How I'd Rebuild My Claude Setup From Scratch (80/20 Rule)

0:00 If I had to start learning Claude from scratch today, this video covers exactly how I'd do it. Over the past two and a half years, I've used Claude every day, whether as a lead engineer at JP Morgan,

0:09 the COO of a $25 million startup, or scaling my current business. But if I had to start over tomorrow, I would not do it the way I originally did it. Instead, I'd skip about 80% of the stuff

0:18 that doesn't make a difference and focus on the 20% that actually changed my life. So, in this video, I'll go through these seven pieces of the Claude puzzle I'd learned in the exact order, which is

0:28 the exact process I've used to help hundreds of founders and their employees quickly and effectively become experts with Claude. And the last piece of the puzzle is the one piece of advice that

0:38 changed my life, and if you pay attention, it will change yours. So, the first piece of the puzzle is set up your workspace. You're on this channel to become an expert, so I'm not going to

0:46 teach you things that are not best practice. And what this means is we're intentionally skipping a couple of key tools that you've seen or heard about because it's wasted effort because we're

0:54 not going to use that in the long run. So what are the tools we're skipping? We're skipping clawed projects and clawed co-work. The reality is these are good tools, but we're going to skip them

1:02 because what we end up using is more powerful and will create a stronger foundation to build on. This foundation will make it so that we don't waste time in the long run. So what do we use

1:10 instead? We're going to use Claude Code through the Claude desktop app, which is optimal for both technical and non-technical tasks. And my biggest problem with this tool is honestly the

1:18 name because people instantly get intimidated. So if you are intimidated by the name Claude Code, just rename it to Claude expert mode in your brain. That's it. Of course you want to use

1:27 Claude expert mode. Just think of it like that. There's no reason to be intimidated. To set this up, it has nothing to do with anything technical. Just download the Claude desktop app,

1:34 hit code, and that's it. And if you're wondering where to download it, it's linked below. Then once you're in it, select a folder to work in. Think of each individual folder as a project that

1:42 you're working on. And if you're brand new to Claude Code, create a master project. And that's where you'll do 90% of your work. For me, I call this project my internal OS, which is what I

1:51 use every single day. Once you create that folder, select it and then write /init and hit enter. This will initialize your project, creating a claude.md file, which is a file that

2:00 Claude reads at the start of every session. From there, we've set up our project, but we want to do two other things. The first is create an about this project file. You'll run this

2:08 prompt, which will interview you about the project, your role, your business, your audience, etc. And this will set foundational context for Claude to give you better responses. The second is

2:17 create folders in your project to store important contextual information. The idea here is you want to curate resources that will help get better responses from Claude. So for example,

2:26 you could have a favorite book that you want to ingest because you like how it thinks about solving problems. This is a process of creating what's famously known as LM knowledge bases. And I have

2:34 a video on my channel with over 800,000 views diving deep into this. But simply put, it just creates two folders. The first is a / raw folder which is where the original resource will live. This

2:44 can be the book that you love or the article that you love. Then I'll create a /wiki folder which has files that help AI identify where to look for the raw information. Think of the wiki as the

2:53 table of contents in a book. We don't want AI looking through the entire book if they just need the information in chapter 10. And if you're working on a technical project and wondering if this

3:01 applies to you, well, all of this contextual information can help get better responses in terms of things like landing page copy, branding guidelines, etc. So, this applies to both technical

3:11 and non-technical work. On screen is the prompt to help you set up your project and create this LLM knowledge base. So, at this point, Claude knows who you are and we've created a foundation to build

3:20 on that won't go outdated and won't limit what you can do. Nothing in this video is wasted effort. And this is the foundation that the next six pieces build on top of. The second piece of the

3:28 puzzle is have Claude make your life easier. To make your life easier, Claude needs to work on the actual files on your computer and it needs to make sharing those files super easy. So, how

3:37 do we actually do this? Well, the first step is download Obsidian. This is a free tool that just lets you view files on your computer very easily. If you're a Mac user, this is just a souped-up

3:47 version of the notes app that you have. Don't overthink of it. It's just a file viewer. So, what you can do is you open it, you create a new vault, and then pick the same folder that you're working

3:56 on in Claude Code. Now, every file that Claude edits, you can open it up and edit it right in Obsidian as well. Now, you have this collaboration ecosystem that's super easy to navigate. The

4:04 second is we want to become power users of a claude feature called artifacts. An artifact is essentially a mini website that Claude makes to visualize something. And this mini website can be

4:13 shared to anybody, an individual, whatever. And so Obsidian and Artifacts, this pair empowers a lot of use cases. An example for how I'll use this is I'll work on email copy and claude code until

4:24 it's directionally right. Then I open it in Obsidian and we'll hand type and edit it. Yes, like Stone Age. And once I feel it's ready, I'll create an artifact and share it with a team member to help get

4:33 feedback. And to do this, I'll prompt create an artifact that shows this email visually with the copy we just worked on. Let my team edit it directly in line and add a button that copies all their

4:42 changes so they can send it back to me. I share the link to them. They edit right on the page and their changes come back in a form that I can paste straight into Claude to make the edits that they

4:52 suggested. There's no friction for them. There's no friction for me. This is how you make your life easier and you set up a foundation that can actually produce productive outcomes. And we want to

5:01 become a power user of this artifact feature because it opens up a lot of possibilities. The example that I just shared is more of a singleuse artifact which can be used to get feedback or

5:09 just share information in a format that's visually appealing to somebody else. But another use case that I love is using artifacts to create micro tools. An example here is that one of my

5:18 clients runs five different restaurants and he uses an artifact as the drink menu so that the bartenders can see exactly how to make cocktails. If a customer asks them about a drink and

5:27 they aren't sure how to make it, the bartender just opens up this artifact on their phone. They don't have to log in. There's no app required. It's just a link that they access and it's seamless

5:35 to create. You didn't have to spin up a whole website for it. So, at this point, Claude is an excellent collaborative tool, but it still can't see anything outside your computer, which limits its

5:43 functionality. And I'll get to piece number three which solves this problem. But I want to talk a bit about some of the recent features from Claude Code. Which brings us to today's video

5:50 sponsor, Anthropic, which you guys know I'm absolutely hyped about because of how much I love Claude and Claude Code. And I touched on this in the last part of the video, but Claude Code allows you

5:58 to edit files directly on your computer. It reads your files, writes the code, and actually builds the thing with you. And I was talking to two of my friends literally yesterday, and they had no

6:06 idea that you could use Clawed Code outside of the terminal. Half the time I personally use it on their desktop app or on their phone app when my fianceé tells me that I have to shut my laptop.

6:15 Just don't tell her that I'm using the mobile app. And since their recent launch of Opus 5, this tool's gotten better in three ways. The first is routines. Now, I've spoken a lot about

6:22 this feature on this channel, but essentially you set up a prompt on a schedule and it runs on Anthropics Cloud automatically. And I've got a bunch running, but the one that I love the

6:30 most is a weekly review of all of the feedback and future requests that users have on my product build partner.ai. And so what I do is once a week I have it pull those requests and analyze all of

6:41 them. And then once it analyze it, it brings me to the second feature, plan mode. So once Claude looks at these requests, it uses plan mode to architect a solution without me. Then I'll go in,

6:49 review the plan, and finalize it based on what I think is best. And with Opus 5, it's actually trained to push back when it disagrees with me. And frankly, I'd much rather hear it's a bad solution

6:59 before I decide to build it. And the third feature that's been enhanced is the desktop app. This is the tool that frankly I've seen the biggest improvements on. So I use routines to

7:07 identify the features. Then I use plans to create it. So after the plan is ready, I kick off a few agents in parallel using their desktop app. One building a feature, one testing it, and

7:15 then the other reviewing the design itself. And with a desktop preview, it lets me see the website directly through Claude, so I never have to switch to another tool. So whether you're Tech or

7:24 not, this tool makes it so you can start building. So click the first link in the description to get started, and you are going to be happy that you did. So, at this point, we've learned how to use

7:31 Claude as a collaborative tool, but it still doesn't have access to any information that's outside of your computer. And this limits its functionality, which brings us to the

7:39 third puzzle piece. Connect Claude to the external world. The quality of outputs from Claude comes down to the quality of the information you provide it and the permissions that it has. If

7:48 you restrict Claude to the model and just your local files, you're limiting what it can do. So, the simplest way to connect to the outside world is using Claude's connector feature. And a lot of

7:58 you may know about this, but there's a chance that you're missing one key component to setting it up. But starting from the top, for those of you that have never used connectors, open Claw

8:06 Desktop, hit the plus button, and then go to connectors. Then the four connections that I think everyone should have set up is connecting their email, their calendar, wherever they're sending

8:14 messages. So for me, that's Slack. And then your meeting recordings. For meeting recordings, I use Granola. And if you want 100% off that tool, it's linked below. You guys know that I talk

8:21 about it a bunch on this channel. Then once it's connected, it's about setting the permissions. What you'll notice when you start using connectors is it'll request you for approval for everything.

8:29 If that happens and you're getting frustrated, go to the permissions and tweak things so it's in a Goldilock zone. It doesn't ask you for approval for everything, but it asks you for

8:37 approval for things that are super important to get right. Once you have your connectors set up, they're the permissions that you want. Go to connectors, tool access, and switch it

8:45 to load when needed. This will save you tokens and only use the connectors when they're actually needed. 99% of people don't have this configuration set up, so go and do that. It'll save you tokens,

8:55 which will save you money. And once you set up these connectors, you'll be able to interact with the external world and enhance the data that you have access to. And if there's a specific tool that

9:04 you don't see connectors that you want to connect to, run this prompt, which will see if that tool has any sort of custom MCP connection that you can do manually. In short, an MCP stands for

9:14 model context protocol, and it's how AI interacts with external tools. A lot of the biggest tools have firstparty integrations with Claude, but this is how you can set that up if needed. So,

9:22 at this point, we've established connections to some of the biggest data pipelines that we interact with. We can have Claude read and send emails on our behalf. We can have it checking our

9:30 calendar to see what we need to prep. We can have it looking through our call transcripts to make sure we didn't miss anything. That external data enhances what we're able to do. But, we're going

9:38 to find that we're often repeating oursel a lot or we're trying to do the same things over and over again. And that's where the fourth piece of the puzzle comes in. Teach Claude new

9:47 skills. This is all about creating custom skills that are specific to what you work on every day. And if you master this one piece of the puzzle, you are going to be ahead of 99% of people. I

9:56 can guarantee you that. A skill is a reusable prompt that Claude runs, so it does the same task the same way every single time. Someone on Reddit said this best. Skills are lazy ways to do

10:05 repeated work, so you don't have to retype everything. You may know about skills, but my guess is you're either creating them wrong or you're not using some of the most important features that

10:14 they have. So, first, how do you actually create Claude skills? Yes, you can just say to Claude, "Create me a clawed skill to do this." But that is not what you should do. Instead, first

10:22 do the task by hand once. Go back and forth with Claude until the output is correct. Then say based on this conversation, build me a skill. And for the sake of this process, let's say that

10:30 creates a skill called monthly report. Now, next time you have to run this monthly report, use the skill instead of prompting Claude. And you would do this by typing /monthly report. And it'll do

10:39 exactly how you did it before. Then step four is the output won't be perfect from the skill the first time. So after fixing the output, going back and forth with Claude, say based on this

10:47 conversation, enhance the skill I used. And over time, you'll continue to improve these skills so they get better each time you use them. This skill creation loop is something that you'll

10:56 continuously run, and you'll build out custom skills over time as you need them. My general rule of thumb is if you do something repeatedly, you should make it a skill. And the workflow I just

11:04 covered forces you to only build skills that you know you need because it's centered around a task that you actually continuously do. So, that's how you actually go about creating and

11:13 identifying what skills to create. But, I've been studying how to best use skills. And there are three features that you need to include in your skills. On screen is a prompt that will enhance

11:21 your skills with these features, but going through them one at a time. The first is ask user question. This will force claw to put its question in a clean form instead of a block of text,

11:31 making the interface a lot more enjoyable. On the left side of the screen, you can see what it looks like after you enhance it with ask user question. And then on the right, you'll

11:38 see what it looks like if you don't use this. This is a big quality of life improvement, and you're going to thank me when you do it. The second is using arguments. If you think about a skill as

11:46 the logic that runs, an argument is the variable that changes how it actually runs. Let's slow that down and look at the example of the monthly report. We can add an argument to the skill called

11:55 output. And then when we call the skill, we could say /monthly report PDF. This will make it so that that skill now knows to create the monthly report in a PDF version. Then we can have it create

12:04 other formats like artifact or CSV. And depending on the argument and how we call the skill, it'll change the format of the output. And the third feature is sub agents. Claude has an ability to

12:14 split tasks up so that they run simultaneously. This allows it to speed up the process. So in the example of the monthly report skill, let's say it had to look at Slack, QuickBooks, and the

12:23 internet. Instead of having to wait one after the other, you can run them all simultaneously with sub agents. This speeds up the whole process. And yes, it will cost you tokens. So you do want to

12:31 be a little bit mindful, but I love just launching sub agents so that these skills can run that much quicker. And these features aren't necessarily needed for every skill, but the prompt that I'm

12:40 showing you will help you identify which skills you already have that can be enhanced using these features. Okay, you now know how to create skills, but what are some skills that everybody needs? I

12:48 acknowledge that everything that we're covering here can be a bit overwhelming. So, in order to continue to use what we're covering, the system has to be easy to maintain. Otherwise, you're just

12:56 going to go back to Claw Chat in the browser and you're just going to pretend that you never saw this video. So, I want everyone to create skills that helps them maintain their system. These

13:04 are skills that I call utility skills that help keep the system aligned. The first is /add new resource where you can give it a file, a link, a transcript, and I'll ingest it into your system to

13:13 make sure it goes in the raw and wiki folders that we created in part one of this video. The second is an improve system skill. This will audit your setup and enhance your skills and context

13:22 based on what we've been talking about. You can use the prompt on screen to create these two skills, but if you want to use the exact ones that I use, you can use my free build partner.ai plugin,

13:30 which is linked below. One of the reasons that I love skills is it starts systematizing tasks. So you can just run a skill and it'll automatically do everything there. Which brings us to the

13:39 sexy part of this puzzle and it's the part that I've personally have wasted the most time on. So I'm going to give you what actually works and is all you need. Piece five of the puzzle is

13:49 minimum viable automation. This is the part that everyone loves to flex, right? They'll say, "Oh, I got hundreds of agents working for me every single night. It's doing this. It's doing

13:56 that." that it's like, brother, if you had a 100 person team working for you, I hope you're making $10 million plus a year, which I don't think you are. They love flexing this stuff, but a lot of

14:05 this is just clickbait garbage. So, here's what actually matters to automate the task that you're working on. In Claude Code, or as we like to call it, Claude expert mode. On the desktop app,

14:14 click routines and then new routine, and then select local. And you can tell Claude what you want to run and when to run it. In this case, we pick local instead of cloud because it runs on your

14:23 machine, which eliminates any of the variables that come up when you have to run it on the cloud for almost every task. Believe it or not, this is sufficient. And routines is Claude's way

14:31 of automating tasks for you. It's really that simple. But what are some of the most important things you need to consider? The first is only build routines with Claude skills you've

14:39 already made. In the details of the routine, if you're typing more than a sentence or two, you have failed me. 99% of what the automation does should be done through a skill. on a video on my

14:49 channel, I break this down in detail, and I'll link to it at the end of the video, but this is called skill- driven automations. On the left of the screen, you'll see a routine that's properly set

14:56 up using skill- driven automations. And then on the right, you'll see one that isn't properly set up, and the difference is very obvious. The second thing to do with routines is for every

15:05 automation you run, make sure that you get a recap and the recap message is explicit. So essentially, at the end of every automation, always notify yourself with an update. So that could be with

15:14 Slack, email, whatever you pick, just pick the one that will be the least effort to monitor. Then that message itself, make sure the updates are explicit, not implicit. So in an email

15:24 automation, for example, that reviews your emails, you would have it say, "Analyze 41 of 41 emails drafted to responses." You wouldn't just have it say, "Email analysis complete." The best

15:33 clawed users are building systems that are easy to maintain. And this is a key part of it. The third thing to consider with routines is create self-breaking systems. Before any automation does

15:42 anything, run a pre-check that will check every connection it needs. So, for example, if an automation needs to be connected to your Slack, it'll check and if that Slack connection isn't working,

15:51 it'll instantly fail. This way, it won't spend the first 20 minutes doing something else and then eventually break. Instead, it'll break right at the start. This will save you time, tokens,

15:59 and money if you do this. On screen, there is a prompt that will help you enforce these rules on your existing routines. Now, you have the whole system running, and this is when you start

16:07 feeling the power. And then you'll realize, right, you're using clawed code or clawed expert mode. And the reality is it can write code. So a lot like a kid in a candy shop, you're going to be

16:16 tempted to start building things. And that is one of the biggest traps that the next piece of the puzzle will solve. Now before we get into that, if this is your first video of mine, welcome to the

16:24 channel. But if it's your second or more, here is our anti-slap agreement. The visuals, the testing, the hours of research that went into this video that's entirely built for humans, not

16:32 for AI robots. So, all that I ask part of this agreement is to hit the subscribe button so this content can reach more people. Also, as a thank you, I give away a Claude Max subscription

16:41 and last video's winner is Joe Scene CGI for building a health and fitness dashboard. To enter the next giveaway, comment what you're currently working on and how this video will help you. I

16:51 truly love hearing about what people are working on and I'm here to help as much as I can. Now, at this point, we covered five pieces of the puzzle, but there are two more that are absolutely critical.

16:59 Piece number six is build new things in the right order. You're going to get an itch that's I want to build an AI tool for this. This is a good itch, right? You're enthusiastic and it's not one of

17:08 those bad itches that you have to pretend that you don't have. And that's why we pick it code in the beginning. You can build with it. But this is dangerous. I call this the AI hammer

17:16 problem. Once you learn about AI and you see how powerful it is, everything starts looking like a nail and AI is the hammer. You're going to start reaching for AI for every single problem that you

17:25 face. And then the second you face a problem, you're instantly like, I'm gonna build a skill for this, or I'm gonna build my own app that solves this problem. For my experience, this will

17:33 lead to you wasting time and focusing on the wrong things. So before you build anything, go through these questions in order. The first is, what would I do if AI didn't exist? This will make you

17:43 think, is this really an AI problem or is this an operational problem? You want to dive into the root of the issue before you reach for AI. Two is, is it a skill issue? Before reaching for a new

17:53 tool, are you using the current tools properly? Slow down. And if you're unsure if it's a skill issue, just ask AI if what you're trying to do is possible with the tool that you're

18:02 using. The third is, does the tool already exist? There are a lot of tools out there and is making a new version of this tool the highest and best use of your time. The fourth is, can you make a

18:12 small tool that enhances a tool that you already use? The best way to explain this is through an example that I experience. So, my YouTube video pipeline that my editors use for project

18:21 management lives in Notion. But the problem with this is it doesn't have a good thumbnail visualizer. So, instead of replacing Notion, I just built a micro tool that helps me visualize

18:30 thumbnails. I didn't have to replace a tool. I built an additional tool that enhanced the one I currently use. In my eyes, too many people think they should replace SAS products because they can

18:40 vibe code. Now, SAS is a trillion dollar industry for a reason. Don't rebuild a worse version of something you already use or is already pretty cheap. And the reality is in a team setting, a small

18:51 tool that makes the existing thing better is a,000% more likely to land than switching everyone to something you built in a weekend. And finally, number five, after going through all of those

19:02 steps, only then do I consider making my own product. And at that point, you need to strip down what you want and start with an MVP. This is the smallest thing that provides the concrete value that

19:12 you need. And a lot of the times an artifact through Claude can serve as an MVP. But if you want to go down the rabbit hole of building out your own product, type SLP plan and say, "I want

19:20 to build product." Interview me to identify the MVP and make sure the value is clear. And if you're technical watching this, you likely have this part of the process nailed down. And if you

19:29 aren't, this is enough to get started. The reality is that most of producing highv value work with AI is avoiding distractions. And this last piece of the puzzle will help you avoid distractions

19:37 in your life. And it is the advice that changed the trajectory of my career. Piece number seven is make everything congruent. Warren Buffett has a famous quote. My life has been a product of

19:46 compound interest. Nothing more, nothing less, and nothing brilliant. If you build slowly and work in the right direction without being distracted, good things happen. AI will let you learn

19:56 anything and build anything. And yes, this creates a mountain of opportunities, but these opportunities can be distracting for most people. They'll start one thing, they start the

20:05 next thing, and they aren't actually building anything, and they're going all over the place. There's no congruency. In this video, we created a foundation that can be built on top of without

20:13 becoming obsolete once you hit a certain level. That's the setup. But now you as the user, you need to point it in the right direction. In my life, once I made everything congruent, I started seeing

20:23 the biggest productivity gains. I won't dive into details of my story. You can see that in the description. And so for you, before you start or do anything, and AI will make this hard. Remember

20:32 what Warren Buffett said. My life has been a product of compound interest. Nothing more, nothing less, and nothing brilliant. That completes the puzzle. Focus on doing things that are all

20:41 congruent so that they can build and compound on each other and you're not starting from scratch every single time. That is the whole puzzle. Seven pieces in order. And it's exactly what I would

20:50 learn if I was starting from scratch. And once you're here, you're going to want to automate more and more. And unfortunately, you're going to struggle getting valuable outputs from the

20:58 automations that you built. So, linked here is a video where I go through a deep dive on how to build out automations that are simple and effective. This builds on everything I

21:06 covered here. And if you like this video, you will love that one. I'll see you over there.