Transcription
Creating Claude and ChatGPT skills too early can actually make your AI worse. They don't just save your process, they save your mistakes and repeat them faster every time.
If you're new here, I'm Dylan. I run an AI consultancy and this exact problem comes up in almost every coaching session that I have. So, I'll show you a four-step loop I use to build skills worth trusting and two steps most people skip. So, let's get into it.
There are many different ways to build skills in Claude and ChatGPT and Codex, but this is a specific loop that I recommend all my clients follow. I'm going to walk you through each one of these steps in detail, including prompts that you can copy and paste and use for yourself as you work through each stage.
As a quick overview before we get into the details, we first start with a mapping. So, we need to map our process before we actually encapsulate it into a skill. After you've mapped it, which is kind of an optional step, you move on to the proof step, which is completely mandatory. And this is where we actually prove the work out with an AI and get an output that we're happy with. Once we've done that, we then capture that proof into a skill. So, it's important to note that steps two and three are absolutely critical to have a high-quality skill. After you've done those two, then we move on to the patching piece because even if you follow these steps, sometimes the skill goes off the rails. And if it does, there's a very specific way you want to go about fixing it so your skill doesn't degrade even further. And those are our four steps.
Now, the issue I see with a lot of people is what happens is they skip steps two and four, and they tend to bake in an idea of a process into a skill without proving out the process in detail. This tends to create over-bloated skills that never meet your expectations, and we want to fix that.
But before we create a skill, we need to first ask ourselves, do I even need a skill for this specific process? And there's three questions you can answer before you do that. And you want to answer these questions in succession. So, you go in this order: first, second, third.
So, our first question is simply asking, "Will I do this again?" Is this a repeated task that I'll do over and over and over? If so, it might be a skill, but we need to move on to the next question.
Which is, "Does the quality matter each time?" Is there a really high-quality standard you have for this task? And does it need to be consistent every single time? If yes to this, then you move on to the final question, which is simply understanding, "Is this specific activity something that can be done across conversations?"
Now, what do we mean by that? As an example, every company usually has some sort of branding guidelines. So, we have a logo, we have colors, fonts, etc. So, anytime somebody writes a proposal or creates a presentation, they have to be in line with those guidelines. Well, you can bake those brand guidelines into a skill. So, anytime somebody tries to create a presentation, they're going to follow that skill and call it. So, it makes sure that presentation meets the guidelines for the company. So, that's a skill that works across conversations. Those are three questions. So, if we answered yes to all these, we can then move into the loop process.
Quick pause and then regular programming. This video is brought to you by me, as always. Two quick things. First off, below is a 30-day AI insight series, completely free. You'll get 30 insights in your inbox of how you can apply AI to your business and your work. The second thing is if you'd like to work with me, below are a series of offerings to see if there's a good fit between the two of us. Now, let's get back to the video.
And the first step here is mapping. Again, this is optional, but it is a booster if you want to start with a really detailed process to begin the next phase. And my recommendation for mapping out that process is doing an AI interview. If you've watched any of my videos in the past, you've probably heard me talk about this. Where what you're going to do is you're going to have the AI ask you one question at a time. Every answer you provide is going to inform the next question the AI asks. And we want it to do one question at a time. And I usually recommend 13 to 15 questions. This is the happy medium. It's not too many to overwhelm you and not too few to not provide value.
And here's a prompt that you can use for this. So, in this prompt, we're asking the AI, "I want you to create a skill for this recurring task." Then you fill in that blank yourself. Then we're saying up front, "I do not want you to create this skill yet. Instead, what I want you to do is I want you to ask me one question at a time. Cap the interview at 15 questions. Your goal for this interview is to understand this specific task and detail associated to the inputs for the task, the outputs, the quality standards, common mistakes, edge cases, etc. After you understand this task thoroughly, I want you to summarize it back to me." And once we get this back, this is going to be the starting point of our next step.
And that's when we actually do the task. And this is by far the most important thing you can do to create high-quality skills. Is open up a chat, a fresh chat in ChatGPT or Claude or whatever else, and do the thing. Create the document, do the data analysis, do whatever that task is from end to end until you're happy with the output. And the reason we're doing this is often times what people do is they base their skills on ideas. They don't actually do the task with the AI. And by doing this, there's a lot of nuance that you're missing, and the AI is going to have to fill in those gaps for you, which often leads to poor skills. But if you actually do the task with AI, you have evidence. So the AI knows exactly what you did and how you did it.
And after you've done this with the AI, you can then use this exact prompt to extract that and put it into a skill. And this prompt goes at the very end of that conversation after you've got the output that you're happy with. So here we're saying, "I want you to review this full conversation. After you've reviewed it, I want you to create a skill from the process that I just proved above. Only keep the parts that are reusable." And this is important. We'll talk more about that later. We then state, "Remove anything specific to this one client, this one date, this one data, etc." So we don't want it to be biased to a given example. We want the actual process that can be abstracted. And that's what we emphasize next. We then state, "This skill should work on future inputs with the same kind of task, not the exact same task, but the kind of task, including the steps, quality checks, output format, and common mistakes." So you can copy and paste this prompt at the end of your conversation to create that skill. And ChatGPT and Claude both have skill creators that can create the skills for you.
Now I emphasize the importance of not necessarily allowing the AI to bias to a given example. The way I structured the previous prompt, I did that intentionally because I wanted to make sure that the AI, for example, if we're doing something in relation to bank reconciliation or payroll process. When you do that specific process with the AI in the chat to prove out the flow before you create the skill. You want to make sure the AI doesn't bias towards a given specific time period for that activity. Or a specific file, or a specific client that you did the process for, or even columns if an Excel sheet is involved. Instead, what you want it to do is focus on the abstractable elements, such as the period of the payroll, the kind of file, any client you can use this for. You can use it for any client. This is just a flow associated to one, and a variety of other things. So, it's important that we create skills that are abstractable from the specific topic we used to prove out the process.
And when creating this skill, there's one last thing I'd recommend adding to it to take it from really good to great. And that's when you have the AI create its own criteria at the end of the skill to grade itself before it provides the output back to you. So, if it's going to provide me a document, some sort of analysis, or whatever else, I want it to have binary criteria to say, "If I pass these criteria, I then have met the user's expectations." And it's important that they're binary. The reason that it's important that they're binary is 0, 1, is that if you give it a vague input for criteria to judge itself on, such as "Make sure the output is accurate." Accurate is subjective. It really depends on the person you're talking to. So, we want to take any subjectivity out of the grading criteria. And a good example of this is if the AI's writing something for you, maybe it's an email, a document, whatever else, you can say, "Every action item has to have an owner." This is a binary piece of criteria. So, that's finishing off our third step, which is capturing the process into a skill.
And finally, we have the patching phase. So, even if you follow those three steps beforehand, there's a chance the skill will still go off the rails later on. And if it does, you want to fix it, but you want to fix it in a very specific way. The key thing with the step in the loop is that when you have the AI update a skill, because you can have an AI update a skill for you in a chat, when it makes those updates, you want it to be surgical in its edits. Because if you don't specifically state this in your prompt when you ask it to fix something, it'll do one of two things. It'll either rewrite the entire skill, which is unnecessary, or the thing that it adds will be massive. It'll take up a huge portion of the prompt, bloating the skill. And you don't want bloated skills. Instead, you want a hyper-targeted update that just fixes the issue that you had.
And this is the prompt that you can use for that. It's a very simple prompt. All we're saying is, "I just corrected the output because you just fixed the output in front of the AI." After you've done that, you're saying, "I want you to make a surgical edit to the skill. Remember, surgical is a keyword here. So, it does this correctly in the future every single time. Do not rewrite the entire skill. Only add the rule needed to prevent this mistake going forward." So, we're focusing the AI's attention to make a minor update to fix this specific issue. And those are our four steps in the loop.
Now, you have superpowers to create skills, and you might get overly excited. So, I want to add a quick caveat here not to create too many skills, at least depending on where you're located. Reason being is that if you're in the browser, which most of you likely are if you're using ChatGPT in the browser or Claude in the browser, if you create a bunch of skills, the AI is always actively looking at those skills, specifically the title and the description, at all times, irrelevant of the conversation. So, if you have too many skills that are named similarly, such as these, where you have a client report, a client update, a client summary, a client digest, there's a chance the AI chooses the wrong skill at any given time. So, you want to mitigate that by having a few skills that matter a lot to you that have distinct titles and descriptions.
And now, this is specifically for in the browser. But if you're inside of the desktop using things like Claude Co-work or Codex from OpenAI, you can have more skills because those skills can be specifically tied to a given folder on your desktop. So, anytime the AI is in another folder, it doesn't see any of these skills. So, you can scale the number of skills without overwhelming the AI and it choosing the wrong one. But remember, this is just for the desktop.
With that caveat out of the way, let's do a quick recap of the loop. The big thing to remember here is that skills don't simply save prompts. What they do is they save your judgment. And if you can save more and more complex tasks and the associated judgment with that in AI, you can create more leverage in your job in your company. And the way that we're going to do that with these skills is the four steps.
So, the first one is mapping. And remember, this is optional. You can go straight to the proof step if you'd like, but if you would like to do the AI interview to map out your process before you do the proving, you can do that here. After that, we go to the proof stage, which is completely mandatory and needed. And that's where you show your work in front of the AI. You do it in a chat and you get an output that you're happy with. Once you've gotten to that point, then and only then do you create the skill, capturing that proof, and that's our third step. And remember, in the capture phase, we want to capture the things that are reusable. We don't want the AI to bias towards any example that we used to prove it. We want it to be extractable content, as well as adding criteria at the end of the skill that's binary, so the AI can grade its own work before it gives you a response, increasing the quality of that skill. After we've done the capturing, we then move to the patch phase, where even if you follow these three steps, there's a chance the AI still goes off the rails a little. And when making those fixes, we want to make sure the AI makes those fixes in a surgical way, not bloating the skill unnecessarily. And that's it.
So, as a reminder, two things. First off, below is a 30-day AI insight series, completely free. You'll get 30 insights in your inbox for how you can apply AI to your business and your work. The second thing is if you'd like to work with me, below are a series of offerings to see if there's a good fit from the two of us.
Now, before you build any skill, here's a question worth asking: Should this even be a skill, or should it be a project? Those are two different tools, and most people pick the wrong one. I made a video that walks you through when to use each, and it's right here. So, go and click that video. I'll see you next time, internet.