← 返回视频(38) The Library Meta-Skill: How I Distribute PRIVATE Skills, Agents and PromptsYouTube/2026年3月25日原始链接↗Transcript▲0:00What's up engineers? Indie Dev Dan here.0:02As you go from base agents to better0:04agents to more agents to custom agents0:07all the way up to orchestrator agents0:10and even give your agents their own0:12devices to operate, you'll quickly0:15approach a problem. Where did you put0:18that notion CLI skill? Where did you put0:20that cracked review agent? Where is your0:23new cloud code multi- aent0:25planning/comand? If you're an engineer0:27working on one to two repos, what I'm0:30about to show you doesn't matter at all.0:32If you install skills or plugins from0:35the public internet without reviewing0:37them, this video is not for you either.0:39But if you're an engineer working on 100:41plus code bases with agents, and you're0:44building specialized private skills,0:47agents, and prompts, this video was made0:50for you. I have run into this problem as0:52well. I've got skills everywhere. I've0:54got prompts everywhere. I've got agents0:56everywhere. I want to share a simple0:59meta skill you can adopt to coordinate1:02sharing your private skills, agents, and1:05prompts across your codebases, teams,1:08and agents. So, let me concisely define1:11the problem. Us, the developer, we're1:13deploying skills, agents, and commands1:16all over the place. We have duplicates1:18all over the place in our 10 plus code1:20bases. The problem is simple. We have1:22endless skills everywhere. They're out1:24of sync. They're duplicated and they're1:25hard to coordinate across your1:27engineering team. For example, you might1:29have a deploy skill that works similarly1:31across your codebases, but you've been1:33losing track. You've been copying them.1:35You've been just sharing them with your1:37engineering co-workers and your agents,1:39and they're just getting out of date.1:41And this problem is exacerbated when you1:43have multiple developers working across1:45your team with multiple agents of their1:47own. We're just copying skills. We're1:49copying agents. And there's no real1:51shared access. Now again, if you're1:53working in a single codebase, this isn't1:55really a problem. All your prompts, all1:57your agentics, your prompts, skills, and1:59agents, they're all just right there.2:01But once you start operating in many2:02code bases, you run into a serious2:04problem. How can we solve this? We can2:06solve this with a meta skill, a skill2:08that unlocks other skills, agents, and2:11prompts. I'm calling this the library.2:13Call it whatever you want. In our2:14previous video, we looked at how Stripe2:16handles this problem. They use a meta2:19tool called the tool shed. Here we're2:21going to focus on a meta-kill and2:23broaden this solution a little bit to2:25cover skills, agents, and prompts. So we2:28have one skill to unlock them all. I2:31want a coordinated solution to2:33distribute my agentics over and over and2:35over across my devices, teams, and2:39agents. All right, so that's the idea2:40here. So what does this give us? This2:42allows us to stack our skills, our2:45agents, our prompts, and then if you2:47want to, like I like to, you can put a2:49just file on top. And so this is how I2:51think about building with agents right2:53now. This is how I think about agentic2:55engineering. Everything starts with2:56these four pieces and specifically these2:59three pieces, right? These are the most3:00important pieces. I think of skills as3:02raw capabilities. Agents allows to3:05attain scale and parallelism. And our3:07prompts, just oneoff single file3:09prompts, let you orchestrate these two3:11levels below. A lot of people are3:13overkilling skills doing everything3:15inside the skill when you really have3:17several other key primitives that you3:18can work with. So this solution taps3:20into that. The great part about the3:21library is that we have a single3:24reference point. So you can think about3:25this like a package.json file or a pi3:28project file where we basically are3:31storing references. As you're going to3:32see in a second here, the library file3:35actually just points to existing GitHub3:37repositories or local files. And that3:40means staying synced between all of your3:42projects is super simple. And this means3:44when you have a single source of truth,3:46you have yourself, you have your agents3:48and your co-workers can just sync and3:50manage this one library file. So that3:52means that all you need to do is have3:54this library file and the meta skill and3:56you're good to go because the library3:58file is just going to reference the3:59GitHub repository. So this is exactly4:01how this solution works. We have a4:03reference to private GitHub repository.4:06Very important. When you're building4:07specialized solutions and you're getting4:09paid for it, you're not putting this4:10stuff out in public. Your top-notch4:13skills, agents, and prompts should4:15absolutely be private. I know a lot of4:17vibe coders, you think that everything4:19is just out there in the public. Trust4:20me, it's not. We need a way to4:22distribute this privately. Of course, we4:25can also touch public repos. And with4:27this system that you'll see in a moment4:28here, we can also just reference local4:30file paths. Something super critical to4:32mention here, this is an agent first4:34solution / library. It is a skill. There4:37is no code associated with this. You can4:39see here that is how it's kicked off. no4:41matter who is using it. All right, and4:43that gives us a lot of very powerful4:44capabilities. I'm predicting we're going4:46to see a lot of pure agentic solutions4:49as we move through this year. Just4:51another benefit to mention with a system4:53like this, you of course get syncing4:55across all of your devices. We're going4:56to walk through a clear example of this4:59between myself here and my Mac Mini5:01agent that I showcased in our previous5:03video. I've run into this problem over5:04and over as I'm spinning up agents on my5:07device and I'm spinning up agents on my5:09Mac Mini device. I've also got more5:11devices coming in. There's a big trend5:13going on right now that's been unlocked5:15or should I say revealed by the whole5:18open claw and claw movement. It's not5:20having an agent run rampid on your5:22device. It's having a team of agents5:24that can operate very very well against5:26your specific domain problems. That's5:28where all the value is. More on that5:29later, but you can see how this solution5:31aids that. We need a way to distribute5:34all of our powerful prompts, agents, and5:37skills. That's exactly what we're doing5:38here. And so here's the full workflow.5:39Just a couple more slides here. I just5:41really want to nail the point home of5:42why this is so important. And again, if5:44you're an engineer that just operates on5:46a couple code bases, this isn't for you,5:48okay? You don't need this. Just use5:49cloud code plugins. Just install from5:52wherever. This is about private agentic5:54distribution. So, here's what it looks5:55like. And here's exactly what we're5:56going to do in just a moment here. We're5:57going to build first. We're then going5:59to catalog with the library.add command.6:01And this is going to update our catalog.6:04This gives us pointers, references to6:07existing code. Because the big idea here6:09is that you don't want to interrupt the6:11flow of when you're natively building6:13out your prompt, skills, and agents6:16inside of the actual value generating6:18repository. And then you can, of course,6:20reuse them. All right. So on your agent,6:22your team and onto your Mac Mini or6:24whatever your devices are. And this6:26includes, you know, cloud sandboxes as6:28well. So this is the full workflow.6:30Build, catalog, distribute, use. And6:32here's what the API looks like. It's all6:34based on this single YAML file. So we6:37can add items to the catalog, we can use6:40items from the catalog, we can push back6:42updates to the source repositories, we6:44can list, we can search, and then we can6:46sync. And so sync is super important6:48here. This is not just syncing your6:50catalog. This is syncing the actual code6:52bases to make sure that you have the6:54latest version. In essence, this is very6:56very simple, but I needed to build out a6:58concrete solution for this because I7:00kept running into this problem and kept7:02doing things the sloppy fast way, which7:04is to just copy things and then they go7:06out of sync, out of date very, very7:08quickly. So, let's manage our agentics7:10the right way. Our prompts, agents, and7:12skills. Let's break down exactly how7:14this works. So, here I have my Mac Mini7:16device. And what I want to do here is7:17showcase how I'm now quickly and7:20effortlessly syncing skills, prompts,7:22and agents between my local device and7:25my Asian device. I'm going to use my Mac7:27Mini device here just because it's7:28simple. It's very visual. You can7:30literally see it here with screen7:32sharing and my physical Mac mini device.7:34But you can imagine you can set this up7:35with any cloud-based system. As long as7:38your agents or your team members have7:40access to your team's private git repos,7:43you should be good to go. Here I'm going7:44to go ahead minimize this and let's just7:46start by understanding the skill7:48ergonomics. So if I open up Chrome here,7:50you can see I have the library. And so7:52this is the root template repository7:55that I have cloned and I created the ID7:58library. So NDubdan library. This is my8:01personal library that I'm building up8:02now. It's in its early versions. This is8:04probably going to be the first and last8:05time I'm going to be sharing this with8:07anyone because there's going to be a lot8:08of private value here. The most8:10important file here is this. And if we8:12click into this, you can see the8:13structure of our reference file. And so8:16we have skills, we have our agents, and8:18we have our prompts. And so you can see8:21here, you know, we're not actually8:22storing anything except for the8:24references to where these things are8:25stored. What I want to do here is add8:27some new skills to this. All right, I'm8:30going to be adding my meta aentics, my8:32meta prompt, my meta skill, and my meta8:35agent. And I have these stored in a8:37dedicated repository. You can see my8:39agents have been doing a little bit of8:40work in this repository, but I have this8:42stored here. So, if we go into cloud8:44again, private repository, this is not a8:45public codebase. The library is public8:47or I'm going to make it public here. So,8:49you can clone this and start setting up8:50your own library if you want to. Again,8:53I I'm not trying to sell you on my8:54vision of how to distribute your8:56agentics. All I'm doing here is8:58showcasing how you can. Every company9:01with private specialized agentic9:04tooling, you're going to want to do it9:05your own way. I just want to showcase9:07how I've thought through this problem as9:09I've worked through and I'm using9:10hundreds of prompts, agents, and skills9:13across many, many repositories. I'm9:14going to start building this up. And9:16what I want to do here is add a couple9:18of skills, specifically my meta9:20agentics, my meta agent, my meta prime9:24command, my meta prompt, and my meta9:26skill. So that's what we're going to do9:27right now. I'm in this repository here,9:29and you can see that this is indeed a9:32git repo. If we type get remote-v, you9:35can see we have the origin repo there.9:37Don't try to access it. It's private.9:38And then we can do something like this9:40tree.9:41L2. And so you can just see exactly what9:44I mentioned there. We have the agents.9:46We have the commands. And we have our9:48skills. And so if I boot up a cloud code9:51instance here and I run / library and9:54hit space, you can see our argument hint9:57command and then neighbor details. And9:59what I want to do here is just run list.10:01So list is just going to tell me all of10:03the references that are stored inside10:05the library file. It is going to also10:07make sure it's pulled so that it has the10:09latest version. And you can see that10:11this library is a skill and a repository10:14inside of my global.cloud skills10:17library. I have a few of these10:19installed, but I don't have any meta10:22agentics here. And so that's what we're10:24going to add. So let's go ahead and here10:26start a new session here. And then I'm10:27going to add these new items, right?10:30Because again, uh, if I do10:32tree.cloud/l2,10:33I have a couple new skills I want to add10:36to the library so that they're10:37accessible to all my devices, team10:40members, and agents. So, I'll do the10:41following. / library. Add. And I want10:45meta agent, meta prime, meta prompt,10:49meta skill, of course, and I'll say from10:53GitHub URL. And that should be all I10:55need here. Now, this library skill is10:57going to load. It's going to find the10:58add cookbook. This is how I like to11:01separate the use cases of each command11:03in the skill. Of course, before it makes11:05any changes, it's going to pull. And now11:07it's going to look for these actual11:09items. So, it found them all locally.11:11Now, it's going to actually add these to11:13the library. Let me make this super11:15clear here. I have the ID library11:18engineer. So, I have this and this just11:20contains reusable prompts, agents,11:23skills. And then I have the ID library.11:26And this is my personal reference11:29library to prompts, agents, skills.11:33Really important to distinguish this11:34because you're going to have multiple of11:36these. You know, maybe we have support,11:38maybe we have sales here. And so these11:40contain all the reusable pieces from who11:42knows where, right? This can be private11:44code bases, public code bases, and even11:46file system references. But then we have11:47the library, which is just that YAML11:49file and the skill. So you can see this11:52was added here. And now we can type11:53library use to actually add these. I'm11:56in the library where they were set up.11:59So I'm not going to do that here.12:00Instead, what I'll do here is open up a12:03new terminal here. And let's create a12:05temp directory. So I have an alias that12:06just quickly creates, you know, a temp12:08file here for us. And let's go ahead and12:10fire up instead of claw code. Let's go12:12ahead and use the pi coding agent. So12:14with IPI, I'm going to boot up my own12:17customized agent. You can see I was12:18running with a Quinn 3.5 plus. Let's go12:21ahead and use a smarter model than that.12:23Let's go and use that Sonic 6 model by12:25Enthropic. Okay. And so ping. This12:26should be pretty simple. What I'm going12:28to do here is access the library skill.12:30I'm going to use the use command. And12:32then I want to say uh meta-star install12:35locally into.cloud/skills.12:39This is going to look at that library12:41reference file and it's going to pull in12:43those specific skills. Once again, these12:45are references. These aren't stored on12:47my device anywhere. These are references12:48to a repository. My customized PI coding12:52agent has damage control built in. So it12:54can't run rm-rf commands. I'll say skip12:56rm-rf so that it does not do this. But12:59it's going to write the code to clone in13:01and actually move these items into the13:03rec directory. And that means I should13:05have them here locally. If I open up a13:07new terminal here and I type ls ls-la,13:09you can see I haveclaude. If we do a13:12tree.claude13:13l2, we can see that we have skills. And13:16if we just, you know, drop the level13:18parameter there, we can see all the13:20details of each one of these skills,13:21right? So you can see my meta prompt has13:23some examples, my meta skill has some13:25docs, so on and so forth. So I just very13:28very quickly added references to my13:30library YAML file. And you know, you can13:33see this here. If we go ahead and go to13:35my library file, you know, this was13:37before. And if we just search for meta,13:39you can see that I was testing this an13:41hour ago just to make sure that this all13:42worked. They're not in the library file.13:44Okay. But if I refresh, you can see that13:46they were just added. So if we search13:48meta dash, you can see they're all13:50there. Now this is a peer agent13:52application. The application is the13:55skill. So there's a very interesting13:56paradigm that's emerging where you can13:58encode entire applications into a single14:01prompt into a single skill as anthropic14:04likes to call it and that's all through14:06this library skill here. That's how this14:08works, right? So I have updated a14:10reference file and now we can add these14:12anywhere. And if I do tree.claw14:15and let's do the uh L2. You know, I just14:18added this to the local directory. So14:20again, if I spin up an agent, you can14:21see my customized pi coding agent14:24automatically on bootup is going to show14:26me where it's loading everything from.14:28You can see in my global cloud, it's got14:3042. But here locally, you can now see14:32those new activated meta aentics, my14:35meta agent, my meta prime, meta prompt,14:38and meta skill. So to be clear, we've14:40covered meta aentics on the channel in14:42the past. These are the skills that14:44build the skills. It's a skill that14:45builds the prompt and it's the skill14:48that builds the agent. All right, so14:50these are super super important. Let14:51this be your reminder if you don't have14:54these agentics at the ready to help you14:57build faster and faster. Definitely14:59definitely build this out. I know Claude15:00Code has some, you know, build me a15:02skill type of prompt. I highly recommend15:04you build out your own so that you15:05really understand how to do this better15:07than anyone. And you know, really, if15:08you're building out specialized15:10solutions, you want to be building it15:12your way. A big theme on the channel15:14right now for the year of 2026 is15:16increasing the trust we have in our15:18agents. The best way to increase trust15:19is to know exactly what they're doing15:21and know exactly what they're running.15:22And that means down to the lines inside15:26of your skills, your prompts, and your15:28agents. But so you can see here that was15:30added. Now, what I actually want to do15:32here is add this globally. And so I use15:34meta agents all the time. So what I'll15:36do here is just to kind of showcase the15:38example once again, you know, I'll go to15:40a new temp directory, boot up a new15:42instance, and we can go and just use15:44cloud code for this library use command15:47meta star add globally. And so now I'm15:50going to have this added to my global15:53cloud code directory. So again, we're15:55not adding anything to the library. It15:58already exists, right? I already have16:00all of my, you know, meta agentics here.16:02What I want to do is now on this device16:05place them globally. I want them to be16:07in that global name space. And you can16:08see here at the top we have default dur.16:10They're going to default to skills,16:12agents, and prompts. And actually,16:13here's a bug. I need to update this to16:15commands because cloud code stores these16:17in commands, not prompts. There's a16:19default, there's a global, and I've16:20built the system so that you can16:22actually add whatever name you want to16:24here. There could be default, global,16:26and special. And special could route to16:28an entirely different path. Here we go.16:30We now have these enabled globally and16:32so this means that we have access to16:34this in any directory. All right, so I'm16:35going to once again go into a new temp16:37directory and that'll boot up my16:38customize pi coding agent. So we can see16:41them get loaded in and you can see there16:42we now have 46. And if I type / meta you16:46can see them popping up. So we added16:47meta agent, metarimet16:50and meta skill. Now, of course, the best16:52part about this is that we can now16:54deploy this solution on any device and16:57just connect to this, right? Because the16:58library is a reference file and the17:00reference file can point to any17:02repository you want. So now I'm going to17:04be building up this engineering focused17:06library with all of my prompts that I'll17:08reuse across teams, devices, and agents.17:11And speaking of agents, we can now open17:13up screen sharing and get this installed17:16on my agents device. All right, so now17:17I'm operating on my Mac Mini. I'll go17:19ahead and fire up the terminal. Move to17:20a temp directory. And so, let's go ahead17:22and boot up a claude code instance here.17:24Let's run a ping just to make sure we're17:25all good on the device. Good. And17:27actually, we don't even have the library17:29installed on this device. So, let's go17:30ahead and do that. I'll move to the17:32cloud. And we need to go all the way17:33into skills lsla. You can see that this17:37Asian device has no skills right now.17:39So, let's go ahead and equip it with17:40some new skills. I'll get the uh repo17:42reference to my library. If you, you17:45know, build this out, do it your way.17:46you're gonna have your own library. So17:48I'm going to go ahead and just copy this17:50clone. Bang. And I want this inside of17:53library. So I'm going to give it a17:55another parameter there. CD into17:56library. And now you can see we have17:58this here. And we can do a more on the18:00library file. So you can see what this18:01actually looks like. And it is exactly18:02what you'd expect. You can see there is18:04our meta agent. And then we can just18:05kind of work through the rest of this.18:06Okay. So great. So now it's on my18:08system, right? Because now it is inside18:10of mycloud user directory, also known as18:13just the global skills directory. Now we18:15can get to work, right? So now we can18:16install our meta skill or any skill that18:19we have referenced in our library.18:21Right? So if we come in here, you can18:22see all these and now we can go ahead18:24and add these to our agent device18:27wherever we need them. So it could be in18:29a specific directory or it could be18:30global. In this case, uh my meta agents18:33are really really important. I like to18:34have these installed globally on the18:36machine they're operating on. We can go18:38ahead and use cloud code here library.18:40And I'll go ahead and add use once18:43again. And I'll just say meta-star18:45install globally. And so to be clear18:49right now, if I move to another temp18:51directory and I use ipi, you can see18:53here that all I have is that one global18:56skill, right? The library skill. But18:58after this runs, you're going to see18:59that we're going to have five global19:01skills. There you go. So it's cloning19:03them in. It's installing. It's making19:04sure we're getting the very, very latest19:06version. This is all on my Asian device.19:08This could be on a cloud device. This19:09could be in an Asian sandbox. This could19:11be one of my teammates, one of my team19:12members. So now we have all of them. I'm19:13going to close my customized agent19:16harness here. We talked about the PI19:17coding agent in a previous video. That19:19one went absolutely viral. I'll link19:21that in the description if you're19:22interested on how you can customize your19:24own personal agent harness. Cloud code19:26is great, but PI is the only real19:28competitor because it lets you customize19:31the Asian harness. As you can see here19:32on boot up, it directly tells me where19:35all of my skills are loaded, and we got19:37those five total skills up from one.19:39This is because I customized my Asian19:41harness. thanks to the PI coding agent.19:44Again, link in the description for that.19:45But you can see here we now have these19:46global skills available on my agent19:49device thanks to the / library skill.19:51And now we can run it. We're in a temp19:53directory. I'll just say a meta prompt.19:55So this is my skill that builds prompts.19:58These are very very important. They let19:59you move very very fast in in a20:01structured way. And I'll have it just20:03build something random for us. So let's20:04see what's useful for a new agent20:06device. We can say something like this.20:08uh create a new bash profile bash20:12functions based on an input arg and20:16we're going to go ahead and run the20:17quinn 3.5 plus model here. Let's go and20:19see how this works. Right? So this is20:20going to create a prompt for us. It's20:23creating a custom slash command bash20:24function. There's the argument hint. It20:26knows how I like to create prompts. If20:29you've been with the channel, you know20:30exactly my prompt format. We have the20:32purpose. We have the workflow. We have20:33the instructions. We have the variables.20:35So on and so forth. Looks like Quinn 3.520:37did a pretty good job. Let's actually go20:38ahead open up a new terminal. Just going20:40to quickly copy this move. So we can see20:43ls. We don't want ls. We want tree. And20:45we want tree.cloud since it's hidden.20:47There we go. Cloud commands bash20:49function. And so now we can execute the20:51bash function. I assume this is going to20:54create a bash function in bash profile.20:56There's a function request output20:58format. Nice. Created a couple extra21:00variables for us. Include comments,21:02include examples, instructions,21:04workflow. Looks great. This was because21:06I had my metaprompt written out and my21:09metaprompt is like 800 lines long,21:11right? So, it's very detailed. I've21:13templated my engineering into how I21:15build my prompts and I wanted to scale21:18this across all my devices. The prompt21:20this important needs to be the exact21:22same across all my team members, all of21:24my work on all my code bases and on all21:27my devices, both human operated and21:30agent operated. This is why a library of21:33your agentics is so important. Last21:35thing I want to showcase here is21:37actually updating this. Now we have 1021:40devices using our meta skills. How will21:43they update it and push it back to the21:44source the one source of truth? And you21:47know that's the important part of this.21:48This is very very much like a package21:50system without versioning. I just want21:52the latest version. But this is very21:54much like a package system for prompt21:57skills and agents, right? A library of21:59them because we're storing once again22:00references. And what I want to do here22:02is 3.1 flashlight ipi update the22:06scale.md colon summary make this return22:10a table format instead. And I think it's22:12doing like a yaml format right now or22:14something. I don't know. There we go. So22:15it's reading that file in. There we go.22:17So now it's going to report this in a22:19table structure instead of looks like we22:21were doing YAML before. Fantastic.22:23Change has been made. And now we need to22:24push it back to the actual repository22:27because once again let's remember that22:29the file right the actual library file22:32is a reference right so this references22:35code bases right it itself is not the22:38codebase it's a reference to the22:39codebase we're gambling here if we use22:42flashlight on this I'm going to bump22:43this up to something more stable nothing22:46more stable than called sonnet nois22:48models we'll jump up to a sonnet 4.6 And22:50it'll say library push and then I'll say22:54meta prompt. And so that's all it needs22:56here. I'm using the push command. And so22:58it's going to find the global version.22:59There's no local version here. And then23:00it's going to actually push this back up23:02to the source repository. Move this to23:05the side. And then open up Chrome just23:08so I can show you exactly how this is23:09going to work. You can see my last23:11Metaprompt update. This was coming in,23:13you know, about an hour ago. And you can23:15see this device just finished. So now if23:17I refresh this, we should get an update23:20inside the source codebase, right? So 123:22hour ago. Refresh just now. So this was23:25freshly updated right from our agent23:28device right here. Very very powerful23:29stuff and it's it's a relatively simple23:32idea. It's just too important to miss at23:35a certain scale. That's the whole idea.23:37So now I can distribute my agentics.23:39Agentics are prompts, agents, skills,23:42things that make your agentics system23:44run. So now I can quickly distribute23:46these across all my devices, across all23:48my team members, across all of my23:51agents. And this super super big unlock23:53here is the following. If we close this23:56out and hop back to the original23:58terminal here and go to cloud skill24:02library, this is a pure agent24:05application. It really only runs in24:07skills.md and library.yaml. It's a24:11purely agentic application. And that24:14means that I can do everything I just24:16did very very quickly with an agent. An24:18entire agent could run that entire24:20workflow or chain together pieces of the24:22individual workflows. And you know, one24:24of the ways I like to organize my skills24:27is like this tree dot. The cookbook is24:30essentially individual agentic workflows24:32that the skill describes when to use24:35each one of them, right? And so you can24:36see here very very cleanly in plain24:38English. Great for humans, great for24:40agents. I have them all written out by24:42their exact command name. So, this is24:44the library. This is how I'm organizing24:46and redistributing all of my skills,24:49agents, and prompts. Now, if you want to24:51get started with something like this,24:52you can check out the library, and24:54you'll start with a blank library file24:57just like this, right? There's nothing24:58in here. Update the defaults to whatever24:59you want, and then you can start adding25:01them. All the documentation you need to25:02get started with this is going to be25:03here. Again, I just want to reemphasize25:05this point. I don't care if you use my25:07library. I'm not trying to like sell my25:09repo or anything. It's just a valuable25:11idea if you're scaling to multiple25:13devices, multiple team members and25:15multiple agents. All right, remember the25:17agentic path for all of us now is base25:20agent, better agent, more agents, custom25:23agents and then we start orchestrating25:25things with our orchestrator agent and25:27the orchestrator agent is a big big25:29trend right now. Calling it a trend is25:31frankly stupid. It's a really really25:32important conceptual idea for agentic25:35engineering. Once you get to certain25:37levels of scale, the only way to25:39continue is to create a leader. Is to25:42create a co-worker, is to create an25:45orchestrator that can operate entire25:47devices for you and entire systems for25:50you, entire domains for you. And in25:54order to really build that out and scale25:55that and to reuse your grade a Gentics25:58across all your devices, you're going to26:00want something like the library so that26:03you can quickly organize all of your26:05capabilities into one location. All26:07right, so that's all I've done here.26:09This is the idea. I wanted to sit down26:11and communicate with you today. This is26:13the library meta skill. The skill to26:15unlock and redistribute your private26:19skills, agents, and prompts to help you26:21maintain your advantage in the age of26:24agents. You know where to find me every26:26single Monday. Stay focused and keep26:29building.