Дай мне 1 час и ты получишь 1000 часов знаний о Codex (2026) — Transcript
Full transcript
- 0:00Today we are going to break down 24 key
- 0:02concepts of Codex. This is exactly what
- 0:05you need to know to master Codex by 95%
- 0:08and start using it today, without
- 0:10spending weeks studying it. And it
- 0:12doesn't matter if you know how to
- 0:14program or not. By the end of this
- 0:17video, you will understand how Codex
- 0:19works, and you will have a fully
- 0:20functional system that will work even
- 0:22while you sleep. Let's get started.
- 0:25Before the first part, let's first take
- 0:27care of the installation and purchasing
- 0:28a subscription. To download the Codex
- 0:31application to your computer, type "
- 0:33Codex download" into your browser and
- 0:35go to the first link from openai.com.
- 0:38As soon as you enter the site, there
- 0:39will be a download button on the main
- 0:41block. Click on this button and
- 0:43download it. There will be no
- 0:44difficulty with the installation of the
- 0:46application itself. The first problem
- 0:48that may arise is registering a new
- 0:50account. If you are registering in
- 0:52Codex for the first time, you will be
- 0:54asked for a phone number. And this
- 0:56phone number, naturally, must be a
- 0:59foreign one. If you don't have a
- 1:00foreign phone number, there are two
- 1:02options. The first is to find a service
- 1:05for renting foreign numbers. And the
- 1:07second option, the most reasonable one,
- 1:09is to find an acquaintance who has a
- 1:11foreign phone number. Well, another
- 1:13option is to buy an account on Avito,
- 1:15for example, but that's probably not a
- 1:17very reliable thing to do. I wouldn't
- 1:18recommend doing that. That's it, you
- 1:20verify it once, and as a rule, it
- 1:22doesn't ban you and doesn't ask for the
- 1:24number again. At least, as of the time
- 1:26of recording this video. After you
- 1:28install Codex on your computer, you
- 1:30will need to choose a paid plan. If you
- 1:33need Codex, then only a Plus
- 1:34subscription or higher will work. The
- 1:37Plus plan has all the most powerful
- 1:39models, has Codex, and you can already
- 1:41fully develop something with it. The
- 1:44only limitation of this plan is its
- 1:46limits. If you are going to use some
- 1:49powerful models or handle some large
- 1:51tasks, these limits will not be enough
- 1:52for you. Therefore, for those who want
- 1:56to work fully and powerfully with Codex
- 1:58, you need to buy the Pro plan for $
- 2:00100 or $ 200. On the $ 100 Pro plan,
- 2:02you will have five times more limits
- 2:04than on the Plus plan. Previously, the
- 2:07$ 200 Pro plan had 20 times more limits
- 2:09than Plus, but now it has 10 times more
- 2:11limits than Plus. And recently, they
- 2:14also rolled out a $ 500-a-month plan. I
- 2:17don't know what kind of rich people
- 2:19this plan is for, but if you are
- 2:21working with some large tasks, really
- 2:23very large ones, then you might need it
- 2:25. On this plan, you have 25 times more
- 2:28limits than on Plus. And there's also a
- 2:31fast mode added, so you spend less time
- 2:33on a task. Meaning the model will work
- 2:37up to eight times faster, but you'll
- 2:38hit those limits much more often. We
- 2:41will talk about tokens, limits, and
- 2:43context a bit later. In short, if you
- 2:45are just starting to learn AI coding
- 2:47and trying to develop something, then
- 2:48get the Plus plan first. Now for the
- 2:51next problem. How do we actually buy
- 2:53this plan? To do this, you need a
- 2:56foreign bank card. To get one, you can
- 2:59either ask a friend abroad to make the
- 3:01payment or get yourself a virtual
- 3:03foreign card. I pay for my
- 3:05subscriptions using a Telegram bot, the
- 3:07link to which will be in the
- 3:08description. There are, of course, fees
- 3:10, but thanks to this bot you will still
- 3:12be able to pay for subscriptions. And
- 3:15one more alternative way to buy a
- 3:17subscription is to use the GGcell
- 3:19service or any other equivalent. With
- 3:21this service, you can buy a
- 3:22subscription using rubles. That's all.
- 3:25After we have registered a new account,
- 3:27installed the desktop application,
- 3:29selected a plan, and purchased it, we
- 3:32can finally start using our Cursor. So,
- 3:36concept number one is that Cursor and
- 3:38ChatGPT are different things. You can
- 3:41notice that we have ChatGPT, which has
- 3:44two modes inside: Chat and Work. And
- 3:47right here in the left panel, you can
- 3:49switch from ChatGPT to Cursor. In total
- 3:52, the application has three modes. The
- 3:53first one is Chat. This is our usual
- 3:56ChatGPT. Meaning we can ask it
- 3:58something here, find something, ask it
- 4:00to draft texts, and so on. Then there
- 4:03is the Work mode. This is a tool you
- 4:06can use to work with your computer,
- 4:07perform certain tasks, make documents,
- 4:09presentations, and a bunch of other
- 4:11things. We will talk about this in more
- 4:13detail later. And, accordingly, for
- 4:15working with code, for creating
- 4:18projects, automations, services, and
- 4:20websites, we have Cursor. Just so you
- 4:23know, the Work mode uses up your Cursor
- 4:25limits, but the Chat mode does not.
- 4:27Meaning you can talk to it for as long
- 4:29as you want, and your Cursor limits
- 4:31won't be used up. Concept number two.
- 4:33Cursor doesn't work in thin air; it
- 4:36works with a folder on your computer.
- 4:39The first thing we must do before
- 4:40working in Cursor is to create a folder
- 4:42on the desktop. So, I created a folder,
- 4:45named it Cursor, and after that, I go
- 4:48into Cursor, select the necessary tab,
- 4:50and click New Chat. Now we are in a new
- 4:53chat. And here we click Select Project.
- 4:56Next, click New Project, Add Folder
- 4:58from this computer, and select the
- 5:00corresponding folder. Click Open. Next,
- 5:03click create project. After that, it
- 5:05will ask if we trust this folder. We
- 5:07trust it. Codex will see everything
- 5:10inside this folder. All your files, all
- 5:13your spreadsheets, all your photos,
- 5:15anything at all. Concept number three
- 5:19is that we choose a specific model for
- 5:21a specific task. Choosing a model
- 5:24directly depends on how you spend your
- 5:26tokens and your limits. Let me explain
- 5:29what tokens and limits are, so you
- 5:30understand. A token is roughly a piece
- 5:33of a word. It can be a whole word, a
- 5:35part of a word, or some preposition.
- 5:37Our subscription consists of these
- 5:39tokens. Our subscription gives a token
- 5:42limit for a week or for 5 hours. The
- 5:45Pro plan does not have this five-hour
- 5:47window. It only has a weekly limit.
- 5:49That is, you are given a certain amount
- 5:51of limits for the week. And if you go
- 5:54over these limits, you won't be able to
- 5:55use Codex for the rest of the week.
- 5:58There is such a concept as input tokens
- 6:00. These are the tokens that you write
- 6:01to it. That is, all the text that you
- 6:03send to it. And not just text. Images
- 6:06also consume tokens. Any working files
- 6:08also consume tokens. In general,
- 6:11everything you send in your request is
- 6:13input tokens. And there are also output
- 6:16tokens. This is the model's response.
- 6:18Moreover, tokens are spent not only on
- 6:20the model's output that it writes back
- 6:23to you, but also on its reasoning. That
- 6:26is, while it is answering your question
- 6:28, it goes online, researches something,
- 6:30studies, looks through other files on
- 6:32your computer if they are needed for
- 6:34the task, checks some documentation,
- 6:36and so on. All of this also consumes
- 6:39your tokens. And we can influence all
- 6:42of this by choosing a specific model
- 6:43and a specific level of reasoning.
- 6:45Let's see what models are there and
- 6:47what they are used for. If we take the
- 6:50latest generation of models, that is
- 6:52the GPT6 generation, and 6.1, we have
- 6:55several models. These are the Luna
- 6:57model, the Soul model, and the Astra
- 6:59model. GPT6 Luna is accordingly the
- 7:02cheapest and the fastest. It is needed
- 7:05to work with simple, specific tasks.
- 7:08Reformatting something, renaming,
- 7:10adding a line of code, changing a
- 7:12button color, and so on. So here we are
- 7:15not dealing with any complex, large
- 7:17tasks that require deep reasoning. You
- 7:19can assign it any routine task that can
- 7:22be done easily according to a pattern,
- 7:24without applying any mental effort. For
- 7:26example, sorting emails, analyzing
- 7:29comments. Well, in general, these are
- 7:31not very complex, specific tasks. The
- 7:34GPT6 Astra model is the most powerful
- 7:37and expensive model that currently
- 7:39exists from OpenAI. It is needed for
- 7:41complex tasks where a high level of
- 7:43reasoning is required. For example,
- 7:46developing a project idea, analyzing
- 7:48project files, or finding a tricky bug
- 7:51that has crept into the code. In short,
- 7:53it is needed for complex tasks that
- 7:55require a lot of thought. And there is
- 7:58also GPT 6.1 Sou. It is almost as smart
- 8:02as Astra, but five times cheaper. This
- 8:04is what you will work on every day. And
- 8:06honestly, I use it for 99%of all my
- 8:09tasks. The main beginner mistake in
- 8:12this matter is choosing the coolest,
- 8:14most powerful model and using it to
- 8:15solve all your tasks. Why is this bad?
- 8:18Firstly, because when performing a task
- 8:20with a specific outcome, the Sou model
- 8:22will handle it just as effectively as
- 8:24the Astra model. But Astra will spend
- 8:27more time on it, firstly because it
- 8:29simply needs to think, and secondly, it
- 8:31will spend a lot of money, or rather,
- 8:33your tokens. In addition to this model
- 8:35selection, there is also a choice of
- 8:37reasoning level here. You see, it
- 8:39changes: light, medium, high, very high
- 8:42, and ultra. The higher the reasoning
- 8:45level you set, the more the model will
- 8:47think before answering you. A light
- 8:49reasoning level is when the model
- 8:51barely thinks about your task. That is,
- 8:55we will use it for specific tasks, like
- 8:57changing a button color, replacing a
- 9:00string with this one, making a specific
- 9:02edit, and so on. At a light reasoning
- 9:06level, it will perform such specific
- 9:08tasks quickly and cheaply. The medium
- 9:10level is accordingly higher, meaning
- 9:12the model starts thinking before it
- 9:14answers you. You can already use it for
- 9:17tasks where there is no specific answer
- 9:19and some reflection is needed. And,
- 9:22accordingly, the high and very high
- 9:24reasoning levels are for when you are
- 9:26analyzing code, finding bugs, or, for
- 9:28example, when you want to analyze a
- 9:30huge number of sources, do some
- 9:32research, and so on. This means the
- 9:35model will work longer and spend more
- 9:37of your tokens, but it will thoroughly
- 9:39think through your tasks. And there is
- 9:41also an ultra mode. This is a
- 9:43completely different level of model
- 9:44operation, and we will talk about it a
- 9:46little later. Concept number five is
- 9:48dedicated to permissions. Code-base can
- 9:50do various things on your computer.
- 9:52Create and delete files, rename files,
- 9:55change their location, run programs,
- 9:58access the internet, and so on. In
- 10:00general, it has a complete carte
- 10:02blanche when it comes to working with
- 10:03your data. Plus, if you connect a few
- 10:06more external services to it, it will
- 10:07also have access to them. And, of
- 10:10course, blindly trusting Codecks with
- 10:12your data is, as you can understand,
- 10:14not entirely safe. That is why Codecks
- 10:17has a special setting that is
- 10:18responsible for all of this. It is
- 10:20called permissions. It is located right
- 10:22here to the right of the plus sign. And
- 10:24there are three options here. The first
- 10:26is to request permission. In this case,
- 10:29it will ask for your permission for any
- 10:31changes to external files or before
- 10:33using the internet, that is, if it
- 10:35performs any operations online. Then
- 10:38there is the "confirm for me"
- 10:39permission level, where it will only
- 10:41request permission for potentially
- 10:43unsafe actions. For example, if you are
- 10:46about to delete an important file or
- 10:48make a change to the code that could
- 10:51affect a huge number of functions, it
- 10:53will ask for permission before all such
- 10:55edits. And another permission level
- 10:58that is here is full access. As you
- 11:01understand, it won't ask for any
- 11:03permissions at all here. Deleting files
- 11:06, moving files, removing or adding
- 11:08lines of code. It will do all of this
- 11:11without your permission. By default, it
- 11:13will most likely not be set to this. To
- 11:16activate it, you need to go to settings
- 11:18in the bottom left corner, find
- 11:20permissions here, and they will be open
- 11:21right away. And right here there will
- 11:24be a toggle switch, full access. You
- 11:26click it here and press confirm. And
- 11:28that is it, now you will have full
- 11:30access activated. As soon as we try to
- 11:32turn it on, it warns us that it is
- 11:34dangerous, that something might be
- 11:36deleted, or that a terminal command
- 11:38might be unsafe, but we click confirm.
- 11:41And honestly, I use this mode more
- 11:43often than the others. To be more
- 11:45precise, I only use this mode. Why am I
- 11:47not afraid of this? Modern models have
- 11:49already become quite powerful. And
- 11:51overall, I have never had a situation
- 11:53where Codecks for some reason deleted
- 11:55something important from my computer.
- 11:58But it is important to craft the
- 11:59request correctly here. If you just
- 12:02write to it, for example, "clean up my
- 12:04desktop," it might delete both
- 12:05unnecessary files and important folders
- 12:07you need. Or, for example, if the task
- 12:10is to clean up your computer to free up
- 12:12more memory. Here, too, it might delete
- 12:14both junk and various duplicates, as
- 12:16well as important, useful applications.
- 12:19In short, it is important to provide a
- 12:20correct, specific request without any
- 12:22double meanings, and then you will not
- 12:24have any problems with this level of
- 12:26permissions. Why don't I like the first
- 12:28two modes? Because it will constantly
- 12:30ask you for these permissions. That is,
- 12:33if you are working on a big task where
- 12:34it needs to think a lot and perform
- 12:36many actions, you might give it about
- 12:3820 permissions in a single request.
- 12:40Honestly, that gets tiring. It slows
- 12:42down work significantly, and you
- 12:44essentially can't switch to other tasks
- 12:46while it's running because you always
- 12:49have to click "allow." In short, when
- 12:51using full access, it is important to
- 12:53assign tasks correctly. Concept number
- 12:55six: the limit is a resource. It's time
- 12:58to explain how the context window works
- 13:00and how your limits are spent. We
- 13:02already know that the limit is consumed
- 13:04by the chosen model and the depth of
- 13:06its reasoning. Now let's find out where
- 13:09to see it. This very limit, and we'll
- 13:11learn how not to burn through it. I'll
- 13:13just type a slash into the input field.
- 13:16After entering the slash, a huge number
- 13:18of commands pop up, but let's find "/
- 13:20status." Here it is. We click on it.
- 13:23And you can notice that, again, there
- 13:24are no five-hour limits here. There are
- 13:26only weekly ones. And I've only used up
- 13:293%of this limit. If for some reason you
- 13:32hit the limit, OpenAI sometimes gives
- 13:35you the option to reset these limits.
- 13:38This is done in the bottom left corner,
- 13:40right here in the "Usage" tab. Let's go
- 13:42there. And right here, as you can see,
- 13:44I have two resets available. One
- 13:46expires on the 22nd, the other on
- 13:48October 29th. If for some reason I hit
- 13:51the weekly limit, I can easily restore
- 13:53these limits using this button. OpenAI
- 13:56provides this as a bonus. You can't get
- 13:59these resets just like that. How can we
- 14:01save these limits? The first rule you
- 14:05can apply is to brainstorm in ChatGPT
- 14:08and then work directly in Codex. As I
- 14:11said, regular ChatGPT does not affect
- 14:13the Codex limit. It has its own
- 14:15restrictions, but that's not the point.
- 14:16Again, this doesn't apply to Codex at
- 14:18all. So, if you want to discuss an idea
- 14:21, outline a structure, or ask questions
- 14:24before starting a task, you can do all
- 14:26of that using ChatGPT. After you've
- 14:29discussed the task with ChatGPT and
- 14:31figured out how to solve it, you can
- 14:33turn to Codex. Next, there is another
- 14:36rule: one chat, one task. That is, for
- 14:39example, you work out the design in one
- 14:41chat, you work out something else in
- 14:43another, integrations in a third, and
- 14:45so on. Why is this important? To
- 14:48understand that, we need to look at how
- 14:50context works. Let's say we have a chat
- 14:53. We give it a prompt, and it replies.
- 14:56That is, to answer us, the neural
- 14:58network processed our request. Then,
- 15:00after that, we input a second prompt.
- 15:02And to process it, the neural network
- 15:04must process our previous request and
- 15:07the answer to that previous request. So
- 15:10, in this way, to answer the second
- 15:11prompt, it is already processing more
- 15:13information than when it answered the
- 15:14first. And, accordingly, when you
- 15:18provide so many edits, reason with it,
- 15:20or give it new tasks, not only does it
- 15:23spend a lot of tokens thinking, but it
- 15:26also processes previous answers and
- 15:28requests to address the current one.
- 15:32And all of this accumulates. And, of
- 15:34course, when answering, say, the
- 15:36fiftieth question, it will spend a lot
- 15:38of tokens just processing previous
- 15:39answers. But it's not all that strict.
- 15:42We have a thing called caching. Here is
- 15:45the point. When the model has read the
- 15:47beginning of a chat, OpenAI remembers
- 15:48what it understood from it. For the
- 15:50next message, the beginning hasn't
- 15:52changed, so it doesn't need to be read
- 15:54again. It simply pulls the information
- 15:56from memory. But this caching has
- 15:58conditions. The first is time. The
- 16:00memory lasts for 30 minutes after your
- 16:02last message. Each new message extends
- 16:04it, accordingly, for another half hour.
- 16:08If you, for example, go away for two
- 16:09hours and then want to continue, there
- 16:11will be no caching anymore. The second
- 16:13point is switching models. If you
- 16:15change the model in the middle of a
- 16:17chat, that also resets the cache.
- 16:20Therefore, in long chats, it is ideally
- 16:21better not to change the model. In
- 16:24short, caching is a useful thing, but
- 16:25it still doesn't solve the problem of
- 16:27the context window. By the way, let's
- 16:29move on to the next concept. This is
- 16:31exactly the context window. That is,
- 16:33the model does not have memory in the
- 16:35usual sense. It has a context window,
- 16:38which is a limited volume that holds
- 16:40everything it sees right now. These are
- 16:43the instructions I give it, my files,
- 16:46chat history, command results, and so
- 16:48on. All of this together fills the
- 16:51context window. If you overload your
- 16:54context window, for example, over 70%,
- 16:56there is a very high probability that
- 16:58the neural network will miss something
- 17:00—some important fact you threw in or
- 17:03information you provided—and because
- 17:05of this, it will start to act slow.
- 17:08This is called context degradation. To
- 17:11somehow shrink the volume of the
- 17:13context window you have used, there is
- 17:15a command called compact. This is what
- 17:18it looks like. I click here, and it
- 17:20compresses the context of this chat.
- 17:22Meaning, it creates a separate summary
- 17:25from all of our previous chat. Thus,
- 17:27the context window size decreases,
- 17:29meaning we can work with it fully, but
- 17:31it might accidentally remove an
- 17:33important fact, and your performance
- 17:35will still be much worse. Therefore, it
- 17:38is ideal to assign one task per chat.
- 17:40Actually, there are many life hacks for
- 17:42saving tokens. I have detailed them in
- 17:44this video. Feel free to check it out.
- 17:48You might ask me: "How do I constantly
- 17:50work in a new chat when I keep dumping
- 17:52a lot of context, and then it
- 17:54disappears when I start a new one?""Do
- 17:57I really have to feed it a huge amount
- 17:58of information every single time before
- 18:00I start working?" My answer: no. And
- 18:03now, in the following concepts, we will
- 18:05figure out how to work with all of this
- 18:07. Concept number nine. It concerns the
- 18:10most important file in the codebase. We
- 18:13just talked about how every new chat is
- 18:15a blank slate. Meaning, the codebase
- 18:18doesn't know who you are, what the
- 18:19project is, what the project rules are,
- 18:21what its structure is, and so on. And
- 18:23you explain everything from scratch
- 18:24every single time. To avoid explaining
- 18:27context a thousand times, there is a
- 18:29file called .agents.md. It is a regular
- 18:33text file in Markdown format that GPT
- 18:35reads at the start of every session, at
- 18:37the start of every chat. So that we
- 18:39don't have to create it ourselves from
- 18:41scratch, there is a command called "
- 18:43init". I type it in like this, and "
- 18:46initialize" pops up in the suggestions.
- 18:48This .agents.md file can be global or
- 18:51project-specific. The global .agents.md
- 18:53affects every single chat throughout
- 18:55the entire codebase. So, regardless of
- 18:57which chat or which folder you enter,
- 18:59these rules will apply. And there is a
- 19:01project-specific .agents.md—this is a
- 19:04file that only acts within the folder
- 19:06or the project where it is located. So,
- 19:09while you can put global settings like
- 19:11"answer in simple language,""write in
- 19:14Russian," or anything else in the
- 19:15global one, in the project .agents.md
- 19:18you write specific rules for your
- 19:19project. What is important to
- 19:22understand is that this file must be
- 19:24short. Perhaps this doesn't seem to fit
- 19:26with the idea of needing to load a huge
- 19:28amount of company information there,
- 19:30but I will explain how to handle it
- 19:32best. If it is included in every chat
- 19:34and every session, even unnecessary
- 19:37information will clutter the chat. For
- 19:40example, you might be working on design
- 19:42, but information about the structure,
- 19:44backend, or some functionality is also
- 19:46there. All of that also burdens the
- 19:49context window and wastes tokens. To
- 19:51avoid this, we create several files
- 19:53with different information. For example
- 19:56, one for the design system, one for
- 19:58the backend, one for the structure, and
- 20:00so on. And we use the Agent MD as a
- 20:02navigator, as a coordinator. That is,
- 20:06for example, in the project's Agent MD,
- 20:07you can write lines like: for the
- 20:09design system, go to files/design.md.
- 20:12For the structure of my pages, go to
- 20:14files/structure.md, for example. So, in
- 20:16the project's Agent MD, you can put a
- 20:19few words about what we are working
- 20:21with and the paths to the files where
- 20:23specific information is located, and
- 20:25our chat will load the information as
- 20:28needed. This way, tokens are saved, and
- 20:30we adequately immerse our chat into the
- 20:33context of our project. This is a very
- 20:35important rule that needs to be
- 20:36understood. Concept number 10 is memory
- 20:38. Cursor has a feature called Memory.
- 20:41By default, it is turned off. And to
- 20:44turn it on, you need to go into
- 20:45settings. So, we go into Settings, then
- 20:47Personalization. And right here, there
- 20:49is Cursor Memory. What is it and how
- 20:51does it work? You don't specifically
- 20:53tell it: "Remember this or that
- 20:54information." Cursor automatically
- 20:57accesses a chat you haven't touched for
- 20:59several hours and extracts useful
- 21:00things from it, like your preferences,
- 21:02agreements, how the project is
- 21:04structured, and so on. It puts all of
- 21:07this into a Memories folder inside the
- 21:09Cursor folder. I, to be honest, don't
- 21:11really like Memories, because it can
- 21:12write any kind of nonsense in there.
- 21:15Meaning, you have to keep track of what
- 21:16it writes and then fix all of it. Why
- 21:19do we need this extra hassle when we
- 21:21have Agent MD, when we have all these
- 21:23other project MD files where we record
- 21:25all the necessary context? In short, I
- 21:28don't really like this feature, so I
- 21:30turn it off. Concept number 11 is
- 21:31dedicated to something just as
- 21:33important as Agent MD. That is Git.
- 21:36What are we dealing with? Cursor is
- 21:38constantly changing files, and
- 21:39sometimes not very successfully. For
- 21:42instance, you are working on a project
- 21:44and decide to implement some new
- 21:46function, a new feature. And at the
- 21:49moment you try to implement this new
- 21:50feature into your project, for some
- 21:52reason, everything stops working. That
- 21:55is, for example, this feature affected
- 21:57some important functionality, some
- 21:58important logic, and because of this,
- 22:00everything in your project went haywire
- 22:02. So, if you don't save successful
- 22:04versions of your project, you can very
- 22:06easily lose what you have been working
- 22:08on for so long. To save our versions,
- 22:12there is Git, a program that saves
- 22:14versions of a folder. Each saved
- 22:17version is called a commit. Just a
- 22:19fancy term. In fact, it is just a save.
- 22:22How does it work? For example, you have
- 22:24a project open where you are developing
- 22:25your service, website, or anything else
- 22:27. You just write a prompt for it. "
- 22:30Connect Git to this folder and make the
- 22:32first commit," and you hit Enter. There
- 22:34is nothing here, of course, but an
- 22:36empty folder is also a commit. Git is
- 22:38connected to the folder, the main
- 22:39branch is created, and the first commit
- 22:41is made. The commit is empty because
- 22:42there are no files in the folder yet.
- 22:44That’s it, now our first empty
- 22:46version is saved. Let’s say we
- 22:48implemented some important feature, and
- 22:50we succeeded. After we succeed, we
- 22:52write directly to Codeix to make a
- 22:54commit. A commit saves the state of the
- 22:57folder, the state of the files, and
- 22:59creates a description. That is, it
- 23:01describes what has changed in this
- 23:03folder. Thus, if you want to roll back,
- 23:05that is, return to a previous version
- 23:07of your project, you can do so thanks
- 23:09to the description, and ideally, you
- 23:11should make commits after every working
- 23:13piece. That is, you implemented a
- 23:16feature, made a commit, finally set up
- 23:18an integration, and made a commit. This
- 23:21way, you won’t lose your project
- 23:23during any unsuccessful changes. But
- 23:25there is one important flaw here. All
- 23:27these saves are located locally on your
- 23:29computer. That is, if for some reason
- 23:32your computer breaks or drowns, I
- 23:34don’t know, whatever happens to it,
- 23:36you will never recover this project
- 23:38because all its files are on your
- 23:40computer. To avoid such a potential
- 23:43problem, there is a service called
- 23:45GitHub. This is what this site looks
- 23:47like. Again, you just register on it,
- 23:49there is no problem at all. To create a
- 23:51repository, you press this plus sign on
- 23:53the top panel and click "create new
- 23:56repository." Here you come up with a
- 23:57name for it. You can write anything
- 23:59here in Latin characters. Next, you
- 24:01choose a public or private project. A
- 24:03public one will be available to
- 24:05everyone who can access this repository
- 24:06. Let’s choose private. We don't
- 24:08change anything here. We click "create
- 24:10repository." Done. Now we have a
- 24:12repository where our project files will
- 24:14be stored. We click the "Copy" button
- 24:16right here. We go into Codeix. Let’s
- 24:18ask it to create some simple page. "Can
- 24:21you create a page with two blocks?"
- 24:24Really, any page on any topic. Just not
- 24:27related to artificial intelligence and
- 24:29technology. Otherwise, it usually does
- 24:31the same thing. Let’s see what it
- 24:32does this time. We click Enter.
- 24:34That’s it, it made some thing about
- 24:36the sea. Well, it doesn't really matter
- 24:37what it made. This is how it looks. A
- 24:39three out of five, let's say. But
- 24:41that’s not important to us right now.
- 24:42Paste the link to our repository that
- 24:44we copied, and type the following: "
- 24:47Push all changes to the GitHub
- 24:48repository." I hit Enter. By the way, "
- 24:52push" just means publish. That's
- 24:54another term for you. Now, since you
- 24:56haven't connected your GitHub to Cursor
- 24:58yet, it will likely provide
- 24:59instructions on how to do so, and it
- 25:01won't be complicated. Once you do that,
- 25:04you'll be able to easily push all your
- 25:06project files to the GitHub repository.
- 25:09This way, you'll save your versions not
- 25:11just locally, but also on the internet.
- 25:13There you go, Matvey, all changes have
- 25:15been sent to the GitHub repository.
- 25:16Let's move on. Once again, paste this
- 25:18link into your browser. And now, the
- 25:20index HTML is located here. I'll
- 25:22quickly show you one more thing; it's
- 25:24also a very important story. How to
- 25:25easily and quickly deploy your projects
- 25:28to the internet. "Deploy" is another
- 25:29term, which also means to publish.
- 25:31Except "pushing" is publishing changes
- 25:33to a repository, while "deploying" is
- 25:35publishing to the internet. For this,
- 25:38there is a service called Vercel. It's
- 25:40also quite easy to sign up for it. Just
- 25:42sign in via Google, ideally via GitHub.
- 25:45You'll find out why in a moment. Click
- 25:47"Add New Project" here. You need to
- 25:50link it with GitHub here so it pulls
- 25:52your entire list of repositories. You
- 25:55select the one you need. Click "Import.
- 25:58" Next, consult with Cursor regarding
- 26:00the settings on how to configure it for
- 26:02publishing. Well, my project is simple,
- 26:04so I don't need to configure anything.
- 26:06I click the "Deploy" button, and that's
- 26:07it. Then, click "Continue to Project."
- 26:09Click this link here. And now our
- 26:12project is available on the internet.
- 26:13This way, you can quickly share your
- 26:16projects with clients, colleagues, and
- 26:18so on. This, of course, doesn't replace
- 26:21a full-scale publication on a server or
- 26:23hosting, but for a quick publication to
- 26:24show someone a project, why not? The
- 26:27next batch of concepts will be
- 26:29dedicated to how to write prompts
- 26:31correctly and how to set tasks properly
- 26:33. The main mistake most people make is
- 26:36writing vague prompts. For example, "
- 26:39make me a beautiful website." The AI
- 26:41doesn't know what "beautiful" means.
- 26:43And since it lacks specific references
- 26:46or examples from you, it will use what
- 26:49it knows. And it knows those typical,
- 26:51sloppy AI-generated designs. You can
- 26:53also make it easier for the AI to
- 26:55understand your task by giving it the
- 26:57path to the file it needs to reference.
- 26:59For example, don't just say "study the
- 27:01design system," but "the design system
- 27:03is located here." Type an "@" symbol
- 27:04and tag the markdown file containing
- 27:06the design system. This way, it will
- 27:08specifically understand where to look
- 27:10and what to do. Plus, besides text and
- 27:13a specific prompt, you can add images,
- 27:16that is, references or HTML code. For
- 27:19example, you’ve visited a website
- 27:20that you find visually appealing. Well,
- 27:22let’s say this one. Instead of just
- 27:24taking screenshots, you can do
- 27:26something even better. Let’s copy the
- 27:28link. Paste it into a service called
- 27:30View Page Source. The link will be in
- 27:32the description. Again, you get the
- 27:34HTML code, click the download button,
- 27:36and drop this code file into your
- 27:38project folder. Now let’s go back to
- 27:42our Cursor and type the following.
- 27:44That’s what it was called. I press
- 27:46Tab and write: "Let’s replicate the
- 27:49main block on this site exactly." And
- 27:52let’s also take a screenshot just to
- 27:54be sure, so it understands exactly what
- 27:55it looks like. So, after 5 minutes, it
- 27:58finished and did it almost perfectly,
- 28:01exactly the same. Let’s take a look.
- 28:04There is a slight issue with this black
- 28:06bar overlapping the orange arrow shape,
- 28:08but otherwise, everything looks almost
- 28:10perfect. And, as you can see, because
- 28:13we provided the HTML, it didn't just
- 28:15copy the image, it actually pulled the
- 28:17video from the source. The twelfth
- 28:19concept will be dedicated to plan mode.
- 28:21That is, when the task is defined or
- 28:23the task is small, you can just write
- 28:25about it to the neural network. And the
- 28:27neural network will do it. But when the
- 28:29task is large or there are too many
- 28:31files, for example, and it affects a
- 28:33large number of files, jumping straight
- 28:34to writing code can be a mistake. It
- 28:37might interpret the task in its own way
- 28:39, do it not the way you need, and
- 28:41you’ll have to revert to the previous
- 28:43version. To think through the task
- 28:46before executing it, there is a
- 28:47planning mode. To turn it on, click
- 28:51slash and find planning mode right here
- 28:54. I click here like this. And now this
- 28:56little badge has appeared here,
- 28:58indicating that plan mode is enabled.
- 29:00In this mode, Cursor will not change
- 29:02your project files in any way. It reads
- 29:05the files, sees what you have, asks
- 29:07some questions, and so on. After you
- 29:09answer its questions, it will show you
- 29:12a plan. If you are satisfied with the
- 29:14plan, you approve it, and it begins the
- 29:16work. The next concept will be familiar
- 29:18to those who have been working with
- 29:20neural networks for a long time. When
- 29:21you are working on a task, Cursor works
- 29:23step-by-step. It completes a task,
- 29:26writes you a report, and you tell it "
- 29:28okay, continue." And it starts the next
- 29:30stage of the task. So, while working,
- 29:33you have to click "continue" a certain
- 29:35number of times, which is cool on one
- 29:36hand because you can check every
- 29:38intermediate step, but it's not great
- 29:40in terms of slowing down your work,
- 29:41constantly getting distracted, and
- 29:43typing the next actions for it. But
- 29:46what if the task is large, doesn't
- 29:48require your intermediate checks, and
- 29:49you just want to set it to run
- 29:51overnight and go to sleep? For this
- 29:53purpose, they created the "GO" command,
- 29:55which is the goal mode. Let me type it
- 29:58in. There, we’ve clicked it. Now, a "
- 29:59goal" has appeared in the bottom panel.
- 30:01What is this mode? You give it not a
- 30:03task, but a goal. And Cursor will work
- 30:06until it achieves it. It won't stop
- 30:08halfway; it will go all the way to the
- 30:11end. Well, at least until the point it
- 30:13considers to be the finish. A goal is a
- 30:15final result that can be verified.
- 30:17Therefore, it must contain three things
- 30:19: what should be working in the end,
- 30:21and how can we tell that it works? And
- 30:24what are the boundaries? Meaning, what
- 30:25should not be touched. For example, if
- 30:27we take the development of an admin
- 30:29panel—a tool that allows you to make
- 30:31changes to a project without using an
- 30:33AI—it would be a bad goal to tell it:
- 30:35"Make a cool admin panel." Again, it
- 30:37doesn't know what "cool" means. It has
- 30:40no metric for "coolness," and it will
- 30:42keep improving it forever. So, this is
- 30:45an unlimited, non-stop process. Here is
- 30:47an example of a normal goal. Create an
- 30:48"Orders" section in the admin panel.
- 30:50Include a list of all orders with a
- 30:52status filter, an individual order page
- 30:53, and a button to change the status.
- 30:55Work until the verification passes. I
- 30:58place a test order on the site and see
- 30:59it in the list. I change the status to
- 31:02"completed," refresh the page, and the
- 31:04status is saved. Do not touch
- 31:05authorization or other sections of the
- 31:07admin panel. So, it has a verification
- 31:09scenario. It will place a test order
- 31:11itself, open the admin panel itself,
- 31:13and refresh the page itself. If it
- 31:15doesn't work, it fixes it and checks
- 31:17again. If everything works, then,
- 31:19accordingly, the goal is achieved, and
- 31:21Cursor stops. One of the downsides of
- 31:24this approach, naturally, is that when
- 31:26performing a long task, it will be
- 31:27working the entire time, meaning it
- 31:29will spend our tokens. Accordingly, in
- 31:32"goal" mode, it can very easily hit the
- 31:35limits. That’s not great, but if
- 31:37you're setting a task overnight or just
- 31:39setting a big task and going out for a
- 31:41walk, then why not? Next concept. Trust
- 31:45, but verify. Cursor might say that it
- 31:48has done everything you asked. And you
- 31:50can take this "done" status with about
- 31:5260%confidence. Meaning, it definitely
- 31:55messed up somewhere, overlooked
- 31:57something, or didn't account for a
- 31:58certain scenario, so ideally, you
- 32:00should check everything manually. To
- 32:02verify what our neural network did, we
- 32:05can either use some skills or plugins (
- 32:07we'll talk about those soon), or, again
- 32:09, just manually go through all the
- 32:12scenarios, testing and checking
- 32:14everything—every single scenario,
- 32:16both successful and unsuccessful. And
- 32:19I’m sure that by the end of your
- 32:20check, you’ll have a huge list of
- 32:21comments to send back to it. And
- 32:24unfortunately, if you are working on
- 32:26large projects or a big task, you will
- 32:29have to perform these tests and checks
- 32:31yourself. You can, of course, dump it
- 32:34all on the neural network, but you’ll
- 32:35highly likely end up with errors. In
- 32:38reality, AI coding is one thing, but
- 32:40when you work on large projects or
- 32:42commercial orders, you can run into
- 32:44major problems without understanding
- 32:46how it all works under the hood. You
- 32:50could ruin a project for a client, cost
- 32:52them money, overlook some legal aspect
- 32:54that causes them trouble, and so on.
- 32:58Even though creating projects has
- 32:59become easier with the arrival of
- 33:01neural networks, without understanding
- 33:02the stages and the final outcome, it
- 33:04will be very hard for you to work and
- 33:06monetize this skill. Actually, on this
- 33:08very topic, I recently finished my
- 33:10first training cohort. Somewhere around
- 33:12here there will be testimonials for
- 33:14this cohort, and a link in the
- 33:15description to the results of this
- 33:17first group. There, I was teaching the
- 33:19guys how to create AI-coded projects
- 33:21from scratch, from writing the tech
- 33:22specs to deploying to a server, where
- 33:24we worked through every stage of
- 33:26project creation. That means design,
- 33:29integrations, payment systems, delivery
- 33:31services, security audits, SEO audits,
- 33:34refactoring, debugging, deployment,
- 33:36post-deployment maintenance, the legal
- 33:39side, and sending notifications to
- 33:41Roskomnadzor. Plus, we studied client
- 33:44work in great detail, meaning how to
- 33:46properly sign a contract, how to sign
- 33:48an acceptance act, how to draft a
- 33:49commercial proposal, and what a brief
- 33:51is. In short, it’s a very intense
- 33:53course. I still haven’t fully
- 33:54recovered from it. And right now, I am
- 33:55accepting applications for the second
- 33:57cohort. If you want to be part of the
- 34:00second cohort, check the description or
- 34:01use the QR code on the screen, go into
- 34:03the bot, answer three questions, after
- 34:05which you'll be added to the waitlist
- 34:07and will be the first to hear news
- 34:08about the new cohort, as well as
- 34:10receive a discount. Anyone who isn't on
- 34:13the waitlist won't get the discount.
- 34:14The next concept will cover the "Work"
- 34:16mode. This is the same mode found in
- 34:19the ChatGPT tab. It’s not Claude, but
- 34:22it’s a convenient tool that you can
- 34:24use even on your phone. So, we open a
- 34:27new chat and select "Work" mode. What
- 34:30is it? Essentially, it’s the same
- 34:32agent as Claude. The same engine, the
- 34:35same models, the same limits. With it,
- 34:38you can do more than just chat; you can
- 34:40work with folders on your desktop,
- 34:42create files, documents, presentations,
- 34:44and so on. Unlike Claude, this thing is
- 34:47also available in the web version.
- 34:49Let’s give it a small task here. For
- 34:52example, putting together a
- 34:53presentation. Why is GPT cool compared
- 34:56to Claude? Because Claude doesn't know
- 34:58how to generate images. Accordingly, it
- 35:00cannot create illustrations for a
- 35:02presentation on its own. GPT, however,
- 35:05has no problems in this regard, and it
- 35:07can create much more interesting and
- 35:09engaging presentations for users.
- 35:11Let’s enter a prompt like this, for
- 35:13example. So, I’ve given it access to
- 35:15a folder, and the files it needs are
- 35:17inside that folder. Analyze the files
- 35:19in the product.mstyle.mdstructure
- 35:21folder. Create an eight-slide
- 35:24presentation in the style of style.md.
- 35:27There are no facts about the model in
- 35:29the folder. Find them on the internet.
- 35:30I asked it to make a presentation for
- 35:32Apple Watch. And these files were also
- 35:34generated by the neural network. Well,
- 35:36it’s just for testing purposes. List
- 35:37the sources on the last slide. For each
- 35:40slide, generate one minimalist
- 35:42illustration in the colors of style.md
- 35:44without people and without the Apple
- 35:46logo. Why without the Apple logo?
- 35:48Let’s remove that. Before finishing,
- 35:50check that every number on the slides
- 35:52is in the sources and the titles are no
- 35:54longer than five words. Do not modify
- 35:56the files in the folder; save the
- 35:57presentation alongside them. In fact,
- 35:59thanks to connectors, we can do more
- 36:00than just save the presentation locally
- 36:02; for example, we can have it sent to
- 36:04me via Google Mail. Let’s write that:
- 36:06"Duplicate the file of this
- 36:08presentation to my Google Mail." I will
- 36:10explain how to connect Google Mail and
- 36:12other connectors a bit later. Actually,
- 36:14why wait? Go to plugins. And right here
- 36:17is the popular Google Mail. Go to the
- 36:20plugin and install it; I already have
- 36:21it installed, so there’s no "install"
- 36:23button here. And that's it. Now let’s
- 36:25see what it comes up with. Okay, it's
- 36:27all finished. It took 10 minutes and 48
- 36:30seconds. Matvey, the presentation is
- 36:32ready. Eight slides, eight
- 36:33illustrations, verified numbers and
- 36:35sources. On the last slide, the
- 36:36original files remain unchanged. A copy
- 36:39has been sent as an attachment to my
- 36:41email. And overall, it turned out quite
- 36:43well. Obviously, it’s not a design
- 36:45masterpiece. Well, because there wasn't
- 36:47even a proper task description here.
- 36:49But it generated the images more or
- 36:50less okay. Although, these two slides
- 36:52are super weird. What even is that? It
- 36:54included some links here, which you can
- 36:57actually click on. And overall, if I
- 36:59had given it a proper task and at least
- 37:01shown what the watch looks like, it
- 37:03would have been pretty good. And let's
- 37:05also open my Google mail. Here is a
- 37:08message I received from myself. And
- 37:10there you go, Matvey, the finished
- 37:12presentation based on your files is in
- 37:14the attachment. In short, with this
- 37:16tool, you can work quite effectively
- 37:18with presentations, documents, files on
- 37:20your computer, and so on. You can, for
- 37:22example, ask it to take and sort your
- 37:24downloads folder. You just select work,
- 37:27select downloads here, click new
- 37:30project, add folder, downloads folder,
- 37:32and ask it to sort, for instance, some
- 37:35files. Please sort all the duplicates
- 37:38that are in the downloads folder and
- 37:40delete them. That’s it. This way, you
- 37:42can easily clean up your downloads
- 37:44folder. But since you have Cursor, you
- 37:46don't really need this GPT Work at all.
- 37:48That is, you can comfortably use Cursor
- 37:50and perform all the same tasks. Concept
- 37:53number 16. Turn a process into a skill.
- 37:56What are skills? For example, we just
- 37:59made a presentation together. It turned
- 38:01out quite well, but still not perfect.
- 38:04Somewhere images were small, somewhere
- 38:06we didn't like the font, and so on.
- 38:09Ideally, we should have clarified the
- 38:11style, asked for a certain image size,
- 38:13requested a specific aesthetic, and so
- 38:15on. And let's say in a week we need the
- 38:18same kind of presentation, well, for
- 38:20another product. What should we do? We
- 38:22have two options. Either we describe
- 38:25the whole process to it all over again,
- 38:27what kind of images it should make,
- 38:28what fonts, and all that. Or you can
- 38:31create a skill, that is, an instruction
- 38:34on how to approach creating
- 38:36presentations. Essentially, it's a
- 38:38regular markdown file that just states
- 38:40in text what the neural network needs
- 38:41to do. A skill has its own name. So,
- 38:45for example, we type a slash and the
- 38:47name of this skill. And that is how we
- 38:49use the skill in our work. Then the
- 38:52skill has a description. For example,
- 38:55we created this skill and labeled it as
- 38:57a skill for bringing our presentations
- 38:59up to brand standards. And we will also
- 39:01add that we should use it when we are
- 39:03dealing with presentations. And the
- 39:06moment we ask it to make a presentation
- 39:08, it will read the skill's description,
- 39:10understand that it needs to be used,
- 39:12and will use it. So, we don't
- 39:14necessarily have to specify this skill
- 39:16with a slash, but the description must
- 39:17be such that the skill is definitely
- 39:19pulled in. Therefore, in this regard,
- 39:21the description needs to be specific.
- 39:23Let's not create a new skill right now;
- 39:25I'll just click slash and show you
- 39:26where they can be found. We scroll down
- 39:29like this and find those very skills
- 39:31here. There is a huge variety of them
- 39:33here. Some skills were in the codebase
- 39:35initially, others I added myself. The
- 39:38set of skills, plugins, and MCPs that I
- 39:40use in my work, I will leave in my
- 39:42Telegram channel. So you can go there
- 39:44via the link in the description or via
- 39:46the QR code on the screen and get all
- 39:48the additional materials. You can
- 39:50create these skills yourself. To find
- 39:53these skills, you can either go to the
- 39:56skills tab, this outlet icon, and
- 39:58select everything you need. You can
- 40:01create a skill or record a skill. That
- 40:03is, you can perform an action right on
- 40:05a screen recording, show how it works,
- 40:08and later the neural network will rely
- 40:10on this and, accordingly, do what you
- 40:12need. This is a bit of a hassle, but in
- 40:15certain situations, it can be useful.
- 40:17And there is simply a button to create
- 40:18a skill. You can click on it. The skill
- 40:20creator is pulled up here. This is a
- 40:22skill for creating a skill. And here we
- 40:24write, for example, I want to create
- 40:26such and such a skill, so that it
- 40:28performs such and such functions and
- 40:30uses such and such tools. That is, you
- 40:33can give a skill not just an
- 40:35instruction, but also a set of tools it
- 40:38will use. For example, the Google Mail
- 40:41MCP, to write messages to email or any
- 40:44other tool. There are a huge number of
- 40:47ready-made skills on the internet that
- 40:49you can use in your work. For example,
- 40:51there is a site called skills. Here you
- 40:54can find skills for every taste. One of
- 40:57the important skills is "find skills,"
- 40:59with which you can install skills from
- 41:00this service. So you just click on the
- 41:03desired skill, select the prompt here,
- 41:06copy it, and paste it into the codebase
- 41:08. That's it, click enter. And this
- 41:10skill is installed for you. The next
- 41:12concept concerns plugins and MCP. We
- 41:15have, of course, already tried to
- 41:16connect Google Mail, so you already
- 41:18know what that is. But let's go over it
- 41:20one more time. MCP is a way to connect
- 41:23your codebase with another external
- 41:25tool, for example, Google Calendar,
- 41:27Google Mail, Teams, BRIX, anything at
- 41:30all. Well, anything that has an MCP;
- 41:33for example, let's go back to the
- 41:35plugins tab right here and find
- 41:37something, let's say, for example,
- 41:39something like Vers. This is the same
- 41:42service we used for our deployment. It
- 41:44redirects us to this page. We give our
- 41:47authorization to connect our Codex to
- 41:49Vercel. Then we go into the application
- 41:52, and that's it, it's loading.
- 41:53Everything is ready, the VS plugin is
- 41:55installed. Now let's go to the project
- 41:58we were working on. To select a
- 42:00specific plugin, we press the@sign and
- 42:02type its name. I clicked Tab to mark it
- 42:06. Deploy this project, which is in my
- 42:08folder, and give me the link right in
- 42:10the chat. In the link, use a name like
- 42:14"test project," but in English. Let's
- 42:17go. So, it told us that the project is
- 42:19deployed and verified. Let's open it
- 42:21and see that it's working. Open in an
- 42:23external browser. The beginning of the
- 42:25link has what I asked for. Test project
- 42:27. And, accordingly, the site itself
- 42:30opens up. We can, for example, use
- 42:32several MCPs at once. That is, ask it
- 42:34to deploy the project we are developing
- 42:36to Vercel and, for example, send it
- 42:38somewhere via email. In general, with
- 42:41these connectors, you can create great
- 42:43integrations between Codex and various
- 42:45tools. And besides MCPs, which are
- 42:48these integrations with applications,
- 42:50there are also just regular plugins. A
- 42:52plugin is essentially a set of skills
- 42:54and tools that solve a specific task.
- 42:57For example, Codex Security. There it
- 43:00is. This is a plugin for checking our
- 43:03code for security issues. It has
- 43:05several built-in skills. As you can see
- 43:07, there are 15 of them. You can choose
- 43:09which ones to use and which ones not to
- 43:11use. And if we click, we can read about
- 43:13each skill. Concept number 18 is
- 43:15dedicated to working in the browser.
- 43:17Many services actually don't have MCP
- 43:20or AI, meaning the neural network can't
- 43:22interact with them at all. Previously,
- 43:25this was a dead end, but now Codex can
- 43:27work in the browser. This means it can
- 43:29visit a site itself and click on things
- 43:31. To use Codex in Chrome, for example,
- 43:35we go to the computer use settings. And
- 43:38right here, there is an extension for
- 43:40Chrome. Click "Install." Click "Add to
- 43:43Chrome" and install it. Next, you log
- 43:46into your account and follow the short
- 43:48instructions here. That is, open the
- 43:51extension in the sidebar, pin it if you
- 43:54need to, and try GPT with real work
- 43:56tasks. Then go to GPT. And you can
- 43:59notice that Chrome is installed. Now we
- 44:02can ask it to do the following. I don't
- 44:04know, I haven't come up with any task.
- 44:05Let's make it something like: you can
- 44:08go into Claude via the Chrome browser
- 44:10and write something in the chat, like:
- 44:13"Haha, Matvey's computer has been taken
- 44:15over by ChatGPT." So, as you can see,
- 44:18this glowing cursor has appeared here.
- 44:21That is our GPT. I mean, you see, it
- 44:23clicked the "Send" button all by itself
- 44:24. This way, you can connect GPT to
- 44:27various tools in the browser. This way,
- 44:31by the way, you can test your
- 44:32application, meaning you can directly
- 44:34ask it to go through all the tabs,
- 44:35click all the buttons, test everything,
- 44:37and make sure everything is actually
- 44:39working. Or if something isn't working,
- 44:42it can provide feedback, create a
- 44:44report, and then you can go through it
- 44:45and make corrections. But, honestly,
- 44:48sometimes it's very slow. I mean, there
- 44:52are certain tasks that are easier to do
- 44:53yourself than to wait for the neural
- 44:55network to figure things out and do
- 44:57what's needed. Plus, it can consume a
- 44:59fairly large amount of tokens because,
- 45:01most likely, it will be taking
- 45:03screenshots along the way, comparing
- 45:05things, and all of that will take a
- 45:06very long time. And it can happen that
- 45:08with some incorrect prompts, the neural
- 45:10network might wander where you didn't
- 45:12ask it to go. Therefore, you need to be
- 45:14careful here as well. Concept number 19
- 45:16. Distribute work to copies, that is,
- 45:18sub-agents. Up to this point, the code
- 45:20was working alone, meaning essentially
- 45:22it had one chat, and it performed all
- 45:24tasks sequentially. Sub-agents are when
- 45:27a chat launches several copies of
- 45:28itself; each takes its own piece, and
- 45:30they all work simultaneously. This way,
- 45:33you can parallelize several tasks at
- 45:35once. I mean, if previously three
- 45:37ten-minute tasks took you half an hour,
- 45:39that is 10 minutes, 10 minutes, 10
- 45:41minutes, now it's all distributed, for
- 45:43example, among three sub-agents and
- 45:45completed in 10 minutes. You can launch
- 45:48them simply via a prompt. For example,
- 45:50check our site for SEO and security.
- 45:53Break the task down into two sub-agents
- 45:54. Let one check the site for SEO and
- 45:57the second for security. Well, let's
- 45:59send our site to it. There's nothing to
- 46:01check there, of course, but I just want
- 46:03to show how it works. You see, it says
- 46:05that it is launching two sub-agents.
- 46:07One is called SEO Audit, the other is
- 46:08Security Scan. And here our sub-agent
- 46:11has appeared with an icon. And this
- 46:14pink one is the Security Audit. You can
- 46:16open any agent and see what it's doing.
- 46:18For example, the SEO Audit. It's doing
- 46:20something there, keeping busy. What
- 46:22else is the advantage? Besides speed,
- 46:25each sub-agent has its own separate
- 46:27context window. That is, its work does
- 46:30not burden our shared context window.
- 46:32It works in a separate space, and only
- 46:34the result of the sub-agent's work
- 46:36enters the main context window. Claude
- 46:39has a feature in its skills called "
- 46:41context fork," which means the task is
- 46:43performed in a separate context window.
- 46:45For some reason, Codeium doesn't have
- 46:47this, so only sub-agents handle it here
- 46:48. But one must understand that when we
- 46:51break a task into multiple sub-agents
- 46:53—and there are modes where there can
- 46:56be 50, 100, or 150 sub-agents—tokens
- 46:58will naturally be consumed at an insane
- 47:01rate. And, by the way, the "Ultra"
- 47:03reasoning level we discussed is exactly
- 47:05what forces the neural network to
- 47:07create these sub-agents. In short,
- 47:10thanks to these sub-agents, we can
- 47:12parallelize a task across multiple
- 47:14executors. This way, we will complete
- 47:16our work much faster. The next concept
- 47:19is dedicated to automations. Everything
- 47:21we have done so far was triggered
- 47:22personally by you. That is, you write
- 47:25to Codeium, and Codeium does it. But
- 47:28Codeium has a special feature that
- 47:30allows you to run automations, i.e.,
- 47:32scheduled tasks. It is called "
- 47:35Scheduled," or "
- 47:35запланированные" in
- 47:37Russian. This is how it looks in the
- 47:39left panel. Those little clocks there.
- 47:41Click here. New task. Let's make it so
- 47:44that every day at 9:00 AM, for example,
- 47:47Moscow time, it sends a post to my
- 47:49Telegram with an image and a text
- 47:51description of the latest news in the
- 47:53world of artificial intelligence for
- 47:55that day. To do this, I need to get the
- 47:59bot's API key and my user ID. Let's do
- 48:02it. Go into Telegram and type "
- 48:04BotFather" here. It's a bot; the link
- 48:07to it will be in the description. Click
- 48:08"Open." Let's create a new bot. Let's
- 48:11name it "newsmaker." Doesn't matter
- 48:14here. And here: news_news_matvey. And
- 48:18it must end with the word "bot." So I
- 48:21type "bot" and click "create bot." And
- 48:24here is its API token. I copy it. To
- 48:29get my user ID so that this bot sends
- 48:31messages specifically to me, I need to
- 48:33go to a bot called "userinfobot." The
- 48:36link to it will also be in the
- 48:37description. Send it any message, and
- 48:40it will give me my user ID. Just before
- 48:43creating the automation, you need to
- 48:44click the "Start" button in the bot,
- 48:46because without that, it won't be able
- 48:48to send messages to my personal account
- 48:50. That's it, now we can insert the task
- 48:52for it. Copy it and paste it. Find the
- 48:55top five news items of the week from
- 48:56the world of artificial intelligence
- 48:57and AI coding on the internet. New
- 49:00models: OpenAI, Anthropic, Google
- 49:01releases, and Codeium/Cursor updates.
- 49:04Choose the most significant one from
- 49:06them, and generate an image. A cozy
- 49:08kitten sitting in an armchair and
- 49:09reading a newspaper. The newspaper is
- 49:11facing the viewer. On the front page, a
- 49:13large headline of one to three words
- 49:15about the biggest news. For example,
- 49:17GPT7 is out. The headline must be
- 49:19legible and fit perfectly into the
- 49:21newspaper. The style should be a warm
- 49:23illustration without any company logos.
- 49:25Send it to Telegram via a bot in a
- 49:26single message. An image and a summary
- 49:29in the photo caption. The caption
- 49:30should be up to 1,000 characters. Five
- 49:32news items, each one to two lines long,
- 49:34with a link, in simple language,
- 49:35without long dashes. That's it. Next,
- 49:37we send the bot token and the chat ID.
- 49:40In general, it is better not to expose
- 49:42your keys like this and put them in a
- 49:44separate file, for example. Let's
- 49:47choose GPT6 Soul as the model for this.
- 49:49Let's make it 61. There is an important
- 49:52setting here: "run on this computer."
- 49:55If it is active, this automation will
- 49:57only work when our computer is turned
- 49:59on and the Codeks app is running as
- 50:01well. That is not very convenient.
- 50:04However, in this situation, Codeks has
- 50:06access to your folders and files. If we
- 50:09run it not on our own computer, then
- 50:11the work will be done in the cloud. It
- 50:14will run using our subscription without
- 50:16any additional costs. And we won't need
- 50:18to turn on the computer. That is, we
- 50:20can turn it off, throw it away, or go
- 50:23somewhere. And this automation will
- 50:25still be executed. The downside is that
- 50:27, in this case, it won't have access to
- 50:29your files on the computer. But here we
- 50:31don't need any files on the computer,
- 50:33so we will naturally do this in the
- 50:35cloud. You can choose to repeat hourly,
- 50:37daily, or set it up to your preference.
- 50:39Time and end of repetition. That is,
- 50:42choose the moment when this automation
- 50:44will finish. Start each run in a new
- 50:46chat. Let's select this so our chat
- 50:47doesn't get cluttered. "Effort: medium"
- 50:49should be enough, I think. Everything
- 50:51seems to be set up now. Let's click "
- 50:53create." All done, your task is
- 50:54scheduled. The next run is on October
- 50:577th at 9:00 AM. Now we can click on the
- 51:00three dots here and click "run now" to
- 51:02check how it works. So, everything is
- 51:05ready. It sent me this post. It
- 51:08considered the release of Geminii 4
- 51:10Argon to be the main news of the week,
- 51:11included all the links, wrote
- 51:13everything out, and compiled a top list
- 51:14. And this thing will work
- 51:17automatically every day at 9:00 AM,
- 51:19regardless of whether my computer is on
- 51:22or not. This is a very convenient story
- 51:24. It is especially cool that GPT can
- 51:26generate images and create such
- 51:28interesting stories. In short, with
- 51:32this feature, you can automate a huge
- 51:34amount of your routine work, collect
- 51:37reports, analyze data for a specific
- 51:39period, and so on. Concept number 21 is
- 51:43a pet. To find the command to make this
- 51:46pet appear, we click in the bottom left
- 51:48corner. And right here there is a
- 51:50button to show mini. For me, it's
- 51:52option-space. Let me press it. And here
- 51:55on the right side, this little pet
- 51:57appears. You can talk to it by voice,
- 52:00give it tasks, or type something to it.
- 52:03This thing stays on top of all windows.
- 52:05I mean, see, I went into Claude, and
- 52:07it’s still here. It hasn't gone
- 52:09anywhere. To be honest, I don't really
- 52:10understand why one would use it.
- 52:12Probably just to avoid opening the
- 52:14Codex unnecessarily. But it at least
- 52:16shows the status of the chat. Create
- 52:18the simplest page in the world. Just a
- 52:20white background, black text, in one
- 52:22block. We hit Enter and watch what this
- 52:25pet does. I switched to another window,
- 52:27and you could see it was sitting at its
- 52:30laptop typing something. Let's open
- 52:32this up. And when it finishes the task,
- 52:34it will notify us. See, this green
- 52:37badge has appeared here. And you can
- 52:39see that the task was completed. You
- 52:42can also talk to it by voice or
- 52:43exchange messages and give it a prompt.
- 52:46Open the page we just made. See, I gave
- 52:49it a task, and it's doing something
- 52:51over there. Well, we've opened that
- 52:53page. In short, it can notify us about
- 52:55what is happening in the Codex. At the
- 52:59same time, we can also interact with it
- 53:01, give it certain plugins, skills, and
- 53:03so on; we can talk to it by voice and
- 53:05give it text prompts. In short, it's an
- 53:08interesting, interactive gimmick, but I
- 53:10don't know who exactly it's for.
- 53:12Honestly, I don't really like it, so I
- 53:14usually remove it. By the way, you can
- 53:16go into settings and change its
- 53:18appearance. For example, choose this
- 53:20owl. I think it's cool. Well, anyway,
- 53:22that’s the story. Concept number 22.
- 53:26Working from a phone. Codex runs on our
- 53:28computer, but we don't have to sit at
- 53:30the computer all the time. In the GPT
- 53:32app on your phone, there is a Codex tab
- 53:34, so all chats from the computer are
- 53:36identical. You can see what it's doing,
- 53:38see the progress, see the questions it
- 53:40asks, and so on. To enable this, go to
- 53:43settings, connection. And right here
- 53:45you can add a device. Add a device for
- 53:48remote control of this Mac. Click add.
- 53:51Connect a device to this Mac. Continue
- 53:53from where you left off. Stay informed,
- 53:55start something new. Click start. To
- 53:58enable this feature, you will need to
- 54:00turn on multi-factor authentication. To
- 54:02do this, you need to go to the web
- 54:04version. Let's go into settings here
- 54:06and find security and login. Let me
- 54:08zoom in a little bit. Here we select
- 54:10multi-factor authentication. We turn it
- 54:13on. Here we confirm our email again. We
- 54:16receive a code. Next, we scan the QR
- 54:18code in any authentication app. There
- 54:20will be examples of these apps right
- 54:22here. And we enter the six-digit code
- 54:24that it provides. That’s it. After
- 54:26that, we have enabled authentication,
- 54:28and now we can press add again, click
- 54:30start, and let's click allow. Then you
- 54:33scan the QR code and that's all. Now
- 54:36you can use Kodeks on your mobile
- 54:39device. Concepts 23 and 24. These two
- 54:42concepts will be dedicated to the new
- 54:44features released on September 29th at
- 54:46the event called DevDay. The first one
- 54:49is Spaces. A new icon has appeared on
- 54:51the left here that wasn't there before.
- 54:54What is it and what is it for? Let's
- 54:56click continue. That was a cool
- 54:58animation. Everything that ChatGPT
- 55:00creates for me lives in different chats
- 55:01. Plans, tables, summaries. In a week,
- 55:05all these files can be lost because
- 55:07they are in different places. Well,
- 55:09Spaces fixes all that. Meaning, if we
- 55:12open the Spaces tab, there is this huge
- 55:15thing with a bunch of stuff, like pages
- 55:17, some sites, images, some texts, and
- 55:20so on. This whole thing works not with
- 55:23Kodeks, but with the cloud, that is,
- 55:26with ChatGPT and GPT Work. So, if you
- 55:28go into images, you won't find the
- 55:30images you generated, for example, in
- 55:33some folder in Kodeks. Or, for example,
- 55:35the website files we made together,
- 55:37they won't be here either. But if, for
- 55:39example, you just went into ChatGPT and
- 55:42used it to make a document, file, or
- 55:45generated an image, all of that will be
- 55:47here. And thanks to this collection of
- 55:50all the files you created in the cloud,
- 55:52you can manage them more easily.
- 55:54Examples are shown here of what can be
- 55:57done in a to-do list. A project tracker
- 55:59that can be filled out quite flexibly,
- 56:02a weekly summary. Here, for example,
- 56:05you can click the plus sign and insert
- 56:07some image, file, or anything at all.
- 56:09We can generate it or attach it. You
- 56:12can upload it from your computer, or
- 56:14you can export it from this Spaces
- 56:17media library. You can also click the
- 56:20slash here and insert either headers,
- 56:22some lists, quotes, a piece of code,
- 56:24and so on. In short, a very flexible
- 56:27setup that works almost like Notion,
- 56:29but unlike Notion, it has these kinds
- 56:31of features. For example, creating a
- 56:33task. And here we write something like,
- 56:36hey, write me five topics for an AI
- 56:38video that I can film based on the new
- 56:41releases from the past week. There is
- 56:44such an interesting feature here. You
- 56:46can rearrange all of this. You can
- 56:48simply insert instructions for the
- 56:50agent like this. You see, I wrote to
- 56:52him, and he answered me in this format.
- 56:55If you're interested, again, I can make
- 56:57a separate video where I analyze
- 56:59several use cases for this space. But
- 57:02in short, it’s just a space that
- 57:04holds everything you’ve been working
- 57:07on in the cloud. And you can mix it all
- 57:10up, share it, process it, and so on. Or
- 57:13, to put it briefly, it's Notion with
- 57:15ChatGPT built into it. And the last
- 57:18thing I’d like to discuss, also a
- 57:20novelty released at DevDay on September
- 57:2229th, is Dots. Again, this thing
- 57:25deserves a separate video, but let me
- 57:26tell you about it concisely and briefly
- 57:28. Everything we did before this, we
- 57:30launched personally. Even the schedule
- 57:32—we wrote it ourselves and told it to
- 57:34do this every Monday. Dot works
- 57:37differently. It is a single persistent
- 57:39agent running on the GPT-4o Astra model
- 57:42, which has its own computer in the
- 57:44cloud, a browser, files, and a terminal
- 57:46. Meaning, if you close your laptop, it
- 57:48will still keep working. It is
- 57:50connected to your email, your calendar,
- 57:52your codebase, and it doesn’t wait
- 57:53for your command. It periodically
- 57:56checks on what is happening with you on
- 57:57its own. And it writes to you first.
- 57:59For example, it might write to you: "
- 58:01Hey, I see an advertiser hasn't replied
- 58:03to you in 2 weeks. Should I remind them
- 58:05that they need to respond?" Or, for
- 58:07instance: "A nighttime task in the
- 58:09project failed; I’ve looked into it."
- 58:11"Here is the reason why it failed." So,
- 58:13you closed your laptop, but it’s
- 58:15still working. To open it here, we
- 58:17click on "your dot" right here and go
- 58:19in. Click continue. And we can
- 58:22immediately connect it to several
- 58:23services. Click continue here. So, hi,
- 58:26I'm your Dot. Write or call me anytime.
- 58:29I will help with tasks even when we
- 58:31aren't talking, and I'll come back with
- 58:33news or questions. Do you want to give
- 58:35me a name? Let's give one. Let it be
- 58:37Anatoly. See, he left a heart as a
- 58:39reaction all by himself. Well, that’s
- 58:41already a bit of a strange story. Saved
- 58:43. Now I am Anatoly. I will start
- 58:45looking for ways I can help. Is there
- 58:47anything you want to tackle first right
- 58:49now? Actually, see, I haven't even had
- 58:51a proper conversation with him yet.
- 58:52He’s already found something here. He
- 58:54found two items in my connected email
- 58:56that he can help with. Checking the BAS
- 58:58soup protection and sorting out the
- 58:59late-night check-in in Hong Kong. Let's
- 59:01set some boundaries and areas of
- 59:03responsibility right away. Anatoly,
- 59:05your zone is my tasks in Coda and the
- 59:07daily news summary that comes to my
- 59:09Telegram on schedule. If the summary
- 59:12doesn't arrive in the morning or a task
- 59:14drops into Coda, find out why and
- 59:16message me. Don't send me anything
- 59:17about the email, I'm tracking that
- 59:18myself. Don't send anything on my
- 59:20behalf without my confirmation. There,
- 59:22you see, he has accepted his area of
- 59:24responsibility and confirmed the custom
- 59:26rules. I don't do anything further with
- 59:28him. That's the whole point. He will
- 59:30look into my projects and the summary
- 59:32himself and only message me when
- 59:34there's a reason. The message will come
- 59:36here, into this chat. This feature is
- 59:37not needed for one-off tasks. For
- 59:39example, making a presentation,
- 59:41creating a file, or setting up a page.
- 59:44All of that will be faster and more
- 59:45cost-effective to do directly in Coda.
- 59:47This feature is not available on the
- 59:49Plus plan. It is available on the Pro
- 59:51plan starting at $ 100. Conversations
- 59:53with him don't use up limits, but the
- 59:55tasks he performs in Coda do use limits
- 59:58. You can have one Dot per account, but
- 1:00:01OpenAI promises teams of these Dots in
- 1:00:03the future. Well, I think I've said
- 1:00:06everything I wanted to say. If I forgot
- 1:00:07something, I'll write about it in my
- 1:00:09Telegram channel. Plus, if you still
- 1:00:12have any questions about using Coda,
- 1:00:13again, write them in the YouTube
- 1:00:15comments, I will answer them. And you
- 1:00:18can also share your own experience
- 1:00:20using that space, this space and this
- 1:00:23Dot. It will also be interesting to
- 1:00:25read. Thanks everyone for watching. I
- 1:00:27hope this video was useful for you.
- 1:00:29Check out other videos on my channel
- 1:00:31too. For example, this pretty good
- 1:00:33video. Bye everyone. Yeah.
About this transcript
This page contains the full transcript of Дай мне 1 час и ты получишь 1000 часов знаний о Codex (2026) by Матвей Шульга, generated from the public captions YouTube serves with the video. The transcript has 11,257 words across 1,643 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 choose a caption track available on the video.
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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.