You're Creating Claude Skills Wrong. Here is my 5 step process.

Most people building claude skills are obsessing over the wrong part. They spend hours polishing the brain, the instructions, the steps, the logic. They write 400line documents. They tune in every single word and then Claude never even opens any of it. Because the part that decides whether a skill runs at all is not the brain, surprisingly. It's the face. The few lines that Claude reads before it even cracks the skill open. And by the end of this video, you will know every part of a working Claude skill. the face, the brain, the memory, the spine, and the pulse. And you will see exactly which one your skills are probably missing. Now, I've built over 70 skills in Claude for myself and others over the last 6 months. Some run multiple times a week without me touching them, and some I built and never saw again. And the difference between the two has almost nothing to do with how well I wrote them. It has everything to do, however, with whether I built every part of the anatomy or just some. Now, quick thing before we go in. This is not a developer video. You don't need to know how to code. You don't need to know much about AI. Even if you've ever tried to make Claude do the same task twice and gotten two different answers, then this video is for you. So, here's the thing that is burning your time. Most people think a claude skill is a really well-ritten instruction document, detailed and thorough and comprehensive. And the longer and more careful the instructions, the better the skill. That's not true. A skill is like a body with five parts. The instructions are just one of those parts. Okay, just one. And every other part is doing a job that instructions cannot do. McKenzie found in their most recent state of AI report that regular Genai use inside organizations roughly doubled in under a year. Now the people who are proving to do this well are not the ones using AI more. But it's those who stopped reprompting claude every time they wanted the same task done. They started to build it once and then create skills around that because they were able to build it properly. Okay? And then they walked away and let it run. Remember this, a prompt is rented. A skill is owned. Every time that you copy paste the prompt into Claude or Chad GPT, you're starting from zero. You're basically paying for the same setup over and over and over. A skill, however, is the moment that you put one piece of your business into something that runs the same way every time without you in the loop, but only if it has every part of that anatomy. So, here's what almost nobody is sharing online. The five parts, that's the whole video. Like I said, the face, what Claus sees first, decides whether the skill even opens, actually. The brain, where the thinking happens, the instructions that everyone obsesses over. The memory, that's the heavy stuff pulled in when the brain calls for it. The spine, that's the structure that holds everything up tight. And the pulse, what keeps the skill alive over time instead of rotting in a folder. those five parts. We're going to go and dissect all of them and then you'll know exactly which one your skills might be missing. All right, so let's talk about the face. The face is the description of the skill. Those few lines that Claude sees for every skill installed, every conversation all the time. This is the highest leverage part of the entire skill in my opinion. And almost nobody treats it that way. Now, here's what most people write. Reviews, contracts. That's not a description. That's a label. Okay, here's what actually works. The description has to do three things. Number one, tell Claude what the skill does. Number two, tell Claude when to use it. And three, tell Claude the exact phrases a person would type to ask for it. Because Claude under triggers by default, okay? It will hold back. It will keep going without your skill unless the description grabs it by the collar. So, your description has to be slightly pushy. Always load when so and so. That tends to work for me. So, maybe you can give that a try. Now, a few months ago, I built a skill called red line. It reviews contracts before I signed them or I sent them to my lawyer. Client deals, agency builds, partnership agreements. The instructions were really detailed. The examples were sharp. I was so proud of it. And for the first two weeks, Claude never picked it up. not once. And what would happen would be that I'd paste a contract into a chat and then Claude would just read it and review it politely but generically. So I went back and rewrote the description. I added the exact phrases that I would type when I wanted a contract reviewed. Something like review this MSA or check this brand agreement. Is this safe to sign? Just simple plain language. You don't have to be super sophisticated because it needs to recognize your actual language and you write that at the top and then you don't even have to change the brain. I didn't I didn't change one single line but after that Claude reached for red line every time I dropped a PDF in the same skill the same face. So the rule is that you need to lead with the code name and one sentence of what it does and then give concrete trigger contexts and then five to 15 literal trigger phrases. You should stay under a thousand characters ideally. So don't over stuff it either. But you need to think of the face as rooting logic, not a summary. Okay? If you fix nothing else about your skills today, please fix this. Okay? Now let's talk about the brain. This is the body of instructions. The part that everyone writes for. It's the part that everyone fights over. The mistake here is not writing too little. It's writing the wrong kind of instruction for the kind of task. I need to explain myself. So basically you need to think about it this way. Some tasks need judgment. Writing copy, designing an offer, reviewing a deal, tasks where there are many right answers and the work is figuring out which one fits best. And for tasks like this, you don't give Claude rigid steps. You give it principles, examples, the way you actually think about the problem and then you let it reason. Okay? And then you've got other tasks that need exact execution like formatting a document a specific way or packaging a file or running a validation checklist. There you only have one right answer and variation is failure. Okay? So for those tasks, you need to give Claude precise steps, not interpretation. No room to be creative for those. Okay. The failure mode is mismatching these two situations. I've seen rigid scripts on judgment work and those make Claude brittle because it follows the script and the work ends up feeling mechanical. Or I've seen vague pros on precision work which ends up making Claude inconsistent because you have the same input but different output every time. I think of it as a freedom dial. Every task that you create a skill for has a dial setting. High freedom on one end and low freedom on the other. Your job is to set the dial correctly before you write a single instruction. If you're not sure, you can ask yourself one question. If I gave this task to five smart people, would I expect five different answers or one? Five different answers. That's judgment work. It's a lot of freedom. If you expect one answer, that's precision work. So you need to lock down the freedom because the brain is not about length. It's about matching freedom to risk or fragility. Now let's talk about part number three and that is the memory. The memory is the reference files. The deep stack of your information that your skill might need but doesn't need every time it runs. So here's the principle. Cheap but decisive information lives in the face. Working instructions live in the brain and then heavy, rarely needed material gets pushed down into the reference files. And if you're wondering why, well, it's because Claude's context window is a shared resource. Every word of context that you load is one less word that it has for the actual work, if that makes sense. So the memory lets you store a lot of knowledge, your full brand guidelines, three years of past examples, a 100 sample outputs without paying the context cost on every single run. The skill only pulls the memory in when the brain points to it. So for example, if the user is writing a long form post, then the brain says read the long form examples in references or long form.down. And that's the trigger. Without it, memory sits on your desk and costs you nothing basically. And most people never use this part. Their skills are one fat document with everything crammed in. And it runs slowly and it costs a lot. And it makes Claude less precise, not more, because it has to filter through a lot of irrelevant context every time it runs. The memory is how you build skills at scale. Now, part number four is the spine. The spine is the structure of the skill. the H2 sections, the consistent layout, the skeleton that holds the whole body upright. And I know this sounds boring, but it's the most underrated part. Here's why. A skill with a good spine is addressable. Because you, as well as Claude, can point to the selfch check section of whatever and find it instantly. The skill becomes a building block that other things can lock into. Without a spine, every skill is its own snowflake and that cannot compound. The skeleton that I use every single time is how to use the skill, north star, core sections, one per concept, anti- patterns, examples, selfch check, and known gaps. Same order, same names, every single skill. And no, I don't remember this. I have a skill to create other skills. By the way, I mentioned all of those and I walked through them in detail in this video here and linked in the description down below. So, you can go and watch that as soon as this one's over. Now the other discipline is to keep the body under 500 lines. Under 350 for any skill that gets loaded consistently. Long doesn't necessarily mean better. Long means slow. And slow means claw does less actual work and more reading. Okay? A skill should be skimable. If you can't scan the whole thing in 30 seconds and know exactly what it does, then the spine is probably broken. Okay. Now let's talk about the fifth part and that is the pulse. The pulse is what keeps the skill alive over time. This is the part that most people skip actually because they write the skill, they implement it, and then they never touch it again. And 6 months later, the skill is dead. It's not broken, it's dead. The world moved on, but the skill didn't. So, here are five things that give a skill a pulse. Number one, one term per concept everywhere. If you call something client in one place and customer in another and user in the third, Claude has to disambiguate every single time. So, pick a word and commit to it. Number two, no timestamp language. Never write as of 2025. Use a current method versus old pattern. That's the structure you can use to keep the skill making sense as things change. Number three, real examples. Actual input, actual output. Do not give abstract descriptions. Number four, an honest known gaps section. You need to write down what the skill does not do yet. Pretending that your V1 is complete basically creates invisible debt that surfaces 6 months later when something breaks and you don't even remember why. And trust me, Claude is going to keep looking and trying to do that if you don't tell it that it's not supposed to. And number five, one skill, one job. If a skill does two things, write the script and publish it, split it. Two well-triggered and narrow skills always, always beat one broad skill that fires unpredictably. That's the pulse, okay? A 10-minute maintenance pass that adds 2 years to the skill's life. Now, here's what this actually does for your business. A skill that runs the same way every time on demand without you in the loop, that's infrastructure. Every recurring task in your business has a number associated to it and that's how many hours per week it costs you. A well-built skill can compress that number. Hours become minutes. Okay, every project, every client every week. Honestly, I have a skill that drafts the first version of every client proposal that I send and it pulls my brand context, our pricing logic, my voice. And the first draft used to take more than 30 minutes and now it takes five. One well-built skill can replace dozens of hours of manual work per month. Five well-built skills replace a part-time hire. And then 10 of them, and you're running a business that does more work with fewer people in the loop. Faster delivery cycles, cleaner handoffs, less drift between what's in your head and what actually gets shipped. And obviously, results will depend on what you build and how you use it. But the math on a properly structured skill is real. Now, two places to take this further. If you're just starting and you want to build your first skill alongside thousands of other people, then we have the Trailblazers Hive. The link is here and in the description down below. It's our free community. We run hands-on challenges that are completely free. We've got already 12 challenges that are in there, including a Claude Co one. So, you can go ahead and get started that way. Now, if you need more handholding and if you need our 90-day specific stepbystep structure as well as the upcoming Claude Co-work course, then we are excited to welcome you into our AI founders hive that's also linked here and below in the description. That one is our paid community, but you will get so much value and so much actionable input and a lot of accountability to make sure that you keep moving and you keep up the momentum. In the meantime, thank you so so much for watching. Like this video if you did. Be sure to subscribe if you haven't done so. Share this with anyone in your circle of friends or family or co-workers who knows they need to be building more claude skills and leveraging their time in a different way, but they're not. Hopefully, this helps them as well. And until the next time, I suggest you go ahead and watch this video here. And I'll see you soon.