The argument in one line.
Claude Code with human-in-the-loop iteration beats automated workflows because the system learns your preferences and catches its own mistakes, turning a 20-hour weekly creative bottleneck into a scalable command center that improves every time you correct it.
Read if. Skip if.
READ IF YOU ARE…
- A content creator or solopreneur with no coding experience who spends 10+ hours weekly on repetitive tasks like data collection, formatting, or content scoring.
- Someone curious about Claude Code but intimidated by terminals and technical setup who wants to see a real build from zero to working system.
- A creator already using AI tools who wants to automate specific workflows (like virality scoring or content management) without hiring a developer.
SKIP IF…
- You're a software engineer or experienced programmer looking for advanced Claude Code patterns — this focuses entirely on overcoming beginner fear and mental blocks.
- You work primarily in fiction, narrative writing, or domains where code-based automation doesn't apply to your core workflow.
The full version, fast.
Claude Code becomes a personal command center the moment you stop treating the terminal as developer-only territory and start treating it as a workspace that reads your files, remembers your preferences in a CLAUDE.md knowledge base, and learns your patterns through corrective feedback. The method shown here builds a short-form content engine by wiring Claude Code to the Apify MCP for Instagram scraping, storing posts and transcripts in a free local SQLite database, and ranking hooks with a custom virality score weighing curiosity, emotion, brevity, relevance, uniqueness, and platform fit. The payoff: human-in-the-loop iteration beats fire-and-forget Ralph loops, planning files with checkboxes preserve context across sessions, and saving the polished UI for last keeps you tuning the system that actually moves the needle.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:29
01 · Introduction & Why Claude Code
Brock admits months of avoidance; positions Mark as expert guide; promises beginner-friendly build.
01:29 – 06:27
02 · Claude Co-work vs Claude Code
Mark frames Co-work as gateway drug (30-40% of firepower). Co-work was vibe-coded by Anthropic in a week, full of bugs. Real tool is still the terminal.
06:27 – 07:57
03 · Biggest Misconceptions About Claude Code
The name misleads. Real unlock = file access + local database + desktop. Limit is imagination, not technical skill.
07:57 – 11:16
04 · Non-Technical Success Stories
70-year-old built a mobile app in 3-4 months. The feedback loop (ask Claude Code to fix Claude Code) is the unlock moment.
11:16 – 14:32
05 · The Power of Going Direct to the Source
Mark canceled $1,200/month in SaaS after going terminal-first. Works in Cursor, Windsurf, OpenCode too.
14:32 – 18:37
06 · The Build: Automating Short-Form Content
Brock shows a 825K-view short. Mark designs: one command (make me viral) -> scrape -> 20 hooks -> rank -> top 3 + captions.
18:37 – 25:07
07 · Architecture: Apify + SQLite + Virality Score
Apify MCP for scraping. SQLite free local database. Three tables: creators, posts, hooks. Virality score: curiosity 25%, emotion 20%, relevance 20%, brevity 15%, uniqueness 10%, platform fit 10%.
25:07 – 31:43
08 · Installing & Configuring MCP Servers
Live walkthrough: mcp.apify.com -> one-line command -> API key -> restart -> verify with mcp status. Transcript actor must run one post at a time.
31:43 – 34:31
09 · Testing Scrapers & Building the Knowledge Base
CLAUDE.md as running memory. Every session discovery committed so Claude never re-learns mistakes.
34:31 – 36:01
10 · Encountering & Fixing a Bug
System stored captions in transcript column, trained on comment-X-get-Y patterns. Mark caught it, Claude purged bad data and re-scraped.
36:01 – 42:14
11 · Virality Score System Live
MAKEMEVIRAL command: scrape -> score all posts -> 20 hooks -> rank -> output top 3 + captions. 80% real hooks; fix path shown.
42:14 – 47:26
12 · Future Improvements & Final Thoughts
Build process before UI. Dual-adviser technique: Cursor + different model for fresh eyes. If terminal is blocker, start with Co-work.
Lines worth screenshotting.
- Claude Code reads your files, remembers your preferences, and learns how you work — which means every session builds on the last one rather than starting from zero.
- The CLAUDE.md file is the behavioral contract between you and Claude Code: every preference, every rule, every recurring task trigger lives there, and Claude executes against it automatically on startup.
- Terminal fear is the primary barrier for non-technical users, and the barrier is almost entirely psychological — once a non-technical creator opens a terminal and watches Claude do the work, the fear collapses.
- An Apify MCP connection lets Claude Code scrape live data from TikTok, YouTube, and Instagram directly inside a session, without any manual data export or copy-paste step.
- Storing scraped viral hooks in a SQLite database inside the project folder means the system accumulates data over time — each run adds to a growing dataset the model can analyze for patterns.
- A custom virality scoring algorithm embedded in a command lets you query your entire hook database and rank by your own criteria rather than using a platform's algorithm as a black box proxy.
- Building the system live, with bugs visible, is the honest demonstration that shows non-technical users what iterating with Claude Code actually looks like — not a perfect demo, but a realistic build session.
- The gap between 'Claude gives you answers' and 'Claude builds things and saves them to your computer' is the conceptual shift that separates people who use it as a chatbot from people who use it as a build environment.
- A short-form content command center that used to take 20 hours per week can be reduced to a triggered command — the hours don't come back gradually, they collapse on the first successful run.
- Mark Kashef debunking the misconception that Claude Code requires programming knowledge is important because the misconception is self-enforcing: people who believe it never start, which confirms their belief that it wasn't for them.
- MCP tools extend Claude Code's native capabilities the way plugins extend a browser — a single installed MCP changes what kinds of live-data tasks are possible within the same conversation.
- The virality scoring algorithm is only as good as the criteria the user defines for it — which means the skill required is not coding but articulating what makes a hook work in your specific niche.
- A beginner going from curious to owning an automation system in 47 minutes is not a trick of editing or selective demo — it is what happens when the tool handles all implementation and the human handles all decision-making.
- Bugs appearing and being fixed live during the build is the anti-glamour demonstration that earns trust: the audience sees that the system works in spite of imperfection, not because everything went smoothly.
- Storing content research outputs in a local database rather than a SaaS tool means you own the accumulated intelligence — it doesn't disappear if a subscription lapses or a platform changes its API.
Claude Code's Real Power Is File Access, Not Coding
Claude Code for non-techies
The name misleads — Claude Code's actual unlock is reading your files, remembering your preferences, and accessing a local database, which makes it useful for anyone, not just developers.
01Introduction & Why Claude Code
- Months of avoidance ended by one weekend build — the pattern interrupt is honest resistance followed by conversion
02Claude Co-work vs Claude Code
- Co-work is 30-40% of the firepower — the terminal is the actual tool, Co-work is the onramp
- Co-work was vibe-coded by Anthropic in a week — knowing the product history explains the UX gaps
03Biggest Misconceptions About Claude Code
- The name misleads — file access, local database, and desktop control are the real unlocks, not code generation
- The limit is imagination, not technical skill — the terminal is the interface, the capability is general-purpose
04Non-Technical Success Stories
- A 70-year-old built a mobile app in three to four months — experience is irrelevant when the feedback loop handles the debugging
- Asking Claude Code to fix Claude Code is the unlock moment — the tool maintains itself
05The Power of Going Direct to the Source
- One developer canceled $1,200 per month in SaaS after switching to terminal-first development
- Works in Cursor, Windsurf, and OpenCode — the terminal pattern is tool-agnostic
06The Build: Automating Short-Form Content
- One command triggers: scrape, generate 20 hooks, rank by virality score, output top 3 with captions
- The automation compresses what used to take 20 hours per week into a single command
07Architecture: Apify + SQLite + Virality Score
- Apify MCP for scraping, SQLite as free local database, three tables: creators, posts, hooks
- Virality score formula: curiosity 25%, emotion 20%, relevance 20%, brevity 15%, uniqueness 10%, platform fit 10%
09Testing Scrapers & Building the Knowledge Base
- CLAUDE.md as running memory — every session discovery committed so Claude never re-learns the same mistake
- The memory file is the compounding asset — the system gets smarter every session without additional setup
12Future Improvements & Final Thoughts
- Build process before UI — prove the automation works in the terminal before adding any interface
- Dual-adviser technique: run two models simultaneously with different perspectives to catch what one misses
Terms worth knowing.
- Claude Code
- Anthropic's command-line coding assistant that runs in a terminal, reads and edits files on your computer, and can execute multi-step tasks like building apps, scraping data, or managing folders.
- Claude Cowork
- A newer mode inside the Claude desktop app that wraps common Claude Code workflows in one-click buttons, aimed at non-technical users who want agentic automation without touching a terminal.
- Terminal
- A text-based interface for issuing commands to your computer. Developer tools like Claude Code run inside it instead of a graphical app.
- IDE
- Integrated Development Environment — an editor like Cursor, VS Code, or Antigravity where you write code and often run an embedded terminal.
- Vibe coding
- Building software by describing what you want in plain language and letting an AI generate the code, rather than writing it line by line yourself.
- MCP
- Model Context Protocol — a standard that lets AI assistants connect to outside tools and data sources, such as databases, scrapers, or APIs, through a shared interface.
- MCP server
- A service that exposes a specific tool or data source to an AI assistant over MCP, letting the assistant call its functions during a session.
- Apify
- A marketplace of pre-built web scrapers, called actors, that you rent on a pay-as-you-go basis to pull data from sites like Instagram, TikTok, or LinkedIn.
- Actor
- Apify's term for a single hosted scraper or automation script that you can run on demand against a target site or API.
- API key
- A secret string that identifies and authorizes your account when a program calls a third-party service like Apify, Gemini, or OpenAI.
- SQLite
- A lightweight database that runs as a single file on your own computer, with no server to set up. Useful for free local prototyping before moving to a hosted database.
- Supabase
- A hosted Postgres database service often used as the backend for vibe-coded apps. Charges per project, which is why a local SQLite file is cheaper for experiments.
- CLAUDE.md
- A plain-text memory file Claude Code reads at the start of each session. It stores project conventions, learned lessons, and custom commands so the assistant doesn't relearn the same things every time.
- Context window
- The maximum amount of text a language model can hold in working memory at once, measured in tokens. Once it fills, older content gets dropped or the session must restart.
- Token
- The chunks of text a language model reads and writes — roughly a short word or piece of one. Model pricing and context limits are both counted in tokens.
- Ralph loop
- A workflow where you hand an AI agent a queue of tasks and let it grind through them unattended. The trade-off is that without a human checking outputs, low-quality results can pile up unnoticed.
- Human in the loop
- A workflow where a person reviews and corrects an AI's output at key checkpoints instead of letting it run fully autonomously. Trades raw speed for higher-quality results.
- Warp
- A modern terminal app with built-in AI that can diagnose and fix command-line errors on your behalf, making installs like Claude Code more approachable for non-developers.
- Cursor
- An AI-first code editor built on VS Code that lets you chat with multiple models about your codebase and run agents inside it.
- Antigravity
- Google's AI-native development environment, positioned as an alternative to Cursor for working with models like Gemini directly inside an editor.
Things they pointed at.
Lines you could clip.
14:00
“Human in the loop plus Claude Code is the most OP combination.”
10:45
“kinda like waking up from a dream. You look around at 90% of AI tools and you realize you do not need them.”
11:10
“I canceled $1,200 a month worth of subscriptions because I literally realized I did not need it anymore.”
20:00
“You have one input and you have not infinite, but many outputs as a result of that one -- that is the very definition of leverage.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00I've been avoiding Claude code for months. It seemed too technical and too complicated. However, my friend Mark has been begging me to start using it.
00:07He is hands down the best person to learn Claude code from. So I brought him on this episode to teach me how to use it from scratch. And I'm gonna be honest, after this episode, I went down the Claude code rabbit hole and spent the entire weekend building a system that literally used to take me twenty hours every single week, and I'm using it right now.
00:23Here's what nobody told me about Claude code. It reads your files, it remembers your preferences, and it learns how you work. So the more you use it, the smarter it gets.
00:31And it doesn't just give you answers. It actually builds things for you and saves them to your computer. So by the end of this episode, you're gonna know exactly how to set this up, start using it yourself even if you've never written a single line of code, and begin automating all aspects of your work.
00:44Believe me, if you've been meaning to try Claude code and start using it yourself but it seemed too technical, this is absolutely the best video you should watch on how to get started. So with that being said, let's get into it. Alright.
00:55What is going on, guys? I'm stoked for this episode. Personally, I've been wanting to use Claude Code for quite some time now.
01:02So I wanna bring on the goat of Claude Code in my eyes, Mark. So, yeah, I'm hoping today we could, a, get me up to speed using Claude Code, get my audience up to speed, and build something cool in this podcast. So, yeah, happy to have you, Mark.
01:13Yeah. Of course. Uh, it's a pleasure to be here.
01:14I know it's been a while since we last touched base. We were probably
01:17more in the vibe code world before, and now we're getting into the real code y world now. So that's good. Yeah.
01:24100%.
01:25So I'm stoked to see what we could build with Cloud Code. I've never actually been able to build anything with it. So to kick start this, I wanna ask one question, and that is Claude Cowork just dropped literally yesterday before we were gonna film this podcast.
01:37So for my audience, what do you think? Is Claude Cowork something that replaces a need to learn Claude code? Do you think it's like an entry, like, gateway drug thing?
01:46Like, give me your rundown on that. Because I like to use you as a filter for AI tools to see, like, should I pay attention to this? Is it overrated?
01:52What not? So give me your rundown. Yeah.
01:54For sure. You stole my description, first of all. So let's make that disclaimer.
01:57Uh, it it is the . Up until now, the average person looks at a terminal, and no matter how many times someone like me shouts and screams and cries,
02:06we can't convince people to hop in a terminal because there's, like, this mental block that terminal equals code, code equals I don't know how to do it, so I walk away. So this gives you this kinda, like, top of funnel taste of what Cloud Code can do. So Cloud Code or Cowork rather is is a decent, like, adjustment
02:24to dealing with more hands on agentic workflows without dealing with all the setup, the download, the, uh, blinking cursor on your screen if you're using a terminal. It's still full of bugs, but it's still, like, one big step up from using the normal Cloud AI. Gotcha.
02:40So we're not using inside of an IDE, for example. Where are we using this? We're using this inside of the Cloud desktop app.
02:45Right? So that's something that's, like, more accessible for most people because a lot of people might have Cloud desktop downloaded. You could just toggle between chat and co work mode instead of having to go into IDE.
02:55So I think it'd be helpful to show kind of a little rundown of what the UI looks like real quick, and then that'll segue us into our next section of actually using Claude code to build something. For sure. So this is what it looks like.
03:07So on cloud desktop,
03:09you have chat, you have code, and now you have co work. So the whole point of co work, like we said before, is they basically took the most common nontechnical things coming through Cloud Code and made them available as one click buttons.
03:23And when you click on these buttons, instead of you having to think about the prompt, it does the prompt engineering for you. So if you say organize files, which is the main use case everyone's going off on on x, you could just say go and fix this folder.
03:37You would tell it which folder you wanna work in, and then you wanna specify if you want, how you want it to organize it, or specifically what the name of the folder is. They give you some defaults like my downloads folder, my photos, some old projects, etcetera.
03:50And then when you go through the process, you can pick one of the language models of choice, and everything that's happening will appear on the right hand side. In the world of Claude code, everything is very monotone.
04:00It happens one thing at a time. So if you do need that aesthetic artifact look, then this will help you keep track of exactly what's happening.
04:08Okay. So you're saying we could make edits to our folders, organize folders, all this stuff, create presentations from folders, and then even spin up, a UI, kind of like a vibe coded user interface directly inside of the Clawd desktop app now using Cowork. Right?
04:22Exactly. So before, you could already create the vibe coded apps with artifacts on their own, but now they just made it that much easier to be able to keep track of all the files, all the context being used. Mhmm.
04:31And you can still use connectors and ad connectors. So you don't lose that MCP flare. You can layer it on, and not everyone's gonna need to reorganize the same folder 10 times.
04:42So we need to find better use cases. Right? These are just some examples that people can wrap their head around.
04:48But when you go through the experience, for example, I tested the crunch data. It will go to the folder, come up with a plan, and you'll notice that when you click on them, you'll get this prompt here that says ask any clarifying questions. So they're trying to encourage you, to encourage it, to think and plan before doing things to lower the likelihood that you break something.
05:07Because in Cloud Code, if you don't know what you're doing, sometimes you can completely nuke your desktop if you're not actually prompting correctly. Yeah. That makes sense.
05:15Okay. So now let's segue into actually using Claude code because, obviously, Claude co work is good for super nontechnical people. I've been playing around with it.
05:22It's been great so far. It's been a good entry into Claude code, but I wanna know, like, does this get you, like, 90% of the way? Do you think it gets you 50% of the way?
05:30Where do you think it kinda leaves nontechnical people when it comes to Claude code? Like, should we still learn Claude code? If we extend the analogy of the gateway drug, key emphasis on drug, this to me gives you 30 to 40%.
05:42This is not even close to the firepower of Claude, and it's full of bugs. Why?
05:47Because they use Claude code to vibe code it in the span of a week. So the actual Anthropic team admitted that on x. So with any vibe coded app, as you'll be familiar and unfamiliar, you're gonna have tons of bugs, especially when you go to production, go to thousands and millions of people.
06:03Tinker with it, play with it, know that there's bugs,
06:07but this is only meant to be the taste of what you get with full firepower. Got it. Makes sense.
06:11Cool. So I wanna move on to a couple questions before we actually go ahead, show you how to download Claude code, and then actually build something with it. Um, Mark actually built something for me, a cool use case I've been wanting myself, and we're gonna go through kind of the step by step process how we did that and then also the the finished product of that.
06:27So my first question for you is, like, what's the biggest misconception with Claude Code? Because when I hear that, I think, okay.
06:32If I'm not coding an app or using it in my coding environment, what's the point of using Claude Code? So kind of break down what the biggest misconception or, you know, few misconceptions you see are. Yeah.
06:41So, obviously, it was a platform originally designed for developers, which is why they called it Claude Code. But over time, a lot of nontechnical people and a lot of nontechnical
06:51use cases popped up, which in my opinion, I talked to lot of people. I say they should have called it Claude creates because then it can create code and artifacts. But the misconception is that you only use it for development or building code or vibe coding.
07:04In reality, because it has access to your desktop and your files and you can create a local, uh, database that runs on your computer, which we will show you today, it can do whatever you want. Like, the limit is your imagination and your ability to just, like, learn by osmosis. So when you ask it, here, go to the Nano Banana API, make an a thumbnail of me for YouTube.
07:24It can go research, look for the Nano Banana API, tell you, hey. Listen. I need an API key.
07:30Here's the URL to go with the API key. Give it to me, and we'll go from here. You're gonna see Python on screen.
07:36You don't have to engage in that Python. You just have to be aware that something's happening to make that functionality work. So the less you are worried about what's happening on screen, and you kinda take a peek at it, even if you're not technical, you will start to see little patterns by osmosis, and you'll be surprised.
07:51In three, six, nine months, you'll be able to upscale yourself by accident. Okay. Yeah.
07:55I love that. And then I guess my next question would be, do you have any, like, non techy success stories of, like, somebody because you run a pretty big I wouldn't call it a Claude code community, but your community is built like, a a lot of your users in there or your community members are wanting to know everything about Claude code and all the use cases you find for it.
08:13So do you have any, like, nontechnical,
08:15like, whatever? Just somebody who might be a bit archaic when it comes to technology
08:19going from complete beginner to actually being able to build useful things with it. Yeah. Yeah.
08:23So, obviously, I have a mid tier kind of size community, but we get a lot of nontechnical people coming in that have decided, that's it. I am going to dig my heels in. I'm paying for the max plan, and I'm gonna build stuff.
08:34And we recently had someone that's actually 70 years old, retired, basically built a full mobile app in the span of three or four months using any den and Claude code. And stuff like that gets me as a, like, a super nerd, super excited because I love seeing that progress and people shattering limiting beliefs. Because people will always come in who are not technical and say, hey.
08:55Listen. Do you got something for me that's not Claude code? Can I just use, like, chat you could see better?
09:00Because they're just so afraid of like, leaping in and running into those first issues. But as soon as you get past that hump where, one, you've downloaded it. Two, you go by trial by fire.
09:09You try to build something, and something doesn't work. And then you realize that you could technically ask Claude code in a virgin window to help you fix Claude code, and you understand this little feedback loop, then they're unchained. And then once the imagination is open, they start building from very minute use cases to things that they're proud of, and then just stacking those skills on top of that.
09:27Yeah. I love that. And that you just literally explained me.
09:30I haven't used Claude code because it just whatever. I have the like, this mental barrier about it. And I had that in the beginning with AI in general, and that's why I created this YouTube channel to, like, show nontechnical people you can use AI tools and actually become, like, a speaker of that language, so to say.
09:43So, yeah, I I love that. Going on a little bit, like, I guess my last question here before diving into showing how to download Cloud Code and then actually building out our use case for this. What do you think a downside or the downsides of not using Cloud Code are now?
09:57Because it seems like this is where the AI space is going. And just like if you wanna be able to, like, 10 x your productivity, a 100 x whatever, like, do you think we're if we're not using Claude code, we're missing out on opportunity here? And it might not be just Claude code.
10:09It could be things like anti gravity. It could be kind of like a transferable language we could use for these other tools. So what do you think the downsides of not learning this is?
10:17Yeah. Uh, it's good that you gave that disclaimer because I was gonna give that disclaimer to your audience. One, this is not a Cloud Code chilling video.
10:23I like to use Cloud Code because I've tested multiple products. It's the one that I resonate with the most with my development style, but you do have Cursor. You have windsurf, you have open code now, which is becoming competitive.
10:32I think they have a million active users that are using open source models. You can use whatever you want. It's the idea that you use language models in a terminal where you are devoid of all these little pretty icons on a UI, and you go straight to the source.
10:47That's the theme of this episode. So I think that if you don't opt into the terminal world, not necessarily the Cloud Code world, you are missing out because you are going to the source. So when you go to the source, you realize it's .
11:01You look around at, like, 90% of AI tools, and you're like, oh, I don't need these. I can cancel hundreds of subscriptions. And I canceled a few months ago $1,200 a month worth of subscriptions because I literally realized I didn't need it anymore.
11:14If I wanted to compress videos, I don't need to pay for the service that does it. I can just tell it, and it creates a Python script that does it for me. You can start to tackle each subscription of your time, which is why I can justify my very expensive subscription to Claude Code.
11:28So it's a matter of using it from day one for ROI heavy activities. Got it. Last question.
11:34I lied. My last one is not my last one. So what's, a use case that you've built yourself that you use for Claude Code that, like because your hands off, it saved you a bunch of time.
11:42I know you and I have talked off camera about it. Yeah. So what's, like, one thing you've built where you're just like, this used to take me hours or used to take a team member hours, and now it's just kinda done with Cloud Code?
11:50Yeah. There's one thing that not wasn't possible, but would be super painful for me to do with the little time that I have, and there's one thing that was a major unlock. So the major unlock was I have a YouTube command center, which is unbelievable now.
12:02It's taken me probably I've put in a hundred hours into this command center where I go back and forth, and it goes and looks at, like, things that are trending, things that are new. We have a bit of a conversation. And then once we land on a topic, I kind of give it my ideas where I wanna take it and where I wanna go.
12:17And then from that idea, it goes and creates 15 thumbnails, very high quality. It takes images of me, transforms them to go with that thumbnail, creates a thumbnail brief. It creates an editing and a script brief for my editing team and creates SEO tags, titles, descriptions, everything from one input, which is why for me, once I started on this drug, I got addicted because once you can have one input and you have not infinite, but many outputs as a result of that one, that is the very definition of leverage.
12:44So once I did that, every part of my life, I started creating what I call command centers. So this is my copywriting command center. I started on LinkedIn, so I wanted to understand how do I write copy on LinkedIn without using AI, but, like, understanding the principles.
12:57How do I do a crash course on everything? Let's create a command center. So creating those command centers helped level me up because even as a developer, I get to learn things I would have never been exposed to otherwise.
13:08Yeah. 100%.
13:09Being somebody that makes YouTube videos, literally, I spend fifteen plus hours a week probably on this, and that bottlenecks every single other thing when it comes to, like, my business. Like, my creative mind is not working because I have to go and I have to come up with these thumbnail ideas, title ideas. I think one thing, and I could be wrong here, but what sets it apart is you're not just, like, saying go off and do this thing and it does all of it.
13:29You have to play around with it because if you didn't come to that, like, idea or, you know, whatever the title was gonna be, like, you had to kind of play around with it to get that title. You know what I'm saying?
13:38Instead of just saying, hey. Go create a title for me on whatever whatever you think is gonna work well. So you have to have a little bit of that back and forth.
13:44But I think once you have that solidified, it then could take it and run with it and go ahead and, you know, create those thumbnails for you and whatnot. So It's really important for your audience because you just made an amazing point. There's this trend right now about this thing called a Ralph loop.
13:56Right? It's a very mind numbing thing for most people that are not developers.
14:00Basically this fantasy that you just give tasks and you walk away, go to the beach, and you come back, and a bunch of tasks have been executed. It's one thing to execute a 100 tasks. It's another thing for those tasks to be meaningful and drive actual value.
14:13So if you spun up and you said, make me a thousand thumbnails and a thousand titles, but 80% are garbage, that's like a net loss. You lost the money on the token usage. You lost time and wasted effort and electricity.
14:25Human in the loop plus Cloud Code is the most OP combination. Combination. We're gonna clip that for for the intro video because that yeah.
14:32That's fire. Cool. So now that's a perfect segue into showing what we're gonna build because we're gonna build something pretty similar.
14:38I guess real quick before doing that, we should show how do you quickly download this if you're a nontechnical person. Where do you do this? How do you do this in, like, two minutes?
14:45But show that, and then we're gonna go into the build and show what Mark built us. Alright. Sounds good.
14:49Okay. So when you go to the beautiful domain claw.com/product/claw-code, you'll see something like this.
14:55And, yes, it is a terminal. Uh, resist the urge to cry. All you have to do if you're nontechnical is you just use this one command in your terminal.
15:04And if this immediately gives you an error, most people, many people, paste this. There's some form of error because maybe your computer's different, your environment's different. There's something that needs to be reorganized to make this possible.
15:15If that is a barrier to you, you can install something called warp. Both of us, not affiliated. It is a terminal that you can run pretty much for free.
15:23It has some AI in there with a little a free tier. You can just tell it that same command. You copy paste this specific thing right here.
15:30And then if there is an error, it'll say there's an error. But instead of passing it back to you to figure it out, it will go and figure it out for you on your behalf, ask for permission, and then install Claude code for you. And then once it's done, you will see your little Amigo, and it will pop up.
15:45You'll see something like that, and then you will be good to go. You'll see a screen just like this. Perfect.
15:50Easy enough. Yep. So this will be your highway.
15:53Again, not affiliated, but it's a free way to, one, use a terminal if you ever need to install Cloud Code, remove Cloud Code, update Cloud Code. Really good way, and you can do it for all kinds of things.
16:03You can even ask that, by the way, to organize files on your computer. All of these IDEs, as long as they have access to your computer, you're good to go. So that's the installation part.
16:11Are we good to jump into the concept? Hold on. Before you go to that, let's talk about where users are actually gonna begin interfacing with Cloud Code.
16:18Yeah. So you can use this in any terminal. You could use warp like I showed you.
16:23I like to use it in cursor. You can use anti gravity.
16:26There are so many IDEs that I haven't even mentioned. An IDE is just an editor. You can just open any terminal.
16:32And once you have this installed and you use the magic words Claude, you should see your little crab icon friend pop up, and you're good to go from there. Perfect. Easy enough.
16:39Cool. Now let's move into let's talk about a what I wanna automate first, and then we're gonna show what we actually or what Mark actually built after that. So let me I'm gonna go ahead and share my screen now.
16:50Alright. So looking at my screen, what I wanna automate is my short form content. So for some of those people or majority of you guys that don't know, I've been kind of going hard on short form content.
16:59This is a goal for 2026. I really wanna tap into this platform here, both on Instagram, YouTube shorts, and TikTok. Alright.
17:05So about a week ago, I had this video right here go absolutely bananas. It's got about almost, like, 825,000 views, bunch of engagement, and I've actually grown.
17:14Like, I had about, like, three or 400 followers before this. This singular video is, like, unlocking this whole new channel that I could tap into. And, this is a huge goal of mine going into 2026.
17:24And then here's a YouTube short too. I posted this three days ago. And And this one, I'll just refresh, is absolutely running.
17:29So the goal for what I wanna build is I wanna build out a system that could help me automate this. And so that way, don't need to be super hands on with it. Because right now, I have somebody that's actually writing scripts for me.
17:40I have to approve these scripts. I actually go and I analyze what videos did well so I could try to recreate that success of that video. So if I had a Claude code system that I could just, like, have on autopilot or I guess I'd you know, there'd be a little bit of human in the loop there that could just help systemize this process for me, that would be absolutely, like, the most valuable thing ever.
17:57So Mark actually built that out. So let's go into a little bit more about that. And if you see on screen there, uh, before he unshared the screen, he had 89¢
18:05from that video. So in exchange for 89¢ that he's paying me, I built the system for him and for you. So let's get into it.
18:13Conceptually for the audience, so I don't lose you. Core concept is Brock, like you said, wants to go celebrity mode on short form. We wanted to focus primarily on Instagram.
18:23So the number one way we need to do that is scrape. Now if we just ask Claude Code, go and scrape Instagram, that is a very mammoth task because there are many proxy layer security, all of these things.
18:35And to actually do the scraping, we're gonna use a marketplace called Appify. You can think of it as the app store where you can take tons of scrapers. There's ones literally designed for Instagram, TikTok, what have you, and you can essentially lease them or rent them to execute your tasks for a small pay as you go fee.
18:52So we're gonna use that. Restore it in a database, and then we'll use that to help create those ideas for Brock.
18:59So, we're gonna focus primarily on Instagram. And instead of paying $10 every single time like you're used to doing using platforms like Lovable and Bolt and what have you, we can just use something called SQLite. And what SQLite is is literally a free database you can run locally on your computer.
19:15This frees you because every time I had to experiment before, I would have to shell out another $10 to Supabase. So this lets you build vibe quote for free. If you wanna go to the next level and then upgrade to Supabase, that's always something you can do after.
19:28So that's the database we'll be using, and in there, we'll be posting or hosting things like the creators, the posts, and the hooks that we scrape. Scrape. In terms of the workflow, if you remember before, my YouTube command center, I have one input, many outputs.
19:41In this case, Brock just wants to say the words make me viral as one command, and then after that, write a topic or a link where Claude can go and read about whatever the tool is and then generate 20 hooks,
19:53then rank those hooks by a virality score. That virality score, I'll talk about right below, and then we'll pick a top three and then create captions for that. Brock can have one input, one idea, and then have multiple outputs.
20:05Great question. So go go back up a little bit. So on the topic side of things, context is really important here.
20:11Right? Because if I give, like, a little short paragraph of what I want the context to be about, it's probably not gonna be the best output. Right?
20:17It's not gonna generate the best hooks cause it's not gonna have context, or will it take that, then do some web searching, and then, like, find the right amount of context and then use that, or, like, am I wrong there? Could you break that down for me? No.
20:29No. For sure. So the first time you use this, everything we're using is a language model.
20:33Language models are not magic. They're not super intelligent.
20:36They look for patterns. So if you say, go look at this Claude code or co work feature, and here's the link, and you say nothing else, and there's no preferences or examples of how Brock writes his scripts, it will go and look at the database where we've scraped other creators and be like, okay. I think I'll do a remix between the different creators and my knowledge about writing hooks as of, I don't know, 2024 or 2025.
20:58That will be your first output. Then as you iterate and say, no. No.
21:01That's not how I like to write hooks. Here's the example of this 800,000 view video.
21:06Try to mimic that style more and tell CloudMD, which is like the command center brain, to update that in its memory. So that's how this incrementally gets better. Got it.
21:14Yeah. So it's just gonna literally just get smarter and smarter and smarter the more I use it hypothetically. Right?
21:18Yes. As long as you keep telling it, listen. Um, this time, this was your mistaken flaw.
21:22Let's commit to memory to avoid doing that moving forward. So you have to be proactive in making sure that it learns. 100%.
21:28And when it comes to the virality score, we kinda just made this up, Brock and I, just to come up with something. There would be some criteria. So this virality score is based off of curiosity, emotion, brevity, relevance.
21:40Relevance being, like, how trendy is the topic, which on short form, even though not short form guy, seems to be really important. Right? Extremely.
21:47And then you have uniqueness and platform fit. These ratios could be off. So these weights could be updated.
21:52You just wanna be able to design the system. So if you need to update them, you can. And then this leads to the overall preview of the process, which is you scrape competitors, you get some inspiration, maybe you add your own examples as well.
22:05You have some razzle dazzle. You generate the content, and then the final piece of the life cycle, which we won't cover in this video, but we you could do later on is after you create the thing, tell it to go scrape your own Instagram to see how did it do so it can actually learn from what went well and what didn't go well.
22:21That's where you have that full feedback loop. Yeah. That analyzing part is the part I'm most excited about, honestly.
22:26Like, scraping the competitors is great, coming up with ideas, understanding my hook style, all that stuff. But, like, analyzing what works because, really, what I've noticed at least in short form content, you find the, like, the hooks in the, like, the format of what works and just double down, triple down on that as opposed to trying a whole bunch of different things, seeing what works.
22:44And once you find something, it's really important. That analyzing part of this is what I'm probably the most most stoked about. Yes, sir.
22:50Yeah. So in terms of the links, these are the links that Brock provided to me of different creators he wants to be inspired by on the short form side. So these were we'll revisit these shortly.
23:00And in terms of the actual build, one thing I like to do, again, because I am lazy, is you can use whatever research tool you want. In my case, I like to use Perplexity Labs.
23:10It's this icon right here. They have a pretty generous feature as well. And what it allows you to do is give it an idea.
23:15So for me, just use WhisperFlow. I just go off for, like, five minutes about what I'm trying to build, and then I throw that into something like Claude or ChatGPT and tell it, can you make this into a more refined prompt for something like Perplexity Labs?
23:26And then we get something legendary like this where there's no way I would have actually written it myself where we won't read the whole thing, but it just says act as an expert senior product manager and slash prompt engineer. I need you to write a highly detailed mega prompt designed for the absolute latest, most capable version of Cloud Code available today.
23:45Why do I say this? I wanted to know about the latest advancements in Cloud Code so it can refine the prompt accordingly. So I don't have to worry about what changed in this version?
23:52How should I talk to it now? It does that dirty work for me. And I say, use your browsing capabilities to verify exactly what version is the state of the art.
24:00Then I tell it what the goal is. We wanna do a competitor scraper. We want internal analytics.
24:05We want a command center, uh, dashboard later on, not a priority. And then we go through the workflow. So once I do this, it comes back.
24:12We zoom back out here with a sample prompt, and I tell it, can you just make this into a markdown file? So I can just drag and drop it into a brand new virgin eyes Claude folder. It's not even a Claude folder.
24:24It's a folder that you import and use Claude code to read what's in that folder. So you're just you're just adding that to, like, your folder on your desktop, essentially? It's as simple as that?
24:32Literally, just to show your audience, you'd go open folder. Okay. You'd create a new folder, and then when you open in cursor here, you'll have absolutely nothing except the name of the folder.
24:43This will be empty. So for me, this is where I drag and dropped the exact file I just showed you here. This was the only file in my cursor, and nothing else existed.
24:52Everything else came about from using Cloud Code. Yeah. Easy enough.
24:56Yeah. So this is the foundation. Now I should have been a bit more intentional to go through the prompt and make sure that all the features that AI came up with are the ones that I care about, but I use this to do this quick and dirty.
25:07That's that. In terms of Appify, one of the scrapers will rent is called an Instagram post scraper. Uh, this one scrapes the actual posts, captions, etcetera, but it doesn't take the transcript of the Instagram videos themselves.
25:20If you don't feel like going and shopping around to see which app in this Appify app store can do that, you can use their MCP server that can just go and navigate their database for you. So if you tell it, listen, I need to do x, y, and z for the cheapest cost, you will go look at all what are called actors in the App Store, pick the cheapest ones and the ones that are most highly rated and recommend those to you, and then you build your workflow after.
25:43Yeah. That makes it so easy. Gotta love the MCP.
25:46Yes, sir. And then in terms of using and installing the MCP, this is where nontechnical people might just tap out. They made it easy as well.
25:54So if you just go to mcp.appify.com, exact URL, you'll see something like this. You can pick what matters to you.
26:01I would just click on things related to data storage. I would just pick everything if you're just starting out. You don't know what you're doing.
26:07And then at the very bottom on integrations, all you have to do is pick on something like cursor or Cloud Code, and then it'll give you the one line command. And if you click on add API token, it'll put a placeholder. All you have to do is go and get your Apify API key from your account, which if you go to your account, is literally right here.
26:24You grab this. You paste that exactly in here, right there, instead of these little curly brackets. You take this command.
26:31You open a brand new terminal session with Cloud Code, and you paste it. So I will show you in one second after I block my API key what it looks like, and we'll go from there. Yeah.
26:40I got one thing to say on that because I've used Apify in the past for certain builds in n n n, for example, to try to create the same thing. And finding an actor that actually works sometimes was a pain because it worked in the past, and then something broke and it wasn't up to date. They weren't, like, servicing it, whatever.
26:54I don't even know. But I think the MCP thing makes it so much more simple, and I'm glad that you brought that up because then it'll just go ahead and find which one is working right now so you don't need to go and do that yourself. Or it it will actually just test them.
27:05So it'll run it. It'll be like, this one sucks. Let me go search for another one.
27:09This one sucks. I'll go search for another one. It will do the dirty work for you, which is why I like that.
27:13So this is what it looks like in Kershur. Obviously, I have blocked my API keys. I pasted that exact command I showed you before, so that Claude MCP add Apify.
27:22I just had to add my API key right here, and then it went, executed that command, and then it comes back at the bottom saying successfully added. And all you have to do is restart a brand new session in Cloud Code, and it will take effect. And I'll show you how you can verify that you have that MCP server.
27:40When you have a brand new tab, you just have to write MCP and then put MCP status. You'll see right here, we have this connected, so we know we are in Valhalla.
27:49We are good to go. We are in the promised land. And beyond that, all you have to do is I just said, are you able to find this?
27:56So I told it, go and look for this Appify scraper. I just wanted to double check it's working. You can see right here, as soon as you see this little green icon and you see m c p underscore something, the m c p is working.
28:07So it found it, and it tells me about it. It tells me about the success rate, the total number of users, which is awesome because like Brock said, it's always aware of what is being used so it can rank the actors by usage and success rate. And it tells you what it does, the pricing, the input parameters of what it needs, what you can offer it.
28:24And then I said, can you can I scrape the transcript of videos? And it says, no. You can't, which is where I discovered that.
28:30So it can do metadata performance, but it can't do that. So then it went and found three alternatives with their respective success rates and their pricing, which is why this is so awesome.
28:40It's very hands off. You are the boss, and it tells you what is the best choice and why. Once you do that, I say, cool.
28:45Can we test it on this? We take, uh, Nektarav's short form, and we ask it for, let's say, the first three videos.
28:52It doesn't work, and then it hangs. And then I basically try to understand why is it hanging until it finally comes back with the following. So this is the caption.
29:01This is everything about that video. So I'm like, cool. Now we're cooking with gas.
29:05We try a few more posts, and I I say, can you try three posts at once and see what happens? So the first one works perfectly. You can see right here, this is the source URL.
29:15If I scroll to the very right and I grab this, you can see let me see right here. Zero to forty seconds, zero to eight seconds.
29:23Okay. Stop building your automations from scratch. Instead, use this AI tool.
29:27So now I know, cool. That is a script, and our next job at some point will be how do we extract what we think is the hook?
29:34Because short form is short. So how do you know which one is the hook if you don't see the first five seconds or the first line or the first two lines? So there's gonna be some judgment there, and there's some nuance, which is why this is not perfect.
29:44One thing I did notice is that the MCP had an aneurysm when I tried to make it do three at once. So it turns out, at least for this actor, it needs to do one at a time, not all at the same time. So once it did that, it was able to do one one one, and then we were good to go.
29:59And then you could see right here, this is the transcript that comes back from each and every video along with the original URL. So that's kinda like the end of the first session. And I basically say the following, because every session in Claude code is a brand new blank slate for the most part outside of whatever you have in this file called ClaudeMD.
30:16You can think of clotMD as your memory bank, the same way you'd have memory in chat. One thing that we have to know is I don't wanna have to go through every single time with a blank screen and say, go and give me the transcript of this video for it to have to go in circles to realize, one, it doesn't have one actor to do that.
30:33Has to do two of them. And then two, that it has to do one at a time. So instead of me having to worry about that, I just say, take note on how you're able to scrape these transcripts and what actors you used and create a Claude MD file that would have a full understanding of how to execute this.
30:48So I don't have to repeat myself each and every time. So you didn't even usually, like, if you're interacting with, like, AI, ChaiChukwuTe whatever, you'd say that, like, oh, give me a prompt to use to then always know how to perform this task. But inside of Cloud Code, this is, like, what stands out to me, is it literally went and created that we could call that a skill.
31:05Right? Is that considered a skill in Cloud Code? I won't let you say skill because I'll miss they'll be misleading.
31:09So Okay. It's not a skill. It's just it's creating this Knowledge base?
31:14Knowledge base. Let's call it knowledge base. Yes.
31:16Okay. Got it. Yeah.
31:17That's cool. I literally created that file instead of just giving, like, a prompt that you would usually get that you'd have to u like, save somewhere in your notes, use it when like, wouldn't you know, whenever you're using a custom GPT, for example. But Yeah.
31:26If you click on it, they make it like a hyperlink, especially in cursor. You could see, like, top priority is, like, the make me viral command. That's after we actually built it out.
31:34But if as you go through, it takes notes on the different data models you have, the hook ranking algorithm we ended up making, everything that it needs to use.
31:43So he has a section here on actors used. So now it knows this is the Apify actor name. This is the cost, and the second one we use to actually grab the transcript of the video is this one.
31:53So it keeps that running memory, which is why this is powerful. And, ideally, it's gonna be way quicker ever like, at least after that initial time because now it knows exactly what to do, so it doesn't need to go through this treadmill thought process of trying to figure it out. Yeah.
32:05So if you compare it to something like a ChatGPT, every single time you fail a ChatGPT, if you don't, uh, if you aren't proactive to go and document and avoid it happening, it will keep happening over and over again.
32:15But in Cloud Code, every time something doesn't go as planned, you could say, go and learn from this mistake as soon as you finally get to the point where it did the right thing. That's where you have that feedback loop. Now to make this a lot more tactical, this session was about executing the plan.
32:30So for me, if you go to here, this is the plan. This is an adapted version of what I created in Provexity Labs. This is the original.
32:37And what I asked for in a separate chat was, can you make this thing instead of just a list of feature requests in a wish list? Can you make it into a plan and make it so that there are checkboxes so you can keep track of what you're doing as you accomplish each task?
32:52Now why did I do this? One, it's no secret that Claude has a context window as of this recording of 200,000 tokens. MCP servers actually take a lot of that.
33:01So it can take anywhere between two to 15% of that session. So instead of me running into a max window every single time and worrying about that, if it keeps checking off when it's completed, I can always go back and be like, listen.
33:14Go pick up where you left off on the plan. I So can see like, okay. Cool.
33:17I haven't started phase 2.3. Let's keep going from there. Yeah.
33:21That makes sense. Yeah. So that's super helpful, which is why I ended up creating this plan.
33:26And the plan, you'll see right here, as it completes things, it will tell you, this is completed. This is not completed. And it'll keep editing the source plan until we're good to go.
33:35And you could see here, it's using the MCP server to test out, again, uh, the different Apify posts scraping. And then if we go to the very bottom here, it's using and creating what's called the SQLite database.
33:47So like I said before, if you click on database and you click on short form, this is what it looks like behind the scenes. You don't need to use a super base.
33:54It creates all of your database for free. You see all the columns here.
33:59And if this is too ugly for you, I had a separate chat where I said, can you make it so that we can have some form of UI to look at this? So this was the chat. I said, can you make it so we have UI?
34:08I hate looking at it here. So I talk to it like my friend, and you can talk to it as well. And you get something like this where you can click on it.
34:15And when I, uh, load it back up, you'll be able to see different tables like hooks and posts, etcetera. That's basically you'd be able to interact with it on a browser. Yeah.
34:23Much easier to look at than the IDE that we're looking at here. Yeah. It's basically like having a Google Sheet on steroids, and the most important thing, like I said, is it's free.
34:31So one thing I wanted to do now is put everything into practice, and I noticed that we encountered a bug. I thought that things were running super smooth. We have the MTP.
34:40It learned how to scrape, and I started storing things in our database and started creating the viral hooks algorithm so I could create this post creator. And then I realized that there was one fundamental flaw, and I'm just gonna wait to just pull that up so your audience doesn't have to watch me scroll.
34:55Yeah. Sounds good. Okay.
34:56So here, at some point, I realized that it was making a fundamental mistake. It was scraping everything, but it would put the caption in the transcripts column.
35:07And you might say, why do I care? Well, you care because it was training on these hooks that were actually the caption.
35:15So it learned to say comment x, and I will send you y instead of the actual hook of the videos. And I got confused because I started getting tens of hooks that all started with comment y for x. And I'm like, this is not a hook.
35:27So then I went to look at the database, and I'm like, you literally put the captions in the transcript. You never actually captured and saved the transcript. So at this point, that's where I said, I just noticed that your hooks are the captions.
35:39This is wrong. So it went through. It deleted all the things, uh, that were incorrect, and then it reloaded and went through and just kept only the posts, and then we start adding on the transcripts.
35:52So when we finally go the the next step, which is okay. Now that we have transcripts in there, we come up with the virality score. It goes through and it creates, like, a ranking of all the posts in the database.
36:03So if we go to the database and you check the scores, you'll be able to see that they're actually implemented and calculated on the fly. So that allows us to do the following.
36:13If I go into here so if we say, make me viral, this new feature, this is the command that in CloudMD, our memory bank or knowledge base as we now call it.
36:23I just said go and look at this feature and come up with material. You'll notice that despite that, we still have a couple of them that were biased from before. I think CloudMD did hadn't updated yet.
36:33It says comment co work to get free access, but we still did have some real hooks for the most part. So eighty twenty, 80% were actual hooks, but we still wanted this to be a 100%. So Here.
36:44Go up one more time. Yeah. I wanna point one of these out.
36:47Like, for example, number two, that hook, pretty solid. Claude just became your AI coworker that finishes tasks without you. Like, that's something that I would actually use in one of my short form videos.
36:56So that's I mean, goes to show that it's at least working. Yeah. And maybe you could use the comment, coworker, and they're happy.
37:02Like, this is solid for the caption as long as it knows what to do. But I just wanted to show this because it's important to know that especially if you go on x, especially the dark depths of x, people will say this thing will replace software engineers, like, remove everything from existence. It still makes mistakes.
37:18It should have been smart enough. If you can use an MCP server to see what actor we need, it should be able to look at and be like, oh, this is probably not a hook. But you have to be aware that, like, you can't just go off to the beach and come back and expect good results,
37:30especially if you're not a non techie. So human in the loop watching this is essential. Yeah.
37:34That's like the keyword for this episode, I think, is human in the loop because otherwise, we could build out this system if we're not constantly iterating with it. It's probably not gonna be the best. So that's why this is a good starting point for us and for me to start using.
37:46I'm excited to see what this looks like a month after playing around with this, having to test out different videos that I post and giving my thoughts and, you know, kind of expertise on it. Yeah. And the thing is you can go much deeper than this.
37:57So right now, this is the scoring
37:59algorithm where you could see right like we said before, curiosity, emotion, etcetera. It's using its judgment.
38:05Like, Claude is going through and ranking everything that we have in our database with a score. Brock can look at this and be like, no. You are so wrong.
38:11That last video was a complete flop, or this video did incredibly well. So this will take tinkering. And if you had a vibe coded a front end, every single time you would change the way you the the process flows and the automation flows, you'd have to go and rework the UI to match it.
38:27So that's where, like, the UI and everything that we focused on for the year, which is, oh, cool. I just built a clone of x. It's not about the clone.
38:34It's about, like, what is beneath it? What is the system underneath it? So when Brock spends a couple weeks until it's actually really good, then he could ask, cool.
38:44Can we just spin up a UI? And maybe we'll use Gemini as the LM to go do the analysis or Anthropic or whatever to execute everything that we've been doing in the terminal. So in a way, you you leave the final part for last where if you have a UI, if you have a product of your own, it's now fully polished.
39:01That's actually, like, a super good point because I was gonna say looking at this viral scorecard thing here, I was like, man, this just seems like a lot of numbers. I'd love it love it if it were in a dashboard. I was literally gonna mention that, but that's such a good point because if I don't have this thing dialed in, so to say, like, right now, it's just gonna be a pain in the ass having to go back and forth and try to vibe code.
39:20My audience knows how to vibe code and how it could be such a pain so many so much of the time. So, yeah, that's interesting. I'm excited to see what it looks like when it's in almost like its own little mini, like, internal software that I could use instead of having to use something that might be a little bit more intimidating like Cursor or something like that.
39:37Exactly. So and then the whole point is that it gives you all of these as outputs. So it tells you now the top three hooks.
39:42So it tells you why it's a hook, why it's a good hook rather. So this one, Claude, code just gave non coders. Lily picked the same one, by the way, that you said is a good one, which is interesting that it agrees, and it comes with the rest of them.
39:55Still has this weird comment thing you'd have to remove from it. And then it comes up with the captions, and what's funny is it wrote the captions as as these.
40:06This changes everything. It doesn't have the comment. Even though that's the hook.
40:09Right? That's, uh, wait. Would that be the hook?
40:11Or, like, the script, I mean? I think it could be a caption, but it's interesting that it didn't put the comment x or y even though we it's very prevalent in the small data that we have. So there's a lot of nuance here.
40:21This is super This would be really sorry to cut you off, but this would be super easy for me to give Claude code one
40:27prompt and say, hey. Listen. For every or for every description
40:31or yeah. I guess it'd be considered a description, I wanna chew I wanna use a keyword where it says common x, I'll send you y, and then it'll fix that. Now every single time I use it, it'll, like, understand that.
40:41Right? Yeah. And even better, since we have the Appify MCP already used, you could say, can you go on my Instagram, look at the top five performing ones, go grab my style of captions, and go grab my style of hooks on the ones that performed well?
40:56And in the database, I wanna change the score now. I wanna add a Brock, uh, weighting where if it's a hook by me that did well, I want you to weight it higher than the other creators because the other creators aren't me.
41:09So this is where the nuance comes in. Instead of you obsessing on updating a front end and some button not working and you spend an hour fixing the button even though the process needs to change, you focus on the stuff that actually moves the needle. Yeah.
41:21This is this is great, dude. This is honestly better than I thought it was gonna be if I'm being completely honest with you. Like, I'm stoked to play around with this.
41:27Good. Good. Just everyone in the audience, this is a process.
41:30And the thing is, if you use something like an antigravity or a cursor, this is my special weapon is obviously, you're dealing with Claude here. But let's say you need virgin eyes on your your code or anything that you're generating, you can always bring in cursor for free for, like, a pretty generous free tier as long as you don't use max mode.
41:47And you can be like, cool. What does codex or GPT 5.2 think about my system to generate hooks?
41:54So you can bring in different advisers, if you will, to take over the same code base, whether you're vibe coding something or just producing something at scale. So you're saying as these models get better, you could then use that as almost like the orchestrator to understand how could I optimize this kind of is what you're saying?
42:10Let let's see if actually ended up building this UI. And for whatever reason, Claude code wasn't cutting it. It just wasn't getting you.
42:16Maybe you bring in a virgin eyes expert in the form of Gemini three Pro, and you say, hey. Look. Both of these share the same code base.
42:23So this code base, anything that's in cursor can see this. So Claude can see this, and anything in cursor can see this. So with that, you can go, like, in ask mode and be like, listen.
42:32Can you take a look and give me advice on how I could structure this better? Or even more hacky, you could say, go and search the web for best practices for prompting Claude code.
42:42Can you help me draft a prompt that would fix this problem that's happening in my code base? And then it could go search the code base and then go and search the best way to prompt it, and then you can copy paste the prompt from here to here. So there's so many ways you could use this dynamic duo.
42:57Yeah. I love it. That's super
42:58like, I didn't know you could do that, so that's that's a good tip. There you go. But, yeah, with that, like, obviously, we can go so much deeper, make this infinitely better.
43:06But it's the idea of just, you attach a couple tools, you throw some API keys, you give it some prompt, and then you start the progress. Like, to build upon that, not that we're gonna do this in this video, but a use case I would see for this is I have been using, uh, or I've been creating thumbnails for my short form videos on Instagram just because it looks better on my profile, whatever.
43:24So, literally, I could do what you did for YouTube where you have this using Nano Banana via the API from Gemini, and it goes and actually creates these thumbnails for you using an image generation tool. And that might be the next thing that I personally go in and actually try to build out because that'd be super helpful.
43:39And you built something that's a bit more intricate than, like, short form video. I feel like the system you built with YouTube is, like, five step a couple steps ahead of this, and it's just crazy to see, like, what we could do with Claude Code. So this video is giving me promo mark, so I'm gonna need to hop off.
43:55Like, dive in.
43:57Yeah. On that thumbnail thing, just like to inspire people. So the way I I created it is I have Nano Banana create the concept, and then it drafts it out as a doodle.
44:06And then it takes that doodle, and it takes that, pumps it to Nano Banana again to create version one of the thumbnail. And then it loads that into Gemini to get its opinion on the image as a YouTube expert. And then after it roasts it, it injects its feedback back to the nano banana to to fix it.
44:24So I have this feedback loop where before I take a look at it, it's already preroasted itself, which is why, again, this is such a high leverage skill. If you're non techie, this is the time.
44:34This is the perfect time to jump in and just get your feet wet. Yeah. That's
44:38and that's I love to hear that. One one other thing I wanna mention before we end this episode because we obviously showed kind of what we built here or what Mark built here. Go back to that Excali draw board you showed because, correct me if I'm wrong, you built this with Claude code, didn't you?
44:52You didn't even build this out yourself. No. I have a system that I built in Claude code that helps me make explanatory
44:58diagrams, especially for my YouTube, where it's doodle. So it makes things a lot more consumable.
45:02So I will walk through. Like, I literally just talked to him. Like, listen.
45:05I'm about to hop on Brock's podcast on his channel. His focus is not as developer y. I wanna break this down.
45:13So help me make three to five images where you think this concept would, like, be done justice, and then it takes care it from there. Dude, I love that. Yeah.
45:20That's just hilarious. I love that you did that. Alright.
45:23So this episode has been amazing. I'm excited to go ahead and try this stuff now and actually begin using this system, build other things on top of it. But, Mark, if we could give the viewer of this, a complete nontechnical person, one actionable item to do right now, like, instead of just watching this video, taking notes, and doing nothing with it, what could they do right this second to actually take that first step, begin begin learning this, and hopefully building the skill that could change their life over the coming years?
45:45Yeah. So I'll I'll give the advice with a grain of salt because I know people will look at this. And even if we prove that it's effective, they still won't download the terminal.
45:54So if that is the friction point, you literally have something that's dropped from the heavens in the form of co work, which is not perfect. It's still an imperfect product, but it will give you this little taste of what it could look like to run agentic workflows. You won't get the breadth.
46:09You won't get the depth, but it will get you more comfortable, and hopefully, it will give you FOMO that if you had, instead of a basic engine, you had the full firepower of a v nine, v eight engine that you could run that much farther. So instead of having to go from zero to deep end, you can go from zero to shallow,
46:26get your feet wet, get used to it, and then when you're ready, make a couple commands and then start your Cloud Code journey. And keep in mind, this is a tool that dropped literally yesterday. So this is your sign from the heavens, like you said, to go and try this and test this out.
46:38There's no excuses now. Yeah. Um, Mark, this has been awesome, man.
46:41I appreciate it. Thanks for building this for me. Thanks for explaining it.
46:44Where could my audience find you? For sure. So
46:47mark underscore cashif on YouTube, if you are someone that is interested on the deep end of things, you don't wanna be just non tech, you wanna go really hard, You could look check out my early AI adopters community. That's where I go hard all the time and I build all kinds of things. And outside of that, uh, hopefully, more on Brock's channel.
47:03Yeah. I wanna give kudos to Mark's community too because I've been in there for, I don't know, probably over six months now. Honestly, if I wanna learn something that's a bit more technical, that is my favorite place to go.
47:13It kinda cuts through the noise of everything else. Don't have to go on YouTube or x and, like, get my anxiety up and my FOMO up. Instead, I'll just go straight to Mark's community.
47:20But it's a great place. So, yeah, Mark, thanks for coming on, man. Sure we'll have Have you on a good one.
47:25See you, guys.
The Hook
The bait, then the rug-pull.
Brock Mesarich spent months avoiding Claude Code, convinced it was too technical and too developer-y. Then he brought on Mark Kashef, the person his community calls the GOAT of Claude Code, and spent 47 minutes going from zero to owning a system that scrapes competitors, ranks hooks by a virality algorithm, and outputs a content brief on one command.
Named ideas worth stealing.
12:12model
The Command Center Pattern
One input, many outputs. Build around a single command in CLAUDE.md.
Steal forAny repeating content workflow Joe wants to systematize
14:00concept
The RALPH Loop Warning
1 input x 1000 tasks at 80% garbage = net loss. Counter: human in the loop + Claude Code.
Steal forFraming for JoeFlow or Batch panel explanations
31:43model
CLAUDE.md Knowledge Base
Claude writes its own instructions into a project file. Every session reads and writes back. Incrementally learns.
Steal forDirectly applicable to JoeFlow CLAUDE.md setup
23:00list
Virality Score Algorithm
- Curiosity 25%
- Emotion 20%
- Relevance 20%
- Brevity 15%
- Uniqueness 10%
- Platform Fit 10%
Weighted formula for scoring competitor hooks. Updatable weights.
Steal forAny hook-ranking system in MCN or content pipeline
42:14model
The Dual Adviser Technique
Same codebase in two IDEs with different models. Ask second model for fresh critique or prompt suggestions.
Steal forDebugging any Claude Code project that hits a wall
How they asked for the click.
VERBAL ASK
46:00next-video
“if terminal is the friction point, you literally have something that dropped from the heavens in the form of co work”
Soft CTA -- no subscribe push, no link drop. Mark channels friction into Co-work as a stepping stone.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
Visual structure at a glance.
