The Library Meta-Skill: How I Distribute PRIVATE Skills, Agents and Prompts

What's up engineers? Indie Dev Dan here. As you go from base agents to better agents to more agents to custom agents all the way up to orchestrator agents and even give your agents their own devices to operate, you'll quickly approach a problem. Where did you put that notion CLI skill? cracked review agent? Where is your new cloud code multi- aent planning/comand? If you're an engineer working on one to two repos, what I'm about to show you doesn't matter at all. If you install skills or plugins from the public internet without reviewing them, this video is not for you either. But if you're an engineer working on 10 plus code bases with agents, and you're building specialized private skills, agents, and prompts, this video was made for you. I have run into this problem as well. I've got skills everywhere. I've got prompts everywhere. I've got agents everywhere. I want to share a simple meta skill you can adopt to coordinate sharing your private skills, agents, and prompts across your codebases, teams, and agents. So, let me concisely define the problem. Us, the developer, we're

deploying skills, agents, and commands all over the place. We have duplicates all over the place in our 10 plus code bases. The problem is simple. We have endless skills everywhere. They're out of sync. They're duplicated and they're hard to coordinate across your engineering team. For example, you might have a deploy skill that works similarly across your codebases, but you've been losing track. You've been copying them. You've been just sharing them with your engineering co-workers and your agents, and they're just getting out of date. And this problem is exacerbated when you have multiple developers working across your team with multiple agents of their own. We're just copying skills. We're copying agents. And there's no real shared access. Now again, if you're working in a single codebase, this isn't really a problem. All your prompts, all your agentics, your prompts, skills, and agents, they're all just right there. But once you start operating in many code bases, you run into a serious problem. How can we solve this? We can solve this with a meta skill, a skill that unlocks other skills, agents, and prompts. I'm calling this the library. Call it whatever you want. In our previous video, we looked at how Stripe handles this problem. They use a meta tool called the tool shed. Here we're going to focus on a meta-kill and broaden this solution a little bit to cover skills, agents, and prompts. So we have one skill to unlock them all. I want a coordinated solution to distribute my agentics over and over across my devices, teams, and agents. All right, so that's the idea here. So what does this give us? This allows us to stack our skills, our agents, our prompts, and then if you want to, like I like to, you can put a just file on top. And so this is how I think about building with agents right now. agentic engineering. Everything starts with these four pieces and specifically these three pieces, right? These are the most important pieces. I think of skills as raw capabilities. Agents allows to attain scale and parallelism. And our prompts, just oneoff single file prompts, let you orchestrate these two levels below. A lot of people are overkilling skills doing everything inside the skill when you really have several other key primitives that you can work with. So this solution taps into that. The great part about the library is that we have a single reference point. So you can think about this like a package. json file or a pi project file where we basically are storing references. As you're going to see in a second here, the library file actually just points to existing GitHub repositories or local files. And that means staying synced between all of your projects is super simple. And this means when you have a single source of truth, you have yourself, you have your agents and your co-workers can just sync and manage this one library file. So that means that all you need to do is have this library file and the meta skill and you're good to go because the library file is just going to reference the GitHub repository. So this is exactly how this solution works. We have a reference to private GitHub repository. Very important. When you're building specialized solutions and you're getting paid for it, you're not putting this stuff out in public. Your top-notch skills, agents, and prompts should absolutely be private. I know a lot of vibe coders, you think that everything is just out there in the public. Trust me, it's not. We need a way to distribute this privately. Of course, we can also touch public repos. And with this system that you'll see in a moment here, we can also just reference local file paths. Something super critical to mention here, this is an agent first solution / library. It is a skill. There is no code associated with this. You can see here that is how it's kicked off. no matter who is using it. All right, and that gives us a lot of very powerful capabilities. I'm predicting we're going to see a lot of pure agentic solutions as we move through this year. Just another benefit to mention with a system like this, you of course get syncing across all of your devices. We're going to walk through a clear example of this between myself here and my Mac Mini agent that I showcased in our previous video. I've run into this problem over and over as I'm spinning up agents on my device and Mac Mini device. I've also got more devices coming in. There's a big trend going on right now that's been unlocked or should I say revealed by the whole open claw and claw movement. It's not having an agent run rampid on your device. It's having a team of agents that can operate very well against your specific domain problems. That's where all the value is. More on that later, but you can see how this solution aids that. We need a way to distribute all of our powerful prompts, agents, and skills. That's exactly what we're doing here. And so here's the full workflow. Just a couple more slides here. I just really want to nail the point home of why this is so important. And again, if you're an engineer that just operates on a couple code bases, this isn't for you, okay? You don't need this. Just use cloud code plugins. Just install from wherever. This is about private agentic distribution. So, here's what it looks like. And here's exactly what we're going to do in just a moment here. We're going to build first. We're then going to catalog with the library. add command. And this is going to update our catalog. This gives us pointers, references to existing code. Because the big idea here is that you don't want to interrupt the flow of when you're natively building out your prompt, skills, and agents inside of the actual value generating repository. And then you can, of course, reuse them. All right. So on your agent, your team and onto your Mac Mini or whatever your devices are. And this includes, you know, cloud sandboxes as well. So this is the full workflow. Build, catalog, distribute, use. And here's what the API looks like. It's all based on this single YAML file. So we can add items to the catalog, we can use items from push back updates to the source repositories, we can list, we can search, and then we can sync. And so sync is super important here. This is not just syncing your catalog. This is syncing the actual code bases to make sure that you have the latest version. In essence, this is very simple, but I needed to build out a concrete solution for this because I kept running into this problem and kept doing things the sloppy fast way, which is to just copy things and then they go out of sync, out of date very, very quickly. So, let's manage our agentics the right way. Our prompts, agents, and skills. Let's break down exactly how

this works. So, here I have my Mac Mini device. And what I want to do here is showcase how I'm now quickly and effortlessly syncing skills, prompts, and agents between my local device and my Asian device. I'm going to use my Mac Mini device here just because it's simple. It's very visual. You can literally see it here with screen sharing and my physical Mac mini device. But you can imagine you can set this up with any cloud-based system. As long as your agents or your team members have access to your team's private git repos, you should be good to go. Here I'm going to go ahead minimize this and let's just start by understanding the skill ergonomics. So if I open up Chrome here, you can see I have the library. And so this is the root template repository that I have cloned and I created the ID library. So NDubdan library. This is my personal library that I'm building up now. It's in its early versions. This is probably going to be the first and last time I'm going to be sharing this with anyone because there's going to be a lot of private value here. The most important file here is this. And if we click into this, you can see the structure of our reference file. And so we have skills, we have our agents, and we have our prompts. And so you can see here, you know, we're not actually storing anything except for the references to where these things are stored. What I want to do here is add some new skills to this. All right, I'm going to be adding my meta aentics, my meta prompt, my meta skill, and my meta agent. And I have these stored in a dedicated repository. You can see my agents have been doing a little bit of work in this repository, but I have this stored here. So, if we go into cloud again, private repository, this is not a public codebase. The library is public or I'm going to make it public here. So, you can clone this and start setting up your own library if you want to. Again, I I'm not trying to sell you on my vision of how to distribute your agentics. All I'm doing here is showcasing how you can. Every company with private specialized agentic tooling, you're going to want to do it your own way. I just want to showcase how I've thought through this problem as I've worked through and I'm using hundreds of prompts, agents, and skills across many, many repositories. I'm going to start building this up. And what I want to do here is add a couple of skills, specifically my meta agentics, my meta agent, my meta prime command, my meta prompt, and my meta skill. So that's what we're going to do right now. I'm in this repository here, and you can see that this is indeed a git repo. If we type get remote-v, you can see we have the origin repo there. Don't try to access it. It's private. And then we can do something like this tree. L2. And so you can just see exactly what I mentioned there. We have the agents. We have the commands. And we have our skills. And so if I boot up a cloud code instance here and I run / library and hit space, you can see our argument hint command and then neighbor details. And what I want to do here is just run list. So list is just going to tell me all of the references that are stored inside the library file. It is going to also make sure it's pulled so that it has the latest version. And you can see that this library is a skill and a repository inside of my global. cloud skills library. I have a few of these installed, but I don't have any meta agentics here. And so that's what we're going to add. So let's go ahead and here start a new session here. And then I'm going to add these new items, right? Because again, uh, if I do tree. cloud/l2, I have a couple new skills I want to add to the library so that they're accessible to all my devices, team members, and agents. So, I'll do the following. / library. Add. And I want meta agent, meta prime, meta prompt, meta skill, of course, and I'll say from GitHub URL. And that should be all I need here. Now, this library skill is going to load. It's going to find the add cookbook. This is how I like to separate the use cases of each command in the skill. Of course, before it makes any changes, it's going to pull. And now it's going to look for these actual items. So, it found them all locally. Now, it's going to actually add these to the library. Let me make this super clear here. I have the ID library engineer. So, I have this and this just contains reusable prompts, agents, skills. And then I have the ID library. And this is my personal reference library to prompts, agents, skills. Really important to distinguish this because you're going to have multiple of these. You know, maybe we have support, maybe we have sales here. And so these contain all the reusable pieces from who knows where, right? This can be private code bases, public code bases, and even file system references. But then we have the library, which is just that YAML file and the skill. So you can see this was added here. And now we can type library use to actually add these. I'm in the library where they were set up. So I'm not going to do that here. Instead, what I'll do here is open up a new terminal here. And let's create a temp directory. So I have an alias that just quickly creates, you know, a temp file here for us. And let's go ahead and fire up instead of claw code. Let's go ahead and use the pi coding agent. So with IPI, I'm going to boot up my own customized agent. You can see I was running with a Quinn 3. 5 plus. Let's go ahead and use a smarter model than that. Let's go and use that Sonic 6 model by Enthropic. Okay. And so ping. This should be pretty simple. What I'm going to do here is access the library skill. I'm going to use the use command. And then I want to say uh meta-star install locally into. cloud/skills. This is going to look at that library reference file and it's going to pull in those specific skills. Once again, these are references. These aren't stored on my device anywhere. These are references to a repository. My customized PI coding agent has damage control built in. So it can't run rm-rf commands. I'll say skip rm-rf so that it does not do this. But it's going to write the code to clone in and actually move these items into the rec directory. And that means I should have them here locally. If I open up a new terminal here and I type ls ls-la, you can see I haveclaude. If we do a tree. claude l2, we can see that we have skills. And if we just, you know, drop the level parameter there, we can see all the details of each one of these skills, right? So you can see my meta prompt has some examples, my meta skill has some docs, so on and so forth. So I just very quickly added references to my library YAML file. And you know, you can see this here. If we go ahead and go to my library file, you know, this was before. And if we just search for meta, you can see that I was testing this an hour ago just to make sure that this all worked. They're not in the library file. Okay. But if I refresh, you can see that they were just added. So if we search meta dash, you can see they're all there. Now this is a peer agent application. The application is the skill. So there's a very interesting paradigm that's emerging where you can encode entire applications into a single prompt into a single skill as anthropic likes to call it and that's all through this library skill here. That's how this works, right? So I have updated a reference file and now we can add these anywhere. And if I do tree. claw and let's do the uh L2. You know, I just added this to the local directory. So again, if I spin up an agent, you can see my customized pi coding agent automatically on bootup is going to show me where it's loading everything from. You can see in my global cloud, it's got 42. But here locally, you can now see those new activated meta aentics, my meta agent, my meta prime, meta prompt, and meta skill. So to be clear, we've covered meta aentics on the channel in the past. These are the skills that build the skills. It's a skill that builds the prompt and it's the agent. All right, so these are super important. Let this be your reminder if you don't have these agentics at the ready to help you build faster and faster. Definitely build this out. I know Claude Code has some, you know, build me a skill type of prompt. I highly recommend you build out your own so that you really understand how to do this better than anyone. And you know, really, if you're building out specialized solutions, you want to be building it your way. A big theme on the channel right now for the year of 2026 is increasing the trust we have in our agents. The best way to increase trust is to know exactly what they're doing and running. And that means down to the lines inside of your skills, your prompts, and your agents. But so you can see here that was added. Now, what I actually want to do here is add this globally. And so I use meta agents all the time. So what I'll do here is just to kind of showcase the example once again, you know, I'll go to a new temp directory, boot up a new instance, and we can go and just use cloud code for this library use command meta star add globally. And so now I'm going to have this added to my global cloud code directory. So again, we're not adding anything to the library. It already exists, right? I already have all of my, you know, meta agentics here. What I want to do is now on this device place them globally. I want them to be in that global name space. And you can see here at the top we have default dur. They're going to default to skills, agents, and prompts. And actually, here's a bug. I need to update this to commands because cloud code stores these in commands, not prompts. There's a default, there's a global, and I've built the system so that you can actually add whatever name you want to here. There could be default, global, and special. And special could route to an entirely different path. Here we go. We now have these enabled globally and so this means that we have access to this in any directory. All right, so I'm going to once again go into a new temp directory and that'll boot up my customize pi coding agent. So we can see them get loaded in and you can see there we now have 46. And if I type / meta you can see them popping up. So we added meta agent, metarimet and meta skill. Now, of course, the best part about this is that we can now deploy this solution on any device and just connect to this, right? Because the library is a reference file and the reference file can point to any repository you want. So now I'm going to be building up this engineering focused library with all of my prompts that I'll reuse across teams, devices, and agents. And speaking of agents, we can now open up screen sharing and get this installed on my agents device. All right, so now

I'm operating on my Mac Mini. I'll go ahead and fire up the terminal. Move to a temp directory. And so, let's go ahead and boot up a claude code instance here. Let's run a ping just to make sure we're all good on the device. Good. And actually, we don't even have the library installed on this device. So, let's go ahead and do that. I'll move to the cloud. And we need to go all the way into skills lsla. You can see that this Asian device has no skills right now. So, let's go ahead and equip it with some new skills. I'll get the uh repo reference to my library. If you, you know, build this out, do it your way. you're gonna have your own library. So I'm going to go ahead and just copy this clone. Bang. And I want this inside of library. So I'm going to give it a another parameter there. CD into library. And now you can see we have this here. And we can do a more on the library file. So you can see what this actually looks like. And it is exactly what you'd expect. You can see there is our meta agent. And then we can just kind of work through the rest of this. Okay. So great. So now it's on my system, right? Because now it is inside of mycloud user directory, also known as just the global skills directory. Now we can get to work, right? So now we can install our meta skill or any skill that we have referenced in our library. Right? So if we come in here, you can see all these and now we can go ahead and add these to our agent device wherever we need them. So it could be in a specific directory or it could be global. In this case, uh my meta agents are really important. I like to have these installed globally on the machine they're operating on. We can go ahead and use cloud code here library. And I'll go ahead and add use once again. And I'll just say meta-star install globally. And so to be clear right now, if I move to another temp directory and I use ipi, you can see here that all I have is that one global skill, right? The library skill. But after this runs, you're going to see that we're going to have five global skills. There you go. So it's cloning them in. It's installing. It's making sure we're getting the very, very latest version. This is all on my Asian device. This could be on a cloud in an Asian sandbox. This could be one of my teammates, one of my team members. So now we have all of them. I'm going to close my customized agent harness here. We talked about the PI coding agent in a previous video. That one went absolutely viral. I'll link that in the description if you're interested on how you can customize your own personal agent harness. Cloud code is great, but PI is the only real competitor because it lets you customize the Asian harness. As you can see here on boot up, it directly tells me where all of my skills are loaded, and we got those five total skills up from one. This is because I customized my Asian harness. thanks to the PI coding agent. Again, link in the description for that. But you can see here we now have these global skills available on my agent device thanks to the / library skill. And now we can run it. We're in a temp directory. I'll just say a meta prompt. So this is my skill that builds prompts. These are very important. They let you move very fast in a structured way. And I'll have it just build something random for us. So let's see what's useful for a new agent device. We can say something like this. uh create a new bash profile bash functions based on an input arg and we're going to go ahead and run the quinn 3. 5 plus model here. Let's go and see how this works. Right? So this is going to create a prompt for us. It's creating a custom slash command bash function. There's the argument hint. It knows how I like to create prompts. If you've been with the channel, you know exactly my prompt format. We have the purpose. We have the workflow. We have the instructions. We have the variables. So on and so forth. Looks like Quinn 3. 5 did a pretty good job. Let's actually go ahead open up a new terminal. Just going to quickly copy this move. So we can see ls. We don't want ls. We want tree. And we want tree. cloud since it's hidden. There we go. Cloud commands bash function. And so now we can execute the bash function. I assume this is going to create a bash function in bash profile. There's a function request output format. Nice. Created a couple extra variables for us. Include comments, include examples, instructions, workflow. Looks great. This was because I had my metaprompt written out and my metaprompt is like 800 lines long, right? So, it's very detailed. I've templated my engineering into how I build my prompts and I wanted to scale this across all my devices. The prompt this important needs to be the exact same across all my team members, all of my work on all my code bases and on all my devices, both human operated and agent operated. This is why a library of your agentics is so important. Last thing I want to showcase here is actually updating this. Now we have 10 devices using our meta skills. How will they update it and push it back to the source the one source of truth? And you know that's the important part of this. This is very much like a package system without versioning. I just want the latest version. But this is very much like a package system for prompt skills and agents, right? A library of them because we're storing once again references. And what I want to do here is 3. 1 flashlight ipi update the scale. md colon summary make this return a table format instead. And I think it's doing like a yaml format right now or something. I don't know. There we go. So it's reading that file in. now it's going to report this in a table structure instead of looks like we were doing YAML before. Fantastic. Change has been made. And now we need to push it back to the actual repository because once again let's remember that the file right the actual library file is a reference right so this references code bases right it itself is not the codebase it's a reference to the codebase we're gambling here if we use flashlight on this I'm going to bump this up to something more stable nothing more stable than called sonnet nois models we'll jump up to a sonnet 4. 6 And it'll say library push and then I'll say meta prompt. And so that's all it needs here. I'm using the push command. And so it's going to find the global version. There's no local version here. And then it's going to actually push this back up to the source repository. Move this to the side. And then open up Chrome just so I can show you exactly how this is going to work. You can see my last Metaprompt update. This was coming in, you know, about an hour ago. And you can see this device just finished. So now if I refresh this, we should get an update inside the source codebase, right? So 1 hour ago. Refresh just now. So this was freshly updated right from our agent device right here. Very powerful stuff and it's a relatively simple idea. It's just too important to miss at a certain scale. That's the whole idea. So now I can distribute my agentics. Agentics are prompts, agents, skills, things that make your agentics system run. So now I can quickly distribute these across all my devices, across all my team members, across all of my agents. And this super big unlock here is the following. If we close this out and hop back to the original terminal here and go to cloud skill library, this is a pure agent application. It really only runs in skills. md and library. yaml. It's a purely agentic application. And that means that I can do everything I just did very quickly with an agent. An entire agent could run that entire workflow or chain together pieces of the individual workflows. And you know, one of the ways I like to organize my skills is like this tree dot. The cookbook is essentially individual agentic workflows that the skill describes when to use each one of them, right? And so you can see here very cleanly in plain English. Great for humans, great for agents. I have them all written out by their exact command name. So, this is the library. This is how I'm organizing and redistributing all of my skills, agents, and prompts. Now, if you want to get started with something like this, you can check out the library, and you'll start with a blank library file just like this, right? There's nothing in here. Update the defaults to whatever you want, and then you can start adding them. All the documentation you need to get started with this is going to be here. Again, I just want to reemphasize this point. I don't care if you use my library. I'm not trying to like sell my repo or anything. It's just a valuable idea if you're scaling to multiple devices, multiple team members and multiple agents. All right, remember the agentic path for all of us now is base agent, better agent, more agents, custom agents and then we start orchestrating things with our orchestrator agent and the orchestrator agent is a big trend right now. Calling it a trend is frankly stupid. It's a really important conceptual idea for agentic engineering. Once you get to certain levels of scale, the only way to continue is to create a leader. co-worker, is to create an orchestrator that can operate entire devices for you and entire systems for you, entire domains for you. And in order to really build that out and scale that and to reuse your grade a Gentics across all your devices, you're going to want something like the library so that you can quickly organize all of your capabilities into one location. All right, so that's all I've done here. This is the idea. I wanted to sit down and communicate with you today. This is the library meta skill. The skill to unlock and redistribute your private skills, agents, and prompts to help you maintain your advantage in the age of agents. You know where to find me every single Monday. Stay focused and keep building.