Anthropic Just Dropped Their Internal Skills Playbook (copy this)

So Anthropic just released their internal guide on how they build and scale skills with inside their own business. In this video, we're gonna explore all of the things that they learnt and then we're going to map them to how the everyday user and business user would incorporate this into their business. So by the end of it, you'll have a very clear picture of what you need to know. So one of the first things that they point out here in the beginning is that a common misconception is that people think skills are just a markdown file that do a bunch of things. And if we flip across to my environment, you'll see that that is definitely not true. And so here we are in my environment and you can see we've selected the thumbnail generator skill just for example here. We have our markdown file, which is what most people think this is. And inside here, you can just think of this like a standard operating procedure. It is the knowledge in your head about how to do a certain task that you have fed to Claude and it went and printed it all out so that it knows how to do that exact process in the same way that you would. More importantly though, it doesn't just read this. There are several things that you can read inside there. You have scripts, which are the tools that get the job done. So for instance, when we're trying to reach out to whatever would generate my thumbnail, we would probably use a combination of a Python script and an API in order to achieve that. So by having the Python script have a very specific set of instructions, in this case, go out there and generate me an image via this API, we know that between having our markdown file with the instructions and our generate, which is the exact tool that needs to complete that job, that we're going to get the thumbnail that we want. How we mix the probabilistic nature of AI with the determinism of scripts and that's how we get reliability in our workflows within a business. More importantly though, inside here, we have other things that help us get this done a lot better. The first one being any assets that we might need. These are examples of what good might look like. So it could be a picture of my face that goes into this thumbnail generator. It could be referenced thumbnails in different styles that I use on this channel. It could be best practices around what we would want it to draw when it's looking at a thumbnail. This would obviously change from skill to skill and I'm not going to harp on about this forever because I have tons of other videos going into skills from this point of view that you should definitely watch after we've covered this thing. The point Anthropic is trying to make here is that it's not just a simple markdown file. This is probably the most important thing that you can learn when you're using AI because it's how you're going to get the best results from using AI within your business. This is where things get interesting. So they categorize their skills, which is obviously the right thing to do. Humans need a logical way of understanding what goes where. So because they're obviously a development house and they're building AI, these are all categorized around a more technical nature. So they've got library and API references, they've got product verification, data fetching and analysis, stuff like that. That's obviously not so meaningful for you, the everyday user or business user out there. You need a much more logical framework that makes sense to you in your life to categorize the things that you're doing every single day. And that's why I've been using pods on this channel for the last six months, because to me, they make most sense. And for those of you wondering what those four pods might be, here they are on screen right now. The first one being acquisition. This is how we get our clients. The next one is delivery. This is how we serve our clients with a product or a service, whatever that might be. We then have operations. And this is everything that runs on the backend to support our own business. And we have support, which is how we support the product or service that we've put out to our customers. For me, this made complete logical sense from the beginning, because these are the four things that you're gonna see in every business. There might be one or two other things. And of course, you might have your own logical interpretations. But I think everybody understands this map. And by doing that, we can understand exactly what type of skills need to go in which part of our business. For instance, in acquisition, I know I need to go out there and do lead gen. So I'm probably gonna have some form of lead gen skill in there. I'm gonna have a content generation skill, because that's really important for marketing. I might have skills for outreach and sales calls and proposals and things like that. In the same thing for my internal ops, I might have planning and reviews, a daily brief, AI news, whatever it is that I need. But by understanding this framework, it's a lot easier for us to break things down for our own minds. This is more for the human. But the other benefit of doing this and categorizing them comes into play when we switch over to co-work. Because we can use that exact same model and we can turn it into different project folders. For those of you who don't know what projects are, they're just a logical space where Claude runs with its own specific set of rules or context and things like that. I have deep dives into exactly how to set all of this up and build an AI operating system using co-work. I'll link that down as well. But at a high level, the reason that we map our pods is so that we can map them to projects in here, again, so that we have that logical interpretation. And we can build our own environments in here. So let's use acquisition as an example. We've mapped all of the skills that we need in there. We've brought them into co-work or built them inside co-work. They're sitting in the backend waiting to be used. Because we've broken it down logically, we know, okay, cool. Let's schedule three or four skills that we've identified running a schedule. We would then just add it into our acquisition project over here so that we logically know where everything is. And I think this is one of the main benefits of categorizing everything. And it's one of the most important things that you can do when you're trying to architect your own solutions or solutions for your clients. Because if you do all of this stuff up front from the audit all the way through to the pod mapping, to the point where you can just lift and shift things into co-work because you've planned them, you're gonna set yourself up for success. The foundations are the most important part of anything that you're building because literally everything else is a decision that stems from that. Then next up, they get onto tips for making skills themselves. So don't state the obvious is the first thing that they point out here. And they're kind of telling you that Claude knows how to do what it needs to do on a daily basis. You don't need to give it every single tiny piece of information to get a job done. The example that they give is that Claude already knows how to code. You don't need to remind it how to do that. For you, the everyday user or business, Claude already knows how to write. Your job shouldn't be to tell it literally how to write. It should be you telling it how to write how you want it to write by providing examples or telling it how not to write so that you avoid the AI slop. That's what they're leading with in these images over here where they show you these gotchas sections. It's things that you learn over time that the AI did that you don't actually want to do. They're talking about it in depth on the gotchas over here. The highest signal content in any skill is the gotchas section. These sections should be built up from common failure points that Claude runs into. And you'll see that a lot. If you ask Claude to go and write something for you, even if you've given it examples of what good might look like for you and what your voice sounds like, it's still gonna put some slop in there from every now and then. Very important to have a list of things not to do or a list of gotchas telling it, do not say this list of words. Do not write like this. Make sure you don't end sentences this way. Something very common that AI does when it writes is it says, it's not about this, but actually about that. And that's how you know immediately that someone is full of ---- and whatever it is that they've written is blatantly just AI trash. So your goal is to learn these patterns over time and make sure that your skills get better by having these gotcha sections in for whatever it is that you're building. The next thing that they mention is to use the file system and progressive disclosure. So we touched on this a little bit. Essentially what progressive disclosure means is that it's not going to read every single thing inside our skill directory up front. It loads it on demand. The first thing that it will load is this little bit up at the top, just so that it knows the skill actually exists. Then if you had to say something like generate me a thumbnail, it would read the description, say, hey, this is a match. I'm now going to read the rest of this skill over here. So it's doing that progressively. And then once it's read through here, anything else that you would reference that it might need in order to complete its job, for instance, any of these files in here, any of these logos that I spoke about earlier, it would then go and load that stuff. And like I said, it's important that we have these things in here because it helps Claude understand what a definition of done looks like, what good looks like, maybe even what bad looks like, what your voice sounds like. Whatever we would need to complete this lives in here and is loaded on demand. That is the benefit. We're not eating all of these tokens up front. So the TLDR is to never just use the skill.md. Use every part of the file system that you can because you're going to get a much better skill and workflow and therefore much better results. Next up, they are very quick to point out not to railroad Claude. By that, they mean don't give it too much very, very specific information. An example they're giving here is a very, very specific list of steps that realistically AI can figure out for itself, like run, get, log to find a commit, run, get, cherry pick, and literally things that you wouldn't need to tell it instead of just saying cherry pick the commit onto a clean branch. For those of you who aren't technical, that probably means nothing. But the point that they're trying to make here is that when you're building these skills inside here and you're just using your plain English to tell Claude how the process works step by step, don't be overly verbose to the point where it's just ridiculous, like open Chrome, click here, do this, type that. It knows how to do all of that stuff. What we're doing inside our skill, we're being specific, but we're being the right amount of specific. You can see here with our steps, we say step one, understand the video. Before recommending anything, understand what the thumbnail is for. I'm not telling it exactly how to understand understanding something. That would be ridiculous. I'm giving it just the right amount of context in order to understand what is involved in understanding step one before it progresses onto step two. Next up, they talk about writing descriptions and they should be written for the model, not for the human. So if we flip on back here, we look at this description up at the top. This is not for you to open this up and read it. It's for Claude to understand as a part of progressive disclosure, whether this skill solves the thumbnail problem. So while you can obviously read this in plain English, you wouldn't want to write a summary of what the skill is for yourself. You're telling the AI how to use this thing. And you can see that we're blatantly doing that over here. Use when user says generate thumbnails, make a thumbnail or thumbnail for whatever the concept is. And then we tell it that it can run a command for it as well. Point is here, those are two very different things. If I was just writing a summary, I would say this is a thumbnail generator for YouTube videos. That's a summary. This is the description that is meaningful for the AI. And don't worry, you don't have to write any of this stuff manually. Of course, the skill creator and Claude can do all of this for you. Next up, they're talking about how to make Claude remember. And this is something that I love talking about because so many channels out there lump everything into memory. And I think that is the totally wrong approach because memory is varied things. And they talk about that exact same concept over here. Some skills can include a form of memory by storing data within them. You could store data in anything as simple as an append-only text log file or JSON files or a SQLite database. And of course that scales up and up and up. But the first thing that I like to do when I talk about memory is to break it up in categories because this all forms part of the context. I think context is the overall top branch. And then these three things that you see on screen, they form a part of that. And again, I do this because it's way easier for the humans to understand what goes where by looking at it from this point of view. So when we're talking about a skill or Claude remembering something, there are different parts of that memory. The first one is knowledge. And this to me is something that the humans know upfront. I know what my voice sounds like. I know who my clients are that I'm trying to reach out to. I know what my business is, what it does. All of that kind of information, we can stash that in different places. We could put it in our Claude MD file. We could put it in a rules folder or we could put it in a context folder that lives inside our operating system. As you can see over here, I've got context for all the various parts of the skills that run within my business. But then we can also store this knowledge like I've been saying this entire time as references inside here. So for this skill, this is a form of memory. This best practices MD file over here is a form of knowledge more specifically from my point of view. I'm giving it the knowledge that it needs in order to complete the goal without it having to make this up every single time I run Claude. It is a static form of knowledge that always lives there. Then the second part of our memory is state because of course, if we're running a workflow and something changes between now and later, that is a state that has changed. Let's use lead generation as an example. If I'm running leads, I might start them off as a cold lead. I run a workflow, it reaches out to them. Then I know that their state has changed. They're no longer a cold lead. I have now reached out to them and therefore it could be in progress. So the state changes over time. And for me, state gets stored in a database, whether that is a SQLite database or an air table, even a Google spreadsheet. The point is with this, we are tracking state. So while it is a form of context and a form of memory, it is a very different thing that can live in a very different system. It is not just something that we throw into a rag database because reasons. Then finally on this, we have memory in the sense that I like to understand pure memory. And that is where Claude remembers things about you just by working with you. It is the things that it learns as a result of talking with you over time. So if we come back to our environment over here, everything that I do here, if I start to say the same thing over again, like I really hate X, Y, Z, Claude's gonna make a note of that and it's gonna stash it inside our memories folder over here, somewhere on the left-hand side. If we were in coworker, we'd do that automatically because Claude has native memory per projects. It also has memory inside its chat functionality. They are separated, so currently they're not linked and they don't feed off of one another. But the point is that they're in there and it is very different to both knowledge and state. It is memories that Claude has learned from working with us. You can do it whatever way you wanna do it, but Anthropic recommends that you need to do it and so do I. You're not really gonna get anything done in your business if your AI employee has dementia. Next up, they talk about skill distribution. This is pretty easy. Skills are essentially just that folder space that we've been looking at this entire time. You can easily zip these things up and share them with whoever you want to, as long as it's got everything that it needs for the skill to run. And that's the power of putting all of the context inside here because as long as we have everything we need inside here instead of some elaborate system out there, this skill is entirely portable no matter who you give it to. So all you could do is just right-click and zip this folder up, or you could ask Claude to zip it up and move all of your skills to a larger repository somewhere else to share it with whoever you want. But there is even a more powerful method that you can use for sharing this that is not only secure, but it makes more sense because of the way that we might want to update skills, and that is to have something called a plugin marketplace. To illustrate that, I'm gonna head on over to Cowork. On the left, you can see I've got four plugins here. This is everything that you saw in my VS Code environment, but all of my skills that live inside my plugin marketplace, and they run my AI operating system within Cowork. So you can see it's like a logical way to group a whole bunch of skills. And again, that's why I started off with the whole pod mapping slash categorizing thing because inside here, I know that each of those pods that we've been talking about, all of my skills can be packaged accordingly and I can just import them into the project. I can then also distribute this to the entire team if I want to via the marketplace that I set up, and they can just lift and shift this to whatever it is that they need to use it for. And we can see a further illustration of this if we just go over to plugins and we hit plus and then browse plugins. This is what Anthropic has for a marketplace between itself and its partners. They're all just sharing a marketplace, and that's how we come here, and you can just hit on these little plus symbols to download any plugins or skills that you want from these marketplaces. More importantly for you as a business, if you wanted to have this sort of thing, you can create your own under personal over here. You can see I've got quite a few, and you can just hit plus and then add marketplace. You can then either browse Anthropic sources to add custom marketplaces. You could then add your own repository as well, which you do via GitHub. I'm not gonna cover that in this video, just know that it's down below. Then finally on this, they talk about measuring their skills, and that's very important. Observability in your entire environment is very important. Specifically here, they're talking about this means we can find skills that are popular or under triggering compared to our expectations. So what they're doing here is tracking what people are using, but then they're also trying to see what wasn't triggered properly, which might indicate that there is a problem with the description that they set up for that skill, amongst a few other things. But for your use case, and more importantly, if for your entire business, you need some form of observability, and that's why we have a dashboard like this. You obviously might not need something as complicated as this, but it blows my mind that Anthropic hasn't given us something natively to look at our skills and the activity within our environment, so that we can make the decisions that they keep telling us to make, which is why I built this thing. It ties into exactly what they are talking about there, where we want to measure the activity of our skills, the cost of our skills, how many tokens they use, and therefore how much that costs us, how they are invoked, any security problems that we might come across. We can put all of that into a command center dashboard like this, and more importantly, it gives us the data that we need in order to make decisions as a business, because the absolute worst thing that you can do is to go out there and connect yourself to a repo and grab a hundred skills from some random place. Not only is it gonna bloat your own environment and complicate Claude's actual brain in getting the work done for you, but it's also a massive security risk. Skill injection is one of the biggest attacks right now because of people who don't understand this, and they just go out there and pull all of this stuff down into their environment. And that's why I always say, start with the problem and build it yourself. If you can speak any language, that means you can tell Claude exactly how to go and build your process from start to finish, because you're doing it every day yourself. And if you don't know your own sales processes, that probably means you're not making much money right now, in which case you need to address that first and then come back to automation. And that's it. The next step from them is to go and get started. And if you want help doing that, there is an entire list of all the videos that I mentioned down below. You can learn about everything that I spoke about in great detail, including how to build your own AI operating system and run the majority of your business on it, whether it's in VS Code or Cowork. Thanks very much for watching. If you have any comments, leave them down below and I will get back to you as soon as possible. Otherwise, check out the videos on the screen. They'll definitely help you on your journey. Or you can join my community where I'm helping people and business owners solve their AI problems every single day. See you guys later.