Best Claude Code quotes

There is an incredible piece of software called Claude Code that I've been using for the last 3 months and it has completely and utterly changed my life.

Claude Code is the tool that I've been spending most of my time on practically 24/7 for the last like 2 months.

It's absolutely sick.

I was intimidated by Claude Code because it involved the terminal.

I was like, "Oh my god, terminal."

"I'm not like some hacker from the 1990s."

"Why do I need to use the terminal?"

And then I realized I was a total dumbass for thinking that.

And actually, it's really straightforward.

I've built stuff for my business and for my customers that literally adds value to our customers every single day.

People are raving about the tools that I've managed to build for them over the last couple of months just by learning how to use Claude Code.

Every single day the gap between the people that just use the free version of ChatGPT and talk to it on the web versus people that know how to build stuff using Claude Code even if you don't know how to code.

Every single day that gap is getting wider and wider.

Using Claude Code, every single day I'm learning so much about the world.

I'm using Claude Code this way as opposed to using Manus, because Claude Code is cheaper.

If I were to do this in Manus, it's going to cost me a lot of money.

That's the really big difference.

They both can do this exact same task, but Claude is going to be way, way cheaper.

I use Genspark to create my instructions for Claude Code because Genspark actually knows Claude Code very well.

At every moment that Claude Code sent me something back, I just copied and pasted it into Genspark and said, what do I do now?

What do I pick?

It's just using one AI tool to help us leverage the other AI tool.

This is easily the biggest change to my basic coding workflow in 2 decades of programming, and it happened over the course of a few weeks.

I really am mostly programming in English now.

While I'm talking to you, my computer in the other room is coding.

I'm coding right now, as I'm with you.

I don't even know coding, but I'm coding.

Claude Code's ruined my life.

I mean, it's the most addictive thing on the planet.

One of my good friends is now addicted to Claude Code.

He came to me and said, listen, Claude Code is basically a video game times a hundred for creators and entrepreneurs.

I'm not technical at all, and I'm currently building some of the most impressive tech myself, sitting in bed at three or four in the morning after my wife has fallen asleep.

AI is teaching me about code.

I can sit down now and have a pretty technical conversation.

I know all the tools.

I understand how they interface with each other.

I understand how code talks to itself.

But I didn't understand any of that two months ago.

People would come to me and they're like, I built this on Lovable.

And I look at them and I'm like, amateur.

Once you understand Claude Code, it's like, why would you build it on Lovable?

It's just using Claude Code behind the scenes anyway.

This technology is alien and magical. It makes it so much easier for people to build and create.

Increasingly, code is not the bottleneck.

Invest in your CLAUDE.md.

Claude is eerily good at writing rules for itself.

I haven't written a single line of code by hand since November.

I'll go to the gym and think, crap, this is something I want to build.

I just talk to my phone, open WhatsApp, and tell Claude Code what to build.

It builds it.

I review the PR when I come home and then it's in production.

98% of the work's done.

I just need to show up and see, okay, here's what the problem was, here's how they solved it, here's how they tested that it was fixed.

And I click one button and now it's live in production.

Think of Animus as, it's kind of like imagine OpenClaw, but very custom.

It's basically like my own version of OpenClaw that has a web app attached to it too.

Everything from my daily notes, like my journal and stuff, what I do every day.

It's all auto-generated.

It's looking at what I do every single day.

I can go through and see, hey, what did I get done this month?

We have investors.

So if I do an investor update, I literally just go to Animus and say, give me the last 30 days of stuff that I did for the business and give me bullet points.

I've never been good at note taking or tracking time or anything like that.

So this is the closest I've got to being that kind of organized.

I usually spend most of my day, if I'm at the computer, either in here in the desktop WhatsApp or if I'm mobile doing it through my phone.

I have a WhatsApp interface.

I can, you can see here like I'm asking, I'm telling it what book I just read.

It triages my inbox.

I tell it to review this, review that, build a feature.

It's all in one interface.

Everything from logging gym workouts to getting a book recommendation to fixing stuff in the app or improving the app to looking up customer account data.

Like it's all in one interface.

I was walking to the gym with my headphones on listening to music.

And then I just used Wispr Flow to talk to my phone and said, create a Linear task for this.

And yes, go ahead and build it.

So then it started building the task while I was at the gym.

I'm hardly ever going into code editors.

I hardly even look at my code anymore.

It's creating stuff, it's doing work independently of me.

It's doing pull requests, it's writing tests, it's verifying that everything works.

What used to be for me a very reactive thing, like I'd go and say I want this done, it's now doing things proactively.

If we get an error, it'll go in, create the fix for the error, push it up to GitHub.

I've got a whole backlog of Linear tasks.

Claude just works in the background fixing it throughout the day while I'm doing stuff like getting on podcasts or doing partnerships.

Some of the pull requests I need to review are things I never prompted.

It just got it from people writing into support.

I could give this thing autonomy and say, once it's done and Greptile's cool with it, ship it.

I'm sure some people do that.

I'm not there yet.

I have a thing called slash check which plays devil's advocate.

What are things that for each change we're doing, how could this affect customers?

What are some ways that this could fail?

How can we roll this back?

How does this affect legacy data?

Greptile comes back with a confidence score.

So it was only three out of five.

It said it's mergeable, but there's a potential thing that could cause an issue.

It basically gives me what needs to get fixed.

I don't want it to be that anyone can email in and then now there's a new button in the app because one guy happened to think there should be a button.

I still want to curate.

Building it is effectively free.

It just does it and pushes it up.

I can get rid of the pull request if I don't like it.

If I think it's a dumb feature, I'll just delete it from Linear and then the pull request is gone.

I said, tell me about my RightMessage account and Bento.

Tell me about my accounts.

And then it just went through and here's last 30 days of data, stuff about our Bento data, how many subscribers we have.

How do I make it so Animus is the ultimate employee?

It's always looking at what's going on with all morning.

It's looking at what's going on with support.

It's looking at what are people doing in the product.

I still feel like this thing's not dialed in enough.

It's not yet monitoring LinkedIn and Twitter for mentions of the brand.

Why is it not doing that?

I should have that set up.

I basically built a gateway, which is an always-on thing.

It's running on a VPS on Hetzner.

I used to run it in my cellar, but that's a point of failure if my internet goes down.

I'm using Tailscale, which lets me secure it by creating my own VPN between my house, my phone, and the VPS.

You can't access any of that stuff if you're not on the VPN.

There's a way to run Claude without it actually running the app itself.

Print response and exit.

So whenever I send a message to WhatsApp, it's effectively doing this under the hood.

Claude Code has a new scheduling function.

They're fine with autonomy or recurring things.

Twice a day it goes in and triages my inbox.

It'll go and write draft replies.

I'm not too worried about that because I think I'm still within the terms of service.

Claude Code's really good at being able to say, look for newsletter sponsorships or email sponsorships or any variations of that, and then look through the files.

I find Claude Code by itself is really good at that.

If you have a lot of experience building software, then you use Claude Code differently than other people do.

I talk to it like another developer.

I give it very specific guardrails about how a feature needs to be developed.

The value of a senior developer using Cloud Code is that they have the experience to know what rabbit holes need to be kept out of.

I don't just say build a feature.

I say, here's what I'm looking to do.

Let's converse developer to developer about how this should work.

I have AI interview me.

So that's a big thing.

I don't just say build a feature.

I say let's converse.

That takes a bit of time.

Probably spent 30 minutes talking to it about this.

Before you hit record, I was saying I've got it working on a feature right now.

Because we feel like it can be working 24 seven.

If it's not doing something, that's bad.

This week alone, we've already pushed a few features.

They're small, but new stuff is continuously going out.

The intercom integration for RightMessage was built in four days.

It used to take like a month to build an integration.

We've got 13 or 14 integrations at this point.

If we want to add integration 15, all we need to do is we've established a precedent for how we do integrations.

It looks at all the other 14 and sees how we do it.

I'm on the 20x plan thing.

Amazingly, I haven't actually hit the limits.

If it sees your API key in an environment file, it'll use that even if you have a plan.

So for three days, I accidentally put my Anthropic key in my environment file.

Then all of a sudden it started billing me on the API.

Every hour another $40 charge.

Claude is really cracking down.

They want you to use Claude Code if you're going to use Claude through a terminal or through an app.

What OpenCode did was they changed their prompt to make it seem like it was going through Claude Code.

So it kind of spoofed Claude Code.

And then Anthropic cracked down on that.

I saw a tweet the other day, a lot of their friends aren't going to the gym as much anymore because they're always glued to their Claude Code.

Because this thing can be working all the time.

If it's not doing something, that's bad.

If you're put off by the code in Claude Code, you shouldn't be.

Claude Code is honestly the best deal right now for both pricing and what it can do in one package.

Even if you're not a developer, don't worry about the code.

Just go in, open it up.

If you're using Obsidian, just launch Claude Code in your Obsidian directory.

Start there.

That's a really solid place for most people to start.

I think the next big thing is proactivity.

Right now, people are shifting to routines like automating responses to customer support tickets.

The next step is that Claude understands what you work on, and just sets up these automations for you.

One of the most underrated features from Claude is actually that skill aspect.

If I find a process that works very well, I'm just going to ask Claude to make that into a skill.

Then when I need a specific prompt for a newsletter cover, I can generate that just by asking Claude.

A lot of people think Claude Code is just for developers.

But marketers like me can build landing pages with it.

You can build almost anything with this tool as long as you've got the right context.

There's a huge developer community of these MD files.

One that was very popular was the Remotion one for Claude Code.

You can create motion graphic videos with Claude Code, which is mind blowing.

With Claude Code, you no longer need to know coding to code.

You don't need to write the code.

You just manage the agents that write it.

There is a reason everyone is talking about Claude Code.

It is the most powerful AI tool available.

People are building incredible apps and websites with Claude Code in a couple of hours max.

Yet another seismic moment.

Claude Code is unbelievable.

Anthropic had somebody build it in a weekend, just working on the side.

Claude Cowork is like you think it's magic.

Claude Code and Claude Cowork both connect directly to your computer.

I had when I first tried Cowork I had it go through and organize everything in my downloads folder which is a mess beyond belief.

It created all these different folders, everything.

Skills basically is like a way to never have to explain yourself again.

Claude can pull the skill as well.

So if you saw it was pulling like when it did the email one, it pulled from an email writer that I have installed there.

Which is really cool.

Skills are pretty easy to create because it can just be you talk or you find a YouTube video for example that walks you through something and then you have it create that as the skill.

It's crazy.

They'll learn how to use agentic AI to do marketing and kind of marketing like an engineer.

Using tools like the terminal and programmatic methods for doing marketing, whether it's SEO, ads, even just research that will speed up their workflows by like 10X at least.

Claude Code is the most powerful technology tool that's ever existed.

It doesn't matter what role you have or what you do day to day.

It can absolutely be a game changer for any of your workflows.

Anybody can both build an app that is production grade, scalable, bug free, nicely designed, and they can also market it with zero experience in either.

Cloud code did this in about 30 seconds.

This used to be someone's job that would take them a week sometimes.

The terminal is like your backdoor to your computer.

You're plugged straight into the matrix.

You're not a couple of steps removed.

You're not looking through a pane of glass.

A skill is a markdown file that has all of that information saved in kind of like a system level prompt for you.

People will look at the skill and they're like, wait a second, what?

It's just a bunch of texts.

It's just a markdown file.

That doesn't seem helpful at all.

But this gets referenced every single time you use a phrase like SEO audit, technical SEO, why am I ranking, SEO issues.

It's like an executable SOP.

Without a skill, the AI agents are just using the base LLMs, which are great and are getting better every day.

But if you want AI to follow a very specific protocol, and if you've learned best practices over time, you want to essentially encode every one of your prompts with that context.

An API is an application programmable interface.

It's essentially ways for tools to talk to each other.

It's an authentication method to go and access the data from other tools.

In my mind, it's always API first, then CLI, and then MCP.

Whenever I have a choice between CLI or an MCP, I'm always choosing the CLI every time.

MCPs are like a quick and dirty wrapper around an API.

A browser is just one tool that your computer uses, but it's going to be pretty limited to the confines of that tool.

The terminal is like a backdoor that unlocks any and every capability that your computer has at its disposal.

When you're loading Claude Code into your terminal, you don't have any more limits.

It can process a lot more data faster.

It can chain tasks together.

It can work across multiple files, even multiple projects at once.

I found more success basically telling Claude, figure out how to do this.

Here's my goal.

Here's what I'm trying to achieve.

Now I want you to tell me the implementation plan.

Actually the way that you think might be the best way might not actually be the best way.

So now you're forcing AI to do something suboptimal.

Context is still king for AI.

What you don't want to do is every time you're working with a different task, you have to re-upload the important context about your business.

I created a skill called product marketing context that each one of these skills always references every time you ask it to use that skill.

So it understands my ICP, what my product does and doesn't do, the competitive landscape, the unique differentiators.

This same context is shared between any of the other skills.

So now I could jump to paid ads and then it can read all the same info about my ICPs, product differentiators, competitors.

As each of these updates, the other ones get better and you have this kind of compounding fly loop that improves over time.

I use the plan mode in Claude Code to plan out a feature.

I'll have it play devil's advocate, think through all the different edge cases, think through the user experience.

I'll have Claude Code use Opus to do the initial implementation.

Then within Claude Code still, I'll invoke a Codex skill that loops in GPT 5.4 to do a review of everything that Opus just built.

Codex and GPT 5.4 is really, really good at traversing a big code base and finding a lot of the nitty gritty little detail-oriented things, but it's just kind of slow.

It's a second pair of eyes and a second brain to get an outside perspective.

Then you can use the Agent Browser.

Vercel built this essentially like a CLI over Playwright.

It'll actually click through and even record a video of itself using your app, just like a user would.

Especially when working with Claude Code, whatever tools Claude Code likes the best are the tools that I like the best.

When you make your site in custom code, then it can update things really, really easily.

It knows how to build pages.

In Framer and Webflow, they're all React based.

You can generate React out of Framer and Webflow, but you can't import React code.

So it's kind of only one way.

I would definitely recommend people start with other people's skills.

There's a ton of really good ones out there.

There's a whole directory resource now called skills.sh.

You're copying it to your own local computer.

So you can always grab updates from the latest changes.

But you can also fork it and customize it and personalize it to your own.

Take the programmatic SEO skill that I've created.

If you're like, that doesn't work super well for e-commerce, I'm going to make some changes to it.

Tell Claude, these are the things I need it to do for e-commerce.

Great, now you've essentially created your own skill.

If you can't make use of what's out there, then go and try to figure out how to make your own.

But I would say that's kind of a last resort.

There are a lot out there already.

Most tools you can go to the settings, find a developer settings page.

For example, API keys.

Either grab your key that's pre-generated or generate one.

You're not really supposed to do this security-wise, but you can paste your key into the chat and Claude Code and then tell it to save it in your environment variables.

An environment variable is essentially your credentials that are saved securely and locally on your computer.

The really only risk security-wise is if any one of your chats is getting uploaded somewhere else.

If they all just get done locally on your computer, the risk is very, very minimal.

I became one of the original vibe coders, just button mashing and accept all, accept all, accept all.

I have a fundamental first principles understanding of coding and web apps and how to build those apps, but I don't care about the actual code itself.

Claude Code is an AI agent.

It has control over your computer, your files.

It can use the browser.

It can connect to third party tools.

So now Claude Code becomes the orchestrator and you're no longer the middleman.

You're not limited to just all the time you have to spend in front of a screen talking with AI.

It can do things simultaneously.

It can do things in the background.

It can have multiple tasks going at once and can actually execute for you.

It can also have it plan out a lot of things where you don't have to tell the AI exactly how to do something.

I found more success basically telling Claude, figure out how to do this.

I've been a Max subscriber since last summer and I have never really like, I would have spent probably tens of thousands of dollars at this point.

At some point it won't be as subsidized by the VC's money, but that's okay.

I'm going to make the most of it that I can.

GitHub is essentially a file storage system built on top of a technology called Git, which is all about version control.

Git is actually incredible.

I personally want to bring Git to everything.

Not just coding and programming.

To copywriting, to marketing, to design.

Having a paper trail of versions and version histories and changes and updates and notes is super, super useful.

Your file organization has to be top tier for AI to be able to navigate without getting tripped up.

Anytime you want to use a skill, literally all you have to do is tell Claude Code, hey, let's use the programmatic SEO skill for this.

And then walk me through it.

For each of your clients, you can have a folder.

For each of your own projects, you can have a folder.

You can split that up however you want.

You can have as many terminal instances as you want.

In general, I'm usually just managing multiple agents at once.

That's where I find most of my productivity. bouncing between the different Claude Code chats of projects and keeping the ball moving forward on each one.

Claude Code makes it seem like it's not for marketers or anyone who's not a programmer.

But that's the furthest thing from the truth.

For a programmer who might not be that familiar with marketing, they can use these skills to essentially allow Claude Code to become their CMO.

In the same way, if you're a marketer, there are skills that turn Claude Code into your CTO.

I don't think there's any limitation whatsoever.

Anybody can both build an app that is production grade and market it with zero experience in either.

I use Claude Code for writing and research tasks all the time and it's a game changer.

People are sleeping on it because they think it's only for technical people.

Claude Code might be the most important piece of software in years and is the blueprint for the next $100B+ in software opportunities.

It took the single most terrifying piece of software ever created, the terminal, and made it 10x more friendly.

My favorite tools right now are Claude Cowork and Claude Code.

I'm literally maxing out my limits every day on the $200 a month plan.

A lot of the work that we needed to do previously was client-facing stuff.

But now we've been able to build tools with Claude Code that replace the previous work that we needed to do with the clients.

I have a friend who has a video editor who they tasked with the outcome of 10Xing the number of videos they're putting out.

The video editor built a video editing tool with Claude Cowork.

So they built a tool to do the job.

I dropped my brand guidelines into Claude Cowork and I said, Claude Cowork, I have a problem.

And I just yapped.

I don't know how to do design.

I need you to help me.

Just control my screen.

I don't care what you do.

Fix it for me.

It got a 47% average opt-in rate.

That's insane.

That was more than 2,200 signups.

The show-up rate was 32%.

Because it thought and realized that it should send calendar invites one by one to everybody.

It also added a special WhatsApp message.

And if the WhatsApp didn't go through, it would send them an SMS.

Claude figured out how to add the MCP to GoHighLevel that I'd been trying to do for two weeks by myself.

I never thought just to say just figure it out by yourself.

I used to be paying this girl $5,000 a month.

Now I'm literally saving myself 60K a year paying a couple hundred bucks a month to Claude Cowork.

Locked in with Claude Code rebuilding half the backend of how we create content.

Automating workflows that used to take the team hours.

Ripping out friction everywhere I find it.

Claude Code is going to take everything you've written and go and execute it for you.

This feels like magic to me.

What a time to be alive.

This is how I organize my entire life with just one document.

I use Claude Code to look at every single thing I've built over the last week, and I put it in a PDF.

Then I mark it up on an iPad with my notes on what's going well, what's not going well, who I need to hire, what we can automate, what we can delegate, all that good stuff.

I upload my weekly reflection document back to Claude Code, enter plan mode, and get it to execute all the opportunities I identified.

A beautiful way of reflecting at the end of the week with an iced americano, and a great way to get a lot of shit done.

Claude is now writing Claude.

If you were to just build it inside of Claude, there's zero setup friction.

You can build things so much faster.

You don't have to think about database structure.

It's perfect for solo operators.

The development costs for some of the things you can do on a $200 a month Claude plan would be 10X what Claude actually charges you.

If you're building just inside of Claude, then you're probably going to be building your solopreneur type of business.

Which is nothing wrong with that, but you just need to understand that.

The main thing with skills is that you have to give it quality inputs and quality examples.

That's the problem with people diving into skills too quickly and downloading different skills from everywhere.

You're leaving your business up to everybody else.

Find examples from other people as a starting point, but take the time to go through and add your own examples.

Adapt it to your business so that you can get the correct outcomes.

GPT stacking is when you use the at symbol, you can actually bring multiple GPTs into one chat.

So you could have a whole collection of mentors solving one problem together by using this at symbol.

We had 154 GPTs and Claude turned every single one into a skill.

The skill maker is even more powerful.

And what I love about what Claude does is that it will go through and test it out for you and then come back and do feedback.

I asked Claude, I said, give me a list of tasks that you could do while I sleep.

And it went through these 154 skills that I had given it and gave me specific tasks related to my business.

You can take that custom GPT and put it into Claude and tell it, hey, make this a skill.

I had a team member go into ChatGPT and just back them all up.

And then I just told Claude, hey, go to this spreadsheet and turn all these into skills.

I find that I don't actually need things like n8n all that much anymore because Claude and Perplexity Computer can handle it.

I really think that the most valuable subscription you can have is the Max plan of Claude.

Claude Code is the closest thing creators have ever had to a co-founder.

It doesn't just answer questions. It builds the things that used to take an entire team.

Learning Claude Code and building your own AI Operating System (AI OS) in 2026 is like buying Bitcoin in 2016 or 2017.

Don't miss out on this once-in-a-lifetime opportunity (again).

Before I would say it was almost entirely n8n, and now it's about 50-50 Claude Code and n8n.

Even for n8n, I don't even touch it.

I use the Claude Code MCP n8n tool, and it just creates and fixes all the automations.

It can even check the execution logs, figure out why an automation broke in that instance, and then try to fix it and patch it up.

I can't use the desktop app after being in Claude Code.

It feels very limiting and then the conversation just stops.

I really encourage even non-technical folks to dive into Claude Code.

It's pretty awesome once you get over the initial learning curve.

My process is literally I'll just film the 20 TikTok videos, airdrop them to my MacBook and then I'll do slash cross post.

That knows to find all the videos I just downloaded today.

It transcribes all of them and then it writes platform specific captions for each one.

I look it over and then it will schedule them all out for the week.

You can build the skills really customized to your use case in business.

Every single one of these is highly customized to my actual workflow.

I use Notion for a really long time, but it does not compare at all.

It's so much faster for me to build all of this in Claude Code.

I can drop anything into Claude Code and just be like, update the corresponding templates, and it just knows what to do, and it does it.

My process is very much AI assisted coding.

Using Claude Code to help me code, but it's not AI just coding it for me and I don't check what's going on.

I have thousands and thousands of users now.

The cool thing about the philosophy MD file is I continuously update it.

I have a skill called learn that tells Claude Code to reflect on the conversation and figure out how to update my philosophy file.

They're gonna walk away awe-inspired.

We are gonna show people what is actually possible with Claude Code with agentic workflows.

And we're gonna show people why most people are wasting their time with AI.

I got started with vibe coding long before it was called vibe coding.

I got started in 2022, right after ChatGPT launched.

I was lucky enough to actually see Sam Altman speak at Burning Man like a month and a half before ChatGPT launched.

There was like 70 people there.

Nobody saw this thing coming.

That logic led to people saying, oh, these things are actually pretty good at writing code too and debugging code.

And that immediately unlocked it for me.

I spent the next six months just copy and pasting stuff back and forth between ChatGPT and VS Code.

All Cursor was doing was instead of me having to go back and forth and copy and paste these big files, it just did that automatically.

I could just tag the file.

All of a sudden we created a feedback loop.

At the end of the day, all we were really doing still is copy and pasting stuff around for a big prompt to put into ChatGPT.

By kind of living through this evolution of ChatGPT becoming vibe coding, I look at everything that's happening and understand it in that way still.

I'm still just thinking what kind of context do I want in this message.

This is context engineering.

That's still the key.

A lot of people that get started with this go out and they download a bunch of skills and they just start poisoning their context.

They start degrading the ability of the LLM to actually know what you want, know what your business is about, know what challenges you have in your life.

The more context you give it, the less of that context window is what you actually want.

People get bad outputs and they're like, I don't get Claude Code.

And it's like, well, remember, all we're doing is copy and pasting stuff.

So what you've done is you've copy and pasted 80,000 words and then you said, hey, let's brainstorm some business ideas.

Of course you're not getting a good output.

Don't just go dump a bunch of skills into your system.

Don't just go install a bunch of plugins.

It actually is gonna really hurt your outputs.

All we're trying to do all the time is context engineering.

Context engineering is how these systems become magic.

And the trick with context engineering is generally to keep the context as low as possible.

There's a concept called the lost in the middle phenomenon.

This is basically the underlying reason why LLM performance degrades at around 60 to 80,000 tokens.

More recent models are a little better at this, but for the most part, at 128,000 tokens, the model performance is about half as good as a fresh context window.

We're gonna be starting new chats a lot.

We don't wanna just run the same chat forever because it's gonna start to lose its mind, just like a human.

If you have too much context up here, you just start forgetting stuff.

You're ultimately going to have to find some balance between convenience and performance.

For the most part, you do want to set everything up locally inside of the project.

We're always gonna err towards in this project, because that's gonna make sure. if I'm talking about my marketing SaaS business, I don't even want it to know that my info product business exists.

If you can make it through installing Claude Code, you're 80% of the way there, because that's one of the toughest parts.

Once you get Claude Code installed, you can just tell it that you're having a problem and it'll go fix it.

It's called Claude Code because it can write code.

It can run commands.

And by being able to run commands at the root level of your computer or any computer, we have the ability to control the digital world.

Just because it's called Claude Code and just because it runs in the terminal, it's still the same thing.

All of these platforms are just wrappers around hitting enter on ChatGPT's site.

One way to send an API call, one way to send a message to OpenAI servers or Claude servers, is to type it in ChatGPT and click enter.

That's a call.

Claude Code, Codex, it's all the same thing.

It's just different wrappers.

The terminal-based systems are the best because it can really quickly go search all the files on your computer.

It can really quickly go search every report that's ever been done in Google Analytics.

It can go pull every single server log that's ever happened on your website.

The ability to pull all of that data instantaneously used to be unheard of, unless you were a top tier developer.

But we don't need to be a top tier developer now.

We have one. We all have a top tier developer.

You may start with Claude and Claude.ai and eventually move to Claude Cowork and then go to Claude Desktop.

That's fine.

I personally think even that is a waste of time.

After you're done doing that setting, you can say, hey, go look at every other setting and make sure that everything's like top 1%.

We're one of the best brands in the world, and we need to act like it.

Google Cloud is the worst UI I've ever seen in my life.

Everyone hates navigating Google.

Now it's just code.

For every Claude Code Max subscription, the 200 dollar a month subscription, if someone fully spends that, they're spending about 5,000 dollars in compute.

And that translates to about 12,000 to 15,000 dollars in API credits.

I usually find every month I'm spending about 12,000 to 15,000 of API credits through Claude.

I'm not spending that.

They're losing a ton of money on me.

And they're not going to do that forever.

Claude Code pops up on my phone every time I pay it, and it's the only charge that I get happy when I see 200 bucks a month going out.

Because I am so happy to pay that.

It's the only thing I've ever seen pop up that I'm like, yes, keep going.

My goal right now is to build as many reliable agentic workflows as I can so that when they finally decide to stop losing so much money on this, my agents will all be good enough that they can run on lower cost models.

The more reliable your agents are, the better those prompts and the better the system is, the worse model can run it.

The models are more of a commodity than people have realized.

What's the differentiator is everything around it.

All we're doing at the end of the day is copy and pasting stuff into ChatGPT.

Claude Code is, they call it a harness, right?

An AI harness.

The harness is the product now.

Whoever has the best harness really has the best model because that's what's going to give people magical performance.

Claude Code is where it's at right now.

I have the 200 dollar a month ChatGPT subscription as well because I max out the Claude subscription every month.

So I try them both and it's just not there.

Claude is such, there's such a culture around Claude and it's just where it's at.

I don't know if OpenAI is ever going to be able to catch up.

Codex is more enterprise-focused.

It catches a lot of things that Claude Code doesn't.

Claude Code is better for a solopreneur trying to build something.

Codex is better reviewing than writing code.

When Cursor came out, they didn't just build a whole coding system from scratch.

They used VS Code.

They forked VS Code and they made an AI-first version of VS Code.

When we say VS Code, when we say Cursor, when we say Anti-Gravity, these are all the same thing.

They're all just VS Code with different little modifications.

Anti-Gravity is Google's version of VS Code.

VS Code itself is from Microsoft.

And Cursor is Cursor's version of VS Code.

We can go into VS Code and we can open up a new terminal.

That is literally the exact same thing, but now that terminal window is just running as a part of VS Code.

And that's a really amazing way to run Claude Code.

The wrong way to use AI in VS Code is with extensions.

Your first gut reaction is going to be extensions.

I can run Claude Code here.

It is NOT the same as this.

Not the same.

They even say, hey, do you prefer the terminal experience?

Because you prefer the terminal experience.

Eventually we want to end up in the terminal.

We want to get to a point where this isn't intimidating and we can just do everything here.

That's going to be where we ultimately want to live.

It's just a much more scalable way to do it.

All the different plugins that we're going to talk about, all the different workflows and the really cool stuff in Claude Code that makes it really powerful and really magical, literally just does not work with extensions.

It doesn't work with Codex.

Claude has such a better ecosystem and such a better culture around it.

It is just a better harness.

When you're telling Claude Code, go set up this skill, go set up this MCP, it's generally going to ask you, do you want this set up globally, like for every project ever, or do you want this just in this project?

We're always going to err towards in this project.

I think Anthropic is thinking about this at an enterprise level.

They're thinking, okay, we need a team of 100 developers to work on this project.

In that way they want a lot of the Claude settings and MCPs and skills in the project, both for the context reason and that's like the general business best practices.

We're going to build a context about your business, about your life.

A life operating system that's really lean.

It just knows about you, your goals, your problems.

And it has the ability to go grab information from your Apple Health or whatever.

That's the magic of AI. when we get that set up.

When we get deep enough into Claude Code, you will start to really be disgusted by having to even log into a website and click around on a website.

AI becomes the implementation layer itself and the scaffolding and the rails that run your whole life, honestly.

I started with this putting my businesses into AI and making my businesses AI native.

Eventually I said, I gotta make my whole life AI native.

Open up a new window, type in slash context.

It's going to show you not only how much of your context window is filled with a brand new chat, but you can go through and see what is actually filling it up.

Down to each and every skill or each and every plugin, exactly how many tokens it's using.

Anytime I haven't hit my usage limit yet and it's about to expire today or something, that is generally one of the things I'll do.

I'll go through and. keep in mind you have to do this for different projects.

The Claude.md and the memory. these things tend to get really big and useless.

Big things that you constantly have to go back and kind of trim down.

I'm gonna continue to use dangerously skip permissions.

If you're new, don't do that yet.

Eventually, I had so many different Claude Code windows running, I would find one that I did like an hour ago and I would open it up and it would have never run because it was waiting for some stupid command.

If you did dangerously skip permissions and then go try to speed up your computer, that's where you hear the horror stories of, hey, AI just deleted my whole operating system.

So be careful when you're starting out.

Going through the process of approving and denying commands. if you don't know what a command is, you open up a new Claude Code window, you copy and paste and say, what is this?

Do I want to run it?

What is it going to do?

There's like five commands that you're going to end up knowing, and it's everything you need to know.

MCPs are actually your least favorite way to connect to tools.

MCPs eat up your context window.

They have a very defined set of actions, whereas an API just has everything.

If it's a big company like GitHub, a lot of people put the GitHub MCP on their Claude Code.

No reason to do that.

I promise you the model is trained already on how to use GitHub.

It doesn't need an MCP.

In that case, I'm generally going CLI.

CLI is my preferred way to do it.

API would be my second way to do it.

MCP is my kind of last-ditch effort way to do it.

All an MCP is is essentially just a standardized wrapper around an API.

It's really just a way for AI to understand how to use an API.

But the problem is sometimes many companies don't give you all of the commands in the MCP.

The API will actually have a lot more capability, but the MCP just doesn't expose those actions.

Every MCP that you run is taking up RAM.

And every single Claude Code window you open is launching another one of those MCP servers.

Today, I just had to turn off a bunch of MCPs because I got to a point where I couldn't run more than like nine Claude Code windows.

To run the GitHub CLI, it's like gh push.

That's it.

Two tokens, six letters.

Now we've pushed the whole thing.

If Claude Code is able to use the CLI, I'm always going to use the CLI first.

What dictates if a tool is going to work with CLI or not is how much training data there is in the model.

When the model was trained, was there a bunch of GitHub CLI tool usage in the training data?

If so, we don't even need to tell the model about any of this.

The model just knows.

One of the first commands that I built is my morning health command.

It literally just goes and looks at every single thing I've been able to think of over the course of a year and a half.

So it goes out and looks at all my logs.

It goes out and looks at Supabase, the database logs.

Every single logging thing I can possibly think of.

It goes and looks at that and then it just triages and figures out if there's any problems.

I've had websites down for a month that I haven't known about, you know?

So it's incredible.

I have an agentic system that looks at my incoming emails for people that want link insertions or guest posts and then negotiates with them and sends them a PayPal invoice and basically pings me when I have to.

Over time, the system needs to realize when it's already replied to someone.

So it created this little SQL database for itself.

I don't even know how to access it.

So don't be intimidated.

I have a human in the loop on this system right now.

There is a bunch of the emails that auto-send, but for anything that it has below a certain confidence percentage, it makes me approve it.

Over time, I'm just going to slowly turn this over to the agent itself.

I have a prompt I usually send that says: now I wanna zoom out and do a 30,000 foot meta analysis.

Let's see how this agentic workflow works and how we can get it closer to full autonomy.

What went right, what went wrong in that last run?

Go look at the chat we just had.

While we're on this presentation, don't just watch it, don't take notes.

Open up the terminal, go try to install Claude Code, go work through it.

Start small.

Tools, little internal calculators and stuff like that.

Don't try to do the app right off the bat.

Don't try to launch a SaaS app or an iPhone app right off the bat.

Building with AI is like building with Legos.

You want to start one Lego at a time.

Can I get my Claude Code to talk to my router in my house?

Okay, now I have it talking to my router.

Maybe I can set up an agent that looks at the router.

It all kind of builds on top of itself.

There's this upper class of tech people that can control the entire internet through writing commands.

We want to do that too.

And we can do that faster than they could ever dream.

I'll talk to people that have been in the computing industry for 50 years that watch Claude Code run these massive numbers of commands back to back to back, and their jaw will drop and be like, do you know how long it used to take me to just type that in?

Half of my day was just typing in these commands.

Claude Code is, I think, over 90%. call it 95%. written by Claude Code now.

Even coders aren't writing code anymore.

No one's writing code.

I eventually got to a point where the main productivity holdback for me was dealing with my Claude Code setup on different machines in different environments.

I was so sick of this.

I literally want to go back to the 50s and have a mainframe.

I basically said, I want to have my main home computer as my Claude Code machine.

I want to figure out how to log into that from my phone and run Claude Code on my computer so that all my authorizations, all the cookies, every little thing is just exactly as it is if I'm sitting at my computer.

The words I'm looking for are Tailscale and Termius.

The 80-20 of Claude Code: make sure you're token maxing however it works for you.

From there, work towards Claude Code and terminal-based AI harnesses.

Then install one of these workflow pipeline tools. compound engineering, GSD, GStack.

Now you've adopted a true top 1% power user way to use this tool.

You will get so far and just skip so much of the learning process if you just get one of these workflow pipeline tools installed inside of Claude Code.

I was not the type of person to really dive into the terminal a year and a half ago.

I knew a little bit about the terminal.

I knew I could ping a website and see if it was up.

I did like three commands in terminal.

And this is where I'm at now.

The system will teach you how to use it.

The key is just get Claude Code installed, and it will help you from there.

Last time we went from zero to Claude Code hero, we went from zero to Token Maxer.

We basically got the philosophy set up.

In this session we are going to show you those concepts actually applied to some things.

Compound engineering, GSD, G-Stack. these are all different plugins that we're going to put into Claude Code.

When we go to build things with Claude Code, they're going to deploy hundreds of different best practices.

And we're basically going to become top 1 percent users just by default.

I try to start with low-hanging fruit.

I try to start with stuff that I knew I should be doing, but never did because it just was low-value work.

To build confidence, I always talk about building with AI is like building with Legos.

One Lego at a time.

You start really simple and then all of a sudden after a few hours of work, you zoom out and you see this beautiful thing that you're building.

Claude Code becomes very powerful when you give it the correct feedback loops.

Hey, go take a screenshot of the site as it is right now, change the color of the button, and then go take a screenshot of the site again and make sure you didn't screw anything up.

Those types of feedback loops are what's going to separate you from truly agentic set it and forget it workflows versus you sitting there chatting with AI all day.

When you realize that your job is pretty much just running the same prompt every day, now you're like, okay, why am I even in the loop?

I want this to be a fully agentic process.

And that's. you've built your first AI agent.

My first AI agent was very simple.

My first iteration of this morning health routine was literally like the prompt that I kept on sending.

It was like two lines.

And then I slowly built up and changed it.

To tell you the truth, I actually don't even know how this is built anymore.

AI has iterated beyond my expertise, of course.

How can we make this system better?

Go use web search tools extensively to identify the current cutting edge best practices in 2026 for the top 1 percent Claude Code power users.

I say it that quickly because I say it so often to AI.

This system now has state management built in.

I don't know how it did it.

It just told me that it wanted to build state management.

You have to know how to use Claude Code to make sure it doesn't just create a bunch of garbage.

You got to make sure you're really fighting back.

One of the great kind of power moves that you can build into your skills. this is a serious power move.

You are going to have a main agent that is doing the coordination of the work.

This is called your agent orchestrator.

If I go do slash context here, the messages have about 23.5 thousand tokens.

Just these sub-agents did over 25,000 tokens alone.

So how am I at 24,000 tokens?

That's the magic of the agent orchestrator using sub-agents to do its work.

My main agent orchestrator still has a very low context usage.

If I go and have one agent, one chat, just go do all this work, it fills up the whole context window, it starts to have to compact, and it just loses its mind.

All that's happening with this is, all that this did was it opened up Claude and it said slash sb-health prod.

It said that, and then it gave it that text.

That is quite literally what this is.

Anthropic's on a mission to bring Claude Code features to their other tools.

But ultimately Claude Code is it.

That's the power move.

And everything else is trying to emulate some of the magic you get with Claude Code.

I can't believe there's still people logging into Google Search Console and Ahrefs to see what is wrong on their website.

That just gets fed into Claude Code and it goes and fixes it.

It'll do 100 pages in 10 minutes.

You could not take this skill and drop it into your business because it's made for my business.

It literally talks about my business nonstop.

This is a great example of you can't go download a plugin to become great at AI.

These skills have been built by me, Lego by Lego.

And that's how you build with AI.

You could not drop this in.

The right way to do this would be to take my skill and put it into your Claude Code and say, hey, this is interesting.

Let's build something like it for me.

That's how you win in this game.

The people going out there and just downloading a bunch of skills or, God forbid, buying skills. no.

I can take this workflow even further if I give it little thresholds.

I tell it, hey, I actually don't even want to think about managing the server resources.

If the hard drive fills up, go buy more hard drive space.

I don't want to be involved.

Everyone knows that you should scale your subscriptions up and down based on how big your site is.

No one actually does it.

But with AI, I don't even know this stuff happens now.

Every day it just runs and it will literally rebuild the server resources.

The goal is to eventually get a lot of these systems to full autonomy.

The morning health system is my most reliable AI employee, essentially.

That guy works like a workhorse, never fails.

I'm trying to get all of my agents to that level because I think this gravy train doesn't run forever.

I think eventually these models are gonna be quite expensive.

And when that happens, I want my agents to be so reliable that they don't need the most expensive model.

The agent you see is so unbelievably specific.

It's not go check the database.

It's go run this command to look at MariaDB.

The more specific your agents get, the more reliable they're going to get, the less they can get confused about what you mean.

I can have someone join or purchase a product of mine and I can literally open it up and be like, hey, reconstruct this entire purchasing journey.

Because AI has set up my marketing attribution and tracking so well, I'm able to usually pull together a 20 to 30 touchpoint list of what everyone has actually done on my website leading up to the purchase.

I have a marketing insights command that goes out and looks at every single marketing source of information.

It has the sub-agents that go out and analyze Google Analytics and PostHog and Google Search Console and Ahrefs and essentially creates a marketing report.

I built my own content management system for my members area because KeyStatic and WordPress wouldn't do what I wanted, so I just rebuilt it.

And it's awesome now because I can just create Markdown files on my computer and they automatically become properly formatted on my actual website for my members.

The first thing you do with MakerKit is basically pull it down from GitHub.

With that demo site comes all the authentication, all of the dev server stuff, billing, Stripe integration, documentation center, multi-tenancy.

All of that stuff is kind of like the heavy lifting work of building any enterprise SaaS.

Almost everything I've built that's actually making good amounts of money and have real paying customers flowing through it is all MakerKit.

I could in theory build my own entire platform.

But again, it's just not worth it.

One of the things I did was I moved off of WP Engine with Claude Code onto a Hetzner server, which is literally a server in Germany, like a computer that I rent.

Cloudflare then serves it off of there.

It is not. it's like 20 percent of what I paid WP Engine.

And my site now is so much faster than it was ever on WP Engine.

I paid 400 dollars a month for a dedicated server, standalone server from them, and it was trash.

It was so bad.

And they charged me so much money.

And yeah, me, like some random server in Germany that I pay like 9 dollars a month for is so vastly superior.

I thought it was going to be so difficult for me to do the migration off of WP Engine onto my own infrastructure.

It's legitimately like, you'd have to hire a 200,000 dollar a year DevOps architect to do that.

So expensive.

This is all custom made in Stripe.

I custom made this entire thing.

It knows exactly how many spots are left at each tier.

All of this is dynamic.

I got split pay with Klarna, crypto payments.

I got everything.

Did Claude Code build this?

Kind of.

It put the components on the page for me.

But ultimately it is me going through and saying, huh, I kind of want this far down the page.

Maybe I'll put a grid in here.

I don't like that.

I wanted this little CRT effect on the text.

I went out and said, hey, I've never seen this before.

I want the CRT effect.

At one point, this page had 600 different forms of this logo with different animations on it.

I was able to just go through and slowly whittle down and be like, that is the exact vision I had when I said I wanted distortion.

Go with version 492.

I definitely think frontend is harder than backend.

I've built incredible actual systems that produce incredible value.

To get a pretty landing page though is still. I would definitely only use AI to do it.

If you're getting into this and you want to spend extra time on security, the concept that you're looking for to ask Claude Code about is secret management.

These are API secrets, secret keys, and there's a lot of different ways to manage them.

I think I want to say Facebook invented React as a solution to how bad all of the programming languages were.

It's literally made by Big Tech for Big Tech.

This is modern design.

WordPress is a PHP platform that's just not. I don't know the details as to why, but you can't get it to look this pretty easily at all.