Transcription
I kept adding more detail to my Claude and Chatbot projects. And the more I added, the worse it got. Formatting still wrong, sections still missing, and it turns out there's a point where more instructions actually hurts the AI's output. The fix is one text file and a folder. I'll show you what it is, when you need one, and three ways to build it without writing anything yourself. So, let's get into it.
Now, here's the hidden problem that a lot of people don't realize is that the longer your instructions get for a given prompt, the quality of the output of the AI eventually peaks. And as those instructions get longer, the quality of the output from the AI degrades. And there's a very specific reason for this. And it really comes down to what the AI can hold in its head at any given time. Because if you're using simple instructions, so if you're using Claude, Chatbt, Gemini, and you're using projects inside of those tools where they have system instructions, the AI loads up all of those instructions in the associated knowledge files every single time. The issue here is that when it does this, it takes up a bunch of space in the AI's head, which leaves very little space for it to actually think about what you asked it to do. And that's what causes the degradation of the AI's output.
And the big AI labs know this, OpenAI, Anthropic, Google, etc. So what they did is they created a thing called skills. And skills are a simple but effective way to work around this limitation of AI today. And all skills are is simply a file in a series of subfolders inside of a folder. So this is an example of a skill. So we have a skill which is a folder. We call it my skill. Inside of this, there's one thing that's mandatory, which is this file here. And all it is is a bunch of text telling the AI what it needs to do. In addition to this, there are three optional things that can be added here. A reference subfolder which gives templates and examples of what good looks like for the AI to follow. It can even have scripts which actually ensures the code follows a specific format that you care about when it comes to the aesthetics or something else. And we can even add assets which is images, logos, etc. that the AI can utilize when creating proposals, contracts, whatever else. And I don't want you to get overwhelmed by this because AI actually creates all this for you. I think it's just good to know that all skills are are just some files and some folders.
Quick pause in your regular programming. This video is brought to you by me as always. So, two quick things. First off, below is a 30-day AI insight series, completely free. You'll get 30 insights in your inbox so I can apply AI to your business and your work. The second thing is if you'd like to work with me, blow a series of offerings to see if there's a good fit between the two of us. Now, let's get back into the video.
And one additional thing I want to call out that's really important is at the top of the skill.md file. So this MD file I showed you here, there's one section that matters most, and that's the description section. And I did say the AI is going to create this for you. But knowing that this is important is going to come in handy in the future when you have to manage a bunch of skills that you've created over time. And the sole purpose of the description is to do two things. First, it tells the AI what the skill is for. And the second is when to use it. That's it. And we'll talk about this later, but just know that this specific section is important.
And you might be wondering, how can I take all the instructions and the skills and give it all to the AI without it overwhelming the AI's head and its memory? How are we working around that limitation? Well, with skills, what's happening is we're only showing the AI what it needs to know at that point in time and not everything at once. Simply said, we're loading these skills on demand. So, let's say that we've given our AI the instructions and in those instructions, we've told it that it has access to 10 skills that it can use. What it's going to do is it's only going to look at that description that I just told you about. That description is one to two sentences. That's it. It's very succinct and tight. Once it finds a skill that it cares about and it needs to use, that's when it only actually loads the other information. So, it'll first load the skill.md file. Then, once it loads that, it'll determine if it needs to grab anything from the references, the scripts, or the assets if they exist. So, it's progressively loading the information only when necessary. And that's why they're so powerful when you use them with the instructions.
Now, what is the difference between the instructions and the skills and how do they actually work together? Well, we already know that the instructions that we give the AI, the prompt, they're always on. It's always going to see this every single time. So, it loads up the entire thing. And we know that skills are on demand. The best way to think about these is the instructions act as the manager. The manager has the overall big picture here. So, the instructions tell the AI what its purpose is, the overall rules it needs to follow, and embedded within those rules are some callouts to skills that the AI can use when necessary. So think the instructions are high level big picture manager and the skill is the specialist. So maybe one skill is really good at writing SOPs and another one's really good at reviewing contracts and all other types of skills. And that's the clear separation between instructions and skills.
Now when do we actually use skills versus instructions? Because it's important to note that you don't always need skills. Honestly, most of you will use instructions more than you'll use skills at least right now. So it's important to know when to use which. So there's a simple rule of thumb for skills. If you have a task that's very repetitive, something that you do more than three times a week, and it has an extremely strict format or process it needs to follow to give you a high quality output, that then qualifies for a possible skill. And here are simple four quadrants that break out the different areas. So we have the skill that I mentioned previously. So something that's repetitive and has a strict output. So that could be proposals, SOPs, and reports that have a very specific brand design that you care about at your company. For instructions, these are repetitive, but the formatting and process is a bit loose. So this could be brainstorming with AI or doing different types of research. And below these two repetitive options, we have one-offs. So a one-off with a strict format could be just a detailed prompt for one special deliverable. Another option is a one-off that's loose. So this is basically quick questions to AI doing basic research. And that's how we know when to use skills versus instructions.
After this, we have creating skills. So as I mentioned, AI is going to create the skills for you. All you need to do is if you're using CodeX, which is from OpenAI, this is already baked into the tool. So skills are readily available. All you have to do is use the dollar sign. So here I have CodeX that I've opened up. If I go to the chat window here and I do a dollar sign, you'll see I have a series of skills that are available. Some are personal and some are system based. So you know it's system versus personal based on these items on the right hand side. So you can see all these say personal and these here say system. The system means it's from OpenAI. It's a it's a skill they provided. Personal means it's a skill I've created or AI's created on my behalf. And here you can see there's a skill from OpenAI that is a skill creator. And that's what we're going to use. Claude provides the same exact thing. It's a skill that creates skills. Very meta, I know.
And now we have Claude open. And what I want to show you is the importance of turning on skills and ensuring skills are enabled. So you can get the benefit of having the skills creator create the skills for you. I just said skills a lot. The way you're going to do this is on the left hand side you have the option of customize. So you select customize. Under customize you'll go to skills. Under skills you have my skills. So these are skills that I've created personally. And then below that we have example skills which are the skills that Claude provides. In here, you can see one that's called skill creator. And as long as this is enabled and skills are turned on, and you can see that it's enabled by this little uh button up here, you can then have AI create skills on your behalf. That's really important. You need to make sure that's switched on when doing these methods I'm going to show you.
And there are three ways we can use the skills creator to create skills for us. The first one is probably the easiest. This is method one. In this method, I'm assuming you've already had the conversation with AI and it's already succeeded in providing you the output that meets your expectations. At the very end of that conversation, all you have to do is copy and paste this prompt at the bottom with the skills creator enabled. And in this prompt, all we're saying is I want you AI to use a skill creator and take everything we just did, the steps that we followed, the format that we used, the decisions we made, etc. And I want you to turn this into a reusable skill that I can use in the future. It'll then go through the process of looking at the entire conversation and then creating a skill from that. This is our first method of creating skills.
Our second method is a reverse AI interview. So, let's say that you've not had a conversation with AI yet, but the process is in your head. We need to get it out so the AI has it. And all we need to do is again copy and paste this prompt into a blank chat, a new chat. Give it some context to what the skill is for. And that's what this highlighted purple section is for. So, you're basically going to say, "Hey, I want you to use the skill creator. I want you to create a skill for and probably use dictation here. So, turn on dictation, talk to the AI, tell exactly what the skill is for." And then after that, we're going to ask the AI to interview us to get more nuance around the skill. So, it's going to ask us one question at a time, ensuring every answer we provide informs the next question that it asks. It'll ask between 15 and 20 questions. At the very end of that process, it'll spit back to us a skill that's ready to use. This is our second method.
Our third and final method is that if you want to start from scratch without using the first two methods, you want to answer three questions very specifically. And here again, I'd recommend using dictation so you give as much context as possible. So, the first question we're going to answer is, what does this skill do? want to be as specific as possible around the process, the steps, the formatting, everything. After that, we're going to answer the question of when should I use this skill? Specifically, are there any trigger phrases that the AI should look for when you're talking about a task to know when to call the skill? And then finally, we want to give the AI an output of what good looks like so it can store this example for future reference when running through the skill. So, you'll simply copy and paste this prompt here. You'll then dictate everything I just stated to those three questions below that. Hit enter and let the AI create the skill for you. Those are the three ways we can have AI create skills for us.
So, now that you're confident in creating skills, you may create too many skills. And that's something I see a lot of people do and they manage them poorly. And this matters because the AI starts to call the wrong skill at the wrong time and give you bad outputs. And this brings us back to the description. Remember I told you about the skill.md file with the description and I told you how important that one to two sentences were. If you have too many skills and the descriptions start to look similar, the AI gets confused and it doesn't know which skill to call. For example, let's say that we have a client proposal skill and a client quote skill. The AI looks at both of these and it's not sure which one to call, so it calls one of them and may get it wrong. Same thing applies to a project scope or project plan. It's important that we avoid this so we don't degrade the AI's output.
So having too many skills is a problem that a lot of people run into and I want to show you how to avoid it. So the first thing you need to know is how many skills you should to create for any given task. So, if you have a very specific task, I'd recommend the sweet spot is between five and 10 skills. If you have any more than that, you're going to overwhelm the AI likely because too many skills look too similar. And there are two primary ways we can mitigate this specific issue. The first one is an AI audit. This is where we simply have the AI audit our skills and look for skills that are too similar and helps us clean them up. Here's the prompt that I recommend you copy and paste. So, here we're saying, "Review all my skill descriptions. I want you to flag any that could be confused with each other, meaning they're too similar. And then after you've reviewed them, I want you to tell me which skills I should merge, which ones should be rewritten, and then for those that should be rewritten, I want you to provide clear descriptions that makes each skill more distinct." So the AI will then provide those suggestions. After it's provided the suggestions and you're happy with those, then you can respond back to the AI, go ahead and make the changes for me. So it'll update the skills for you. That's our first approach to mitigating this issue.
The second approach is actually using agents instead of chat tools. So if you're using ChatGPT, Claude, or Gemini in the web browser, you're using the chat tool. But if you use the desktop tools such as CodeX from OpenAI or Claude Code or Claude Co-work from Claude, then you have this second approach available to you. And the second approach is basically having your skills broken into two areas. We have global skills. So these are skills that the AI always sees at all times in any conversation. And these skills should be generalized skills that you use across projects. So things around writing emails effectively, recapping stuff, etc. And then we have project skills. So a project skill are skills that are dedicated to a certain folder. So say you have a series of folders on your desktop. You have a folder called Client A, you have a folder called Content, and you have a folder called Operations. Each of these folders can have their own special skills. So when the AI opens this folder inside of CodeX or Claude Co-worker, Claude Code, it only sees the skills in that specific folder and not the other skills in the other folders. So it's not distracted by these and it's not going to pick the wrong skill. And this is another way to mitigate and manage many skills you've created.
Okay, so we've talked a lot about skills. Now it's time that you actually get your hands dirty. So I want you to do this right now. I want you to choose a task that you want to apply a skill to. And the question you need to answer here is for that task, have you already done it well with AI? If you've done it well with AI already, you're going to choose path one. That's where you're going to basically apply the complete and capture approach where you go back to that conversation where the AI provided the output you wanted and you're going to copy and paste the prompt I provided previously to have the skill creator create the skill for you. If you've not done it already with AI and it's in your head, you're going to run through the reverse interview where you have the AI ask you a series of questions. At the end of that, it produces a skill.
So, as a recap, a few things to remember. First off, more instructions doesn't always equate to better outputs. Actually, if you have too many instructions, the intelligence of the AI degrades because it's filling up the head too much. Next is the difference between instructions and skills. So, you want to think of instructions as the manager. So the high-level perspective focusing on what's the purpose of the AI. What are the general rules that it needs to follow? And within those rules are call outs to specific skills that the AI will call and have that specialist skill take that work over for them. Next is what use cases are most suitable for skills. And these are going to be use cases where it's a repetitive task, something they do more than three times a week, as well as has a strict format on either the design or the process associated to it. And then finally, there's three ways we can create skills, all of which AI can do for us. We start with a complete and capture approach if you've already had success with AI in the past. We then move on to the reverse interview if it's all stuck in your head and you've not already interacted with AI on this. And finally, if we want to start from scratch, we just answer three questions. And those are what should this skill do? When should I use it? And what does good look like? And that's it.
So, as a reminder, two quick things. First off, Blow is a 30-day AI insight series completely free. You'll get 30 insights in your inbox so I can apply AI to your business and your work. The second thing is if you'd like to work with me, blow a series of offerings to see if there's a good fit between the two of us. Now, imagine you drop a meeting transcript into a folder and your skills kick in automatically. One pulls out action items, one drafts the follow-up email, and the third updates your CRM automatically, all at the same time. No uploading, no copying and pasting. And I watch my clients do this every week. And it all starts with agents. I'll show you exactly how to do it right here. So, go and click that video. Go ahead, click that video, internet. I'll see you next time.