Claude on Steroids: Skills + Sub‑Agents + MCPs for 7‑ and 8‑Figure Amazon

Claude on Steroids: Skills + Sub‑Agents + MCPs for 7‑ and 8‑Figure Amazon Speaker 1: Hey guys, welcome back to another Seller Sessions. Doing a special today. We're going to look at Claude skills. I'm going to do some demos. What I'm seeing at the moment, obviously we understand that AI is moving fast. We have to keep our feet on the ground, of course. But what you're about to see is I've been doing things that would take me a day, two days, three days, and I'll show you the difference between the two. So down to 10 minutes. And I think that the skills have been slept on. So I'm going to cut back on the talking. No more need to intro. I'm going to get straight into it and share my screen. And then we'll start walking you through all these different factors just to give you a bit clarity because I think. Now, with what I've been talking about with MCPs, and I know it seems to be above people, and where we're going with what I like to call snack apps, right? They're single feature apps. We're not talking about full-blown production with grown-up engineers, security, authentication, and safety of the code and running a proper business. And we're not talking vibe code in here. But these will give you an opportunity to build stuff rapidly, fast, and present them in a good way. And reduce your learning curve in reaching to that point. So without further ado, I'm going to get in and share my screen. So first thing I'm going to do, I'm just going to run this in the background because it's going to take 10 minutes. And if it doesn't work, because the course is live, doesn't always work. I've got one that did work before and we do a blue Peter, you know, something that something that we can show from earlier. So can you analyze this Amazon PDP to give it a URL? And then build me a report. Use the skills for building the dashboard with the analysis and the MCP optimization for token efficiency. Now, if you understand anything about this is one of the most annoying things about Claude Desktop is it burns through the conversation fast. MCPs are good, a lot of them do work, but certain ones are very hungry on tokens, it maxes out the conversation. Now, if you're building, as I said, snack apps, not like real applications, then you're still going to burn through two or three conversations, then you have to bring the context over. Now, this has been greatly reduced because you're able to create. What are called skills, which I'll go into in a minute. So let me just set this off. As I said, it's going to take a little while. I'll let that run in the background. We'll keep coming back to it. So here's an example of a search term report. So what I've done is an unbranded one instead of using my curve design system. I've generated another one so that you guys can have a copy of this. But basically what this is, this is a search term Performance report and all we said was build a dashboard and analyze this data and it built that. Now, can you change this? Absolutely. But the magnitude of this is that let's say you haven't built anything. You know what Claude's like. It will give you pretty much these kind of grayish, purplish, washed out color. Think something like I mentioned the other day in a post. It reminds me of kind of a washed out Purple Rain by Prince original press cover from 1982 and then you get the understanding. So this is just a little bit more structured but also If you know anything about building out any of these apps, the fact is you can see it's pretty much pixel perfect. And that's because of the way that skills work, which we'll get into. Here's another one for keyword ranking. Just gave it a file, a .csv file. Again, can you analyze it? Now, this is what we're going to be seeing shortly once Claude's done its work. So, to give you the magnitude of this, all I did was, because I've set up the skills, I said, could you analyze this PDP? And build me a dashboard using my design system. It hadn't done it properly because I could have give it a better command, but hey, it looks all right. Yeah. And when we look at this, you've got all of these tabs. Now, if you look, it's not washed out. It's pretty good, right? It's just a report. Like these are what you can just quickly generate and look half decent in your own brand. Now, this is one that I built that I spent 40 hours on. I did it on the show last week. This is more advanced because I've got my Rufus intelligence agents that are built into it and there's some competitor comparison, right? And there's some stuff that we've got coming with this which will be working with A-plus content analysis and some of the image models that extracts information to give you a bit more data. But the idea of this is to show what's leaking on the PDP. Now what I've done is I've kind of done a snack-like version, if you like, with this one. The difference is this is 6 to 10 minutes work. This took me 40 hours because I'm going back and forth and there's a lot of stuff going on now and I'll cover because of the competitive landscape on your PDP. I'm using 600 plus data points and capturing all the competition on the page and in breaking down who is the highest threats and stuff so you can put together defense campaigns and stuff like that. So you've got Claude desktop skill system. Now if we look at what skills are. Now skills A lot of people are used to using Claude projects and in the Claude projects, let's call it niche based. You might do a load of research and it's based around a certain thing. It could be around ranking. It could be around NA10 and things like that. You're picking a niche basically. You've got a project base that is focused and then it has all of its instructions, behaviors and everything else. Now skills are different. These are skills a bit more closer to a Claude MD. So if you ever use Claude code, you have like your working document with all your behaviors and you know, It depends on how you like to work. There might be certain scripts and stuff in there that are generated. There's a way that you communicate formatting from programming languages if you're a programmer. But if you're a marketer or a business owner, what Claude code is and what people don't realize is basically a computer Automation tool but it's called Claude code because people think you've got to code which isn't the case. I don't code at all. So let's have a look what we got here. Skills, skills of behavior modification and this teaches Claude desktop how to respond. You also what's interesting about this is that Claude will also automatically select this. Versus, let's say I'm in Claude code and I run an agent, I can say, can you run agent 1 or I use a slash command with the agents because I've got them in the markdown file in a way that it's structured so I can access them quickly. Or I can use the microphone and ask, can you do agent hyphen 1? Depends what it is. It could be agent 22, et cetera, et cetera. So let's look at the skill ecosystem. Some of you might find problems with them. I know I did. In terms of the formatting, it can be a bit tricky because it can say that the files have got errors, but they must be contained in a zip file and then you drop them into the skills, you go into your settings and then you'll find the skills from there. So let's look at the ecosystem. So a skill file is what's called a YAML file for pronunciation with markdown instructions. So here we've got Claude desktop settings and then you install your skills and this will auto-activate and detect patterns and starts to apply rules by the commands you give or the prompts that you're giving. And then in this case what we're doing is we're executing on token efficiency. So if we look at component, purpose, benefit, it's YAML formatting, the skill, in this case, metadata, name, description, the benefit, Claude desktop recognizes and loads at skills. The component, markdown instructions. So you've got your behaviors with rules and patterns and then you've got persistent access across all the conversations. So you've got crossed context across the conversation. Then you've got trigger conditions, when to activate automatically, and then you've got no manual activation needed. It says that, but sometimes it might just mean you need to rephrase something. So for instance, I showed you earlier on, it didn't have the full curve design system. It looks similar to it, but it didn't have the header and the footer and the definitions that we have in terms of our style system. Response templates, predefined output formats and consistency and done at speed. Now what's the difference between skills and versus sub-agents? So to understand a little bit more about the architecture and I've only, I built my first agents about a month ago now. I think I got about 33, 34, but there was a period that I really had problems with the agents because I couldn't work out what do we do with the context. So I did put them into SuperBase and put them into a structured format. But when you've got so many different agents, you either have to go to the markdown file and look up the different agents, right? So you can see only it jumps there. It's listed, individualized these agents here. But then I've got a document. This one says it goes to 31. You've got here skills maker. So I've actually built an agent To build skills documentation, right? Which we'll get into a bit later. So the difference between the two, right? Skills, what are they? Behavioral instructions like we discussed, response styles and modifications, water detection patterns, and then design system enforcers. Sub-agents, what are these? Well, these are complex automation systems, multi-tool workflows, specialized processes, business logic engines. MCPs stand for Model Context Protocol. So, what are they? They're external services that basically are connections, for a better word of it, I'm just being very broad, it connects One AI to another AI. This was developed by Claude back in November last year. If anyone is old enough to remember, years and years ago in a recording studio, if you understand a bit about music, in order to link the drum machines and the computers and synthesizers together, they had something that was called MIDI, which stands for Musical Instrument Digital Interface. It was the only way to connect all these different Components together that weren't from the same company. So today if we look about our the way we've got pretty much universal connection and charges for mobile devices, the ecosystem at the moment is MCPs. Now MCPs like Many connections back in the day. There is a thing where they're a bit sloppy. There are other risks involved if you haven't got security put into place. But what you need to understand is that. Having the skills and the MCPs and having a good handle on safety and security, these are going to make a massive difference, right? And what's happened now, MCPs are now becoming much more useful because you can control the efficiency through skills. Now, if we come back down to here, so let's look at some of these examples, right? When to use, Skills, Repetitive Task, Brand Consistency, Token Optimization, Auto Generation Workflows Okay, the sub-agents. Multi-step operations, data transformation, system integration, and advanced automation. MCPs, when to use, external data needs, service integration, real-time updates, and cross-platform sync. So an example would be, is I've got the full taxonomy database of the whole catalog structure. And so I'm able to marry all those things and their components together because I have them in a structured database in SuperBase. And I can connect to SuperBase using the SuperBase MCP. So that gives you an example there. Now, if we look at some of the characteristics as we come down with skills, characteristics, always active in the background, auto detect and apply, persistent across sessions and zero, Invocation, overhead, then sub-agents, manually invoked for code base access, state management and complex decision trees and then on the characteristics of MCPs, real-time data, external services, persistent connections and a bi-directional sync. So let's look at something in terms of the problems we have to resolve with Claude. And the context window is the amount of tokens that get used. In here, the user requests a simple task. Claude response, you know, we're using, I don't know, horoscope numbers here. They're just here for the definition to give you an example. Claude response might take 800 tokens. And then the follow-up may more explanation that's another 600 and then you start to burn through your context window availability of the conversation. And so once you start to break all of these down, if these are running in the background and they are structured and efficient, The means that you're going to get an increase in capacity through efficiency. So if we were to look at this here and then we look at skills after maximum efficiency, user request, upload CSV, skill activates, auto detect, zero questions. Then execute, 150 tokens done, context saved, 8% reduction, so you can keep working. If you've, like me, I've had to go from a stage where One, even though I'm using Constraint prompts, I know I don't prompt, I built in a bot to do that. So Claude does them, I just record into the microphone and it converts them. But the thing I'm trying to say there is that I was always trying to manage that window, it became impossible and then I moved to Claude's code. Now I can move back to the desktop because it's a lot easier to work. Now we have a bit more welly to use and a bit more room to do more things than a handful of requests, especially When you're using a persistent memory, you've got a design system, you've got things that are set up that eat up a lot of your tokens at the very beginning of the conversation. So if we look at the traditional workflow, yeah? We've got convert CSV into interactive dashboard. We have to go through all of these stages, right? Now, if we're doing this where it's powered by a workflow that is a combination of different skills, you're using a lot less and you become more consistent. So, if we look at this as a complete stack, we've got skills, MCPs and sub-agents, okay? If we look at here, the layer components, layer, right? So, an example, automate content pipeline. So your layout, there's your input here. Component is the MCP for Norton API. Function, fetch, latest issues of the podcast of show notes. Then you've got here processing, analysis. I won't read all through it, it's just giving you a bit of an overview because we want to go back and see what's going on with the cooking on the other side with Claude. Production Skills Library, let's go through this, what we've got here. Yeah, these are just user cases. I'll release these, I'll probably just whack it out as an article so people can get a better idea if we come down to here. Now, again, when I say the words app, I'm using the apps words very, very lightly. So I call these single feature snacks. Okay, so they're actually do something useful for what you do, but you can build them pretty quickly. If you think what pretty quickly is two or three days is pretty quick and eight-hour days pretty quick. But now we've got to the stage where you can build almost the same thing within reason outside of writing additional logic and stuff because you've got the skills in 10 to 15 minutes. So building apps at light speed, right? So traditional app development with Claude or building artifacts because I know there'll be some people in the comments who will pick it up on the language because they're engineers and they're very clever. Concept, user explains 500 tokens. Question, Claude asks questions. Dash is taking 800 tokens and then you've got the design back and forth. You're going to explain all the logic you want as part of the structure and the output of the app and then you've still got to debug etc etc but then the likelihood is that you're going to have to take it over to a new conversation, burn up the front end of that conversation with the context to bring that over. This way It's still not perfect. What you can do, let's give you an idea. Are you going to run out of context window if you're sloppy? Absolutely. Are you going to build Facebook or YouTube or LinkedIn inside an artifact? Absolutely not. And will you blast through the conversation? Yes. If you're building simple things, what you'll find is with the uploader, layout, the logic, the components, the changes that you make, the refinements, the prompting over and over again, the testing of the app as well, working on logic, what will happen is that part will probably take three to six conversations depending on what you're doing. Now you can do all of it in a single conversation as long as you've put what you want in the skills. So if we come on down to here, what we've got here, understanding context windows. So Claude this desktop limits what we've got here. Yeah, it's just repeating. I won't bother going over and over on this bit. Managed MCP strategy. What we've got here Smart Contacts Preview 90% token reduction on large pages which is why we're able to run to the PDP. If I give this one here this gives a bit more of a visual style. What we've got in here is there anything you should know? Now it's just more of a visual version that I ran out earlier on. Building traditional apps, production, dashboard, auto generation. Right, so what you'll be able to see in a moment when we say to build a dashboard there are parts of the description of what is in that auto dashboard because the way that I've built it The goal was, let's say someone wants to use Claude. They want to take any kind of data. They're not sure what to do with it, but they want to make it look good. They want it to do some analysis of some kind, because then that encourages them that, okay, this is good, because normally people are able to adjust through seeing and learn through the seeing part. So if you ask someone to use Claude and give them a .csv file, they might not know what to ask. But if you've got a skills which says whatever the data is, do a number of these different things, use these components and adapt to the data to produce useful information and make it look pretty. What would then happen is that they can look at that and go, do you know what? I'd like to make changes here, here and here. And then that encourages more people in the community To build these and not see it as a way that is off limits. All of the stuff I'm showing you here, everyone I know as Amazon sellers are smart, right? Because you have to be smart to be an Amazon seller. And what I'm showing you here is all cutting and pasting. It's not coding, right? And if you can cut and paste, you can do what I'm doing here, right? So, let's go over to Claude. Let's see if he still doesn't want to do it. Right, so what I'm going to do, as Claude won't show me the love, we're going to go into the chat here and I'm going to show you another one that was done, right? Because I can't put out a thumbnail and say it's done in 10 minutes and then it didn't work. So, I installed the skill here, right? This skill is to optimize the MCPs. The problem with MCPs, they're hungry on tokens. So we run through some tests and stuff here, right? And then in this one here, how about go to amazon.com and do an analysis of this product detail page because that tends to really eat up a lot of tokens and then maxes out the conversation. So we did it on this here, which I'll show you in a minute. So then devises the strategy. And he goes through, he navigates with the URL, you can see here, okay, he ran Playwright, yeah. Anyone who listens to the podcast that does use Playwright knows exactly what I mean when they see this, how impressive it is to be able to do this in Claude desktop without it maxing on the conversation. So the evidence skill worked because this is just a quick test. So without this skill, the Amazon PDP snapshot would have returned a massive Response consumed a load of tokens. So what did we achieve here? The skill basically truncated down so the token usage was down and reduced against potentially 50 to 75k. Again Claude can't count in a lot of cases so don't use these numbers as literal examples but do understand that even if they're wrong And you express them as a benchmark to get a general idea. You're in the conversation, basically. And then he goes, what has been captured? So he's gone through and he's captured all this information, right? Then I'm like, okay, well, if you've captured all this information, where is it? Let's have a little look. So what have you captured? You mentioned bullet points above, but have you captured an analysis and then put it in the artifact window as a markdown file with specific information, right? So here it's gone through, it's done this, and then it's produced. The information that is collected from the PDP. Okay, as you can see, so there is quite a bit of information here. Okay, then, as you can see here, the next step. Why Extracted and Organized, Product Intelligence, Market Intelligence, Optimization Intelligence, Key Findings. Now you've got to remember I haven't actually give it a logic or a task to do and spent time of saying I want to break it down like this. The goal of this and to give this to people is that it It shows you what it can do. It presents it in such a way that you can then work off of what you've got in front of you and then make the adjustments to exactly what you want. Right. So it's a bit more like a this is a show and tell. Now, look at this. Pixel perfect. There's no fucking around trying to style all the cards, issues with colors. It doesn't use the proper design system. But in the same conversation, was able to go and capture a load of information from the PDP and then generate that, which took about 10 minutes. So, what I said was, now can we deploy the proper curve system, because the header and stuff was wrong. And then at that point, it did hit the conversation window. So, what we've got here is what we've captured using Claude Desktop with the MCP that's been, the set of MCPs that have now been efficiently optimized in a conversation that would normally max out the conversation. We've done two things. We've built a full dashboard, pixel perfect. We've had all the analysis done and we've used MCPs all within one single conversation. So for me, this is a big step forward in Claude. Let's have a see what we can do with Claude. I want to see if we can go back and work out why. Let's just change the gist. One, two, one. Oh, see. We're going to ask Claude again. We'll do a new chat. And we're going to take this URL. I'm going to ask you one step at a time like before, as I've asked for two things. So let's just drop this in here. Claude, can you use the Playwright MCP and do an analysis of this product detail page, but also make sure you use the skills. To manage the token and the efficiency of the MCPs. Let's see if we can get that going. And then we build the dashboard after. I'll analyze. Yeah, ready to be filed. What we've got here? Let me scroll down. Proceed with the optimized playwright analysis. Okay. So it looks like I didn't break it down by the two steps. Maybe gave it too much information because obviously we tried it twice. And it wasn't having any of it. Excellent. Now we've got the product specification and additional detail. Now I can give it direction. I've chose not to because it's purely for demonstrating skills. Let me go to here now. Let's go to. Where do I want to go? I want to go to. Go to comment and we're going to do Claude because I'll show you where to drop the skills into. So I am on the max plan but you don't have to be on that. I think I'll pay £200 or £180. I'm not sure exactly. Oh yeah, you've got the usage here. I'm using 15% I've got a spending cap, but actually I use another tool for when I'm running Claude code and I'll show you in a minute. So if we look at capabilities, there's two things that are going on here. You want to use the code file creation tool, right? Because what we're doing here is, and you can use this for presentations, but this is helpful because it writes script. Kind of help shape some of the logic and the structure from the information that it is capturing already. And then what you do is you drop in your skills. So you can see, look, auto-dash build generation client. So I've did this so it's got no curve branding, which I've showed you the black and white ones there. And then as we can see here, this one, this prevents massive MCP responses for WordPress, Playwright Notion and other MCPs. Now these can be quite heavy so for instance WordPress will always going to be heavy. Imagine if like the podcast is got maybe what 1100 and some change in terms of pages. If there's no control over that it's going to just do a full-on search because it hasn't I've got the specific instruction or there is no way for me or I haven't set up a way to control the MCP. You can do slides. I'll set this up for Cheryl, my wife, and then share that with her. Dashboard auto generation. This is the one that I use for my design system. I've also got the design system outside the dashboard generation because it's a different project. Then you've got the Claude, sorry, you've got the concise execution mode. So this is optimizing responses. These come with Claude, so you can test these. Algorithmic Responses, Artifact Builder, similar to kind of what I'm doing here. This will determine the components, the stuff that you may not need to know anyway at this stage. You can do your brand guidelines. I've not used any of their standard ones. They do have a skill creator, but I actually built a skill agent, which is more sophisticated because The skill agent I've built will allow me to take SOPs and then turn them into skill MD files. And because I can turn them into skill MD files, we'll be able to do and use the SOPs The work to be done, depending on what it is, by Claude, because it can use the MCPs. So where you'd normally give, say, your VAs, the SOPs to follow, some of the SOPs are going to be in a way that it can be automated. Now another thing on top of that I'm converting some of my agents with the skills maker in Claude Code. I'm going to take some of my agents and then some of those will be able to be turned into skills which of course we'll be able to use in Claude Desktop. All right, let's see if we've got anywhere with this. Right, so this is still going. Okay, so it's at the end now. So it's gathered comprehensive product data analysis and targeted JavaScript instructions, which is much more token efficient. This actually took longer. I don't know whether because there's more data that was collected this time around because I didn't give it an exact information. But this shouldn't take this long or it might just be speeds, might be server things related to Claude itself. What we've got here? There we go. So it's now given us a list. Okay, this has all been extracted. This is what is taken. All right. Now, can you use the full design system and then take the data that you've collected and build a dashboard and run me an analysis? Now that should give me enough context based on the skills to capture the information. Okay. I'm not sure how much is left. Reducing token usage by 85%. Let's see. I build a comprehensive dashboard with Curve. So it's using my design system. While that's doing, let's go back into Comet and we'll go back through Right, so as I was saying to you before, now this is something that I spent more time on. You can see with the header and it has the footer there and so this is based for conversion leaks. If you're spending money on PPC, which of course most of you will be, and you've got high click-through rate but low conversion rate, it's always good to understand what's going on on the PDP, right, where possible. And you're going to have to extrapolate for that information. This one here is where I'm bringing in the content quality. I'll go into this on the next show because I've still got some tweaks to do and this is the Rufus intelligent warehouse of my agent to conversate with Rufus to work out the it's framed as finding the objections of why people wouldn't buy as some data. See a lot of people when I've said this before Post purchase reviews are for post purchases. If someone's not buying, you want to understand why they're not buying, not why they bought it and either liked it or pissed off and wanted to return it. So you don't learn from the post buyer the information you often would need to understand the barrier that is stopping them from purchasing in the first place. It is good data, no question, but you want to work out why they didn't buy it, not why they bought it, didn't like it and then returned it. They're two different things, right? So it's collected all the information here, 13 sections. So this is the information that it collected. And then as we go down to here, can we take that same data and build the dashboard using the curve design system? And we did and we got this. And then if we look at some of the other reports. I won't do that one. That belongs to someone that everyone knows who wouldn't be happy if that was visible. Ah, there we go. So what I've given to people, your search term report would lay out like this. You upload the file. Claude, can you use the skills file to build me a dashboard using the attached search term report? Ensure that you use the full design system. Yeah, the idea is of what I've given to people outside the PDP that obviously that clearly I need to do some tweaks even though there is a variance in the first one, got it all in one conversation. That's the extraction managed of the MCPs and to build the front end of the report app or whatever you want to call it, snack app. And then on the second one obviously it's the token limit. So what I'll do is I'll debug that after the show and see if I can get that under. It just might be that we reduce some of the technical debt in the design system so we're using less tokens. We've got here, let's have a look at this one. Oh, that's the search term performance. Let's go down and chats, 18 dashboard, dashboard design project, What's this one? Okay, hidden complexity. This is the building of, this is one I did as well. So this is the search results page. So .csv, product, search results on the keyword Beardol, right? Just as a quick test. So let's say you're using one of the software companies. They've got the .CSV download the report. I'm sure you can do that. Or there are some other tools that you can use for this as well. But as you can see, same principle again, it's I've just give it a file and it's built a report. So it works on different data. Most you will get that in your inbox at some point or through connections of people. So there, as you can see, I did that all on the fly. You know, it was a bit of a rush there, but you've got the general idea how simple this is. You should be looking into using skills if you use Claude, you may be over OpenAI on the dark side with Andrew Bell. I've been trying to get him over to Claude. He's not been having any of it. But all this stuff around Erotica, apparently with OpenAI, they're introducing that, right? And I mean, Let's be honest, there's a lot of money to be made in that because there's a lot of, and I say this respectfully, there's a lot of what we call thirsty men, right? So there's always going to be a market for it. But being that Andrew is a man of the cloth, he's not very happy about that. But we'll see if we can get him over to Claude. But look, joking aside, I want you guys to win. Tough out there for a lot of you. But there's the opportunity. The opportunity isn't, I would say, AI powered and AI this. It's taking AI and then fixing the broken parts of your business. And if you can do that, I think that's where the win is. It's not pivoting away, it's innovating and then integrating AI to make your business move faster. So, without further ado, I'm going to head out. I'll speak to you guys next week. Take care of yourself and your family. Much love and I'll see you again soon.