YT YouTube2Text

YouTube2Text → Transcripts → УЛЬТИМАТИВНЫЙ ГАЙД НА ВАЙБ-КОДИНГ

УЛЬТИМАТИВНЫЙ ГАЙД НА ВАЙБ-КОДИНГ — Transcript

by OneDray · 6,166 words · 954 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:13So,
0:14you
0:16think you need to be a programmer for
0:18this?
0:19I don't have time. I need to program
0:20and make money.
0:21No.
0:22By the end of this video, you'll
0:24understand how to put it all together
0:25without knowing a damn thing. In this
0:28video, I'll show you how to use neural
0:30networks for any purpose and field,
0:33starting from real architectural
0:35vibe-coding to, yes, generating anime
0:38girls.
0:39Smile.
0:41Oh, don't smile anymore, you're scaring
0:43us.
0:44This video will be useful for both
0:46beginners and seasoned IT folks on
0:47antidepressants. After this video,
0:50you'll know the entire path from
0:53creating a virtual phone number and
0:56paying for services yourself so that
0:59you have exclusive access to your AI.
1:02Safonov pay.
1:04And I'll explain how not to get hit
1:06with a ban, for example from Anthropic,
1:07so you don't have to create new
1:09accounts every time. all the way to
1:12deploying click-clients, creating
1:13skills, using multi-agents, and
1:15integrating into almost any app using
1:17MCP bridges. And listen closely now.
1:20All this is much simpler than you think
1:23. People love to fill these complicated
1:26, damn names with hot air, which they
1:28then try to sell you as super-important
1:30expertise. I'll just explain all these
1:33things so simply that even a monkey
1:35could understand them. Important note:
1:38there will be no ads in this video, and
1:40I will show the tools that I have been
1:42using myself for quite a while,
1:44naturally with all the proof. Well, at
1:47least I'll try to provide it. So,
1:49without further ado, let's get started.
1:51Step one. We need to get to the
1:53resource in general. As we know, in the
1:56Russian Federation, the giants of the
1:59AI industry, like OpenAI and Anthropic,
2:01don't provide their services, which
2:03means we need access to those three
2:05letters that are dangerous to say out
2:07loud these days.
2:11But considering you're watching this on
2:13YouTube, I think you've already solved
2:15that issue. But if for some reason you
2:18need it, you can get it from our team
2:20by subscribing to Boosty or just by
2:22following the link to our bot.
2:25Speed, I am speed. Faster than fast,
2:31quicker than quick.
2:34Pupkin SBP. Your guide to the global
2:38internet. Important point: you should
2:43ideally have a stable IP address and
2:45country of connection.
2:47Stable my ass.
2:49Let's start, of course, with getting
2:52access to Claude, since it's a bit more
2:54of a hassle, and you'll need a foreign
2:57phone number. Getting a number for
2:59registration nowadays isn't difficult
3:02at all. Easy. Take literally any
3:05virtual phone number service. For
3:07example, this one. I repeat, any
3:09service, I don't give a damn which one.
3:13For those who had the thought that
3:14maybe my phone number will be given to
3:16someone else and they'll hack my
3:17account. No, guys and ladies, that’s
3:20not how it works. We choose the SMS
3:22service we need and select the required
3:24country. Personally, I recommend
3:26matching the country to your VPN
3:29location. This will reduce the risks
3:31and hassle from Claude, albeit only
3:32slightly. And here is an important
3:34point. Once you secure this number, it
3:37will no longer be offered to anyone
3:40else for registering with this service.
3:44That’s it, enter the number, enter
3:45the phone number, and we’ve got
3:47Claude in our pocket. Essentially, the
3:49registration principle for ChatGPT is
3:51more or less the same, but there is an
3:53even easier way, so we will look at it
3:55using the product from OpenAI as an
3:57example. We need a smartphone,
4:05basically any, but since the process is
4:07a bit more complex for Apple users,
4:08among the new things—I got a haircut
4:10today, and we set up an actual camera
4:12now, damn it, not a webcam.
4:14I will show you using the iOS system as
4:15an example. The first thing you will
4:17encounter is that the app is missing
4:19from the store. For this, you need to
4:21change your App Store region. Here is a
4:24quick guide on how to do it. In short,
4:26just buy a Google phone. Google is OP.
4:29Alright, I'm kidding. Open settings,
4:31tap on your name, Media&Purchases, View
4:34Account, Country/Region, and select the
4:37country we need. I personally use the
4:40USA. Google an address and paste the
4:43data. And also a phone number. Feel
4:45free to just tap random digits. No
4:47confirmation will be needed. And you
4:50don't need to specify payment details
4:52for now either. Before changing the
4:54region, make sure you've spent all your
4:56cash and canceled all active
4:57subscriptions, otherwise the system
4:59won't let you change the country. Oh,
5:01and almost forgot, also disable it; if
5:02you have a family and Family Sharing,
5:04that will also need to be turned off.
5:06And, actually, after these simple
5:09manipulations, you can easily install
5:11the ChatGPT app on your device,
5:13authorize via Apple ID, and you're all
5:15set. Access is obtained on Android. For
5:19you, in principle, it will be enough to
5:20just download it. To launch the apps,
5:23you will also need a VPN. So, that's it
5:26. And now we have the question of
5:28paying for this good stuff. Take the
5:30minimum plan for testing, then adjust
5:32it to your specific needs and tasks
5:34later, and we will talk about that soon
5:37too. And when you finally choose a plan
5:40that suits you, when upgrading the
5:42subscription, you will only pay the
5:44difference between them, rather than
5:46making a full payment a second time.
5:49This is an important point for many, I
5:51assume. So, there are literally two
5:54options for international payment from
5:56Russia. It’s a foreign card, either
5:59real or virtual, and payment via Apple
6:02ID and Google Pay. Let's look at both
6:05options. I'll tell you which one is
6:07preferable, and naturally, I'll explain
6:09how to get them. There are quite a few
6:11services for issuing instant cards
6:13right now. There are instant cards that
6:16only work for a month; they don't
6:17support additional top-ups. Roughly
6:20speaking, here is how it works. You buy
6:22a $ 100 card, and you have a card for a
6:26month with a $ 100 limit. Whatever you
6:29didn't manage to spend, naturally, is
6:31gone. And then there are yearly instant
6:33cards. They are at your disposal for
6:35the entire year. You top it up through
6:37an operator. And by the way, this is my
6:39choice. That's how I make a hell of a
6:42lot of international pur-pur-damn
6:44purchases, goddamn purchases. There is
6:46also a third option. You can issue a
6:48real plastic card in your own name.
6:50Almost every other bank provides this.
6:53Even MTS does this, among others.
6:55Yearly instant cards have an additional
6:58positive side. Complete anonymity.
7:01There you go. I personally use services
7:03from USA Virtual Cards, but I honestly
7:06don't know if they take new clients, as
7:08it's not a mass-market service; I've
7:10been working with them since, I think,
7:13February '25. They actually helped me a
7:15great deal with registering a Steam
7:17page. If you happen to see this video,
7:19then thank you very much once again.
7:22The kid needs help, otherwise he'll
7:24grow up to be a gangster. Actually,
7:27having a foreign card in your holster,
7:29paying would seem not so difficult. But
7:33before I break down why paying just
7:35anywhere is bad, let's look at the
7:37second option. The second method is
7:40ridiculously simple. Since you have
7:42already changed your region on your
7:44iPhone, you simply buy App Store
7:46balance top-up cards. And so you
7:48downloaded ChatGPT or Claude, logged
7:50into your account, and selected the
7:52necessary subscription option. I will
7:55add from myself here, you can buy cards
7:57on the Plati Market resource or, again,
8:00damn it, on MTS. I look at where the
8:03offer is more profitable each time, and
8:04I pick based on the situation. And I
8:06also recommend using the US region,
8:08since you can buy larger denominations
8:11in one gift card there, instead of
8:13punching in five to seven codes. Now
8:16the most important thing, listen
8:17carefully. It doesn't matter if you
8:20have an Android or an iPhone. It is
8:22advisable to always pay for the account
8:24via a subscription from your phone. Let
8:26me explain why. Those jerks at OpenAI
8:29cannot see exactly which card you paid
8:30with if the payment is made through the
8:32App Store or Google Play. And it's
8:35harder for their algorithms to detect
8:36your pattern for a ban. Translating
8:38into plain English. When you use a
8:41default user setup—like opening a
8:43browser from Latvia, entering a German
8:45phone number, and then using a new card
8:48from, say, the US, while your
8:49connection country keeps changing—
8:52that is a clear red flag for an
8:54Anthropic ban. But a regional VPN tied
8:57to the same country as your phone
8:59number makes you look slightly more
9:00legitimate. And if you also pay
9:03consistently from a single source,
9:04getting banned becomes, to put it
9:05mildly, quite difficult. Personally, I
9:09linked a virtual card for a year to my
9:10Google account on my phone, and
9:12that’s how I pay for both of these
9:14services via Google Play. I’ve
9:16explained how to get the chatbot. So,
9:19you can already handle trivial things
9:21like finding a barbecue marinade recipe
9:23or generating a greeting card for your
9:25beloved grandma. And if that was all
9:28you needed, you can close the video.
9:30But what if we want something more than
9:33just a smart Google?
9:36Mount Everest in the background behind
9:38me
9:39is [ __ ].
9:40For example, teaching it to work with
9:42large datasets, write code, create
9:44interfaces, find bugs or
9:45vulnerabilities on your servers, and
9:47patch them on the fly. And doing all of
9:50this autonomously, too. You can
9:51literally teach it to do anything, and
9:54you won’t even need top-tier hardware
9:57for it. You just need a pinch of
9:59perseverance and desire. I hope that
10:02after this video, the number of "
10:04wipe-coders" will decrease. Okay, you
10:20stayed.
10:21My respects. And here is the first
10:24question likely spinning in your head
10:25right now: "I’m already chatting with
10:27it in the browser, and I’m perfectly
10:28happy with that." Why would I need some
10:31terminal? Remember that sketch where a
10:33guy uploads a project archive to the
10:35chat and then copies the answer back?
10:37Essentially, you’re just a biological
10:38bridge between the chat and the project
10:40. You carry a file over, take the code
10:42out, paste it, something doesn't work,
10:44so you take the files back to the chat.
10:47It’s basically just being a monkey
10:48with two windows. A terminal on your
10:50computer breaks that whole cycle. It
10:53sits right inside your project folder,
10:55sees your files, reads them on its own,
10:57edits them itself, and runs commands.
11:00You don't copy-paste anything. You
11:02literally say, "Fix this crap for me."
11:04And it goes and fixes it right on your
11:07drive. A more advanced user might have
11:09another very smart question. "But
11:12there’s a desktop version, a window
11:14with buttons.""Why do I need a terminal
11:16?" And yes, it is literally the same
11:19Claude Code or Codex, just wrapped in
11:21an interface. If you want to start with
11:23that, then go ahead. There is nothing
11:25shameful about that. But I am showing
11:28you the terminal. And that is why the
11:30terminal is for when you want
11:32everything to be everywhere, completely
11:34on a server with no window at all,
11:36where you can do everything via SSH
11:38from your phone in another country. And
11:41it will all be wrapped and covered in
11:43your own scripts. And everything we'll
11:46be running later? MCP, DCP, skills,
11:49agents. For the terminal, this is a
11:51native environment. Once you master the
11:54console, it works anywhere for you
11:55without being tied to a pretty
11:57interface window. Let's move on to the
11:59installation. First, I'll show you the
12:02whole pack we'll need so you don't run
12:04around in a panic later, shouting, "
12:06Where did this come from?" And the list
12:09is quite short. First is the terminal.
12:11It’s already installed on Windows and
12:13Mac. Next, install Git from git-scm.com
12:17and just click next, next, next.
12:20Actually, the Claude Code itself. And
12:23here is a vital note. You don't need to
12:27install NPM or any of that other junk
12:29mentioned in old guides. Before, yes,
12:32it was necessary, but not anymore. The
12:34native installer pulls everything
12:36itself without dependencies with just
12:37one command. And fourth, you need to
12:40have your VPN turned on before you even
12:42start the installation. Because it
12:44downloads from Claude AI, and you know,
12:47that’s blocked for us. For your
12:50convenience, I will attach a file with
12:53all the links and all the commands that
12:55will be mentioned here. Basically,
12:59hover your mouse over Start,
13:02right-click, and select Terminal as
13:04administrator. PowerShell will launch.
13:09We enter this line there or copy it
13:11from my file. A couple of seconds and
13:15it's done. And here is the main trap
13:17that many people stumble over. We close
13:19this terminal window and open a new one
13:22. And in the new one, we type the
13:24cherished word: "claude". We agree to
13:26the authorization and log into the very
13:29account we created and paid for
13:31together. As for the Code app.
13:33There is a simple rule for Code.
13:35You just need to open the site, log in,
13:37and download the desktop version. Or,
13:39if you’ve already installed Claude
13:41Code, you don't even have to get off
13:43your ass; you can just ask it to
13:45install the desktop app for you. And
13:47now, the most important part. Why will
13:49this desktop version be more than
13:51enough for you? You might never have to
13:55open the terminal for Code, because
13:57their desktop version can actually work
13:59with both MCP and the browser. Meaning
14:02you don't have to write a single line
14:03in the terminal. It handles agents, too
14:06. Sub-agents, multi-agent delegation.
14:09Tasks in the cloud. All of this lives
14:10right in their windows. You toss a task
14:12at it, it processes everything on its
14:14servers, and spits out the finished
14:16result for you. Which means you can
14:18throw three tasks at it in parallel and
14:20go grab a smoke. It can also use your
14:22computer, literally look at the screen,
14:25and click on various buttons with the
14:27mouse. And I'm not telling you not to
14:47install the Cursor terminal because
14:50it's weak. On the contrary, they have a
14:52damn impressive desktop version. I'm
14:54saying don't install it because if you
14:57need a terminal, you should already
14:59have a different terminal pet. That's
15:01Claude-Code. Why would you need another
15:04mutt digging around in your console?
15:06The logic here is simple. Claude-Code
15:09sits in your system, in the terminal,
15:11digs through files, runs commands on
15:13the server, and writes code. These will
15:16be your working hands. And you keep
15:18Cursor in a nice window for other
15:20things. You throw tasks at it to
15:22generate something, look at diffs, send
15:25big files to the cloud, audit Claude,
15:27or make a clear plan and project
15:29architecture. In other words, these are
15:32two different agents. They have
15:34different roles, and each of them is
15:35good at its own. But the Cursor
15:38terminal exists too, and it's also
15:40powerful. And for those who still want
15:43to install Cursor, all the console
15:44commands will be in the same file in
15:46the description. The process is
15:49essentially exactly the same, but I
15:51personally use both of these models and
15:54recommend you do the same. So, which
15:57model brains should you use? Because
15:59both Claude and Cursor are, well, not
16:02just one neural network. They are
16:04multiple models that will eat up your
16:06budget in different ways. First, I'll
16:09explain the main principle, as it's the
16:11same for both tools. The smarter the
16:14model, the slower you get an answer,
16:16and the faster it burns through your
16:18balance. Therefore, the rule is simple:
16:22use defaults for routine, and top
16:25models for heavy tasks. In Claude, you
16:29switch models with the ` model` command
16:31. There is Sonnet, a real workhorse.
16:34You can run audits and pick out
16:36articles. OPUS is a real professional
16:40tool. It handles complex architectures,
16:44bug fixes, major refactoring, MCP
16:46bridges, creative work, and coding.
16:49Actually, the biggest chunk of tasks in
16:52AI coding is done specifically by Opus.
16:55And Sonnet 3.5 or similar is top tier.
16:58It’s like, above Opus, designed to
17:00work with large contexts and really,
17:02truly super-heavy tasks. I personally
17:06advise using it exclusively as an
17:08orchestrator for other models. In
17:10Cursor, everything is basically the
17:12same, just with different names: Haiku,
17:14Sonnet, and Opus. Opus is the flagship;
17:17it's the most powerful thing for coding
17:20, long autonomous tasks, and handling
17:22tools better than anyone else. After
17:25choosing a model, you need to decide
17:27how hard we're going to push it. And
17:30this is where the tools differ, so pay
17:33attention. In Claude, it's a separate `
17:36eff` command, and in life, three
17:38settings will be enough for you. Normal
17:40mode is set by default. You don't need
17:42to touch anything. On it, Claude
17:44already thinks, well, quite reasonably.
17:46That is enough for many everyday tasks.
17:49If you need to turn up the heat, then
17:51set it to Haiku. When you've gotten
17:53into a real mess, found some bug or a
17:56wrong architectural decision, you type
17:58Ford X, and it starts thinking
18:00noticeably longer and deeper. It will
18:03be more expensive in terms of limits,
18:04but, believe me, it is worth it. And
18:06the ultra-code mode is, you know, an
18:08analog of "do everything yourself."
18:10This is a mode for multitasking. For a
18:13big task, Claude, strictly speaking,
18:15stops working alone. It starts breaking
18:17it down into subtasks on its own and
18:19runs several agents in parallel. One,
18:22for example, digs into the code,
18:24another fixes something there, the
18:26third checks after the first two, and
18:27so on. I hope, damn it, I explained it
18:31clearly. Well, with Codex, it’s
18:33organized a little differently. There
18:36is this slider, like easy mode, medium,
18:41high, very high, and ultra. Also, we
18:45need to configure access permissions.
18:48This is the last toggle switch, and
18:51it’s purely about trust. By default,
18:53Claude is quite the paranoid. Before
18:56every file write and every command in
18:58the terminal, it will ask for your
18:59permission to perform a given action.
19:02And on one hand, that’s probably
19:04right, because that way it won't wipe
19:05out half your system while you sleep.
19:07But there is a catch. You can’t throw
19:10a big task at it and go drink your
19:12macchiato or, god forbid, vodka,
19:14because it will keep bugging you. "Can
19:17I do this? Can I do that? And this?"
19:20Every [ __ ] 10 seconds. And the
19:23solution to this headache is a simple
19:25key combination: Shift + Tab. Press it,
19:28and Claude switches between trust modes
19:31. By default, it will ask, as I already
19:33said; it’s a good mode when you’re
19:35doing something really important and
19:37want to track every step. There is also
19:40auto-approval of edits. It will write
19:43and edit files on its own without
19:44questions, but it will still ask for
19:46your permission for terminal commands.
19:48This is a happy medium for active work.
19:52There is also a planner mode and the
19:55choice of real men: auto-mode. There it
19:57is, this miracle. It makes decisions
19:59itself, writes itself, executes itself,
20:01and goes straight to the result. You
20:04don't touch a thing, but you understand
20:05that if something goes wrong, it’s
20:07not Claude who will [ __ ] their pants,
20:08but you. Think about it. With Codex,
20:11strictly speaking, the logic is the
20:13same. In the desktop version, we just
20:15choose the permission level, crank it
20:17up to full access immediately, or
20:19you'll find the terminal commands in
20:21the file. They’re just called
20:23something different there. And let’s
20:25clear this up right now, seriously.
20:28Auto-mode is not a money button.
20:30In short, I need a company tomorrow. I
20:32want to be in Silicon Valley recording
20:34podcasts, not doing all of this.
20:36Let's go.
20:38You can't just turn it on and forget
20:40about it forever. The more freedom you
20:42give it, the more carefully you should
20:44watch what it does. Especially at first
20:46, and especially on a production
20:48project. Yes, it's a powerful feature,
20:51but it’s not without its flaws, I’m
20:53telling you right now. It will very
20:56often try to screw you over. So, for
20:59everything important, unfamiliar, and
21:01especially production-related, run it
21:04in standard mode. If it’s a routine
21:06task where you clearly understand
21:08what’s happening, choose auto-apply
21:10edits. Or if it’s a big task in an
21:13isolated sandbox and you don't mind
21:15rolling things back, then auto-mode
21:18will really suit you. And the golden
21:21rule: keep version control, make
21:23backups, always, always. Do you hear me
21:26? Okay, we’ve learned how to set all
21:28this gear up, but Claude out of the box
21:31is, you know, a generalist who can do
21:33everything, but only a little bit.
21:35What can you do? Everything. Shoot,
21:39make halva, shoe a stallion, crack a
21:41safe, forge documents, deliver a baby,
21:43write an article, sing in a choir. And
21:48skills are when you give it a memo.
21:51This is how this specific task should
21:54be done. A technical skill is just a
21:56folder with a skill.m file, and inside
21:58there's plain text with the
22:00instructions. Not so difficult, is it?
22:03And here is the first thing important
22:04to understand. Skills can be very
22:07different and for completely different
22:08professions. These aren't just tools
22:11for programmers; they are a set of
22:13rules. For example, a writer's skill:
22:16how I structure a chapter, my editing
22:18style, script format, or character
22:21prohibitions. Whatever your imagination
22:24allows: for coders, for instance,
22:26it’s how to do reviews in our team,
22:29our style guide, how to format commits,
22:32or what not to do. 3D model skill for
22:34the Blender workflow, especially in
22:37connection with the MCP bridge. But
22:39more on that later. You see, for a
22:42marketer, a lawyer, an analyst—
22:44everyone will have their own specific
22:46skills. And there are two sources to
22:49get, pay attention, ready-made skills.
22:52And the difference between these two
22:54sources is just a difference in trust.
22:57There is an official store from
22:59Anthropic: Claude Plugins Official. It
23:01is already connected from the first
23:03launch in the terminal. Nothing extra
23:05needs to be set up. Essentially, all
23:08the top-tier skills vetted by Anthropic
23:11are there: working with documents, PDFs
23:14, docx, presentations, spreadsheets,
23:16web app testing, frontend, and so on,
23:19and so on. It’s like the App Store:
23:22clean, safe, and everything works. And
23:24there are unofficial community stores.
23:27These are just git repositories
23:29gathered by enthusiasts. There is a
23:31wild amount of stuff there. Hundreds of
23:34skills for any profession and taste,
23:35but you could load up total garbage,
23:37blow up, and splatter the curtains.
23:39Installation is quite simple.
23:41Everything revolves around one single
23:43command: slash-plugin. A plugin is just
23:46a wrapper containing one or more skills
23:49. On the screen, you can see both the
23:51breakdown and an example of installing
23:53the Code Review plugin. This is a real
23:56first-party plugin from Anthropic. Well
23:58, naturally, for code review. By the
24:00same principle, any plugin or skill
24:03from the store is installed. And if
24:05it's a third-party store, say, from
24:07enthusiasts on GitHub, the principle is
24:10the same. Only the repository address
24:12changes. Or you can just download the
24:14skill folder from GitHub and drop it
24:16into the right place. As for the right
24:19place, that's a key point many people
24:21skip. There are two places, and where
24:25you put them determines where the skill
24:29will work. There is the root folder of
24:32Claude. Whatever you put there will be
24:34available to you in all projects. Put
24:37only universal things here—your style
24:40, your habits. And there is a Skills
24:43folder inside the project itself.
24:45That’s a project-specific folder. The
24:48skill stays right in the project
24:49repository and travels with it. And if
24:51you commit it, it will appear for the
24:53whole team. Put things here that are
24:56specific to a particular project. And
24:58following the previous point, there's
25:01another common mistake beginners make.
25:03Don't launch Claude from just anywhere;
25:06launch it from your project folder.
25:08Press Windows + R, type the CMD command
25:11, and enter the path using CD. If your
25:15project isn't on the system C drive,
25:17you need to type the drive letter in
25:19the terminal, like drive D, and then CD
25:21. Or, if you have Windows 11,
25:25right-click the folder and select "Open
25:26in Terminal" from the menu. Why is this
25:29so important? Project skills reside
25:32inside the project. If you launch
25:35Claude from your home folder or the
25:36desktop, it simply won't see them.
25:39They're in a different place. And in
25:42general, Claude launched from the
25:44project folder sees your files,
25:46structure, and context, whereas one
25:48launched from anywhere else sits in the
25:50dark, not fully understanding what it's
25:52working on. But all that is just
25:55messing around. You need to create your
25:58own custom skills. You see, someone
26:01else's skill is someone else's way of
26:03working. It’s average, just like
26:05everyone else's. But your value should
26:08lie in how*you*do things specifically.
26:11Your skill is your experience, your
26:13rules, your lessons learned, all
26:15written down once. And it’s literally
26:18just a text file. You don't need to
26:20know how to program; you need to be
26:22able to describe how you work and know
26:25how to think. And just ask Claude to
26:27format it all into a skill. That's it,
26:30damn it. Judge for yourself how simple
26:33it really is. Naturally, it’s, well,
26:35a pretty primitive method. For example,
26:37how do you do this correctly? You open
26:39a Word document and write: "I know
26:41there are IDOR vulnerabilities, I need
26:44to check the server for IDOR, check the
26:46server for Nmap, and
26:48blah-blah-blah-blah-blah-blah-blah-blah-blah
26:51." Well, if you're really, if you're
26:54just too lazy to format it properly,
26:56you just feed that Word file to it and
26:58say: "Create a skill for such and such
27:00tasks or for such and such a project."
27:03That's it, feed it a pipeline for an
27:05MCP bridge, for example. Well, I'm
27:07getting ahead of myself. Alright, let's
27:08get back to the topic. So don't rush to
27:10install 100 skills right off the bat.
27:13Start with the built-in ones. Add the
27:14official ones, run them, and test them
27:17out. And write your own when you've
27:19done the same task ten times and are
27:21sick of explaining the same thing to
27:22the AI over and over. That’s when you
27:25write a skill file, and you never have
27:27to repeat yourself again. And it's all
27:29the same for Code. You can even migrate
27:32these skills between them. Code
27:34understands Claude's skills and vice
27:36versa. So, this principle of operation
27:40works for both. Okay. Up to this point,
27:43we've been running Claude in a sandbox.
27:45It read files, wrote code, and ran
27:48commands in the terminal. That’s all
27:51powerful, but it's still locked away.
27:53MCPs are bridges that give it hands to
27:55interact with real programs. For
27:58example, Blender, Maya, Unreal, Excel,
28:01GitHub, browsers, databases, and much
28:04more. But most importantly, this isn't
28:07a toy. MCP is a standard that Anthropic
28:09came up with, but everyone has jumped
28:11on board. OpenAI, Google. One bridge
28:15works with a bunch of agents. And this
28:17is no longer just an enthusiast's
28:19experiment. That is the true power of
28:21this crazy piece of tech. It used to be
28:24like: "You want to automate Blender,
28:26you sit down and write a Python script
28:27for every single task. You change the
28:30Blender version, the script might break
28:31, and you have to fix everything." New
28:34task, new script. You’re your own
28:37programmer, debugger, and tech support.
28:39It’s literally just busywork for the
28:41sake of busywork. With MCP, everything
28:44has changed. The bridge gives the agent
28:46a ready-made set of tools for the
28:48program. And which tools to pull and in
28:50what order, the agent decides for
28:52itself. You write to it in plain human
28:55language. Like, for example, place
28:57lights along this spline and make the
28:59key light warmer. Say, 3200 Kelvin, and
29:03it just goes and does it. So, you
29:05connect it once and can use it with any
29:08agent. So, what are some of the most
29:10popular MCP bridges, for instance?
29:13Unreal Engine 5 version 5.8 has a
29:15native one from Epic, but it's still
29:18total garbage. There's a paid bridge
29:20with 1,000 tools. It's called Ultimate
29:23Copilot Engine. And for 3D modeling,
29:26for example, there is an official
29:28bridge for Blender. Blender MCP. The
29:31agent moves objects, applies shaders,
29:32sets up lighting, renders, and models
29:34everything live right in front of you
29:35in an open Blender window. There's one
29:38for Maya with essentially the same
29:40principle, but be prepared to write a
29:43damn wall of instructions, and it will
29:46be working from scratch. That’s
29:49dead-on-arrival crap. Anyway, there are
29:53also these MCP bridges for Excel and
29:55Chrome, so it can Google whatever you
29:58need, work with Excel spreadsheets, and
30:01even for that, what is it, for CG
30:04graphics, damn, I want to say Painter.
30:09Damn, the word Photoshop completely
30:10slipped my mind. There is an official
30:13bridge for that, too. And that is just
30:16a drop in the bucket. There are
30:17literally thousands of bridges. What if
30:20there’s no bridge yet for the
30:22software you need? No problem. MCP is
30:25an open standard, not some closed-off
30:27gadget. You can write any bridge
30:29yourself. And you'll be writing it with
30:32the right tools, like Claude or Codex.
30:35Just think about it. The agent is
30:37essentially building itself new hands.
30:40You tell it: "I have this software with
30:41such-and-such API, but there’s no MCP
30:43bridge for it.""Let’s make one." And
30:46you sit down and do it. Since there are
30:49official SDKs for MCP, Python, and
30:51TypeScript, Claude and Codex know them
30:53perfectly well. It spins up a server
30:56skeleton, describes the tools, connects
30:59to your API, and you get a bridge for
31:01your niche software that didn't exist
31:03before, all without you writing a
31:06single line of code yourself. That is
31:08the level that makes all this fuss
31:10worthwhile. You stop being a hostage to
31:12what someone else has already written.
31:14Need to integrate custom software? An
31:17internal studio tool, some crazy
31:19machine API? You tell the agent, and it
31:22builds the bridge together with you.
31:24There is no ceiling. And one of the
31:27most important things is memory. The
31:30agent has two types of memory, just
31:33like a human.
31:34It has a 16 MB context window.
31:38This is the short-term session context.
31:41It includes everything you’ve talked
31:42about in the current conversation.
31:45It’s dynamic, but finite, and it
31:47degrades as it fills up. When the
31:51context gets clogged, the agent starts
31:53acting dumb, asking things you've
31:54already told it a thousand times,
31:56forgetting its own decisions, mixing up
31:58file paths, and contradicting itself.
32:01It’s not that Claude is broken; its
32:03brain is just overflowing. And there is
32:06long-term memory, strictly speaking,
32:08that very clot.md file. The Codex has
32:11the same thing too, for your
32:13information. Just a reminder, we are
32:15simply analyzing this using Claude as
32:17an example. So, long-term memory is
32:20memory that lives forever. Claude loads
32:23it at the start of every session
32:25automatically. It's like his project
32:27passport, his own rules you've set with
32:30him, commands, agreements, and those "
32:32rakes" you’ve already bumped your
32:34head on plenty of times. So, what is
32:37clot.md, anyway? It's just a markdown
32:40file in the project root. It records
32:44what Claude shouldn't forget between
32:45sessions and shouldn't have to guess,
32:48like how to build or test the project.
32:51The architecture, in a nutshell, your
32:53agreements, your style. Always do it
32:56like this, or never do it like that. In
32:59short, whenever you start a new project
33:02, be sure to create long-term memory
33:04files. And a primary clot.md file. And,
33:08naturally, a file for the "rakes"
33:10you've hit and any critical decisions
33:13made. There should be several files,
33:16ideally about four of them. The main
33:18rule for working with memory is to
33:19remember the following. Your favorite
33:22command is/clear. It wipes the context
33:26completely, after which you switch to a
33:28totally different task. Once you've
33:32finished a task, recorded it in
33:34Claude's memory, run/clear and move on
33:38to the next task, because an
33:40overflowing context is a guarantee of
33:43poor work. The main thing is to
33:46understand that the agent has a leaky
33:48memory; within a session, the context
33:49will fill up and that's it—he gets
33:51slow, asks redundant questions, and
33:52confuses paths. This is fixed by
33:54keeping important information not in
33:57the chat, but in files he picks up
33:59himself. So, what kind of files should
34:02you maintain? Well, clot.md, as we
34:04already said, the project passport. And
34:07the main thing is to keep it short—
34:09under 200 lines is ideal, otherwise
34:11you'll stuff in a huge sheet, and he'll
34:13simply start ignoring it and drown in
34:15it. Next, you should create separate
34:18files for things that accumulate. For
34:20instance, you should create. Those same
34:23"rakes"—things you stepped on and
34:25where you went wrong. For example,
34:27library X conflicts with library Y, or
34:29never touch this config manually under
34:31any circumstances. /clear, oops, wrong
34:34place. The next file is bugs. There you
34:36keep known bugs and their status,
34:38whether they've been solved or not.
34:40Because if you run into something that
34:43has already been solved, or something
34:45similar, you'll deal with the problem
34:47quickly. The next file I would
34:50recommend keeping is decisions. For
34:53example, these are your reasons why
34:55things are done this way and not
34:57otherwise. Without this, in a week, he
34:59will "improve" your created solution
35:01right back into garbage. Believe me,
35:03damn it, this will happen more than
35:06once. And the next file is a to-do list
35:09, meaning what you have in progress and
35:11what you intend to do next. Another
35:14file is style.md. Style is strictly
35:18about how you write code or text for
35:21articles, for instance, like that bold
35:26dash which gives away your use of AI.
35:30In the style guide, you can write: "
35:31Don't you dare, you dog, use this
35:33symbol" or other special characters.
35:36And most importantly, in the main file,
35:39specify that these files exist and
35:41should be referenced under such and
35:44such circumstances. Now let's talk
35:46about multi-agents. Multi-agents are
35:49when Claude builds a micro-team of
35:51micro-people and distributes chunks of
35:54work to assistants who grind
35:55simultaneously. In short, multi-agents
35:59give us a speed advantage, because
36:02three agents working in parallel are
36:04faster than one performing three tasks
36:07sequentially. Secondly, it gives us
36:10resource allocation. Meaning, a simpler
36:13model is sent to a simple task, and a
36:15more power-hungry one to a complex task
36:18. And each model will use its own MCP
36:22bridges, tools, and skills, but don't
36:24overdo it either. No more than three to
36:27five assistants at a time; that's the
36:30ceiling of common sense. Any more and
36:32you'll just be exhausted stitching
36:34together what they produced and you'll
36:35lose more than you saved. You can only
36:38parallelize independent tasks. If tasks
36:40are linked—first this, then that—
36:42they cannot be parallelized. Roughly
36:45speaking, you wouldn't send two
36:47employees to the same store, one to buy
36:49bread and the other to buy milk.
36:52Exactly. And the logic here is
36:53essentially the same, so it won't be
36:55difficult to remember. In this short
36:58video, I've given you all the basic
37:00tools. This is literally all the
37:03knowledge you need to start out better
37:05than 90%of other beginners. I tried to
37:08explain everything as simply as
37:09possible, but if you're too lazy to
37:11delve into it and want it all broken
37:13down for you, then welcome. A "
37:16Financial Masochist" tier subscription
37:18on Boosty will let you pick my brain
37:20for 2 hours a month. Not necessarily
37:22all at once. You can break it up into
37:24half-hour or hour sessions. This video
37:26was about what tools exist and how to
37:29set them up. But don't think that
37:31knowing all these tools will
37:33automatically make you stronger than
37:34everyone else. The biggest difference
37:38between a person who finishes a project
37:40in a week and someone who struggles
37:42with an AI for a month isn't Claude,
37:45Cursor, MCPs, or the model—it's how
37:47they think and define their tasks. If
37:51you're interested in a dedicated video
37:53on thinking alongside artificial
37:55intelligence, prompting without the
37:57snake oil, rules for working with
37:58agents, vibe coding in general—when
38:00it actually speeds things up and when
38:02it just turns you into a drone—or
38:04just the philosophy of using AI in
38:06general, let's hit 3,000 likes. You
38:09won't do it, which is why I'm feeling
38:11bold, and the next video will cover
38:14exactly that. But if this video was
38:17actually useful, leave a like,
38:18subscribe, drop a comment, and if you'd
38:19like to support the project for my
38:21future game, all the donation links are
38:23in the description. Thanks for watching
38:25.

About this transcript

This page contains the full transcript of УЛЬТИМАТИВНЫЙ ГАЙД НА ВАЙБ-КОДИНГ by OneDray, generated from the public captions YouTube serves with the video. The transcript has 6,166 words across 954 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.