[0:00]With AI agents on Claude Code, Cowork, and Codex becoming more powerful by the day, building good skills will be one of the most important things to get good at in 2026. Building agent skills allows anyone to automate any workflow by simply prompting it and can even become self-improving, and good skills will likely become monetizable soon. So in this video, I'll show you why you need to learn skill engineering, show you what skills actually are and when to use them, and a simple framework to build better skills than 99% of people. Now, this guide should help you whether you work with Claude Code, Cowork, OpenAI, Google, as they all have integrated skills by now. Now, before we dive into what skills are and how to build them effectively, let me quickly explain why learning skills is a big deal. Now, we're all seeing that these AI agents are getting really powerful, but no matter how good they get, they still need that specific, those specific guardrails, context and SOPs around all the unique ways you and a business do things and use software. Now, we've tried to solve this before, of course, we have projects and custom GPTs where you set a system prompt and add context files, but generally the problem with these is they're isolated. You're hopping between different windows, they don't really self-improve and can't handle a lot of context. On the other side, of course, we have NN and automation platforms that hard code the guard rails and make a system deterministic. And for fully deterministic, non-human in the loop workflows, these are still great, but most day-to-day work isn't actually deterministic. Uh, the processes aren't always clear cut, that it's context-dependent, it requires judgment, and that's why human in the loop is often really important to get good outputs from AI. And skills sit in the middle, they're essentially instructions for an AI agent on how to do a specific process, they can self-improve, they have human in the loop, but unlike a project or custom GPT, thousands of skills can be accessed by the same agent. And the game-changer is that they can be created and updated by simply prompting, and that means anyone can start automating their tasks and workflows through a single interface on Claude Code, Cowork, or Codex. They're also shareable across your business. So, one person's process and domain expertise can instantly be used by the entire team, which, of course, has huge implications for onboarding, consistency, and how a company operates. And if thousands of these specific capabilities can be given to one very capable general agent, it seems likely that AI agents will slowly become the single interface for doing work. Now, of course, we are very early with this, but I predict that building the skill infrastructure for these agents to do specific tasks well will not only allow you and your company to become far more productive, it will also become monetizable. Sort of like a new software layer, but just like with software engineering, prompt engineering, and AI automation engineering, you need to actually get good at it, even though anyone can create skills through prompts, the same skill can produce completely different outputs depending on how you build them. For example, this is an infographic that I got out of an infographic skill that I quickly created. And this is the output of the same input of an infographic skill I put a lot more effort in. As you can see, it's a lot clearer and a lot more aligned to my brand. I'll show you a full demo of the infographic skill later in this video. Also, if you want access to that infographic skill and all of the other skills me and my team are building out, you can also check out my AI accelerator where we list them all for you to download or customize. Now, before showing you how to build these types of skills, it's good to understand what skills actually are. And simply put, agent skills are folders of instructions, scripts, and resources that agents can use to do things more accurately and efficiently. And at the core, we have the skill.md file, and you can basically see this as an instruction on how to do a process. Think of it like a system prompt, a Claude project, or a custom GPT, but besides that process instruction, we can have additional instructions on how and when to use knowledge files, how and when to use tools, how and when to spin up sub-agents, and how and when to use code executions. So the skill.md, you can basically see as the SOP. Now, that skill.md is just the core, what makes skills more powerful are those additional reference files or the context we add. Now, what kind of reference files can we have? Now, firstly, we can build really simple skills with no reference files at all. It's just a skill.md or a process instruction. For example, we have a sales account research skill here from Anthropic, which has no additional reference files, just the skill.md here, and the skill.md, as you can see, includes an execution flow, right, step one parse request, step two web search, Right, step three enrichment and step four CRM check. But commonly, when we build our skills, we want to add some reference files to provide the skill with more information. Firstly, we can have text files, for example, common ones you might want to include are example outputs, style guides, contact around your ICP, your background, et cetera. For example, in my newsletter writer skill, you can see I have multiple reference files here to give the skill more context on my background, what we do, my ICP, and voice personality. Another common one to use is an MCP instructions file, which basically lays out to the agent how to use a specific tool efficiently in this specific skill process. For example, the way you navigate or use your CRM for this specific task. Now, don't worry, your agent can build these MCP documents. Now, a second type of reference file we can have in there are assets or non-text files like images, presentations, videos, or even binary. For example, to give it good output examples of a presentation layout that you want, and then lastly, reference files can even include code scripts, Python or JS functions that can take actions like doing API calls in a software or executing functions.
[5:14]For example, in my infographic skill, we have a script to do an API call to the Nano Banana Google API call. So you can see that these skills can be very simple or can become quite complex and can look a lot more like a software. Except, skills, of course, are software for AI agents. But how are we able to give one AI agent access to thousands of these skills, feed it all of these documents without overloading it with context? Now, this is done through a process called progressive disclosure. Now, it sounds complicated, but it's not. Basically, when we create or upload a new skill and give it to our AI agent in Claude Cowork or Code, only the metadata, which is the description and the name, are stored in the agent memory. In my infographic skill, this is the name and this is the description, and this is how your agent knows when to trigger or use a specific skill. And then only when the skill is triggered, the skill.md or the process will be loaded into the context window of the agent to understand how to execute on this specific process. And only when the skill instructs to use a reference file, it loads the reference file into the context window. And because of this progressive disclosure of context, we can give one agent access to thousands of potential skills. Now, I see a lot of confusion between plugins and skills, too, so let me explain it very quickly. Plugins are essentially packaged or bundled sets of skills, commands, agents, and connectors, and this basically means and does three things. First, it adds a layer of complexity on top of skills because plugins can include multiple skills, it can include commands, which serve as workflow triggers, they can trigger multiple skills in sequence to automate more complex tasks, it can have specialized agent teams, and a preset set of connectors. So we can add a lot more functionality to these plugins. Second, through plugins, these bundled packages can easily be shared and divided between company departments. Sales teams can get a sales plugin, marketing teams can get a marketing plugin, each with the connectors and skills that each department needs. And thirdly, these plugins are also becoming versionable, skills too, by the way, which means they can be updated at any time across any account that uses them. And because of these characteristics, plugins start to look a lot more like software, of course, or SAS, and I predict that SAS will start to launch their own plugins with their specific functionalities. Now, if you want to learn more about plugins, I recently did a video on it, so I'll make sure to link it in the description below too. But skills are really the key aspect even inside of these plugins, and even though we have plugins, we can still directly build, trigger, and access skills even without these plugins. Now, skills and plugins are relevant for any industry, business department, and any niche workflow, and we're seeing three layers of skills and plugins appear. We have the general plugins and skills, for example, the ones built by Anthropic or OpenAI, but we're already seeing marketplaces of skills like Skills and Peace myth, where people are building and potentially selling them, and potentially SAS businesses will start creating their own plugins. Now, as these will have to be general-purpose skills in order to be relevant to multiple industries and businesses, companies will want to build their own ones based on their unique processes or customize some of these general skills. But even within a company, it makes sense for different employees and different people to customize skills to their specific way of working and their specific workflows. For example, an Anthropic or third-party, let's say sales outreach skill, can be useful, but a company probably wants to customize and improve it by adding the company brand tone, the ICP and the business context, and one sales reps copywriting style will be different than another, so they can even start customizing those skills for their specific copywriting style, for example. Of course, if you as an individual start to get good at it and you have specific domain expertise and build a skill out of it, you can potentially monetize it by providing it to the general public. So, how do we actually build these skills ourselves? Now, we can first of all customize Anthropic or the third-party provider skills. There are multiples of these marketplaces now, but the real unlock is really building them yourself because every person and every company has their specific ways of working and their specific expertise. Now, there are two main ways to build your own. First way is you do a task once manually with an AI agent and then ask it to save it as a skill. And the second way is to instruct it on how to build the skill right away. For example, here I downloaded a third-party provider meta ads skill from one of the marketplaces, imported it, and used it here in my own account. And then I asked it to use this skill combined with my specific knowledge sources to create a new skill that helps me create ads. And then it created me a new skill which I can add to my library. Customizing Anthropic's built-in plugins and skills is also very easy. You can just go to skills, click on edit and customize with Claude, or even the plugins by just clicking here on customize. And here I created my own skill by just prompting it and giving it some reference files. Now, whatever the method, applying some best practices is important. So let me start by saying that building good skills is kind of like an art, it's the art of putting your domain expertise of a process into something productizable and usable for an agent. And the more I do this, the more similarities I see between skill engineering and software engineering, but it's sort of like software engineering for an AI agent. But you have to think about UX, right, when to add human in the loop, you have to think about context engineering, right, which balance of context produces the best outcomes, you're adding features, removing things, handling edge cases, et cetera. But the beautiful thing with skills is you can easily update them with prompts, so skills are never finished, and what I've learned is the more you use them, the better they get. I've probably updated and iterated on my infographic skill around five times to get to the outcome that I get now. Now, the first step is the step that most people skip, but what makes the biggest impact, before prompting to build or customize a skill, you really do want to take a step back and think about the ideal step-by-step process to get to a good outcome. You want to think about what knowledge sources or additional information you can give your AI agent in this skill to do a better job. Now, generally, some context or reference files that I recommend you have, especially if you're going to use these skills more for marketing or sales-related skills, is a what we do, basically a description of your business, a description of your ICP, of your voice personality, if you're using this for LinkedIn, et cetera, newsletter strategy, in this case, I have this also for my YouTube, for my LinkedIn, et cetera, and also writing framework. These are some of the documents that I reuse across different skills, too, so I highly recommend this. Now, if you don't have these reference files yet, you can do it together with Claude, for example, here I asked it, I want to create a YouTube strategy document that I can later give you when I'm building skills for research, ideation, scripting, et cetera. And then I basically asked it to ask me some questions about my YouTube strategy, and through the planning mode, as you can see here, ask questions, et cetera, it will get more and more context. And at the end, here, you can see Claude generated a YouTube strategy document for me, which I can then reuse in multiple YouTube skills that I create. And once you have a few of these, building skills will get a lot more efficient. Now, once you have that, of course, you want to think about which tools or software the agent needs in order to do this task, and then if possible, you always want to prepare good output examples, and this is generally the thing that impacts performance the most. Now, once you have that sort of clear, we move to the building, which, of course, in this case is prompting. Now, I just go through a framework that I like to use, which is not hard science, but I think can help you to think about what to include to get to a good first outcome of your skill. Now, first, you want to define the name of the skill and how the skill should be triggered. So, in this case, the name of the skill should be Infographic Generator, and it should be triggered anytime a user mentions he wants to generate or create an infographic, and this is what the agent uses to write the meta description. Then, just like with prompting, you want to define the goal or the objective for the skill. Uh, it can be quite short because we dive a lot deeper into the process later. Now, in this case, an infographic skill that creates quality infographics according to my brand style for LinkedIn and newsletters. Then we want to give Claude context on connectors, APIs or MCPs that it needs to use in this skill. Now, if there's a specific table, maybe a page or a process, it should take in that software for this skill, you can also describe that here, and then you want to lay out the process step by step. This is one of the most important things, and it's good to actually spend some time on it. Now, a good way to think about what to include in each step is what does it need to do? First of all, when do you want human in the loop, what kind of human in the loop do you want, with the dynamic Q&A boxes of Claude, we can now choose between checkboxes, open fields, single select, et cetera, which kind of becomes like UX designed. And for what additional context should be used to do this task better? If it's very short, you can add it in the prompt, but if you have a specific file it needs to read, you want to specify that for each of the steps. Even if you don't have the reference file yet, you can just paste the context here in the prompt and ask it to create a reference file for that skill with that context. Generally, want to try to keep the skill.md very clean and focused on the process, any additional information should be in the reference files, that's how your skill is going to perform a lot better. And lastly, you want to think about what you expect as an output for each of the steps. Something that I like to do and has made a big impact on my productivity is when you have human in the loop steps, ask Claude to always give you multiple variations or options, which you can choose from instead of just one off outputs. So I usually say something like that, it should give me five different ways an infographic could visualize this idea or concept. And lastly, you want to think about rules for the skill. Here you can basically predict what could possibly go wrong when executing this skill and write it down as a rule. Now, this is also the section that you continuously update when improving the skill, and two tips here in the rule section is, first of all, to double down on instructing it to use the knowledge files as obligatory steps in the process when needed. Otherwise, I I notice it tends to skip over some, and second, I usually double down on the multiple variations with human in the loop, and lastly, a really good thing to include is progressive updates. Basically, instruct the skill to be automatically updated and improved when using the skill so it becomes self-learning. For example, every time I define a clear thing not to do, anymore in the skill, update the rules section. And what can be very powerful too is instruct in the process that if a user approved the final outcome to save that as a good example, so it learns and builds more data around what good looks like automatically. Now, let me show you a quick demo of how my infographic works, so you get a better idea of what you can get out of it. So, I'm in this case I'm using the code tab here because with API calls it can be a little bit easier. Uh, so I can just trigger it with a slash command, and this is already an updated version from this initial prompt that I gave it. First, it asked me what content should I turn into an infographic, so I'll just paste in a part of my video where I talk about what skills are. A later update I made from this first prompt is to ask it first for which platform it is to understand which format it has to generate it in, so in this case LinkedIn. And as you can see, it uses these QA boxes because I instructed it to do that. And then what type of visual should it be? Infographic. Then I later adjusted it and iterated on it to get a checkbox out so I can actually start generating multiple variations. But here it suggests me what we should visualize. In this case, I'll go with Anatomy of a skill. And then it suggests me ways to visualize that concept. I can choose what makes most sense and generate those. I'll just do all of them. Now, again, the first type, I like this prompt, wasn't this good yet. I had to iterate three or four times, adding extra rules, et cetera, so it became an iterative process. And it's really about that, the more you use it, the better it becomes, but these are quite powerful and really follow my brand style and guidelines. And you can see I added this box at the end. If I click keep all, finalize, it will save it as the good output examples, so it gets trained on what good looks like. It's now, it is important how you instruct the model to change things. An easy rule of thumb you can keep in mind on improving and updating is if it doesn't follow the process correctly, ask it to make changes to the skill.md. The key in the skill.md is that it doesn't get polluted with a lot of additional information, it should mostly focus on the core process. Now, if you need additional information, you want to ask it to add a reference file. Now, if it does something wrong, not just for this specific task, but you never want it to do. For example, I had this with my infographic, I never want the black background because the shadows are not visible, then you can ask it to add a rule or update the knowledge file. Now, if it struggles to use one of your softwares, MCPs or tools, you can first guide it manually to do the specific action necessary, and then ask it to create an MCP reference doc on how to do the task in the software. And you can see that my first version is completely different from my last version. Now, lastly, once your skill is performing well, you maybe want to share it on a marketplace or maybe to a team member. The easiest way is to ask Claude to give you a zip file for that skill, you can then just share that, and someone else can go into the settings, the capabilities and upload the zip file. You can also deploy a skill through GitHub, I'll put a link in the description below on how to do that. Now, if you have multiple of these skills, you maybe want to combine different skills, maybe commands, add in agent, specific connectors, you can bundle them into a plugin, which again, can be done by just telling Claude to build your plugin. It will ask you for all the skills, et cetera, you want to include, and then you can add it to your library, and if you want to share it, again, you can do it through a zip file or by uploading it to GitHub. And if you're a business, and you have multiple plugins across multiple departments, you can create an entire plugin marketplace, which is a bundle of all your plugins. Now, you'd have to create this together with Claude Code and upload it to GitHub in order to get a link. Again, Anthropic released a full guide, which I'll share in the link in the description below. Now, if you want access to all of the skills, plugins me and my team are building out, including the infographic one, you can also check out my AI accelerator, my AI community where we list them all. We also do weekly AI workshops where we dive a lot deeper into these types of tools. We are blueprints on how to start your business, how to find your first client, so if that's potentially interesting to you, uh, you can check it out in the link in the description. Thank you so much for watching, and if you want to learn more about Claude Cowork, you can check out the video here.
