Transcription
Skills are one of the most powerful things that you can set up in Notion right now. A single well-written skill page is the quickest way to go from AI as a nice party trick to, "Wow, this is actually insane. If you are not using them yet, you're repeating yourself to AI every single day, and you're barely scratching the surface."
So, in this video, we'll go over why skill engineering is the most important thing that you can learn right now. The five levels of skill writing and why most people get stuck at level one to three and how to build a scalable Notion AI skill system to really operationalize AI for your work and for your team. And before we dive into the tactical parts of this video, it's really important to understand why this matters. I did add timestamps though of course to this video, so if you want to jump ahead, feel free to do so, but I highly recommend that you go through this in order.
First, what is a skill? The term comes from Claude Anthropic and basically just describes a reusable prompt. In Notion, it kind of works the same. A skill is a Notion page, a simple Notion page that contains instructions for your AI.
[music]
When you want to use it, you add mention it in the chat. The AI reads it and executes it. One page, one skill. Now, that sounds simple, and it is. But this isn't just a random term. Using skills well will redefine how you use AI. Here's why they matter so much and why this might actually be the most important video that I film all year.
There are three reasons why skills matter a lot more than most people realize.
First, they force you to uncover your hidden assumptions. When you say to AI, "Create a report," you have an intuitive understanding of what it means, what material to read, what the output format is, right? What good looks like, and what edge cases and issues it might run into along the way. AI doesn't. AI has to assume. And sometimes these assumptions will go catastrophically wrong. Now, this isn't just an AI problem. It's also very common in human interactions, right? You have one way of understanding the process, I have another one. But the thing is that a human typically pushes back, right? Or has colleagues that they can ask or a lifetime of experience. AI doesn't. AI will just silently guess. The process of writing a skill and really writing it, not just blindly copying a random template from the internet will force you, right, to engage with these hidden assumptions and actually write down how something is supposed to work. It's one of the most important concepts for AI right now. The more hidden assumptions you can uncover and make explicit, the better your AI output will be. And skills train you to do that.
Number two, skills solve for context overload. At this point, we've kind of all understood that context is important for AI. So, naturally, what we tend to do is to create these giant master prompts, right? Where we try to cram in every rule, every bit of potentially interesting information, edge case, etc. into it. Right? If you were listening to the first point, right, I hit an assumption and like, you know, no problem. I've basically downloaded my entire brain, right, and just gave it this giant wall of text. Well, that is a problem, too. AI, despite its vast capabilities, also has a finite context window. Think of it like working memory. AI research confirms this, right? The longer your set of instructions, overall, the worse the performance gets. It's called the curse of instructions. Basically, the more parallel rules, right, you try to force upon AI or ask it to follow, the worse it will perform across the board. It's a little bit like multitasking, right? If we as humans want to be as effective as possible, we should focus on one thing. If we're trying to hold too many simultaneous things in our brains at once, right, things get worse. And with AI, it's exactly the same.
This again is where skills come in because they allow your AI to focus on one thing and to put a very important concept into practice. Progressive disclosure. This is what progressive disclosure effectively looks like. At the very top of your Notion AI setup are your personal instructions, right? If you're talking about your personal agent, the one that lives in your sidebar, if you're talking about your agent, right? It would be the agent instructions. And there are a bunch of different terms flying around for this, right? In Notion it is the instructions page. I call it the CLP, right? A context layer prompt, which is a specific type of these personal instructions where we say, "Hey, we want you to go as lightweight as possible of those, right? We want to share the context for the overall Notion workspace and what it interacts with, right? That's the purpose, not to cram as much as possible into it." And then sort of like borrowing a term from the world of Claude, right? Particularly Claude bots, the sole MD file, right? Which is basically uh describing to the agent, right? Like who he is, right? And what he does, right? Sort of like how does he respond like the general acts of interaction rather than again, right? And if you need to draft up landing page, this is what I want you to do. No, like big strokes and then that's what's always loaded, right? Every time you start in chat with your AI agent, this is part of it. Which is why you want it to be lightweight, right? Because if this is like worth half a book, then you're already starting every conversation with a ton of baggage.
Beyond that or below that sit then the different skills, right, that get loaded in on demand.
[music]
And there's different options. We'll touch on them later, but most of the time, right, you probably call on a specific skill. You're telling the agent, "Now we need to work on the landing page." So read the landing page skill, right? "Now it's time to work on X." Let's read that. And then these skills later we'll break down further in additional reference pages because one of the things we'll have to touch on in a moment is that of course also a single skill shouldn't be worth a whole book. Also within a skill you want to still follow the principle of progressive disclosure, right? So that that skill only reads a specific reference page when that becomes relevant.
And now number three, skills help you to operationalize AI. With that, AI goes from "well, nice productivity hack" to "this is really driving our team forward." Because first of all, skills are outcome-driven. A skill will always lead to a specific result. A summarized report, a processed meeting, or a triaged inbox. That means when you're building a skill library, you're naturally answering the question, "What does our organization use AI actually for?" And not just in the abstract, right, with random buzzwords, but with concrete, specific, measurable outcomes. Second, because skills make outcomes reliable and repeatable, you can slot them into different work streams. The same skill might be used in three different, you know, processes. Or you might take a skill and later turn it into an agent, right? That then operates these things on its own. So by creating your skills, you create the framework to use them like modules wherever you need them. Third, and that's huge for teams, skills are sharable. Instead of this, "well, everyone just figure out their own prompts and how to use AI," you're creating this library of shared resources that everyone can draw from. Knowledge transfer becomes a lot easier because you're externalizing things. You're writing them down, and when someone new joins the team, you can point at this library and are like, "You know, this is how we use AI here."
Now, you might be thinking, "Isn't that just good prompting?" And the answer is yes, absolutely. But it's only part of it. Write good prompts is advice. Skill engineering is a discipline. It adds a framework, common language, and a
[music]
system to go from the generic prompting process to specific business outcomes for your team. In some way, prompt writing is to skill engineering what bike coding or code writing is to software engineering. Same medium, different level of rigor and process. And that's what I want to share with you in the rest of the video. The frameworks, principles, and tools that we at our consultancy have developed so far to teach people to go right from simple prompt writing to proper skill engineering.
Now, if you're still with me after this lengthy explanation, then now let's take a look at the five levels of skill engineering.
Level one is the easiest level. It's the one where you basically don't use skills. Instead, every time you interact with AI, you explain it what it's supposed to do. And that's totally fine for one-off things, right? And if you have very varied tasks throughout the day, but chances are there are always patterns, even if you have to write a different type of email to every client, right? There will still be some kind of style guidelines that are the same across the board. So now at this point, right, you're just repeating yourself over and over again, and you're not really building leverage into your work. You're not really using one execution to make the next one easier.
Level two is the most dangerous one because it's the one that most likely will lead you to abandon skills, uh, and sort of go back to level one, and that's blind delegation. Now, blind delegation can mean one of two things. It either means you just download a skill from the internet, you use it, and you realize it doesn't do what you wanted it to do, obviously, because the skill is all about you telling the AI how it should perform a task, right, for your specific situation. So a template likely not going to do it. The other issue is if you ask AI to write a skill with no context. Basically, you go in and ask, "Hey, can you help me set up a marketing skill so we can create better landing pages?" What a good landing page is, right, or what it means. Well, there's nothing here to be found. This is where the problem of a hidden assumption, right, is most obvious because at no point, right, when you ask the AI to, hey, write a skill for X, did you explain what that actually is supposed to mean. So AI will take a completely wild guess and come up with something very generic that likely won't solve your task at hand.
Level three is slowly getting there. Still suffering from the hidden assumptions problem, but it means that you're manually writing the skill. It means that you're giving your input, right, on how this is supposed to be done. This is basically the classic SOP creation that we had pre-AI, right? If you want to delegate a task to someone, you sit down and you explain how it's done. Pretty good. But still chances are that when you explain how something is supposed to be done, that there are a ton of things that you assume the other person, or in this case the AI, knows, right, and will follow. And particularly with AI, right, chances are that it doesn't have that, that it will still make some guesses, and sometimes they will be right, but often times they will deviate, right, meaningfully enough from what you would expect. Here, and unlike with human interaction, right, you often don't have this natural feedback loop, right? This natural, "Okay, you know, how do we improve this SOP going forward?" And that's exactly what level four is trying to solve.
Level four is about co-creating this, so making sure that your domain knowledge is there, but AI gets to ask questions. And we do this through a structured four-step process that we call AC/DC. So the question becomes, "How do you surface all the stuff that you already know, your domain expertise, your lived experience, and all these unspoken assumptions?" That's exactly what the AC/DC framework is supposed to solve. AC/DC is the framework that we developed and presented at the last big Notion conference in Munich. It's supposed to help you discover and codify your workflows for skills. And it stands for Assess, Collaborate, Draft, Certify. Let's go over what it means and why it helps with hidden assumptions.
Step one, Assess. Here's where you basically want to try to collect as much information about how you're currently doing something. That might mean examples of previous work, some meeting transcripts where you talk about how it's done, documents, right? Preps, SOPs that you might have written for humans, uh, everything that you can think of. And you all want to add that basically in one AI chat. Ideally, you then also start adding a lot of context on the fly, right? That's why I love AI dictation tools so much because you can just press a button and start talking, which means it's a lot more likely that you add a lot of, you know, nuance and color to this rather than if you had to type it all down. And your goal is to create this loose stream of consciousness, this
[music]
brain dump where you just share with the AI as if you walked, like a coworker, over lunch, right, through this process. And feel free to just throw everything at it, right? The structuring will happen later. But here, you basically want to try to get as much of how this process works initially into the AI chat.
Step two is Collaborate. Now, after you've dumped all this information on the AI, ask it to review that, analyze it, and then ask it to ask you questions, right, based on everything that's unclear, that it doesn't fully understand. And the goal is here to get into a conversation, into a back and forth where again, right, it comes in really handy if you can just talk to your AI because it's just so much quicker than having to type it all out. And what will happen here usually, right,
[music]
is that the AI is really, really good at structuring existing knowledge. So it will go through that. It might uncover things that you know don't fit together. Maybe in one document you say A, in another one B, right? So it will highlight these to you. It will ask you questions to close gaps, and more things will naturally come to your mind, right, as you answer these questions. Things that you didn't even think about before, that the AI hasn't thought about before. So really take your time here, right? Of course, this is true for the whole process. If what you're trying to create, right, is a very small and quick skill, then this might be just a thing of five minutes. If you're trying, right, to condense sort of like a reporting process that usually takes several days, you might take a few hours, right, to get this right.
Now, after you're back and forth with the AI, you will naturally come to a point where you feel like, "Okay, everything is said and done." At that point, you will ask the AI to basically execute, right? That's the Draft step. Let's try this first time. And what will happen, right, is that the AI will go in now with a lot more context than had previously, and will make a first attempt at executing the task that you gave it.
After the draft run, you go to Certify. You look now at what it produced. And spoiler alert, this is probably still not where you want it. Again, it will depend on the complexity of the skill that you're trying to create and how much time you spend on the previous steps. But chances are, right, you're sort of anywhere between 70 to 90% of the way. And the important thing is to not stop here. Now, the important thing is to understand this as a circular process where you now go in, you look at the things that it did, and you basically kind of go back to the beginning, right, to the one where you give it the context. You look at it and you share with it, "Okay, now that I see the output, here's what I'm thinking, right? Actually, this is supposed to be that. I didn't explain that before, right? Really sorry. Can you keep that in mind for next time?" And you're restarting the process, right? You're restarting the conversation now. Sort of like every loop will probably be quicker than the previous one until you get to the end, right? Until you get to the certify step again and you look at it and say, "Yes, this is pretty much the output that I would expect from this skill."
This exploratory process is the whole point of AC/DC because it's very, very unlikely that you can uncover all your hidden assumptions with just a bit of reflection. Maybe you're incredibly self-aware and incredibly organized, right? And then this might be super quick. Most of the time though, as you go through this, right, you realize all the well-hidden assumptions that you had about this process because you're looking at the output that you expect and the output that the AI produced. But instead of turning it into a frustrating experience where you say like, "Okay, this clearly doesn't work," you turn it into a feedback loop to refine the process and to, you know, really codify all these elements. That's how you go from level three in skill creation to level four. And you could stay at level four, right? This is really the 80/20 jump going from 3 to 4. And asking the AI at this point to know, "Now, simply write everything down that you learned about this process so you can remember it next time." This will vastly outperform most skills that other people create.
But there is, of course, still like the last 20% that we can push by going to level five and not just asking AI to create a skill, but instead by teaching AI first what a skilled skill looks like. Just to restate this again, right? Level four gives you the knowledge, that's 80% of the work. Level five gives you now the context and the structural principles to turn that into something that's repeatable and based on current research and best practices.
There is, of course, one caveat. AC/DC works best if you know what you're doing, right? You need to have previously done the thing, build intuition and learnings around it, figured out what good actually is supposed to look like, and now we use AC/DC, right, to make sure we extract all our information and teach that to the AI. That will only be part of the situations when you want to create skills still, right? Naturally, you will come to situations where you want to create a skill for something where you don't quite know how it works. It doesn't mean that you can't create skills in this situation. It's just important to know, right, that you will start much lower in the progress. Whereas this might get you, you know, like to like an 80% result. If you have to guess what good looks like, right, you're probably more starting at a 20 or 30% of the outcome, right? You're starting from zero, after all. You still would want to use the same iterative process, right? Basically guessing what good looks like and then going through it.
And I think nowhere is this currently as apparent as in the whole AI report thing, right? It's sort of the most obvious, most apparent use case for anything from agent to skill to AI, right? "Give me a weekly report," right? "Sort of give me a weekly briefing." But on what, right? Previously, we've never had this kind of report, probably, unless we're an executive, right? With a great personal assistant. So for us now, all of a sudden to figure out, "Well, what would a good report look like?" That's pretty tough. And we see this play out in action all the time, right? Also personally, right? My first attempts at creating a morning briefing or like a weekly information, right, for my project management in the company were horrible. It didn't give me anything that was actual. It was all noise, no signal because I couldn't tell it what good looked like, right? But I couldn't tell it, "Hey, I have a lot of project management skills and I know the kind of report that I need on my desk every Monday has A, B, and C." No, I'm kind of learning as I'm going, right? Which means of course the whole process of iterative feedback with AI will take a lot longer than if you know what the whole thing is supposed to be in the end.
[music]
And just as a side note, if you go through this process, chances are for a while, maybe even a long time, you will invest more time in this than you get out, right? If you just continue doing your work without ever playing with AI, you have your usual linear growth, right? You will improve at your normal rate. If you play around with AI, right, and sort of like try to figure out where it can help me without knowing, right, what good looks like, then you very much for a while will be in this investment phase where this takes longer, right? Just the other day, I invested several hours trying to figure out whether we could finally get Claude co-work with the MCP to build, uh, make automations for us, right? And after like 3 hours of playing around, I built a whole thing in 10 minutes myself, right? That's sort of this investment phase to collaborate and figure out, "Is it working? Right? What do we need to do?" But slowly, right, you will reach the break-even point and then the compounding returns set in. It's just important that you're willing, right, to go through this investment phase because otherwise, you're never going to get here.
And now you might be listening to this and think, "So is creating a skill also a skill in itself?" And the answer is yes. We've actually created a series of three skills for you, like a skill creator, a skill reviewer, and then sort of a skill improver over time that I'll share with you later in this video. But that's exactly the idea, right? We teach AI a specific outcome, how to produce a good skill, and then we can use it over and over again as you produce new skills. This principle is nothing new, right? Claude already has this. Claude has a skill skill. So if you ask it to help you turn a specific process into it, it can help and assist you. This approach, right, ports Note that principle to specifically Notion.
This, in essence, is the skill production pipeline, right, or lifecycle that we recommend. You use AC/DC right to figure out all the context and co-create the initial draft for your skill. Then you use the skill creator, which is a skill inside of Notion already, to teach the system of how to turn all your findings from AC/DC into a best practice skill. And then, since this is sort of one part of the job, right, we want a skill reviewer. We want another skill to check this and make sure, "Hey, is this actually up to par?" And again, the goal is sort of like this progressive disclosure.
[music]
Because the skill creator, right, at this point will have had a ton of context on how this whole thing is supposed to look like. So we kind of want a set of fresh eyes that just looks at the thing in isolation, right, and makes suggestions based on that. And then as you use it, one of the big compounding things over time, right, is to whenever you interact with the skill and you learn something new, right, you figure out, "Actually, we might have to tweak our instructions here," or "Now that I see this output, we need to add that now." This will over time, right, add more and more stuff to your initial skill document. So you need this compounder that regularly runs over it, depending on your changes, right? Maybe once a week, maybe once a month, maybe once a quarter, and condenses it back to the essentials.
So, let's go over the principles, right, that power these last 20% of skill creation.
Principle one is "Show, Don't Tell." If you can show AI what good looks like, it's really good at figuring out how to get there. You should know at this point, right, thanks to AC/DC, how this good output looks, right? You should know what "done" is so you can teach it to the AI in the skill. Interestingly, while we were doing research for this video, one of the things that I found is that apparently it doesn't really matter anymore that you instruct it to take a specific role. Remember 1-2 years ago when pretty much every prompt started with, "You are an expert marketer." So apparently currently this mostly leads to the AI being more confident but not really changing anything in its response. So you want to really lean into your domain expertise, right? And sort of like telling it, "Hey, this is what great looks like. This is the process that we follow," rather than, you know, generic situation prompting. Another key improvement is to explain the why. Similar to what we tried to do in this video to show you why skills matter, right? The same is true for AI. So instead of saying something like, "Never use passive voice," right? You would say, "Use active voice because it makes the audience feel more engaged." By giving it the reasoning, right, AI can then play to its strengths and apply that same reasoning to other aspects of the process.
Principle number two goes back to something we talked about earlier in this video, the curse of instructions. It can be very tempting, right, to now create an incredibly long skill, particularly if what you're trying to explain, right, is very complex and long. But as we learned, right, the more rules AI is supposed to follow in parallel, the worse it performs. Which naturally begs the question, right, for like complex multi-step flows, "How do you kind of square here the circle?" We kind of know the answer to that already, right? Progressive disclosure. Skills are a core component, right? The AI reads the skill only when it needs to care about the outcome. The second level to this is that you want to keep the main skill page fairly lean, right? And focus really on the high-level and the goal and the output and the why. And then when you have very detailed instructions for specific situations, for example, maybe, um, the result, right, example is super long, you want to outsource this in specific reference pages so that the AI looks at it when it matters. For example, right, in a report creation process, the AI might go through four steps of gathering information, you know, taking some conclusions, then, um, cross-referencing it with previous elements, and then putting it into the specific result at the end. By making sure that how the specific result looks, right? By making sure this lives separately and by instructing the AI to only look at these instructions after it has done all the other steps, you make sure that it stays focused and doesn't get derailed by, you know, that end thing before it has even started with the process.
Interestingly, again, one other aspect of AI research shows that elements that are at the beginning of the prompt and at the end of the prompt get the most attention from AI, and what's in the middle tends to be messier. That again means, right, for your skill writing or for the skill skill that we need to instruct AI to put the most important things, right, in one of these two places to make sure they are for sure followed. I think it's actually quite funny how similar AI seems to behave to human brains, right? We basically read the first introduction of a page, skim the rest, read the end, right, and then assume for the rest and hope that this will turn out somewhat right. And it seems like AI is kind of following into our footsteps here.
And principle number three is to treat your skill as a living document. It's very unlikely that you get it 100% right the first go. Instead, you want to approach this as something, right, that gets constantly refined over time. We already naturally built in one review loop in our skill skill, right? We have one skill that creates it and a second one that reviews it afterwards to see whether there are any obvious immediate gains. But then you also ideally, depending on the workflow, want to build in a little feedback loop, right? Where sort of based on the output of the skills over time, it learns and adjusts its prompt. That will mean that the prompt probably gets longer over time, right? As it adds more edge cases and more learnings. So that's where prompt number three comes in, right? The one that regularly reviews the whole thing that you have and condenses it back to what really matters.
One important note, as you might guess, this process takes time and effort. And that time and effort will scale, of course, with the outcome that you're trying to achieve. If this is something super simple and banal, right, an easy task, then you can probably condense this whole thing into the process of half an hour. But again, if it's something that usually takes humans days to put together, that means also, right, the time that we need to allocate to prepare and improve that process will scale. One way to sort of address this problem of, "Ooh, this is a very big thing," is to actually chunk a process into its individual segments because if you look at sort of maybe the whole, "Hey, how do we solve a bug?" or "How do we create this client report flow?" you have a very big thing with a lot of sequential tasks, and maybe AI isn't actually good at delivering this end-to-end yet. So thinking about skills also means breaking down a process into the individual skills required to deliver an overall outcome. We're going to talk about this in a second. Right? Basically, collection of skills to make sure that you can break things down into these individual manageable chunks that you can develop individually, that deliver value individually, and that then over time, right, you can combine to sort of not not automate, right, but improve the whole process. That's very important, right? Don't look at basically, "Hey, how do I run my business?" as a skill. No, skills naturally have a certain maximum size, right? Also based on the context text window of AI. So there are always areas where you need to chunk something into two separate skills. That's, after all, the whole point of progressive disclosure.
So now you have a whole production line to create skills. AC/DC to uncover, right, all the context and hidden assumptions. The skill creator skill to turn this into a first draft. The skill reviewer to improve that. And then that potential loop, right, with a skill compressor to over time make sure that it doesn't explode too much. But how do you now take all these skills that you might build for your company and turn them into an actual system? How do you manage this over time so it doesn't just turn into a giant graveyard of knowledge? Well, if only there was a tool that would help us organize knowledge in organizations, right? Make things findable, categorize them, and put them in the right context.
[snorts]
Jokes aside, Notion is of course a great medium, right, to organize them. We just need again a framework to build this. And here we can borrow another concept from Claude plugins. In Claude, a plugin is basically just a container for a collection of resources of specific skills, specific tool calls, MCPs, resources, context. And it flows from this natural insight, right? That skills have a maximum ceiling. Since the AI needs to be focused on certain parts of a task, right? If it's very, very big, that means we need to combine several skills to create a larger outcome. In Notion, plugins map to a series of interconnected databases, what we call the Notion plugin system. It contains at its minimum of the plugin database, which is sort of the big grouping element, and a skill database for the individual executable items, both with a description, so it's also easy for your AI to find the right one, and of course, with a relation, so you can group them together.
The next logical add-on is the agents database, which again connects both to skills and to plugins because as you keep developing these, one thing you'll realize is that a skill is just a fixed set of instructions. You can either run this manually on demand by interacting with your personal agent, or you can give that skill to an agent so that it runs off and does that on demand or automatically, which means, right, in our system of plugins that work towards larger business outcomes. Agents slot in very naturally. They take over part of, right, this longer process that we've identified and run with that autonomously. But this is just the beginning. Notion is expanding the platform like crazy, and when you watch this video, it might very well already have much more advanced capabilities. Currently, they've already announced custom MCPs and the ability to execute code through Notion, right, through what they call at the moment workers. So again, MCPs and these tool calls again slot into our plugin system. They will be their own databases. They will be sometimes related to agents, sometimes related to the overall plugin. So that again, right, you can build this army of different tools that then slots in to push one business outcome forward. All with the general battle-tested best practices of building Notion systems, one database per object, linking between them to make sure context flows naturally, and a system that overall is scalable and can grow with your team.
And it matters because it slowly but steadily can give the AI more autonomy as it gets better at making these decisions. Right? As you know, we very much are in favor of what we call a context layer prompt. This very minimal, lightweight prompt for your personal AI to explain it generally how to operate in the system. We don't want to overload it with stuff, right? And you know now why progressive context disclosure. But what of course we can tell it is that, "Hey, when we're working on some larger things, you might want to scan the plugin database to see whether one of these descriptions, right, fits what we're working on. And if that's the case, you might want to go one level deeper and just quickly check the skills, right? Again, just based on the description, not based on these, you know, page-long instructions, to see whether one of them might be helpful." And then based on your preferences, you might ask it to automatically invoke them, or what I would recommend, right, to merely suggest them, so that the human still has a lot more driver influence over what happens. But this will also help with your team because realistically, not everyone in there will be an early adopter and sort of push Notion to its limits. A lot of them just want to do their job. And by making skills, plugins, all these things generally available and searchable with the general Notion principles, it means it's much easier for someone to discover something that helps them in a given moment without knowing that you've created this skill.
Last but not least, this Notion plugin system allows you to assign skills to multiple work streams because, at the end, that's what's very likely going to happen. This one skill, right, for producing a certain output will not just be relevant for one workflow but potentially across teams. And this way, you make sure, right, that this one skill always stays in sync. It might be used by three agents and seven different teams, right, but if you make a change in one place to improve it, everyone immediately profits.
Now let me show you quickly what these resources all look like put together. So here's a very clean and simple first implementation of this plugin system. And of course, right, you want to take this as inspiration and properly integrate it in your system. This is just the isolated component looking at it. And at its core, right, we have the three databases: plugins, agents, and skills. So if we look at plugins, right, we see this comes pre-loaded if you download it, um, from the newsletter resource list with, um, the first main skill collection, right? Skill Engineering. Uh, so if I open up the plugin, right, I have a short description. This is the one that would be exposed in the CLP, right, that the other things can read to find this. Um, I'll see, "Okay, what is this? It's a tool for workflows for creating, reviewing, and maintaining AI skills." And then I can tap over to skill view and see we have three skills in here: the skill creator, the skill critique, and the skill compressor. We can check those out in a second. And we see one of those is actually linked also to an agent. And we see that agent also here, right? And you see how this uses Notion's native functionality of simply being good at organizing stuff. Now, specifically designing it around this AI challenge, right? Where you have so much information, different levels of things, right? Skills, tools. Uh, you need to, uh, combine these for like larger workflows, right? To make sense of that all.
All right. So, let's have a quick look at one of the skills, right? Let's check out the skill creator. And we see that this is sort of written with the idea, right, of our best principles. It's a fairly small page, right? If we scroll over this, this is not a ton of context, right? Purposefully keeping it lean. You could supplement this with a reference page if you wanted to, but as it is, right, this should be good enough to help you turn all that amazing context that you've collected into a dedicated skill page. And it takes and touches on a lot of the things that we talked about, right, with surfacing assumptions, with making sure, right, that things are written in a specific way. You see that this kind of also like already builds for the future. As of today, right, Notion skills don't auto-trigger unless you tell your CLP to look for it. So this is something that you might want to modify. Um, and then also, for example, right, we have then towards the end, step eight, right, very importantly, to review the architecture to make sure that if you're working on things, right, that are super long and take a lot of stuff, that it automatically will start suggesting splitting that, right, into multiple skills. That's actually what we've done here, right? Um, of course, right, you could say, "Well, um, creating a skill, reviewing a skill, and compressing a skill could just be one document," but we would sort of violate our own rules, right, and make it mean that like at each individual skill would be fairly large and big. So, this kind of is an example, right, of that principle in practice.
The skill critique, right, is then sort of like the second thing that runs. It is complementary, right? It looks at very different things than the main creator. Not just checks, "Hey, did the skill creator do what it said it would do, right?" That would be pointless. It's a chance to introduce other elements, right? Like a cold start reader simulation, right? Or a failure mode analysis. And then same for the compressor, right? The compressor does not try to do the role of the critique, right? Um, it explicitly calls it out that you might want to run the critique on something first, depending on what the situation is. The compressor really is designed to take something that has gotten overboard, right, and trying to refocus it again on something narrower or potentially suggesting to split it into multiple parts.
And then if you look at the agent, just very briefly, right? Um, the agent instructions, uh, that you would add here, right? I've actually added for the final template. Give me one second. They are at the wrong location. So here they are, right? Was creating a different page for that. Um, but it basically instructs, right, a small and simple agent to, on a schedule, run the skill compressor, right? So you don't have the same information in your agent and the skill. The agent simply reads the skill whenever it needs to, which means a human can run it on demand, and the agent can run it on a schedule.
All of them are available for free for the subscribers of the newsletter. You can simply click down below, right, to sign up, and then you get that and much more stuff around Notion AI and Notion. And if you're a little bit overwhelmed by all of this and wonder, "Hey, how can I roll this out to my team, right? I need some help with AI enablement." Or maybe you're at a point where you're like, "Woah, this is amazing, but we still haven't figured out the basics of Notion." Well, just let me know. My team and I are certified Notion consultants, and that's what we do all day long. Think about Notion, think about AI, and help teams implement both. We do that through a structured 8-week process to get you the quickest time to value. Just check out the link in the description.
All right, so there you have it. A lot of thoughts on skills from the AC/DC framework to uncover hidden assumptions to the three principles, right, that actually make good skills perform. You know now why skills matter. You have three skills that help you create skills, varimeter, and a plugin system to organize them as you scale this and roll it out to your organization. But this is just the beginning. Over next week and months, we plan to release a lot more frameworks and thinking on how to operationalize AI. It starts with things like the unofficial Notion AI dictionary, that's already live on our website, or the ultimate guide to Notion mastery, which is available for all of our newsletter subscribers. And we'll continue with building block on building block to make sure that AI isn't just this nice party trick, but something that makes you go, "Wow," every single day.
If you want to go along for the journey, you know what to do. Subscribe to the channel. Leave a comment. Let me know what you think about this. Right? I'm very curious to your perspective. We don't have many of these long talking head style videos. So curious whether it's helpful or not. Right? Both fair points and very curious to hear about it. And otherwise, right, if you want to continue learning, then I highly recommend this video next. Just click here, and I will see you in a few seconds.