so much for joining. Let me know what you're where you're joining from. Let me know what you're working on. If this is your first time joining Rubber Duck Thursdays, it is a very chill stream where we just we learn, we build, we vibe, answer questions, and just learn and grow together. Today I want to talk about There's an echo. Let's see what mic is being used. Oh, not sure. Is there is there an echo reverb for anyone else? Sound issue. Really? Oh, that's quite sad, isn't it? Okay. Can you hear me? Okay. Okay. So, it was too many videos playing at once. Okay. Gotcha. Gotcha. Hopefully the sound is better. Um, I didn't change anything, but hopefully hopefully it doesn't sound uh crazy for anyone else. Like I was saying, welcome to Republic Thursday. I'm Kata. Have a great stream for you today. Today, I want to talk about agent skills. So, I've been learning more about agent skills, just like really digging into it to understand what it is, how to build agent skills, how to use them because, you know, like these things aren't going anywhere, folks. So, we have to upskill and have a really good foundational understanding. And in my learning, you know, I went back to using the remote best practices skill from the remote team to programmatically create videos. And as someone who creates videos for work and also for fun, um this idea was just a very curious idea to me. And you know, in playing with Copilot CLI and all the tools that I have access to, I was able to create a really great um video agent that I want to show off today and just tell you how I built it and how you can also create one, too. So, first thing I want to do is go through the change log, which I always do. So, awesome. I'm glad that the audio is great. We have Greece. We have Australia, Morocco. Wow, you guys are joining from all over the world. Welcome to the stream, Pakistan. So happy to have you here. North Dakota. Yes, yes, yes. Welcome, welcome, welcome. So, I'm going to share my screen. Let me pop you guys to my other screen that way I can share this screen and then we can get into the stream. How are you guys doing? You know, I always love these stream. This is one of my favorite things to do in the week is to just come on here and just chat with y'all. So, I'm just getting everything I'm just getting my screen blown up right here. Boop. Also, let me know in the chat, have you been looking into agent skills? Have you used it? Have you built it? What are some interesting use cases you found for agent skills? I find myself using skills more often than like, you know, fine-tuning, prompting these days, but let me know what your experience is as well. Boop. Okay, now we're on the change log. So, let's take a look at what we've shipped at GitHub recently because literally we ship every single day. Like, we're shipping so many features, it's insane. And not just copilot features, we're also shipping core GitHub dev features that a lot of you have been loving. So let's take a look at what we've shipped in the past few weeks and then I can show you all the thing I've been working on. All right, so looks like just today we shipped an improvement to copilot uses met uses metrics. So it now includes per user GitHub copilot activity and organization reports. I'm sure uh enterprises and their admins will love that release. That's awesome. GitHub actions had a few updates just shipped today. So, let's see. Customizing entry points for service containers actions. OIDC tokens now support repository custom properties. It's awesome. Azure private network networking now supports V-Net failover. Incredible. I also know that the GitHub actions team just shipped support for time zones. So you don't need no longer need to manually calculate time zones. GitHub actions will now know where you are in the world, which I think is really really good. We have some security improvements to the security tab. So the security tab on GitHub is now security and quality. Oh, so the top level security tab across repositories, organizations, and enterprises has been renamed to security and quality on github.com. This change restructures the navigation to to colllocate code quality findings alongside security alerts, making it easier to triage all code da da. All right, so that's a name change issue on github.com. So, I like to go through the change log with you because I want you to realize how much we're shipping and how important it is to go through the change log, you know, like if you're trying to keep up with GitHub because things are moving so fast across the industry. It's insane. It's insane. There's improved search for GitHub issues that's now generally available. Incredible. So, projects and issues got some love. Security C-pilot actions. So, right across the platform, uh, stuff is being shipped. So, but this one I'm super excited about. So, you if you've ever used Copilot coding agent, you know how it, um, it typically loves to open up pull requests even if you don't ask it to and even if you're trying to have a conversation with it or like looking into your repositories on.com without like pulling everything down. So now you can do researching, you can plan and you can code with copilot cloud agent and now it's copilot cloud agent instead of copilot coding agent just to make the name a bit more reflective of what the tool does. So this means copilot cloud agent will no longer automatically open up a PR. You can chat with it. You can ask it about your project. you can do planning and then you can tell it to implement the plan and then you can tell it okay great I want to open up this PR thank you so much for contributing to my work so I think that's pretty pretty cool and if you've never used Copilot cloud agent if you go on I have a lot of tabs wow but if you go on github.com you see this agents tab here if you click on this agents tab and I'm just going to open it in a new browser I don't want to lose lose that repo here. You'll see right here. Huh? And this is essentially where you can interact with Copilot Cloud Agent. And so you can, you know, ask questions, do all sort of things, right? Do all sort of things. It's it's incredible. It's incredible. If you haven't used it, give it a go. You can assign issues to it using at Copilot. You know, it will respond with like a little I emoji and it will get to work on your behalf. But that's that's something that's really really interesting. And let me know if you want me to go into that some more. But today, I really want to talk about a agent skills, okay? I really want to get into it. I have something really cool to show you. And I want you to tell me if you think it's cool, too. But this one is a really great release. Go check it out. Let me pop up the um the change log link here. If you go to gh.io/rdt/change io/rdt/changel log, you'll be able to see all the things that's been shipped on GitHub recently. And as you see, just this week, we had so many releases. All right, so I would love to know in the comments, how many of you have been digging into agent skills? I know that I have been for the past few weeks, and this is the open standard for agent skills. uh for the past few weeks I've been trying to learn as much as I can about agent skills so I can come and you know teach you the knowledge that I have grasped. So yes, welcome. Welcome Gerald Dean. I didn't know anything. I didn't know about AI agents before the streams. Very interesting. Yeah. So welcome. So AI agents. Okay. So you know how you have large language models, right? Large language models, that's the the thing that's trained that can, you know, collate text in put all sort of things together in order for us to use chat bots to talk to it, right? So chat bots like um like Claude chatgypt github copilot on github.com uses LLM under the hood to respond to our questions and to respond to our inquiries. When we go a step beyond that we have AI agents and AI agents they are able to do things for us. So, you know how you send a prompt to like Copilot or, you know, one of the chat bots and you're just like, "Hey, what is 1 + 2?" And then it responds and says 1 plus 2 is 5 or 1 plus 2. 1 plus 2 is three. 1 plus 2 is three. Please don't laugh at me. But with an AI agent, you're able to say, "Hey, can you go out and buy some apples for me?" and it will literally go on the internet, click in your browser, add some apples to the cart, and have shipped delivered to your house. That is an AI agent. And I like to talk about these tools and technology in a way that that's not so overly technical just so that you know there's clarity among everyone who's joining us here on stream. So that's an AI agent, right? And then when you get into agent skills, agent skills enable AI agents to be properly tuned to have those specialized expertise and how to do something. So let's go back to the agent buying you an apple, right? An an agent skill would allow us to give that apple buying agent the skill of a personal shopper, right? So we could create an agent skill called personal shopper wherein that agent will know exactly what type of apples you like, what type of pricing you like to, you know, to do, where you like to buy it, how you like, you know, your app, you know, your apples delivered, all that jazz. So the agent skills allow you to give your agents specialized skills and specialized capabilities to go forth and do stuff. And I think that's incredible. And so the agent skill and so the agent skill that I've been looking into this um this week is the remotion agent skill. So Remotion is an open source project that allows you to programmatically write videos like to programmatically build videos. Isn't that incredible? Like you can literally build videos using code, JavaScript, React, CSS. Like I think I think that's incredible. And the Remotion project has been around for quite a few years. It's it's it's not new, right? It's been around for a few years, but it's increased in popularity recently because they created what an agent skill. And so the agent skill they created teaches or enables AI agents to be able to create videos just by us prompting it. And I think that's incredible. And so as someone who creates videos for work and also for my personal socials, I found it this skill quite interesting. And so this week I've been working on building out a um an an internal tool for the team to see if we can, you know, create videos a little bit faster for GitHub releases. Cuz as I just showed you on the GitHub change log, we are constantly constantly shipping, constantly shipping. every day we're shipping multiple things. And so it's kind of hard to keep up with all the changes. And so and so we thought, hey, what what if what if there was what if there was a an agent that can build videos for us for these releases? And I was noodling with it this week, and y'all, it it works. It works a little bit too well, but it it works incredibly well. And I didn't come up with with this idea. A few weeks ago, I saw Debbie uh who previously worked at Goose. She did something similar where she created like a release agent and I was like, "Ooh, ooh." So, I kept it in the back of my head and this week I was finally able to sit down and get something working. So, I want to show you what I have. I want to show you what I have. Let me know what questions you have so far before I go for go forth in my excitement cuz as you can see, I'm so excited. I'm so excited. So, let me know what questions you have. Let me scroll through and see. Nice. I don't like using cameras. Exactly, Thiago. So, imagine if you wanted to create videos. You can create videos and you can create videos programmatically with your agents here. Wow, that's what I said. Joshua, wait until you see the video. Okay. So, why use agent skill instead of an MCP server? Great question. Right. So an agent skill like I said gives your agent specialized capabilities to do stuff and so the stuff that you can you can enable it to do can be you know like procedural knowledge company specific knowledge team specific knowledge user specific context um that they can load on demand an MCP server give your agents access to tools right so let's say there is an MCP server for um Adobe Premiere Pro. Adobe Premiere Pro is a is an editing a video editing tool, right? So, let's say my agent has the motion video making skill and then it also has access to the Adobe Premiere Pro MCP server. What I can do is I can say create a video for XYZ and then using the Adobe Premiere Pro tool edit that video and add XYZ to it. Right? So the difference between MCP servers and agents and skills is that MCP servers give your agent access to external tools. So tools like your database, tools like your um external documentation, GitHub, Sentry, observability, it gives your agent access to tools that you're currently using. A skill allows your agents like empowers your agents or enables your agent to do a certain task more efficiently and better. So let's say me for example, let's say I want to learn how to draw. Drawing is a specialized skill, right? So, if I want to learn how to draw, I'm going to go I'm going to learn, watch videos, take some drawing courses, and then eventually I will have the skill of drawing, right? But when it comes to tools, I need to draw, I'm going to need a pencil, right? I'm going to need a sketchbook. Those are tools. So, you can think about MCP servers and skill kind of like that. So, if you want to learn how to draw, the skill that you need is learning how to draw, like the skill of learning how to draw. you're going to go and go ahead and take courses for that and then you're going to have all that context in your brain. But if you actually want to draw, you're going to need drawing tools. And so the MCP comes in with the pencil and the sketchbook. Hopefully that makes sense. Um, Pi Miracle. And so that's why um I'm using Skills instead of MCP servers. Though, if I wanted to, I could use the GitHub MCP server, uh, for example, to, um, add the video to whatever issue is on GitHub. That way, it's just programmatically there and the team has it already. Hopefully, that makes sense. Let's see if there's any more questions. Yes, really interesting. Extending capabilities. Exactly. Extending your capabilities. Welcome. So agent is like a digital human and you know you know I don't like to liken AI to anything human. Um I I don't I I wouldn't I see where you're going but I I wouldn't put it I wouldn't put it that way. But but for understanding sake yes for understanding sake yes. So what is the difference between this agent and chat GPT codeex? Great question Muhammad. So um openi's codeex is an example of an AI agent. So it's a coding it's a coding agent right so you have openi codeex you have google's anti-gravity you have github copilot cli you have um a lot of other ones right cloud code those are all examples of coding agents so um the difference between this agent and traptex agent is that this is a custom agent I built to use with tools like codeex and copilot cli hopefully that makes sense All right. Can you use your own voice and add to the video? Yes. Yes. Yes, you can. So, let me pull it up right now and then we can Let me pull it up right now. Um, wow. Thank you so much for the explanation. Technology and AI is really changing and coming up with new things every day. It's hard to keep up sometimes. Listen, Geraldine. It's not hard to keep up sometimes. It's hard to keep up all the times. Every day there is a new tool, new model, new ship, new new this, new that, this happening, this happening. Honestly, the name of the game these days is to focus. What do you want to learn? What do you want to be? Where do you want to go? And you focus on those lanes. And so, yeah, that's why I've been digging into agent skills cuz I'm just like, let me focus, narrow the path because there's just so much noise. There's so much noise. All right. So, I see your question, Diego, and I'm going to pin it. Boop. Um, I'm going to pin it. And I'm actually going to keep this page open because this page is going to answer that question for you. Let me actually link this. It is on GitHub here. So, let me actually link this for you. So, you can go forth and do learn about agent skills. I always have typo. Anyone else always have typos? At this point, it's just a part of my brand. Like, it's crazy. Even if I even if I review something like 20 times, there's always going to be a typo. I've accepted it, you know, but sometimes it's a little annoying. All right. So, go to this repository to learn more Go there to learn more about agent skills. Um, all right. So, let me get into the the demo. Right. So, let me pop let me pop my co-pilot over here. Boop. Oh, no. Wait. I have to um remove this from being full screen. Okay. So I have copilot here running and what I'm going to do is so I want to show you how the agent work and like while the agent is working in the background I'm going to walk through how we build this and how the tool works. All right. So this is my custom release video generator agent. So if I go to the change log here, there was one that I wanted to use specifically. Where is it? One second. I think I lost it. Ah, okay. So let's say I want to create a video for this change log, right? For this release. I'm going to copy this linky here and I'm going to pop up my copilot CLI. Boop. And I already have this enabled. So, I'm just going to paste this here. Boop. Hey, I forget that whenever you whenever you highlight something in Copilot CLI, it automatically copies it for you. So, great great feature, but when I forget, it's all right. So, I'm just going to pop in the link. I'm going to send it to Copilot CLI and I'm sending it to the agent right here. And I'm using Opus 4.6 for this. And as you can see, it's initiated phase one fetching the change log. Okay. So, let's walk through how this works while Copilot is working in the background. Unless you want to see it actually doing it here. No, we can go through the we can go through the output together later. So, let's go to my goodness. Let's go to here and let's look at the read me. Okay, so how this works. Oh, I forgot to hold on. Let me just stage this first before it builds new videos on top. Perfect. Okay, so welcome to the GitHub video release generator. Um the GitHub release video generator, right? So this is using Remotion React um I fed it the GitHub brand um and it and the fonts that we use at GitHub from our brand system. And so everything is in here. And so how this works is uh it's a four-phase pipeline. So the first agent uh fetches the change log information. So, it's going to grab go to the link, go to the site, grab the text, grab any images that's on the change log, and also grab any videos that's on the change log and pop it into the project, right? It's going to grab all the media assets. The second agent is then going to script a video. So, it's going to write a text script, and then it's going to also generate a voice over script for the video. I know it's wild. It's going to have intro, outro, features, all the things. The third agent is the B-roll agent. So this agent is responsible for creating um animated assets for the video. Um and then the last agent puts everything together. So the last agent puts the the media assets, it puts the B-roll together, it puts the voice over, and it also adds captions following the GitHub branding. I think that's so cool. And so this custom agent is a markdown files cuz so to build custom agents you would use markdown files. And so this release video generator agent is the main orchestrator agent. So it's the agent that tells the other agent you do this now. You do this, you do this. Right? So it's going to call the B-roll designer agent, the change lock fetcher agent, the script writer agent, and the video assembler agent last. and it tells them what to do, right? So, here's what the release video generator agent custom agent looks like. So, in the front matter here, you have a name, a description, and you it gives access the agent access to all the tools. And so, these are internal tools that it will use um that it would need access to on the computer. This is the instructions. So, you are a main orchestrator for the release video production. Um you will do this this this right input format. So it tells the agent what type of format to expect. So it's going to expect a like a change log video. I also added the ability to add multiple URLs just in case there are any ex other sites with more images or more media assets that can be added to the video. So that's that's also here. So you can add multiple supplementary sources. Then it's going to parse the primary and supplementary URLs to get all the information. And it's using a free voice from 11 Labs. It was just very easy to use the 11 Labs voice. It does sound very AI, I will warn you. But imagine this video is done. And now with the script, I can go ahead and and do the recording for the script and post the video. That's going to take about maybe like 15 minutes to record and, you know, add the the voice and stuff to the video, which is a massive productivity boost and a massive timesaver. What do you think? Am I making sense here? And so this is what a custom agent typically looks like. This is how you will write one. Let me know what questions you have. And let me pull up copilot here to see the progress. So copilot here is still working. So it's it's finished with phase three. So it's now at phase four. So now it's assembling the video, right? That fourth agent is coming in to do the video assembly. Okay. So, let's see what questions you have. This is amazing. I'm happy you think so. Welcome. Okay. So, you're saying this thing can do anything, but how can we trust this thing with our sensitive information? And so, I'm presuming this thing refers to AI agents, right? So, you have to be careful. You can't you can't just be telling it everything and you know giving it access to your personal computer. Okay? You can't you know you have to you have to be careful because the technology is new. It's evolving and it can make mistakes, right? So you have to practice proper judgment in terms of what you will allow it to access and not access. Resources to learn AI agents. That's a really good question. Honestly, lately I've been using some resources on Oh my goodness. What is Hold on. Deep learning deep learning. Deeplearning.ai is an incredible free resource for learning and that's what I've been that's what I've been using in addition to uh some other things. But go to deeplearning.ai and get your hands on some free courses to learn all about AI agents. Whether it's like agent memory skills, um honestly all the things, all the things. It's such an incredible resource. Is this agent for developing a specialized skill or solution or can it be dynamic? Joshua, tell me what you mean. Tell me what you mean here. Uh, Thiago, I have typos a lot. I write it in three different languages every day, you know, and I'm only writing in one language, Thiago, and I I just be having the typos. So, like, I can forgive, you know, I can forgive you cuz you're you're writing three different languages, but I just have the one. I just have the one. I don't take typos as errors again, but as a reminder that it's human programming. Yes, I've been leaning into a human wrote this. Um, what skills are needed for a design AI agent? You would need to create a custom design skill, right? And there are also you can find um agents. If you go to the awesome Copilot repository, it has some really great agents that you can pull from. It looks like Copilot is finished here. So, I want to I want to show you the video that it made, but um let me show you this and then let's look at the look at the video together and then you tell me what you think about the video. Okay. Uh no, awesome copilot GitHub. Okay. So, Awesome Copilot is a repository you can go on to find skills, plugins, custom agents that you can use in your projects. And there may be a design agent. I'm not entirely sure. There's a front-end specialist, not a design specific agent. Let's see if there's a design one. DevOps.net. There are no design agents here, but there are there are quite a few like open-source agents that you can find on GitHub, and there may be a design agent somewhere. I I'm I'm pretty sure I have one bookmarked somewhere um that you can use. Thank you. I'm wondering if I can use it with open router.AI API. Uh I don't Does open router support AI agents? I think so. Not agent skills. Uh uh yep, open router does support agents agent skills and tool calling. Uh so you can definitely use AI agent like agent skills with open router. Okay. Okay. Okay. Okay. Is this agent just for video generation? Yes. So this specific agent that I created, the release video generator is just for uh videos, but you can create agents for anything that you want. Right? So let's take a look at the video that it created. So the video is here. So I'm just going to I've been using this um this window to pop open the video. And here is the video. Oh my gosh. Okay. So I want to make sure I'm sharing my sound. How do I make sure I share my sound? Stop sharing. Okay, hold on. Also share tab sound. H also share system sound. Yes, there we go. Entire screen also share system sound. Okay, let me know if you hear it when I play it. One second. Let me turn off this background music. Okay, so let's take a look at the video that Copilot just generated. Let me know if you can hear it. The Copilot coding agent just landed in your terminal. GitHub Copilot CLI is now in public preview, giving you a local synchronous AI agent that understands your code, your repos, and your GitHub context. No browser, no context switching, just you and C-pilot command line native. This isn't just autocomplete in a terminal. Copilot CLI plans and executes complex task. Okay. Okay. So, you can see that it grabbed this image from the change log. This image came right from the change log. Um, where is it? Here. So, you can see it's it's literally doing the thing that we asked it to do. Though, you know, it needs some improvement, but you know, look at the the captions were put there by Remotion. The look at this. This is the GitHub font, the Mona Sands font that we use. GitHub colors. I think it's incredible. Are all accessible through natural language, authenticated with your existing GitHub account. Plus, Copilot CLI ships with GitHub's MCP server builtin, and you can add custom MCP servers to extend it even further. Getting started takes seconds. Run npm install, authenticate with GitHub, and you're ready to go on any C-Pilot Pro, Pro Plus, Business, or Enterprise plan. Whether you're exploring a new code base or implementing features from issues, Copilot CLI meets you where you work. Install Copilot CLI today and bring the coding agent to your terminal. Head to gh.io change log for the full details. You guys, isn't that crazy? Like we just did that video literally like it and it didn't even take long. like we did it right here together. And there's also a short a short form video, but I'm working on the short form video because it doesn't look very good. So, I need to um I need to iterate there to make sure that the short form version looks appropriate for a short for the platform. But that first video, so this is what the short form looks like right now. It's not readable. You see the text is too small. So, it needs to just be bigger for the short form. But, um, I can I'll have to, you know, fine-tune that. But, this video is incredible, right? Doesn't this look like a GitHub video you've seen before, right? Like, I am I'm so impressed. I think we're like 60 70% of the way there. Um, there's still some fine-tuning that's needed to be done in order for this to be something that we would actually ship from GitHub. But what do you think? I think this is so cool. I think this is so cool. Okay, let's see what you're saying in the comments. Let's see. Is there a replay of this session? Yes. So, the session is recorded. LinkedIn, YouTube, um Twitch, all the places you can go back. You can go back and rewatch the session. Yes. Uh it's audible. Okay. I'm glad you That's awesome. Yes. Great. V you like the voice. So, I think the voice is a little bit AI robotic. Um I'm using one of the free voices from 11 Labs. So, how this works is you feed it your Let me show you. So there's APN. And so Remotion is automatically using 11 Labs internally. And so I just provided it with my 11 Labs API key. And then I chose a voice from the platform and I copied the voice ID. Bada bing, bada boom. That's that's all it takes to get this to work um properly. Wow. Okay, great. I'm happy that you think this is as impressive as I think it is because I'm currently working on a blog post for this to teach people or to teach you how step by step of how you can do this as well. The voice is so real. Wow, I'm impressed. Okay, the voice doesn't bother at all. Okay, so maybe the voice isn't so bad. Hey, Jeremiah. Only one camel case co-pilot. That's the most impressive part to me. I agree. I did peep that. I did peep that. Um, so I think I will have to tune like just like have the agent add a set of instructions in the markdown like in here that says, you know, always use co-pilot one word because currently I have it saying always um the CTA should always be gh.io/change log. So the CTA will always be gh.io/change io/change log. Um, I didn't I opted not to have it do like slash name of feature because then we would have to go in manually and you know create those short links. So I figured go to the change log and you can you know find the changes but it's pretty cool right and sometime we when we ask questions to the agent they are very confused and in previous question they are lost in memory. So how do so you're asking about agent memory? Yeah, honestly, agent memory is a really big issue right now in the space. I haven't dug into it myself, so I don't have enough to say on the topic, but I do know that it's an it's an issue, right? Because for example, if I close if I end this session with Copilot, I can resume the session, right? GitHub Copilot has the ability to resume sessions. And so it will pick up. So like right here this co this thing here I can say copil it resume session and it will have the context from the session so I can carry on like I you know like I didn't end the session that's great but usually when you end the session and you start a new session it doesn't have the context and so that's a that's a problem that the industry is working on solving honestly. Great point. This looks great. I'm you know I'm so happy you like it. Not bad. I think so too. Uh, okay. If we're using the clo skill, can we create an agent based on vibe coding? Okay. What does this question mean? Right. if you're using claude skill, do you mean claude code? If you're using cloud code, you can use agent skills and all cloud code skills and rules works with copilot CLI as well. So, as you can see, I have a cloud code and I have agitub here and get up GitHub understands everything that's in here, which is awesome. Uh, but I'm not understanding your question. I'm so sorry if you don't mind rephrasing. This is a game changer. But if it was a man voice, it would be it would have been great. Really? Really, Bara? Really? That's what you're going to say? If it was a man's voice, it would have been great. I don't know if you're pulling pulling my u my lick here. The presenter sound effects are the best. Do you mean me or do you mean the video? You're probably talking about the video. If you're talking about me, thanks. Uh, a step further. If you would make this agent available for use by other folks so they don't have to write one, how would you do it? Ooh, great question. So, if I was supposed to make this um this custom this custom release agent available to all, honestly, you would need to just copy these markdown files. Um, but I would have to I would have to turn this because this is currently in the remote project, right? And so there are custom custom scripts for captions, composition, fonts. Like this is not just this is from the motion, but copilot CLI edited. So it's specific for the release um the release that we need. You know what I mean? So, in order for me to give this to you to use, I would have to turn it into a what? A skill. Into an agent skill. Because skills allow you to have markdown files. It allows you to have markdown. Wait, sorry, hold on. I have a blog post that I'm working on here. So skills allow you to have like scripts, rules, markdown, um, TSX files, um, and so that way I can package it all up and just send it over to you for you to use it. I haven't got I haven't gotten to that part yet. I've been testing it out. The background music, please. I agree. There you go. Um, so yeah, super happy that you guys like like it as much as I do. I have really really been enjoying it and I'm so impressed by it. And you know, a lot of people will see something like this and say, you know, Cadia, this makes me so scared off the future. And valid, valid, because things are changing so fast. But also I think it's a really great opportunity for you to learn how you can use these tools to you know like make yourself better to enhance your life to enhance the life of people around you and it's a responsibility of all of us to upskill and to keep up with the pace of the rapid um change of technology. Um this is this is so interesting to me. This is the blog post that I'm working on. So, you'll probably see this on the GitHub blog or dev.2. You'll see it somewhere in the wild. Make sure you're following me on LinkedIn so you don't miss whenever I uh post about it. But yeah, this is what I have been up to. It's been super cool, super great. And yeah, I forgot to ask, is it open source? It's currently not open source. I'm building it as an internal tool for the team um because it's it's spec it's it's very specific to the GitHub change log release videos process. Can we get this code? Is any possible? You you you probably will be able to get the code. I'll I'll get everything um tuned for you. It will take some work for me to do that because the the skill is very specific to GitHub, right? It's using the GitHub branding. So, one thing I wanted to tell you is when you're building up your agent skills, you can context, right? So, I sent it this link and this is the GitHub Mona font that we use. So, I pointed uh Copilot CLA. I said, "Hey, here is the font that GitHub uses." And you know what Copil CLI did? It came here and it says, "Oh, perfect." And then it downloaded the font, the actual web font, and it's the thewf. And then it saved it in the repository as the font. Where is my VS code here? So if I go to where did I see I think it's in public. Is it in public? Yep. So it saved the font here as the font. And so now every video generated is going to be references GitHub's Mona Sands font because that's just the the font that we use. Something else I did was I pointed it to the GitHub brand guide. So this is our brand guide. It's a public it's a public page here. So I told it to read the color. I told it to read logo. I told it to read typography. I told it to read voice and tone. Iconog iconography. Sometimes that word is hard for me to say even mascot, right? I Well, I didn't give it mascot. I probably I probably should, but I told I gave it context on how what what it what does it feel to be GitHub? Like what does it mean to be GitHub? And that's why the video turn looked so like GitHuby. Um yeah, so that's also another way that you can give your agent context. I just gave it the link and it's able to fetch the link and fetch the information that's on the site to get things going for you. And so if you want to build your own agent skill, here's how I would start. So what do you want to build an agent skill for? Tell somebody to give me an idea in the comments and then let's do it together. And I don't want to close this. So, I'm going to start a new. No. No. Um, no. Takers. What type of If you were to build a custom custom agent, what type of agent would you build? Um, as I think as well, my brain is a little mush. The idea The idea is the hardest part. You don't say. Yes. The idea is the hardest part. Okay. So, let's think. Let's say we want to build Let's say we want to build an agent that I don't know. My brain is my brain is gone. Hold on. co-opilot d-allow all. Have you been using Copilot CLI by the way? It's incredible. It's incredible. If you haven't been using it, it's incredible. Make sure you're using it. Um, make sure you're using it. It's so good. It's so good. I promise. I'm not I'm not even just saying that because I work at GitHub, okay? I'm saying it because I use it every single day over other tools on the market. Um, and it's it's really great. Oh, look at that. Okay, so let's say we want to build an agent that aggregates all the comments off of GitHub live streams um from different platforms, right? So, currently the way that we stream is we use a tool called Streamyard. And so we're able to stream to LinkedIn, YouTube, Twitch, and um Twitter simultaneously. I still say Twitter, and I will still say Twitter. I'm sorry. Um and so what if there was an agent that could scrape all the comments? Ooh. Okay. So, if I wanted to build this agent, what would I do? I would, you know, use voice mode. I use a tool called Handy Handy.comput to um to kind of talk. Let me show you. So, this is a free open- source tool um similar to tools like Whisperflow that allows you to talk, but this one is free. It's open source and it uses local models. So, this is the tool that I use to kind of talk to my tools. And so, if I wanted to build something like this, I would come in here and say, um, I want to build a custom skill. Let's see if it's even working. That didn't work. Hold on. Let me um Is it this? Is it this? Okay, never mind. Okay, I want to build a custom skill or custom agent that allows me to aggregate all the comments from GitHub live streams. Um we stream on LinkedIn, YouTube, Twitch, and Twitter using the StreamYard tool. Is this possible? And then I would actually like shift to plan mode. And so the right here you'll see that I'm now in plan mode. Uh okay. So you have autopilot mode. I don't know. Oh, you can't see. Hold on. Let me hide the comment here. Let me hide everything. Hold on. Banner hide. Okay. So you'll see autopilot mode. If you look down here, if I do shift tab, it goes into just like regular mode. And then you have plan mode. So like I would start in plan mode. Boom. So I would just tell copilot my idea just like that. And always I always like to choose the model. So I'm already on quad opus 4.6. And I like using quad opus 4.6 when I'm ideulating because it's just it just has very high reasoning ability. You can also use uh sonnet 4.5 or even GPT 5.4 for is also very good at reasoning. But you want to use a model that's um has like very good reasoning capabilities and with co-pilot you have access to all the models which I love models. Oh no sorry don't do that. Don't don't send another instruction while it's doing something. But um yeah so let's see what it does. So as you can see it says okay the user wants me to create a custom skill or agent that aggregates comments from GitHub liveream across multiple platforms. Let me explore the codebase to understand what we're working with. Oh no. Oh no. Oh no. So it's currently it's going to explor explore the current codebase. And the current codebase is the release video codebase. So that's not really going to help us here. Um I'm going to say do not explore the codebase. help me think through the um the problem please. But it's already exploring the codebase. So let's see. While copilot is working, you can send a like another thought to it that you have. It's going to show pending, but eventually it's going to pick up that that thing that you sent it while it's working and it's going to apply whatever you said. you'll see. But ideally, I would start by creating a completely new project, probably like an X.js project or just like an empty directory and then that way Copilot wouldn't have would have a blank slate and then would be able to just like go ahead and help me create like think through the agent. Okay, so now you see that it's it picked up what I said. do explore the codebase. Help me through think through the problem, please. Perfect. So, let's see if it's stop exploring and um help us think through here. But that's kind of how I started um when I was creating this release video agent. I said, "Hey, I have this idea for this agent. Here's how I wanted to here's what I wanted to do. Do you think this is possible?" And then I had it ask me questions back and forth. And then I said, "Yes, go." It drafted a plan. I adjusted the plan and then I said, "Okay, go forth and build." And then within like I think like 20 if not less than that 20 30 minutes I had like something to work on. The first video was very bad but um I think by the eighth video or so I I had something that was that I was able to show the team. Okay. Yeah. So it stopped exploring the codebase and now it went into sequential thinking and so it's using an MCP server called sequential thinking. interesting that it pulled for that tool, but you know, not surprising. Now, it's using the GitHub MCP server to do some web searching. And so, here you can see the differences between an agent skill and an MCP server. They work together, right? The skill gives your agent the ability to do something um more efficiently, more effectively, just better output. And MCP server gives your agent access to external tools that it can use to pull information. I'm totally new to this. This looks very fun. Honestly, I want to learn everything. It's so fun. It's so fun. I promise. It's It's so fun. It's so addictive. Uh follow me on LinkedIn. I post about this stuff all the time. Talking about GitHub, talking about agents, talking about fun projects I'm building. Um, it's it's actually really good. Oh, look at all these idea idea coach. Plan my day. Yeah, those would be really good agents, too. I'm using Gemini CLI. Is it better? So, I don't I don't like the word better, right? It's it's it's a matter of choosing the tool that you like to use. It's a matter of choosing the tool that aligns with how you work. I like using Copilot CLI. You like using Gemini CLI. Use the tool that you like. Use a tool that you like is what I will say. I wouldn't say better not better, right? We do have a lot of features though. Um, but use a tool that you like. Uh, I saw a bunch of you ask me who am I on, what am I on LinkedIn. Uh, just like Cadeshakur. You can look for my name Cadeshakur on LinkedIn. I don't even use CLI. Do you use GitHub compiler in VS Code? because I still use GitHub compiler in VS Code. It's also very good and you can also use custom agents with GitHub copilot in VS Code. Did you know that? So like if I go to VS Code for example and if I open up Copilot here. Okay, let me let me start from the beginning agent. So if I open up Copilot, I can choose these are all custom agents I have. And I can choose my release video agent. Isn't that incredible? And then I can choose Opus 4.6. Of course, I like using Opus 4.6 to build out the videos because I just get better results from it. I've used quite a few models and Opus 4.6 does the best job. So, I've been using it. And then I can pop the URL in here and hit enter. And GitHub Copilot in VS Code will also be able to create the videos for me. So, choose where you like to work. If you like to work in the editor, use the editor. If you like to work in the terminal, use the terminal. If you like to work on github.com, use copilot cloud agent. At GitHub, we want to make sure that we meet you where you like to work best. And so, that's why that's why we um have GitHub Copilot kind of everywhere that engineers work. We just want to make sure that you're happy. But um anyway, so yeah, if you want to build your own custom agents, just have a conversation with um your coding agents and say, "Hey, here's what I'm thinking to do. Here's what I want to do." And have it have it go in plan mode first. Don't just jump into building. Okay? You have to plan first and then build because what are you if you don't think through the plan, what are you building? You have no idea. So plan it out and then build it with your agents. Try it out. If it doesn't work the way you want, you you tell the agent, "Hey, when I do so and so, so and so happens, can you help me fix it?" It will say, "Hey, I would love to help you with that." And it will go forth and do. But right now, it's researching live stream comments APIs. All we need is human creativity and imagination to produce anything. That is 100% correct. I love that. Um, we are still very much needed. Copilot CLI is a part of GitHub Copilot. So, you do need to have a GitHub copilot subscription. You do get 30 days free on GitHub Copilot, though. So, if you've never used it before, you can start using it right now for free. I use copilot chat. Yeah, exactly. So you can use copilot chat and this is copilot chat. So within copilot chat you have access to agent mode, ask mode and plan mode, right? And then if you create your own custom agents like I have, you then have access to your custom agents that you can use and then you have access to multiple models. It's great. I love it. All right. So now it looks like our agent is coming back with something. So Jeremiah, this agent looks like it will be like a beast to build because we have to plug into public APIs. So Streamyard doesn't have a public API. I know the Twitter API is not free, but there is a Twitch and YouTube API that we can use to pull in comments. The LinkedIn API, I've used the LinkedIn API to build to build products before. It it is a little complicated and you do have to be a part of the LinkedIn partner program for approval. Yeah, but we could start with Twitch and YouTube. Okay, let's see. Bottom line, YouTube and Twitch are fully doable. LinkedIn is gated behind partner access. Uh, Twitter and X live chat is not possible via API. Before I create a plan, I have a few questions. Look at this output, right? And this output is coming from me saying, "Hey, I have this idea for a skill to aggregate all comments from live streams." Um, and so because of this, I can either select one of these options that it's giving me or I can type in my answer. But let me see what what it actually is saying. Do you want to collect comments in real time or after the stream ends? Post stream batch collection. This significantly changes architecture. I would say after the stream. Let me just do that. Compile CLI is like cloud code or open code both. So compiler CLI is like cloud code and it's like open code. So open code is open source. Cloud code is closed source and GitHub copilot and you know GitHub copilot you know works with GitHub and all that jazz. So yes, they're all coding agents. Great question. So, best to use voice to modify. Not necessarily. Not necessarily best. It's just sometimes I just don't want to type. Sometimes my brain is a little bit faster than my fingers can move. And so, I just find speaking to be nicer. What do you want to do with all the aggregated comments once collected? Great question. What do you want to do with the aggregated comments, Kadisha? I didn't even think of that. Export to a JSON CSV. I think I think Yeah, I think JSON CSV would be good like as a start cuz that would be easy for um the agent to do. Can this agents developed by us be monetized? I love that question. Yes. Yes, they can. Have you not seen the videos, you guys? I see so many videos of people building agents for businesses and selling it to them. Um, and it works. Like, for example, I'm I'm building an agent for my brother's uh business, and I'm not going to sell it to him, obviously, but people have been interested in it, and I'm just like, "Oh my gosh." So, you can build agents for anything. Go on YouTube and look it up. Okay. Do you remember I wasn't I wasn't in the tech industry in the '9s. Uh but now feels very much like the beginning of the internet where it's like the beginning of AI in business. You can absolutely um build these agents and you know monetize it. Great question. I love talking about monetization, but I talk about it a lot on my personal platforms, not on a like GitHub official live stream. Great questions are coming in. Bring keep keep the questions coming. Okay. Given the LinkedIn and Twitter X API limitation, are you okay with starting with just YouTube and Twitch? Yes, let's just start with YouTube and Twitch. I wonder if it will be able to build it while we're here. I can stay um for about like 10 more minutes. I do I have a meeting at 2 2:30. Okay. What form should this tool take? Should it be a custom agent you invoke through? Interesting. You say cloud code a standalone script or both? I would say like a custom agent. That's what we want to make. Okay. So, does the CLA work with multiple agents? Yes. Let me show you. So, I'm just going to go to a different tab. I can do slashmodels. Oh, hold on. You can't see cuz the thingy is blocking it. Hold on. Let me move hide your question. So yes, if you do slashmodels, we have a bunch of slash command slashcomands. And now you see we have all these models that you can use. Sounded 4.6, 4.5, 5.4, 5.3 codeex, models galore, folks. Models galore. So yes, copilot CLI has multiple models. And I think that's the beauty of one of the beautiful things about Copilot CLI because like for example I like to use GE the Gemini models for front-end work because it's it's a very good at front-end work. I like to use Opus 4.6 for like deep work kind of like like building out uh a complicated a complex agent such as the video release agent. I like to use Opus 4.6 for that. But if I'm just like, you know, asking a question about the codebase or asking a question about something not like super simple, but something that doesn't require such intense um thinking, I'll use like slotted 4.6 or GPT 5.2. GPT 5.4 is also very good at reasoning. So that's another model I pull for sometimes when coding. But having access to all these models in one place, it's so good because I don't have to go back go between different tools. It's one tools, multiple models, one flat fee, you know. It's great. Do you have developer API credentials for YouTube and Twitch? I don't I don't have it. I would need to get that. So hopefully that answers your question about working with multiple Oh, did I I misread your question? CLI can work with multiple agents. What do you mean by that? So you can have multiple custom agents in the CLI. Yes, you can have as many agents as you want. And actually you can install plugins. Um so plugins is a is a way for you to install like a group of agents. For example, for example, some of these some of these agents I have here are custom agents. Like the release video generator is custom, my security reviewer is custom. But a lot of these are installed through a plugin. And so how a plug-in work is, let's go to a different thing here. If I do slashplugin then I can do marketplace then I could do list I can see the marketplaces I have access to in the CLI and hopefully you can see here hopefully it's not too small. So I have the copilot plugins which is um automatically included and I have awesome copilot plugin which is also automatically included but within my marketplace I installed the anthropic agent skills and I installed the everything cla code uh marketplace here. Everything cloud code is a repository I came across the other day where this engineer literally put together so many awesome skills and so many awesome agents in one place that I can use. And so I installed it, right? And so from here I could do slashplugin. And I admit I'm still learning how to navigate this plugin slash command in copilot. How do I do it again? It's going to tell me. So you can do /plugin marketplace add. And then if you want to add this, you do. I should probably do a video or a blog post on how to use plugins and copilot CLI, right? I think yes. I think yes. But yeah, plugins allow you to install a bunch of stuff. And then you can do slashsklls once you install all those things. Don't mind don't mind this big arrow block. I need to look into it. Um, but you can see I have access to so many skills here. And then if I do slash agent, I have access I have so many agents, guys. It's It's kind of too much cuz then when I want to go down to the the video release agent, it's just me going down. But um yes, you can use so many so many um so many agents with Copilot CLI. It's an awesome tool. Let's see where we are here. All right. So now it's it's creating a plan. So the process is not fast. Yeah. 213. I have a few more minutes, maybe like five minutes. So, when using Copilot Agent on the cloud, can I set it to switch between AI models? Yes. Let me show you. So, if I go to GitHub here and if I go to agents here, boom chicka boom cha boom cha boom cha boom. Um, no active sessions. I should be able to create a session. Excuse me. Hold on. Let's go to Hold on. Let me stop sharing. Hold on. Hold on. Hold on. I don't want to show you anything. I'm not supposed to be showing you. So, look at Copilot working here. Let me go to a repository. Oh my goodness. Have I been muted all that time? What was the last thing you heard? Oh my goodness. Okay, so yes, I was saying that yes, you can you can use um multiple models on with um agent on the cloud. So, I was saying that you can choose between all these models on github.com. Now, now Streamyard is sending all your messages that you've been sending me. But, um, you can choose all these models on github.com, which I think is so great. Or you can choose all these, you can choose between these agents on github.com as well, which is also incredible. So you can use copilot coding, copilot cloud agent, you can use cloud code and you can use codeex to um do stuff for you on github.com. Did you know that? And this is all included in your GitHub copilot subscription. Mind blown. Mind blown. So yes, watch the terminal. I know, right? Let me go find it. Let's see what's going on there. What time is it? It's 2:17. Okay, folks. We're going to be wrapping up soon. Okay, so live stream comments aggregator. We have a plan. Okay, we have a plan for this custom agent. So, the approach build a custom agent. Copilot. You are copilot. Build a custom agent. is going to be called stream comments backed by TypeScript that fetches post stream comments from YouTube and Twitch normalizes them and exports it to JSON CSV. YouTube, Twitch, these platforms are not supported right now. It gave us some key deliverables like things for it to do and then dependencies to add would be Google APIs or raw fetch TypeScript. Key consideration. Twitch doesn't have a clean get all VOD chat API plan. Use Twitch downloader CLI as the most reliable approach. Okay. So now I can say accept the plan and build it. Accept the plan and build it on autopilot. Exit plan mode and I will prompt myself or suggest changes. Right now I'm not going to accept this plan because I'm in the release video repository and I don't want to mess it up. You know what I'm saying? I don't want to mess it up. But but this is how you can build your own custom agents using Copilot CLI. Tell it your idea. Go into plan mode. Tell it your idea. Answer the questions. Ask it questions. It's going to generate a plan. And then you're going to say boop, accept the plan, and it's going to go forth and build it for you. And then you can iterate from there. Folks, don't be scared of these things. You know, learn it. Um, use it to your advantage. Um, and have fun with it. It's a great time to be learning. It's a great time to be learning. So, I think I'm about to wrap up. I think I'm about to wrap up. I hope this session has been incredible for you as it's been for me. I have had a really awesome time. Operation cancelled by the user. Yes, I aborted because I don't want to mess up my repo. Okay. Okay. I don't want to mess it up. I hope this has been enlightening for you. I've had a lot of fun just hanging out with you and showing you this video that was created with AI. You guys, what a time to be alive. What a time to be alive. Thank you for joining. I hope you learned. I will see you next time for rubber up Thursday. I hope you have the most beautiful day. All right, see you on the internet. Bye bye y'all.