§ 01 · The Hook
The bait, then the rug-pull.
A $4-million-a-year business run primarily through AI agent skills sounds like a demo hook — but Nick Saraev spends this course making the case that most people are using skills wrong, chasing personal-assistant novelty while the real money sits in three unglamorous places: lead generation, follow-up, and cold outreach.
Where the time goes.
00:00 – 01:05
01 · Introduction and credibility hook
Business does over $4M/yr profit via AI agents. Course preview: five demos, then how to build.
01:05 – 01:37
02 · The business case for skills
Skills are the evolution of SOPs — for agents, not humans. Most demos miss the real value.
01:37 – 04:31
03 · Demo 1: Lead follow-up nurture
Reads CRM pipeline stages, pulls full email chain history, writes personalized follow-ups per prospect. Runs in one shot each morning.
04:31 – 05:39
04 · Demo 2: Thumbnail generator
Superimposes host face onto reference viral image; generates multiple variants for selection.
05:39 – 07:21
05 · Demo 3: Lead scraper
LinkedIn Sales Navigator query in natural language -> Google Sheet with full names and email addresses.
07:21 – 09:35
06 · Demo 4: Cold email campaigns
Reads existing high-performing campaigns, rewrites them for a new client niche, sets up in cold email platform.
09:35 – 11:30
07 · Demo 5: Website builder
One-prompt website generation deployed to Netlify; used as outreach value-add for prospects.
11:30 – 13:07
08 · Bonus: WeWork booking
Chrome DevTools MCP books coworking desks 30 days ahead automatically on trigger.
13:07 – 15:26
09 · Bonus: Amazon shopper
Chrome DevTools MCP compares products on Amazon, returns structured cost/quality recommendation.
15:26 – 20:49
10 · How skills work mechanically
SOP to markdown translation, front matter, progressive disclosure, skill matching. Spec sheet walkthrough.
20:49 – 21:05
11 · Skills are self-annealing
Agent patches its own errors: identifies failure, fixes script, rewrites skill file automatically.
21:05 – 34:21
12 · Building three skills in parallel
Meeting notes extractor (pure markdown), invoice data extractor (markdown + Python), content repurposer (parallel sub-agents). Live builds in three simultaneous Claude Code instances.
34:21 – 42:51
13 · File structure and hidden folder walkthrough
.claude/skills/<name>/SKILL.md structure. How hidden folders work on Mac and Windows.
42:51 – 47:44
14 · What skills are worth building
Front-end first: revenue-touching tasks only. Lead gen, outreach, follow-up beat internal fulfillment pipelines on ROI.
Visual structure at a glance.
Lines you could clip.
19:23
"If you have an SOP, you have a skill."
20:49
"Skills are self-annealing — they heal themselves, they get better, they improve constantly."
43:23
"You can do anything — but that doesn't mean you should do everything."
44:26
"I was doing somewhere between 50 to 100 cold calls every single day. When I was going door to door, it was something like 80 knocks on physical doors per day. I wish I had a hundredth of the technology we have today."
Things they pointed at.
01:05toolAntigravity (Quad Code agentic IDE)
05:39toolLinkedIn Sales Navigator
09:35toolNetlify
23:00toolWhisperFlow
11:30productWeWork
How they asked for the click.
45:45
subscribe
"If you guys like my content, please subscribe. 69% of the people that regularly watch my content are not subscribed."
Mentioned twice (5:39 and closing). Frames it as a channel growth mechanism rather than personal ask. Effective because it gives context (algorithm impact) rather than just requesting.
Word for word.
HOOK opening / re-engagementCTA the pitch
00:00HOOKHey, welcome to the definitive resource all about skills. I currently run a business that does over $4,000,000 a year in profit and I manage it primarily through AI agents and skills. And I teach over 2,000 people how to do the same thing.
00:11HOOKUh, I think a lot of demos and walkthroughs of skills right now are, like, really flashy and they're typically centered around the personal assistant angle. But a lot of people are leaving tons of money on the table because they're not applying them to specific business use cases that actually tend to produce large returns on investment.
00:27HOOKAnd so what I wanted to do in this resource is I just wanted to walk you guys through what that actually looks like. I'm gonna start by showing you guys examples of a bunch of skills that I currently use in my $4,000,000 a year business. Skills that other people are currently using across a variety of different industries like e com, service businesses, consulting, and so on.
00:44HOOKAnd then I'm gonna show you guys how to build them. And we're not just gonna build cookie cutter skills that are glorified personal assistance, but skills that you guys could actually implement in your business to do the work of dozens of people in just a few minutes.
00:54HOOKOkay. So everything you need is in the link in the description. It'll give you guys these skills, so you guys could use them across your own business.
01:00HOOKIt'll also give you guys the framework that you could use to to build new ones regardless of whatever use cases. No fluff. Let's get right into it.
01:07So if you don't know what I'm looking at right now, this is anti gravity running five different quad code plugins. If you guys are new or unfamiliar with this sort of interface, just check the link that I'm pasting right over here.
01:19It'll walk you through everything from like what the icons here mean to, you know, how to communicate with Claude code and Gemini and and other, you know, agentic coding platforms literally from the ground up. So you'll you'll know everything after that. But assuming you guys are somewhat comfortable, I wanna show you guys five different skills today and then I'll throw in a couple of additional bonuses.
01:37The first one over here is gonna follow-up with all of my leads. So if you guys are running a business at really any level of revenue over a few thousand dollars a month, you probably have something akin to like a list of leads or a list of contacts. This type of thing is typically referred to as a customer relationship manager.
01:51And in general, what you do is you store leads at different areas along the pipeline from when you first met them all the way to close deals and money is in your pocket. And so I have an example one here with a bunch of leads at different stages. Meeting booked, Jimmy Clinton, proposal sent Sarah Chen, Jessica Park, Aisha Mohammed, and so on and so on and so forth.
02:09And basically, right now, in order for me to move this pipeline forward, every day I just have to manually check-in with all the leads. Well, I built a skill that effectively automates the entire process. In order to trigger this, all I need to go and do is go backslash follow-up nurture.
02:24Now what this skill does is it'll goes through my pipeline. It identifies, it captures all of the conversations I've ever had with any of these prospects. So it actually goes through all of my email chains and so on and so forth.
02:34And then it uses a couple of templates to personalize check ins, um, depending on where they are. So, you know, if a meeting has been booked but we haven't actually attended it, it might check-in with Jimmy and say, hey, Jimmy. How's it going?
02:44Just wanted to check-in on our meeting in a couple days. Really excited to have it. You know, if it's to Sarah and maybe we send a $2,500
02:51quote for a specific type of product, I'd be like, hey, Sarah, you know, hope you've had a lovely week so far. Just wanted to circle back on x y and z proposal. Let me know if you have any questions.
03:00The whole key here is we do this really informally and then really casually, so they think it's us. I think a big problem that a lot of these automated follow-up services typically run into is they just use really like nonhuman language.
03:11It's very clear and, you know, I think prospects can tell when an LLM does the following up for them. But I also like the fact that this just does it all in one shot. And so the workflow is literally like you just wake up in the morning where your salesperson wakes up in the morning.
03:22Um, they go slash follow-up nurture. They immediately clear out their whole pipeline of follow ups, and then they can just focus on acquiring new business. What's also cool is it does all this in the tone of voice and in the same chain of conversation as the initial email.
03:35And so in this case, this is an email that I sent to mister James that said, hey James, circling back on the proposal, let me know if I can answer any queues. This is another one that I sent to Priya that said, hi Priya, hope you had a great week checking in on that test brief, let me know where we're at. You know, this just continuously customizes based off where the person is, and then it also gets to pull in, you know, context and, uh, more or less everything that you currently talked about.
03:55So you're not just repeating the same thing. This also goes through and then it lists all of the follow ups that I've made to every single person. So you could see here, we're signing off slightly differently between them.
04:03We're using, uh, reply chains and stuff like that to stay within the same thread. Um, and I'm gonna give you guys that for free. I use it basically every day.
04:10The second flow is basically a one shot thumbnail generator. Um, I have this funny image of Jon Hamm or Don Draper in Mad Men, if you guys are familiar. And this image went viral a little while ago.
04:20It's all about vibes and whatnot. And so I figured, you know, as somebody making videos on vibe coding and agentic platforms, I would try and reproduce the thumbnail. But instead of doing it manually, I'm actually just going to say reproduce the John Hamm thumbnail,
04:31use similar lighting ETC. And then what it's going to do is it's going to take this thumbnail or this image, which has already been created, then it's going to go get my face and then it's just going to superimpose my face onto that image in a really realistic looking way. It's also instead of just generating one variant, it's gonna generate multiple so that we can pick and choose which one looks the best and also because not all AI outputs are perfect.
04:53This is an example of the finished product and we did that just in a few seconds. I also ran it again so you guys could see even more. But I mean, just like going back and forth between this, pretty freaking close.
05:03Right? However, there are some important differences. The people in the background are different.
05:07The lighting is just a tiny bit different, and so on and so forth. And I mean, we've we've generated more. I have this running again in the background.
05:13This one's with like a darker lighting and different tone. I don't know. This one here is sort of like matrix style blue and sepia.
05:19And so all these are just a little bit different. Ones that it's not just like a one to one rip of the source image, but also because, know, I get to pick and choose variants that I like the most. And so in my case, really like this one.
05:28I think this like did the most justice to my face, so this is probably the one I'm gonna use for my video. By the way, despite the fact that I hate when people do this to me, I'm gonna do it to you. If you like this sort of thing, please subscribe.
05:38HOOKSomething like 69% of you guys aren't subscribed for some reason, so that's two out of every three people that I'm talking to right now. And it just significantly impacts my channel's ability to grow.
05:47HOOKSo if I've given you anything, any value at any point in time over the course of this video or any others, just click that button for me. You'd doing me a big solid. Okay.
05:55HOOKBack to the video. This next one here is all about scraping leads. And so I could actually just say, scrape me 50 management consultants
06:02in, uh, I don't know, let's just say Arizona. And what this will do is it'll go on LinkedIn Sales Navigator, which is currently the highest quality place that you can get information about, uh, real human beings. It then creates a bunch of LinkedIn search URLs.
06:16And in case you guys didn't know, LinkedIn Sales Navigator allows you to filter people really granularly based off things like what their job title is, um, where they live, you know, their various seniority levels and other industry.
06:29And then basically, just get a list, not within Sales Navigator, but in a Google Sheet with these people and then all of their email addresses. And, like, preexisting solutions exist to do this for Apollo and Apify, but I don't think there's, like, currently a big one for LinkedIn Sales Navigator.
06:45And the reason why that's valuable is because there is no better source right now for b to b lead data than LinkedIn Sales Navigator. It's the most current. Most other services are scrape from there anyway.
06:54And as a result, I got to go straight to the source and then just ask a system for what I want in natural language. So within, you know, three or four minutes, I end up with a nicely optimized list of high quality leads with probably like the highest deliverability out of more or less any source that you could get today.
07:08And then you end up with a really high quality list of people, literal or full names, first names, last names, email addresses, and so on and so forth that we've sourced basically directly from LinkedIn Sales Navigator. Um, this is done pretty effectively via cost as well. And then the fact that it just are gonna occur in the background while I do other tasks, it can retest its leads, It can figure out like the best filters to use and so on and so forth.
07:33Eliminates a fair amount of my own day to day work as somebody that runs like a growth management agency. This next one here writes my cold email campaigns. So for those of you guys that don't know, a cold email is something that goes to somebody who you haven't actually talked to before, hence why it's cold.
07:47And then it tries to sell them on some service. And so I run cold emails just as my day to day for a couple of clients. Essentially, what I commonly have to do is I have to iterate and change cold email campaigns
07:57based off of a new client's information. So And what I wanna show you guys is basically how you can automate this process. I am now gonna use the backslash convention.
08:05I'll go cold dash email dash campaigns. Sorry. Think that's actually a forward slash.
08:09And this will basically ask you, hey, what are the details of the client? And now I'm just gonna voice transcribe and say, the client is one second to copy. Their offer is they want to write free 500 word sample articles with no strings attached.
08:22Find the similar campaigns and then use those as inspiration. And then gonna feed that in, and then my skill is going to go through and then essentially duplicate high performing pre existing campaigns within my cold email platform, uh, and then rewrite them so that they're as similar and high quality as humanly possible for that given niche.
08:42Then sets it all up in my cold email platform. So, again, a step that previously might have taken me thirty to sixty minutes just logistically is now taken care of. And then obviously, all the offer management and stuff like that is also handled entirely autonomously.
08:54So, I mean, takes something that might be three or four hours into something that takes minutes. And then afterwards, we end up with a bunch of different campaigns. This is an example of one of them that says for, I don't know, Nick.
09:03Just wrote you a blog post. Hey, Nick. Big fan of your work.
09:06Was on one second copy site earlier. Thought I'd write you a free 500 word sample article with a twist. I know giving value upfront is how you form connections, so I thought I'd start with that.
09:13If have any interest, just reply and I'll send it right over. If you want more, just say the word. Got a few other interesting pieces of content for you.
09:18It also gives you different variants that you could use to split test and then also even like a little follow-up. So in that way, you know, me just templating out like 90% of my work significantly streamlines the actual offer building process.
09:29And nowadays, because the models themselves are getting really good, like I was doing this with Opus 4.5, now I'm doing it with Opus 4.6. Because the models are getting really good, like half the time I don't actually make a change. I just say, okay, sure.
09:40Let's give this a test and we'll see how this performs relative to some of the other ones. The last skill I wanna show you is a simple website builder. Basically, make a long story short, you know, we do a fair amount of outreach.
09:49My goal here is I wanted a way to quickly and easily whip up high quality, reasonably templated, but also pretty unique websites for clients or prospects. And then just send them over whether or not I have their business. I was thinking about ways to offer more value while reaching out to people, cold.
10:05And, um, me literally building like end to end high quality websites and then giving them a link that they could use to access them in one shot. It's probably like the biggest value add that you can get so far. So this is like very, very low hanging fruit.
10:17And I would also say that there's quite, uh, room right now to do knowledge arbitrage, where basically, because you have access to tools like this and then most prospects don't, you could offer either completely free websites or you could offer like a really simple or templated website or something like that in order to really blow their socks off.
10:32When it's done, it actually pushes this to Netlify, which is a hosting service that lets me get a website like this. Then I can actually just click this link and then I have the website right over here. And I don't think you guys could see it so super clearly, but, um, it's pretty high quality.
10:45You know, it uses its own customized prompt in order to build slightly different navigation bars and menus and then, uh, different content types and stuff like that. And then you end up with, I think, what most people consider to be pretty high quality. In case you guys are interested, I showcased a really similar builder on my Claude code course a little while ago and then my Gemini course as well.
11:04Um, you know, this is pretty high quality. Right? I think if an average person were to get a website like this for free, they'd probably be like, no way.
11:09Really? There has to be a catch. But nowadays, you can generate these deliverables for basically cents on the dollar and then get like high quality animations and stuff like that, um, with literally you just putting one prompt.
11:21So obviously, I wanted to build a bunch of economically valuable skills and I have dozens more, probably close to 30 or 40 right now that manage most aspects of my business. But I also wanted to show you guys sorry, that's my alarm here. I wanted to show you guys how easy it was also just to build skills that handle like really nuanced one off things.
11:38And so, um, in addition, I also built a couple of cool ones that I wanna show you. One called WeWork Booking, another called Amazon Shopping. And I mean like, these aren't even really as business y as they are me just like dealing with a problem that I had.
11:49So for the WeWork one, um, WeWork is a shared coworking space, which basically allows you to book and then show up sometime in the next few hours. Uh, you have access to things like kombucha on tap sometimes. Uh, I don't know.
12:01Free coffee. Well, it's not free. I guess you're paying for it.
12:03Looks like their marketing has worked on me. But to make a long story short, in order for me to access WeWork, I need to like book ahead every day.
12:10And one day, I tried booking the same day, and then I showed up and my card reader didn't work. And I was like, why? And they're like, well, it's usually good to book ahead because sometimes the card reader takes time to sync.
12:19And I was like, well, sucks. I don't wanna freaking do this. And then I realized, wait a second, I have Cloud now.
12:23So why don't I just write a skill that just automatically books me every single day for like the next thirty days? And so that's what I did. I basically just took like thirty seconds of my day, then I said, hey, write me a skill that books me in for WeWork.
12:35So when I wanna access it, I just go book me into WeWork. Then this goes and books me automatically for, the next thirty days ahead using some credentials that I've already set up, and then it also, like, actually literally opens up a tab. And so you could see this is doing it right here at the one that I currently frequent, Steven Avenue Place.
12:53So I'm just gonna say, hey, hit me, give this a a book, and then it'll go through and actually make, like, HTTP requests behind the scenes in order to push. If necessary, it'll actually also open up WeWork, so you could see it actually, like, do the freaking desk bookings, which is wild. And, you know, watching this thing go is pretty neat.
13:09And then after I built this and then, you know, did all of the WeWork bookings, I also naturally got more interested in, what could I automate with my browser? And I ended up creating another skill. So that was a pretty interesting one off build.
13:20Takes just like thirty seconds as you see there. You just have to log in to whatever service and then say, hey, I want you to automate this with Chrome Dev Tools MCP. I'll show you guys how to do that as well.
13:28But I was also kinda interested in what else I could automate that was kinda like a silly personal task that I usually have to do. And one thing that I do nowadays quite a bit is like I just buy stuff off Amazon and then I have it shipped directly to my door. And I do that because I think, know, I'm making a fair amount of money and I don't really wanna drive all the way over to Home Depot or whatever because it'll just consume a lot of time and energy and ultimately be unnecessary.
13:47Why not pay for the premium of having somebody else do it? The only issue is, as I'm sure you guys know, anybody here that shops on Amazon, there's like a bajillion bullshit listings nowadays with like crappy SEO and they try and use every term to rank for everything.
14:00So what it is, I just built a really simple Chrome DevTools MCP called Amazon Shopping. And what I do is I simply say, hey, find me the best floss on Amazon. It then loads up the Amazon skill right over here, And then it opens up Amazon, in my case, CA.
14:16And I'm just gonna make that a little bit smaller here so you guys can't see my postal code. And then it automates the process of like going through Amazon and then looking for whatever the heck I want. So in my case, I'm looking for dental floss.
14:27Right? It's just gonna open up one little thing here and then it's going to go through and it's going to compare all of the dental floss with each other to find basically the best combination of cost and then effectiveness that does whatever I want. And at the end of it here, you could see we have all of the products that it's recommending, um, literally in just like a little sheet, so I can give this a quick click.
14:44We also have the cost, the popularity, the why buy, and so on and so forth. And then, you know, he basically uses all these things in order to recommend them.
14:53So I basically built like a personal shopper app. I've done the same thing with like grocery delivery, which is pretty cool. So I just one shot my grocery delivery.
15:00I go on, you know, Instacart or another related service, click a few buttons and then boom, you know, I'm basically done for the rest of the year. And, you know, you could build these sorts of one off systems at this point now with Opus 4.6, you know, GPT 5.2, 5.3,
15:14or Gemini 3.1, whatever model you wanna use, whether or not you're using Cloud Code or or some other platform. Okay? Now that you guys have seen at least a little bit about what I think, like, actual helpful useful skills that realistically move the needle, not just glorified,
15:26you know, personal assistant demos look like. Let me show you guys how to actually go ahead and then build them. Okay.
15:31So first, if it's not already abundantly clear, skills are basically the evolution of standard operating procedures just for agents. Whereas SOPs
15:42and checklists are for human beings for the most part, skills are for these new AI agent intelligences. And just as such, you wouldn't really just give a checklist over to an agent and say do it. You'd have to translate its language just a little bit into sort of like the native format that the agent understands.
16:01So a brief example here, if I have maybe a checklist and it's make p b
16:08and j sandwich and you know, I work at a company where all we do all day is make p b and j sandwiches. Boy, would I love to live there.
16:16You can imagine how new hires go their very first day and then they're given a checklist of tasks. And it's basically, take bread
16:26out of basket. Put jam on one side.
16:32Put peanut butter on the other side and then maybe one final one that just says stick them together.
16:39So this might be for a human and essentially as a human goes through it, what do we do? We obviously check all of these tasks off. That's just how standard operating procedures are in business.
16:47They just for the most part standardize things, hence the name. Because this is organized a little bit differently for agents than you know, for human beings, we organize them into what are called markdown or dot m d files.
16:58And these markdown files are basically just text files with a little bit of a twist. For instance, I have a markdown or dot m d file open over here. If I try and create a new file here and then I call it skill spec dot t x t and then I paste everything in, you notice how everything is the same color?
17:14Whereas if I go back to that markdown variant, we all of a sudden have blue, we have orange, we even have some green. Well, the way that markdown files work is they take base data and then they also take a little bit of formatting and then they apply the formatting to the base data, just so that it's a little bit easier to see, a little bit more interpretable.
17:32And so if you guys could tell, the data that we're marking up here, k, funnily because it's called markdown. But the data that we're changing the the color of and stuff like that is usually prepended with a couple of symbols. The first symbol here is this little hashtag.
17:47This always refers to like a header. And so, can kind of think of it like a chapter heading in a book. Another one are these sort of back ticks here, which typically refer to files.
17:56There are a couple of other little tiny formatting things here like links and so on and so forth that you don't really get with TXT files. But hopefully, you guys understand that it's not really that big of a deal to memorize everything in a markdown file. You just have to know that they're slightly different formatting compared to text files.
18:13The reason why is because you're never really gonna have to make them yourself. All you're gonna have to do is just give your agent some bullet points, usually a brief description, which I'll show you how to do. And then it'll combine all of that into the optimal
18:25markdown skill format for you. Now, just a few months ago, today, every major provider on Earth basically, so our OpenAI's,
18:34our Googles, our I wrote Claude here, but I really meant Anthropic. They're all moving towards harnessing and then making skills generally available,
18:43just as part of like their their their LLMs. And so here, you could see Anthropic has a page on their Claude code docs that say, extend Claude with skills. These are the guys that sort of made skills out first.
18:55But because of popularity, a bunch of people that were open at developers and stuff like that started talking about it and asking for it.
19:02And so they've created their own skills format. It's very very close to the Claude code skills format. And then also, over here, the Gemini CLI, you could see you can add agent skills.
19:13Uh, basically, no matter what framework or LLM or platform you're using, you can do the exact same thing. So my rule of thumb is, if you guys have a standard operating procedure, if you have an SOP, you have a skill.
19:26Okay. All you really have to do is just copy that SOP from your business and feed it into Claude or Gemini or Codex and ask it to make a skill.
19:36As of the time of this recording, the skill spec or definition is, uh, it's not like baked directly into the models.
19:43So the one thing that you have to do in addition to that is you basically have to give it the skill spec and then ask it to turn your task list into a skill. But assuming you can, you know, drag a file into your freaking working directory and assuming you have any SOPs at all, even poorly written ones,
20:00you now basically have an agent capable of doing a big chunk of your knowledge work. Once you're done with this, you just ask the agent to run it at least once because you have to test it end to end. On the first run, if the agent needs something else, it'll ask you for it.
20:14Otherwise, it'll build the asset yourself itself. So maybe like an Excel file or something like that. I'll actually put that together for you.
20:21Although, obviously, you have assets like some context or PDFs or whatever, can provide them. No problem. And then assuming that you've prompted it right with the spec that I'm going to provide you, every time the skill runs after that, if the agent finds a mistake while doing it or maybe a service is out or it doesn't have the knowledge it needs, it'll automatically figure out how to solve it and then it'll patch the skill for you.
20:41And so in that way, here's the real power of skills, they're self annealing over time. They heal themselves, they get better, and they improve constantly. Much like an ambitious intelligent staff member who sees a checklist, notices that there's a gap and then chooses to fill it.
20:56Skills are the same thing, just with agents, which is what makes them so incredible. So why don't we get started and produce ourselves some skills? And then as we produce and then shortly after,
21:06I'll also run through the special directory of dot claud slash skills, which with claud is how you organize these things. I'll also touch on just a couple of other ones in case you're using a different LLM.
21:18But for the most part, they're gonna be exactly the same. So the very first thing you need to do is you need to grab the skills specked on m d. And you can grab that down below from the Google Drive link.
21:27It's free. As mentioned, you don't need to sign up or give me your email. I don't need it.
21:31I have enough of those. And what you're gonna see in here is basically the condensed spec or definition of skills,
21:38as well as the file structure, all the locations, all the different formatting options and stuff like that.
21:45And what we're gonna do is we're gonna give this to our agent alongside our request, so that it can just figure out how to put together the highest quality skills for us. What I did in order to put this together, by the way, is I just went over to Anthropix Claude Co.
21:58Docs with skills page. I copied the entire thing, fed it back into the LLM and then just said, hey, I want you to really, really shorten this, compress this, significantly increase the information density I don't want this to be 500 lines. I want this to be, you know, a 150 to 200.
22:13Okay. So once we have this, basically what we do is we just make this our claud dot md. You know, just so we're on the same page here, claud dot m d is just the prompt that's injected at the top of every new conversation.
22:24So And what I'm gonna do is I'm just gonna rename this and just call this old claud dot m d. And this one here is gonna be the new claud dot m d or just claud dot m d. Any file in your root directory that contains the capital c a c l a u d e dot lowercase m d will automatically be prepended and and added to the prompt.
22:43And so what we do is basically by by doing this, every workspace where it is just automatically going to know what skills are. If you have other stuff in your Cloud MD, you can also just copy this and then paste this, um, as well.
22:54Although, you know, at least for the purposes of creating skills, I recommend just being as short and and straightforward and basically having this in as few lines as humanly possible. Next up, to build your skill, just open up a new agent window.
23:05So I've just done that here with Claude code and then just tell it what you want it to do. I'm gonna use a voice transcription tool, WhisperFlow, just to put in my own SOP. What I wanna do, at least for this demo, is just build a simple inbox cleanup tool that just goes through my inbox.
23:19It then opens up all of the, um, unread messages and then just, uh, immediately reads all the ones that are pointless and sort of, I don't know, invoice reminders or follow ups about whatever or simple templates. And then I wanted to highlight, um, like realistically the high priority emails in my inbox.
23:37I just have a short list of everything. So it's pretty simple and straightforward. I'm just gonna hold this little key down and then say, I'd like to build a Claude code skill
23:46that goes through my inbox, identifies all unread messages,
23:52reads all of the unread messages, and then marks as read all of the ones that aren't important to me or high priority.
24:01By definition for important is basically anything that isn't automatically generated.
24:07If the email is automatically generated by like a service or it's a simple notification email or whatever, it's not important. If the email is written by like some cold email copywriting fellow who's trying to reach out and get my business, it's not important.
24:22The only stuff that is important are emails that are personalized, customized
24:26and they give off the vibe that they were written basically just for me. Once we're done, we're going to feed that right in, then we're going to go ahead and build it. The very first thing it's going to do is look at all of the existing skills and any gmail code inside of my workspace
24:41to see if there's anything that it could use. And it's doing this because, know, as human beings go, we don't really want to rebuild the wheel if we don't have to. Right?
24:49So this is it just reading everything right now. You can see that it's also changed to plan mode, that's why that little thing in the bottom left hand corner is changed alongside that red bypass permissions color to blue. This happens more or less anytime you try and build something that it would consider to be even slightly complicated.
25:06It wants to go into plan mode, so that it can plan out the build of the skill. Now, it correctly went through and found that I had a couple of configuration files that allowed it to use different emails here. And for the purposes of this demo, I'm just gonna use this one.
25:19Obviously, if you didn't have this, it would just ask you what email do you wanna set up. Now what it's going to do is actually create the skilled on m d. It's also gonna create a script for me, which looks like it's going to call inbox underscore cleaner.
25:32Then afterwards, it's gonna test an end to end. Why? Because that's just what's in our spec sheet.
25:36We basically tell it how to do all of this stuff natively. Now what it's going to do is create the script. You could see that because I have a pre existing skill, that follow-up one that I was showing you guys, it gets to borrow based off of that.
25:49And you could see that it's actually gone ahead and wrote the script. If I go to inbox underscore cleaner here and actually look at the skilled on MD, you can see it right over here.
25:58If we just make this a little bit bigger, let's move this here so we could focus on this. Basically, we have is a step by step guide
26:08for a future agent on how to do the task that I just asked it to. And you'll see that it's now been organized in a very particular way. For instance, there are these three dashes up at the top, three dashes underneath the section.
26:20There's a name key that says inbox cleaner, a description key that says clean up Gmail inbox by reading all unread emails using and identify which ones are generally important and marking the rest as read. Use on cleaning inbox, triage an email, or clearing unread notifications, and it also allows it to use some tools.
26:36Now you may be wondering, you know, I think these two are pretty self explanatory, but why is it allowing certain tools? Well, Claude and all other LLMs basically have access to some little pieces of software under the hood, like they can read files, they can grep, glob or bash.
26:51And these are just various like command line utilities that allow it to look through files, look for specific sections, and so on and so forth. And so we've done with the skill, k, it's almost like a sub agent. We've basically given it the ability to do these sorts of things when we invoke or instantiate this.
27:06Aside from that, we have the title, we have the goal, what counts as important, what gets marked as red. Notice how it basically took everything that I have provided it and just a few lines of voice transcription
27:17and then turn this into a pretty coherent and then consistent SOP. Now going back to the actual building thread here, it's actually gone ahead, wrote the script, and now it's doing the classification.
27:28We give this button a quick little click, you could see that it's now doing it in batches of 10, which I didn't even ask it to. K. It's identified which ones they consider as important, which ones are not important.
27:39And now it's saying, hey, do you want me to mark these 97 emails as read now? I don't actually want it to mark the 97 as read. What I wanted to do is I wanted to show me what they are so that I could tell it which ones are right.
27:49So I'll say, no, I'd actually like you to show me all of the 97 that you marked as red, so that I can take a look and tell you if there are any exceptions. Now what it's going to do is give me a big list of all of the emails.
28:01And as you could see here, it is correctly identified that 79 of the emails I got were from a broken make scenario. So it's pretty interesting.
28:10Um, obviously, these are more or less exactly the sorts of emails I don't really want to. One is some normal commission notification from referral stack. Another is a subscription renewal for that voice transcription flow,
28:22another is from Bright Data, another is from whatever company. I have a bunch of cold outreach over here. K.
28:28I have sponsorship pitches, one word yes reply to lead magnet. I guess what I'm trying to say is this did a really good job.
28:35There isn't a single email here that I actually really care about. So I'm just going to say, nope, mark all as read. That's pretty sweet.
28:42Right? Now at any point in time, I could open up a new Claude instance, and we have to go new just because of how skills work. K?
28:50Then we can go slash inbox cleaner, and then we can run it. And because we are running based off of this SOP, it's always going to run-in the exact same way.
28:59It's gonna be repeatable, consistent, and dependable. Also, if it runs into any issue, let's say the script doesn't work, let's say there's an API rate limit or something like that, it will identify the problem and then automatically go through fix it and then rewrite its own skill.
29:14In my case, I've got a lot of unread emails. So I think what I'll do is I'll say, hey, I've got a ton of unreads. Could you actually go through and fetch a thousand instead of a 100?
29:23And now I have the ability also to modify the way that the skill works just a little bit. So that instead of the base 100, maybe I can do a thousand. Maybe I could pass in a flag like dash
29:33a and that stands for amount and then I can do you can see that's actually what's going on under the hood. It's passing the the a number 1,000 to said script. Now that you guys see how these skills actually work in practice and how the easy it is to build one as long as you have the right spec sheet, I just wanna really quickly deconstruct the various parts of this file to make sure that you guys understand
29:53how all of this stuff works and what each of these parts are for. Okay. So you guys have probably already noticed these three little dots here are really different from all the formatting elsewhere.
30:04What this basically is in at least markdown speak is something called front matter. Front matter is a highly optimized way of basically putting a tiny summary at the beginning of any Markdown file. Remember how there's different fields and colors like keys,
30:23headings, there's sort of more headings. I mean, there's a lot of different things here.
30:28We can just think of front matter as being the same idea. It's just a different like variable or type. And the way that you start it is with these three dashes and then you end it with these three dashes.
30:38And so this little summary is basically doing you a massive service by significantly reducing the number of tokens that you need to load into your contacts window at any point in time. If you think about it, what we could be doing every time we load the skill in, if I just copy this whole thing and go to a free word counting service,
30:56is we could be adding 373 words, which is approximately 500 tokens into context. Okay.
31:02But with front matter, what we do is instead of storing the entire script into context and making it accessible to the model, what we do is we store a much smaller summary, is closer to like 60 or 70 tokens. And then we say, hey, if the user calls inbox cleaner specifically, then and only then do I want you to load the other, uh, you know, 350
31:21words or so. K? And so what this means is in practice,
31:25if I were to show you guys all of the different skills that are currently loaded into context, they basically all look like this. And so my prompt at the very beginning, just assuming I have these three skills,
31:37is always basically going to be my claud dot m d up here. Underneath, we're gonna have our skills. So there's gonna be one,
31:45two, three. And then underneath, it'll be your prompt, which is basically whatever the heck you wanna say. Hey, I want to do x y z.
31:54And so this is more or less the structure that every conversation with Claude in my case, but obviously any LLM in yours, whether it's gemini.md, agents.md, so on and so forth, is going to realistically look like.
32:06You'll have your prompt down here, but before then you'll have a bunch of hidden stuff. And then what's cool is, you know, if you say, hey, I want to do follow ups.
32:14What it's going to do is it's gonna say follow ups. Wait a second. I know that we have some follow ups mentioned earlier on under the skill section.
32:21Okay. Great. That means this must be a standardized skill.
32:24And then it'll go and actually load the whole follow-up nurture directly into its context, which will be the rest of this. Then it'll basically stick the rest right here and then resubmit the prompt and now it knows everything that it needs to do in order to obviously get that right.
32:37So the agent would go, okay, great. I have the skill spec right here. Let me follow each step and then it'll go on, you know, start with number one, which is load pending leads.
32:48Number two, for each pending lead and so on and so on and so forth. Now, this actually has a term. This is called skill matching.
32:55Anytime a user types a request like, hey, get me 50 dental leads, it'll automatically scan those descriptions or that front matter for whatever the most relevant description is. It says find business leads here.
33:08It says find business leads here. It'll say, oh, this must be the one that they want. Then it'll load the full skill .md in the context before finally calling some scripts.
33:17Okay. And then giving you guys an output. And where you put those scripts don't really matter.
33:22The basic skill definition, think just bundles it all into one folder. You could put the script somewhere else if you wanna keep them organized. And if you have other assets like PDFs or whatever, you can build whatever scheme you want to keep those files handy.
33:32The reason we do this of course is both for efficiency because the less context in a model's context we know at any point in time, the higher the quality of the output. But also for costs, Providers like Anthropic and so on and so forth actually don't want you to spend an arm and a leg, believe it or not. Um, they want you to continue using their software and in order to do so, they need to make it like reasonably cost efficient for you and deliver a high ROI to you, um, at the same time.
33:57And so this concept of progressive disclosure is now not only really being applied to skills, but a lot of other things as well. And all this is again in the pursuit of reduced cost, reduced overhead.
34:07Okay. Let's build three more skills with what we know now. And just for simplicity and to make this as fast as humanly possible, both for myself and for you guys.
34:15What I'm gonna do, is I'm gonna open up three separate Claude code little plug in widgets here. And in the first, I'm gonna switch all these to bypass permissions. And make this a little bit smaller so you guys could see.
34:29And then in the first, I'm gonna paste in skill number one. In the second, I'll paste in this skill number two. And the third, I'll paste in this skill number three.
34:38What I'm gonna say is create skill one, create skill two, and create skill three. And we're just gonna trigger all three of these and see how they operate. So what skill one is, is it's a simple way to turn meeting notes to action items.
34:51We basically just paste in a meeting transcript and get a structured action item with owners and deadlines and so on and so forth. The idea here is to show you guys that a skill can be really nothing more than a well written SOP, which doesn't even include Python scripts or or anything like that. The second one is an invoice data extractor,
35:08which is basically gonna be skilled at m d plus some Python script. So we're gonna take a PDF invoice, then return structured JSON in the form of a vendor, amount, date, line items, and tax. And this is to help you guys see how skills can delegate deterministic work to code, aka we could use Python scripts and and whatnot like you guys have already seen.
35:26And then the third skill is gonna be a content repurposer, which is gonna let us take a transcript and then output a tweet thread, a LinkedIn post, and then a newsletter draft all in parallel. And the idea here is we're gonna have templates.
35:37I'll show you guys some brief examples and I'll also show you guys how this works with some instructions. And so before I was even finished reading this freaking thing, we're almost already done two out of the three. Isn't that wild?
35:48Anyway, the first skill, pretty straightforward, just called meeting notes. It just processes meeting notes. So what we can do here if we want is we could say, hey, I want you to output specific
35:57elements of the meeting. In this case, it has some decisions, open questions and so on and so forth. But I also want you to find and identify any contact information.
36:06So I'll say, hey, this is great. I want you to update the skill also to extract any contact information that's mentioned. I'm looking for things like email addresses, websites.
36:16I'm looking for phone numbers and so on and so forth. Anything that somebody offhand mentions during the call. Sometimes we also specifically ask for this information for follow-up reasons, so it would be nice to have it in a structured sort of way.
36:28Over here, it's now created both files of the invoice data extractor. So we could see there's a skill dot m d here, which basically takes the PDF invoice return structure JSON with vendor. So in order for me to really appreciate how this works, I'm actually gonna go and grab myself a brief little invoice.
36:42I'll just feed in in an example, and then we'll see how it works. I see a couple of services here that have sample invoices. This one looks pretty good.
36:49Why don't I download this? I'll say example underscore invoice. What I wanna do is I just wanna feed this in right over here.
36:56And I'll say, great. Use this on example underscore invoice dot pdf. Then over here, it looks like this is now done.
37:03And what I'm gonna do is I'm gonna go find some of my own content and just feed that in, just as an example. So we'll go on my YouTube channel and I'll find something that is probably not four hours because that would take a lot of tokens for no purpose. This ten minute one looks pretty good.
37:16Thanks. Let's get my fat ugly head to shut the heck up. I'll scroll down here to where it says show transcript.
37:23And then right now, just because it's an example, I'm just gonna copy all the stuff in manually. But obviously, you guys could do whatever you want with this. Going back here, I'll then paste this in.
37:31And what I wanna do is I basically just wanna see whether or not I can do all of those three things. I'm also gonna take a quick peek here at the script. And this is what I wanted to happen actually, which is nice.
37:42It ran into an issue with one of the summary rows and the subtotal. So what it's doing now is it's actually going to fix the script's filtering logic so that it doesn't catch the subtotal or the credit terms or something like that. K.
37:53Cool. On the right hand side, it's extracted this transcript. Looks pretty good.
37:57And now what it's going to do is spawn all three formats in parallel, which is its way for saying it's going to rewrite my main transcript and turn this into like blog posts and Twitter threads, whatever the heck it's gonna do. While all of this is happening, this is me pictured.
38:10I'm just manifesting all of the skills. I actually have to do too much. I just kind of close my eyes and wait and then they appear.
38:17Taking a look at this tweet thread over here, let's just make this look a little bit prettier. We now have a bunch of tweets that are divided. This is automatic, which is pretty cool.
38:26A 16 year old posted a prompt on Twitter, former professional web designer copy pasted in the Gemini. The AI produced a better site than he could in five hours in minutes. Design is commoditized.
38:35Sure you guys have seen these sorts of threads all over the Internet. Let me tell you why. It's because people have built very simple repurposing pipelines like this.
38:43And although that one just took me five seconds, the content's reasonable. I'm not gonna say it's gonna win any awards. It's pretty LLME,
38:48but hopefully you guys could see with at least a few neurons between your two ears, you guys could repurpose, change the tone of voice and so on and so forth and make it something actually pretty cool. Likewise, we also have LinkedIn posts and then we even have a newsletter draft for me. So that's pretty sweet.
39:04You could see that its design decision was to do this via Sonnet sub agents. You guys can modify this SOP at any point with literally just a line. Finally, in the middle here, skill two is all good to go.
39:14Great job. It's gone through and it's actually updated the main skill as well, which is quite neat. If you guys were paying attention, you would have noted that every time that we add a new skill or change skills or whatever, um, there's a top left folder over here that's dot claud slash skills
39:30that changes. And this is essentially the format and file directory structure of a skill. So it's worth us pointing it out.
39:37Now, in case you guys have never seen this convention, in the top left most corner, you see how there's a little dot in front of a Claude folder? Well, that's basically just the kinda universal file explorer convention for hidden.
39:48Now, if I open up this exact folder, the skills example folder in my macOS native file explorer, you notice that it looks pretty different from what we're seeing over here.
39:59Notice how on the left hand side, have dot claud dot t m p, then we have one second copy config data. And notice over here that the list actually starts at one second copy, then goes claud dot m d, then it goes config, and then it goes data. So I mean, aside from the fact that this is organized a little bit differently, generally speaking, we always have folders organized first in the file explorer here.
40:16So if we kind of match that up one for one, you notice that the only things that are missing are basically the files and the folders with little periods in front of them. So the first thing to note is, like most people just have no idea where this thing is because when they go to the file explorer, like it's always hidden.
40:29Well, in macOS, it's pretty easy to reveal. You just go shift command dot period basically, and then you can show all hidden folders.
40:36They'll be present. It's just they'll be a little bit more translucent.
40:41And then we can actually open it up and then go through all the skills. Okay. So that's the very first thing to know.
40:45On Windows, which you guys may be using, it's it'll be a different hockey. Just Google it. There's plenty of information on there about hidden files.
40:51Think you can also just right click and go like show hidden files or something because I used to have a PC and that was pretty straightforward. But, uh, yeah, essentially, inside of every main workspace in any sort of IDE, if you do have skills, they're probably gonna be hidden from you, which is kind of difficult and intimidating to to kind of understand.
41:07Okay. So then after that, the way that it's organized, which is not super plainly clear here. So to make it a little bit clear for you guys, is you have the dot clod folder.
41:14Right? And then underneath the dot clod folder, you have the skills folder. Then underneath the skills folder, you actually have the specific skill folder that you want.
41:24So let's just go, you know, inbox cleaner. That's another folder, and then inside of that folder, you actually have the skill file itself. So then you have the skill dot m d.
41:36So notice how nested this is. I mean, it's a massive pain in the butt. I don't know why they do it like this, they make it pretty difficult for most beginners to understand.
41:42And they do so in this sort of like hierarchy that begins again with dot claud, then goes skills, then goes in, you know, the specific name of the skill and then the skill MD itself. And then over here, okay, and this is sort of optional, but this is how Anthropic typically does it. Here, you can put your other stuff.
41:57What is your other stuff? Well, your other stuff is gonna be things like your scripts. So hypothetically, if you had, like, scripts that helped you do the invoice cleaner, inbox cleaner, or whatever, maybe an inbox, uh, -cleaner.
42:08Py, You know, you'd probably put it somewhere in here. I personally like doing it in a different folder called execution.
42:14Sue me. I just think it makes more sense, a little bit cleaner. Uh, likewise, if you have, I don't know, like your Gmail auth tokens or something,
42:21you know, you might include them over here. If you have, I don't know, some other file, like example dot TXT, um, you'd include it over here.
42:29And so this is just like the base way that all skills are organized. And, yeah, I think it's just unfortunate that it's usually in a hidden folder, which is kind of intimidating. I should note that the skills live right alongside a bunch of other hidden folders like agents and stuff like that.
42:39I'm not gonna talk about those now, but if you want more context on that, just check out my full length Cloud Code course. I got a ton of juice for you in there. Okay.
42:45CTALast thing I really wanna do is just talk a little bit about what skills are actually worth making. You know, I do a lot of business consulting now. Somebody that started in door to door sales, I saw the importance of focusing on the levers that actually increased revenue first and foremost above all else.
43:03CTAAnd so for me, it was always door to door. But then transitioning into marketing and then broader through things like, uh, content creation, articles, and now videos. I think a lot of people are very misaligned on what like a high revenue
43:17CTASOP even looks like to begin with. And the reality is, with agents nowadays, you can do anything. But that doesn't mean that you should do everything.
43:26CTAThese things are so damn good. They can of course design whatever crazy back end fulfillment pipeline you want. But I find most of the demos on the internet that talk about stuff like that really miss the mark because we've been able to design these pipelines forever
43:42CTAand most people that do these sorts of things still don't make any money. The difference between people that use new upcoming and trend breaking technology to make money and actually have big, you know, success outcomes
43:55CTAversus people that play shiny object syndrome all day and just collect a bunch of Pokemon cards Our the former group focuses primarily on the front end of their business, so they focus on the acquisition of opportunities.
44:07CTAThey focus on the generation of leads. They focus on mechanisms that improve the quality of the sales experience for their leads.
44:15CTAThey focus on selling, closing mechanisms, and so on and so forth. And then the latter people just over complexify the hell out of everything and then make every excuse under the sun not to focus on sales and marketing. So this is all just like a pretty meandering and long way of saying, if you're gonna design skills, I highly recommend you focus the design of those skills primarily on front end tasks.
44:36CTALike most of the people that are gonna be watching this video, you know, they're probably gonna have some sort of side hustle going on or they're going to be looking to get into that realm. The simplest and easiest way that you can move the needle in your own business is if you're gonna design skills and stuff like that, spend your time designing them on things that help get you sales, on things that help, uh, you know, you do your marketing faster and better.
44:56CTAAnd then don't just have them sit around in your folder, actually use them. You know, when I was growing my business, I was doing somewhere between 50 to a 100 cold calls every single day. When I was going door to door, it was something like 80 knocks on physical doors per day.
45:09CTAI wish I had a hundredth of the technology we have today, and I'm not even that old, to go back and then automate that process for me. I don't.
45:17CTABut it blows my mind how many people do and then they choose not to use it on things that actually make money like those laborious door to door sessions or or cold calls or whatever. Okay.
45:28CTASo ultimately, things that I have seen work really, really well. Content repurposing pipelines. These genuinely can significantly improve your reach assuming you're growing a brand.
45:37CTAThat's for inbound. Lead generation systems. Things that do stuff like scrape leads for you.
45:42CTAThose are really big. Skills that use Chrome Dev Tools MCP to post on forums automatically for you, uh, to, you know, do product research quickly and then tell you which threads are saying bad things about your product. These things are valuable.
45:55CTASkills that, you know, quickly go through a bunch of books and identify which things are business expenses and can be written off versus things that aren't. These sorts of skills are valuable. The skills that are not super valuable are like skills that help you build other skills or skills that
46:11CTAinvent some new design framework that assists you in the building of AGI. Right? These things are not ultimately the stuff that I want you to take away from this video.
46:21CTAI want you to focus on real practical use cases that actually move the needle for your own business or the businesses that you're working with. Because the end of the day, that's the stuff that's gonna make the economy more efficient. That's the stuff that's gonna make you more money as, you know, AI swallows up more and more knowledge work.
46:40CTAIf you guys like my content, please subscribe. I realized the other day that, uh, 69 something percent of all of the people that regularly watch my content are not subscribed.
46:48CTAThat means about two out of the three people that I'm talking to right now for whatever reason haven't. If you like the sort of content that I make, you know, the algorithm really does like it if you subscribe, and then it pushes my YouTube channel to a bunch of new audiences.
47:01CTABut if I don't get enough subscribers, then none of that stuff happens, and then most of my videos bomb. So I really appreciate anybody that has taken the time and energy to subscribe, and I'm not gonna do a big cry session here. Although as a newbie YouTuber, I always wondered why everybody was pitching me on subscribe this, subscribe that.
47:16CTAAnd now I think I get it. Most people just never do. But, yeah, I I would really appreciate it.
47:21CTAAside from that, if you guys like this sort of thing, just leave a comment down below with any sort of question or recommendation as to future content, and I will happily take that into consideration. I have skills that now scan my YouTube comments and then use that to help me ideate new videos to make and so on and so forth.
47:36CTASo, yeah, a lot of cool stuff in the works for you guys, but I definitely love to know what direction to take it. Thank you very much for watching. Have a lovely rest of the day.
— full transcript
§ 05 · For Joe
The only skills worth building touch the front end of your business
WHAT TO LEARN
Skills are just SOPs in a format the agent understands — and all the ROI concentrates in lead generation, outreach, and follow-up, not in automating the work that already exists downstream.
- Any existing checklist converts to a skill in minutes: provide the skill spec, describe what you want, and the agent writes the markdown file and any supporting scripts automatically.
- Skills self-patch on errors — the agent identifies what broke, fixes the script, and rewrites the skill file without human intervention, compounding quality with every run.
- Front matter limits token overhead to roughly 60 tokens per skill; the full spec only loads when that skill is explicitly invoked, keeping context clean across large libraries.
- Skill matching is triggered by natural language: casual phrasing like 'follow ups' is enough for Claude to locate, load, and execute the correct skill from front matter descriptions.
- Reading full email chain history before generating follow-ups — instead of using cold templates — produces replies that prospects cannot distinguish from human-written messages because tone and thread context are matched.
- Going directly to LinkedIn Sales Navigator for B2B leads outperforms Apollo and Apify because those services re-scrape it anyway; the skill eliminates the intermediary and writes results to a structured sheet.
- Running multiple Claude Code instances in parallel to build several skills simultaneously is a production-viable workflow, not just a demo shortcut.
- The skills that fail to generate ROI are focused on internal complexity — back-end fulfillment, meta-tooling, or building frameworks — rather than on acquisition and outreach where the revenue actually lives.
