AI intro 1 - Introduction

Normalized prose based on the original transcription. The content was not intentionally shortened; call artifacts, timestamps, service speaker labels, and obvious speech glitches have been removed.

We have already talked with many people here, and not with some others yet, but it is time to talk about AI. Raise your hand if you are already actively using AI. I have absorbed it by about half so far. Do not worry, everyone will now become professional AI users—after today and probably a couple more calls.

Overall, the reality is that everyone in IT is now starting a major AI automation project. There is always a lot of interesting stuff that can and should be automated, but essentially everything is divided into four blocks that we need to figure out.

The first direction is a unified AI knowledge base. The second is skills for IT processes, meaning specifically for IT: who makes decisions, what the task-setting flow is, what the priority is, and so on. The third is AI skills for general operational automation, meaning project and product work, some administrative tasks; we need to determine where we can implement some processes in all of this. The fourth block is frontend.

But to begin implementing this, we first need everyone to learn how to work with AI so that we are on one page, on the same page. This is very important, because AI right now is truly high-tech, high-tech. The processes we will be implementing are based on the ChatGPT 5.5 model, and this model was released only two weeks ago. All these processes that we are about to come up with are literally two weeks old. A month ago, it was impossible to do this. The reverse is also true: what we come up with now may completely change again when a hypothetical ChatGPT 6 comes out, for example, in July or August.

We need to learn to stay precisely in the moment of current AI technologies, not get stuck in technologies from a year ago, and not tell each other fairy tales that AI generates bad code, but also not invent something from the future. We need to understand exactly what can be done now, what cannot be done, how to do it, and so on.

By the way, the processes are drawn here as they are now: a bunch of red lines going in different directions. We need to organize these chaotic lines. To get everyone on the same page and learn AI, we created a basic AI guide. It consists of three points.

How do you learn to use AI? You need to believe in AI. The first point is that we all need to believe in AI. This is a somewhat ideological point, mythology. I am currently talking to many people and hearing dozens of opinions: one person does not believe because they tried it and nothing works; another does not believe because AI is dumb and they are smart; a third does not believe because they are a super-genius and read an article somewhere saying that AI cannot produce repeatability; a fourth has something else. In short, there is a huge amount of all kinds of mythology from every direction.

We all need to arrive at a shared opinion—simply believe in AI. I can say that believing is very easy: you need to start using it and do so in the right way. If you start using AI correctly, that belief arrives literally within a couple of hours: a person sees the magic AI makes possible, and they immediately experience this life change.

In fact, all processes will change by 100%. What you were doing half a year ago, or are doing now, is already outdated—and hopelessly outdated. Almost all processes will need to be changed. That is exactly why we gathered here.

Now I will say right away, regarding developers: developers have split. There are experienced developers with extensive expertise who are now telling stories that AI writes code poorly, that it needs to be fixed, checked, and done as they were taught five years ago, and all that kind of thing. Therefore, experienced developers—you are immediately in the risk zone.

There are other developers who have mastered how to generate AI code now and understand that if they need to write a script, for example, to find large files in a database, they no longer need to write such a script manually—they write in ChatGPT: "Yo, write me a script to find large files," and in three minutes they get a ready script. Then they write two commands to fix the script, and that is it—you have a finished program; there is no need to spend time and try to remember all the variables, write obvious code by hand, and so on. All these processes have changed radically, and we need to learn how to do all of this.

I would also add one point here: if there are any doubts, then at first it may be worth starting to use AI slightly excessively, especially in development, in order to develop a certain intuition—where it makes sense to do something with an agent and where it does not. It is like learning to work well with a person: gaining experience and understanding where you can do something quickly yourself and where it is better to delegate and not even spend your own resources on solving it. This comes only with experience: if you force yourself to do it iteratively, even sometimes getting a bad result, after several iterations you begin to understand better where it is worth asking the model something quickly and where you can figure it out yourself. On the other hand, based on my experience, I can say that there are parts of code that are still very poorly suited to vibe coding.

The most important thing is that, initially, not all cases where AI can be applied are obvious. Your mind simply does not work in this mode: "Aha, here I am sitting and programming a function, but I could have received the result from an agent in five seconds." This transition to delegating to an AI agent does not happen immediately, so I am in favor of forcing yourself in the initial stages, even slightly beyond what is needed.

AI intro 1 - Naming AI models

AI tools have somewhat strange naming. We need to figure it out.

We currently have AI companies:

• Anthropic
• OpenAI.

These AI companies have web and mobile programs for ordinary users, so-called chat interfaces where you can ask AI questions and receive answers, called:

• at Anthropic, it is called Claude
• at OpenAI, it is called ChatGPT.

Next, there are AI models.

• at OpenAI, the models are called GPT with a number, the latest version is 5.5, as well as prefixes such as pro, mini, nano, and so on. There are also other specialized models: if you work with images, this model is called Image 2.0, or o1 for research, Whisper for audio, and so on; there are many models.
• at Anthropic, model names are more specialized: the main model for dialogues is Sonnet, Opus is for coding, and there is also the lightweight Haiku. Haiku is the lightest.

If you ask how to fry eggs—that is Sonnet; if you do some vibe coding—that is Opus. The third model is Heika, a lightweight version.

In ChatGPT, the issue of a lightweight model is solved differently: as a lightweight model, you can use models from older versions, for example gpt 5.2, gpt 4.1, and so on. For example, for transcribing audio into voice, there is no point at all in using the gpt 5.5 pro model; the gpt 4.1 mini model is used. However, in Claude you can also use this approach; for example, for code review you can use Opus 4.6 instead of the more expensive Opus 4.8.

Next, there are desktop applications for vibe coding:

• at OpenAI, the desktop application is called Codex.
• at Anthropic, the desktop application is called Claude Code, and there is also a Claude Co-worker version—this is not for coding, but more like a companion in day-to-day work.

Accordingly, Codex is the same thing as ChatGPT, only on desktop.

There is also a platform for using AI models via API.

• at OpenAI, the API platform is called OpenAI API Platform
• at Anthropic, the API is called Anthropic API.

On the OpenAI platform, you can work through the API with numerous AI models—text, audio, voice, images, video; there are assistants—an enormous set of tools.

This is the superpower of the AI API: you can use and combine many different models in your applications, solving in a single request tasks that previously required building full-fledged large services and products. Now just one request to the OpenAI API is enough.

This puzzle needs to be remembered in order to understand that when we say we need a Python package for OpenAI, this means that our Python on the computer must be able to access the OpenAI API out of the box.

AI intro 1 - Basic rules of communicating with AI - Rule 1: using a paid top AI model

We had two main guides:

• how to use AI
• and how to set up the environment. Let us make two more guides—how to use coding and vibe coding, and how to write prompts.

Our first guide is how to use AI at all. This is quite a complex topic because, in my experience right now, few people use AI the right way. We collected points on how to use AI correctly. You can add to them—this is a super-dynamic story, and what we are discussing now may possibly become outdated soon.

It is important to understand the basics: if you use AI incorrectly, it will inevitably lead to incorrect decisions. You first need to learn how to formulate requests correctly, and then AI will produce correct solutions. If you formulate requests incorrectly, AI will produce incorrect solutions, and you will quickly join the ranks of adherents who tell everyone that AI is dumb. But in fact, it is not dumb—it simply answers what you specifically ask it.

There is no need to join the ranks of those who say AI lies, hallucinates, is stupid, is not repeatable, and so on. Let us figure out how to use AI correctly.

There is an important point here: we are talking specifically about modern AI that works right now. This AI has a number of characteristics—it is neither good nor bad, it is neither dumb nor smart; this is simply how AI works right now.

The first rule—regardless of what anyone says—is that you need to use paid ChatGPT. This is the foundation of all foundations. ChatGPT currently comes in three versions: the default version for 20 dollars, a light version for 5 dollars, which will generally also work, but since we are developers here, it will not work for you. Your option is either the 20-dollar one or the third version, for 100 dollars.

The very first thing you need to do is install paid ChatGPT and buy the twenty-dollar plan.

The second point is that it really must be ChatGPT. I have never said this before, but this week I met a colleague who began convincing me that ChatGPT is dumb. Such cases are always interesting to me. I asked him to send me a screenshot of his ChatGPT. He sent me a screenshot of a completely different application. I asked where he got it. He said he simply typed "ChatGPT" into Google Play, downloaded it, and uses it. This is the second important point: make sure you are using the original application.

What about using other models?

The AI Elon Musk makes—Grok—has a fairly limited use case. It is worth using if you are looking for some information from Twitter, but using Grok needs to be learned separately. To get a normal result in Grok, you need to put effort into it: set limits on the amount of output, configure additional parameters, and so on.

Is everything clear with the first point? Who else wants to try DeepSeek because it is cool? Nobody—excellent.

AI intro 1 - Basic rules of communicating with AI - Rule 2: the feature of AI adapting to your question

The second point is related to the characteristics of AI itself.

It always adapts to your question. If you ask, for example, "what sad weather today, it is snowing," AI will write a hundred confirmations that snowy weather is sad. If you ask, "what cheerful weather today, it is snowing," AI will confirm that snow is fun.

You need to remember this: when forming a request to AI, do not suggest the answer inside the question. This is fundamentally important. If you say, "vibe-code me a native mobile application," AI will vibe-code in a native programming language. If you simply say, "vibe-code me a mobile application," AI will ask what we are going to vibe-code in, and you get a superpower to choose—and choose, for example, Flutter. Then you will take the right path.

This point is especially important in vibe coding. When you are vibe-coding some function or a small thing, if you suggest how to make it in the question itself, you will always be off target. AI will become not a super assistant, but a colleague who will only agree with you and worsen your incorrect decisions.

How should it be done? If I do not know how to do something, I write exactly that: "what are the options for solving this task?", "how do I write this function?", "how do I make this script?", "why are we making this config?". You need to ask such correct questions, and then you will get the correct answer. This is an extremely important point to understand.

AI intro 1 - Basic rules of communicating with AI - Rule 3: the feature of AI helping you find confirmation for your incorrect decision

The next point is the same thing, but from the other side.

You cannot offer AI a solution you have already invented, because it will tend to confirm it. It is precisely this feature that produces many stupid decisions supposedly based on confirmation from AI. I recently read about a girl who came up with a new theory of relativity, better than Einstein's. She really came up with it. How did it happen? She took ChatGPT and started writing to it: "you and I will come up with a new theory of relativity." Of course, AI helped her come up with that theory. They came up with it together—the girl wrote invented assumptions, and AI invented confirmations for those inventions. Do not do that.

There is a lifehack that lets you solve this task with AI better: you need to always request a different assessment, an assessment through the eyes of another person.

How is this applied in vibe coding? If there is a fork for making a decision, you need to write directly to AI: "what are the solution options?". It will show all the options. You can use the Linus Torvalds approach—the best programmer in the world. You can write to AI: "evaluate my code through the eyes of Linus Torvalds," and you will get an analysis of your function, script, or anything else from that person's point of view. Want another opinion—request the point of view of someone else. This way, you can always validate the process and avoid entering a cycle of inventing nonexistent situations.

These two principles—adapting to the question and confirming an incorrect decision—are fundamental. You need to use them every day when working with AI, otherwise there will be trouble, and perhaps worse. It is true that if you also break Rule 1 and offer an incorrect solution, for example, in DeepSeek, you will get double or triple distortion, and instead of a great solution you will get something incomprehensible, which is usually called AI hallucinations.

Remember: first you need to do research, as in an ordinary task, and it must be impartial. If you initially lean in some direction, AI will immediately pick up that bias and your entire work process will be distorted.

This is the foundation, the foundation of foundations.

AI intro 1 - Basic rules of communicating with AI - Rule 4: applying context

Next—about context.

Next come the details. The details lie in context. You should try to provide context. Yesterday I met a person who claimed that AI is useless without context. That is not true—AI also works fine without context. But for some answers, context is needed. For example, you can ask AI which programming language is better—NodeJS or Python? Without context, AI will give different answers every day: one day it will say NodeJS, another day Python. If you and a friend ask this question simultaneously on different accounts, it will tell one person NodeJS and the other Python. Because without context, AI answers this question randomly: which is better—AI logically thinks, any one, what difference does it make?

Accordingly, you need to understand that AI may produce completely random answers to some questions if they contain no logic known to AI. This creates a major topic for AI haters—that AI "lacks repeatability," that AI "deceives." Screenshots often appear where AI answered some nonsense. But if the question is difficult, unclear, not obvious, then the answers will be like that. If you ask something simple, such as how to fry eggs, you will get an optimal instruction immediately.

But a question such as what to have for breakfast today—eggs or pancakes—will produce a random answer.

Accordingly, you need to provide context. Context is, for example: I woke up today, I want something, but I do not want bread. If you provide that context, then when you ask what to fry—eggs or pancakes—it will understand that a pancake is somewhat like bread, and our context is that the person does not want to have bread for breakfast, so AI will answer—fry eggs. This is how context works.

The exact same thing happens with your software development tasks. No matter how complex the task is, you need to learn to feel when AI needs context.

At the same time, context needs to be applied wisely. On some tasks, context may be perceived, see Rule 3, as a solution you have already invented. Therefore, you need to be careful here. But nevertheless, if you are getting random answers from AI, it means there is not enough context.

For example, yesterday I had a situation: in my program, there are two options for connecting an external service—the first option is via API, the second is via a plugin. It is not obvious to my Codex whether to connect through the plugin or through the API. From its point of view, both connection methods are completely valid, and that is indeed true.

I had to give it context: say that when you connect through a plugin, not all functions are available to you. Secondly, the plugin is slow: a request takes not half a second, but 3–4 seconds. I do not know why it is slow, but that is how it works for me. Thirdly, with the plugin we encounter access problems, because authorization in the plugin is through OAuth, while in the API I have access through a token. When I explained this to it, it immediately said: aha, now I will tell you which functions are available through the plugin and which through the API. And it gives me: 10 functions are available through the plugin, 30 functions through the API. After that I say to it: listen, we need more functions. It is like: okay. And now the situation has become obvious for AI: we connect through the API because we need more functions, faster access, more reliable authorization through a token, and so on.

Another important point: you often encounter the position of AI specialists who try to build enormous context systems with complex relationships, and here you need to understand that everything is good in moderation. Some tasks may indeed require a large context; for example, we use huge questionnaires to solve product tasks. But to choose the right solution for a software development task during vibe coding, as a rule, even using several phrases to clarify context already makes your result many times better.

Does everyone understand what context is? Okay.

AI intro 1 - Basic rules of communicating with AI - Rule 5: ask more questions

The next point is about asking many questions. This point appeared after I talked to one big boss. He told me: why would I ask AI twice? I value my time. Unfortunately, AI will not help such a person. Why? Because it is almost impossible to get an answer to a question from AI on the first try.

I have written here: ask 10–20 follow-up questions. If you have a product task in front of you, for example, how to make a great product, then you should ask around 20 questions. Or 200. If you are simply implementing some function in vibe coding, then as a rule you do not need to ask 20 questions, but 2–3 questions are genuinely necessary. This should become your good AI habit.

The "one question—one answer" model does not work. Right now, it works poorly. I wrote this guide in January, and now I can make an update: it still does not work. Since January, the situation has only gotten worse. Do not do it this way.

A huge number of AI haters, as a rule, send screenshots: one question, got an off-target answer, took a screenshot, went to boast that they caught AI by the tail. Do not do that. You need to understand: you always need to ask many questions.

The dumber your questions are, the more effective the AI answer will be. When you vibe-code your first web application, do not be shy about constantly asking throughout the entire workday how to make the frontend, why it is this way, why not another way, and finding out how and what to do.

Question: give an example of how to ask questions during vibe coding?

For example, you decided to implement a function that displays the USD-to-EURO exchange rate. Then you write in Codex: I need to make a function displaying the USD-to-EURO exchange rate. It tells you: great, let us do it now. But you should not agree immediately; you need to ask questions. Listen, what information sources are there for the USD-to-EURO exchange rate? It will write: Forex, NYSE, Binance, something else. Then you ask: what is the fastest and most reliable way to get this rate? It answers: here are one, two, three. Then you ask: which of these methods is free? It says: this one is free. You ask further: what technology is best to use to retrieve it so the rate on the page changes in real time? It will mention, for example, a webhook or another technology. Then you ask: show me all API functions, all capabilities available in this technology. It shows them. And that is it—you can already make a perfectly precise decision. To do this, you need to ask questions.

This is the correct way to apply the principle of follow-up questions.

Of course, there are exceptions where you definitely do not need 10 questions. This is a range of coding and technical tasks that have already been solved a million times before. Here, with one request to AI, you immediately get an ideal solution. There are a great many such examples: a situation where research is not needed, but rather a solution that has already been worked out somewhere a million times. There is no need to search for anything; everyone has long known how to do it, and generating perfectly correct code works in one request. . In code, in the technical part, there are many repetitions of already solved patterns, and in that case one good request always gives you an excellent result without problems.

Here you need to get used to it at the development level. In my experience, after some time—not a very long time—you will develop a sense of where it is worth clarifying with questions and where you can immediately generate the script you need with one request.

Another common situation is when you work with someone else's previously created code, where some programmer wrote something several years ago and only they themselves and their friend know what they wrote there. Then, of course, to analyze such situations you need to ask as many questions as possible, because we do not know what they did there. Nobody knows what they did there.

But if we are dealing with problems that have been solved a million times, then, of course, you can do everything with single requests, and everything will be wonderful.

AI intro 1 - Basic rules of communicating with AI - Rule 6: use pictures, do not use CLI

The sixth principle, the most important of all six important principles, is to use photos and file uploads. When you install paid ChatGPT on your mobile phone, you gain the superpower to photograph everything around you and get information about that object. I live in Turkey, and from time to time I photograph plaques by ancient Romans and ancient Greeks, and ChatGPT translates perfectly, word for word, what they wanted to say three thousand years ago. Accordingly, if ChatGPT can read ancient Romans, then you can screenshot and upload your program's response, an error in code, or a message from your program into your Codex to speed up your work.

How is this applied in vibe coding? In vibe coding, you can screenshot any response from any program and upload it back into the Codex interface. And this works wonderfully. AI vibe coding is essentially screenshot coding. You constantly send these screenshots and add your requests to them, thereby getting precise information about what is happening and progress on your tasks.

This Rule can be used in many other non-obvious cases that we need to research together. For example, you were sent a 300-page guide on how to learn to code in Python—take that PDF, upload it into Codex, and now you are an experienced Python programmer. Any files can be uploaded to AI.

Exception—you cannot upload confidential information, commercial information, personal data, or other data that your company prohibits from being uploaded to AI. Do not send first and last names, do not send addresses, do not send financial reports, do not send banking information, and so on. That is, we do not use anything sensitive and protected by law or corporate policy in working with AI. We use software solutions, process solutions, and solutions for coding functions and processes.

You need to force yourself to truly get into these six principles and learn to use them despite not wanting to, and then everything will be good; you will become an AI builder.

AI intro 1 - Environment setup - operating system

The next large, important section is setting up the environment for vibe coding. This was news to me: I would never have thought that AI depends on the operating system on your computer and your Windows or MacOS accounts. But it is a fact. When you install Codex or Claude Code, which are desktop applications, they begin working with your code tools, various APIs, mail, Jira, and other things.

And if your Windows or Mac has an outdated operating system—for example, not Windows 11 but your beloved Windows 10, which you installed five years ago and really like—then AI tooling will not work on Windows 10.

Formally, you will be able to launch Codex, formally it will launch, but during work you will constantly run into problems: PowerShell did not start, there were not enough access rights, a folder was not enabled, a package for a programming language did not install. The scale of problems will simply be endless. And Codex will try to work around or solve each of these problems, which will lead to an enormous waste of your time. In the case of Claude Code, this will lead not only to an enormous waste of time but also to a loss of money. Therefore, you should not do this.

You need to strictly install the latest version of MacOS, or strictly the latest Windows 11. If you install Windows, you strictly need Windows 11 Pro; on Windows 11 Home Edition, the problem will be the same: you will constantly experience errors with access to folders, outgoing web connections through the firewall, and so on.

And your Windows 11 Pro must necessarily have an active license in order to receive all the latest updates for all Windows components. Without the most modern operating system, it is impossible to assemble a workstation for AI vibe coding.

In summary:

First step: you need to get the latest version of the MacOS or Windows operating system. On Windows, it must be Pro only. Second step—you need to install the operating system on a completely clean computer so that there is no inheritance of access rights, folders, and programs accumulated over years of computer use. First delete everything from the disk or get a new computer. Third step—be sure to create only one account with full admin rights; an operating system with multiple different accounts will not work, because you will inevitably encounter problems with access rights and programming language packages installed on the computer. If you try vibe coding on a computer with two accounts, the idea initially seems great, but in practice nothing will work. Ultimately, your Python will get confused about versions between the two accounts, start mixing up file storage locations, settings storage locations, package installation locations, and so on.

This is the configuration that will allow you to start vibe coding well. Deviations from these rules will lead to problems.

AI intro 1 - Environment setup - required programs

Now we will go over which programs should be installed on your computer. Developers most likely install the entire set of programs at once, but for product people all this will be new. In general, installation on a Mac is many times easier because many libraries are already included in the operating system; Windows is more difficult, you need to install everything manually. But here I agree with developers: if your development environment is already configured, you do not need to do all these exercises—you can simply install Codex and go. But if the environment is not configured, it is better to do everything from scratch, because there will be problems.

Now what needs to be installed for the environment?

Imagine that you installed Windows, it is clean, you installed all the drivers for various devices, and Windows is ready to work.

After you have clean Windows, you install Codex. Codex is the ChatGPT desktop application. You need to install such a Codex desktop application.

As a second step, I usually install a screenshot program, but in Windows you can use the standard screenshot application; it is excellent, and it now records video perfectly too.

Every year, we will write fewer and fewer letters; vibe coding mainly uses screenshots and voice messages.

Next, we need to install Git. Git is a thing for writing code directly on your computer. AI skills are already required here. When you install Git, it will give you 10 screens, each with 10 checkboxes. You should remember the rule we just studied: if you see text you do not understand—screenshot it and send it to Codex. That is exactly how Git is installed on your computer. Moreover, in Codex you provide context: you tell Codex that we are installing Git and setting the options for those checkboxes needed specifically for working with Codex. Of those 10 screens, it will leave most as they are, but change several checkboxes.

The AI skill of configuring unclear checkboxes in unclear programs will also be applied in all other applications. After Git, you install PowerShell. This is an interface for working with the command line; it is also called a terminal, and the more fashionable name CLI is also often used. PowerShell is what computers looked like 30 years ago. Back then, it was called MS-DOS.

You can write some command in PowerShell, for example the Claude command, and the Claude application will launch in the CLI version. PowerShell can be confusing because I currently have three of them on my computer: there is the standard Windows PowerShell, PowerShell 7, and a regular terminal. Codex says that for full-fledged work with AI, you need to install PowerShell 7, and indeed, when working with it, problems with access to the command line from the Codex interface are resolved. Further in the text, we will call PowerShell 7 the terminal.

Next, you install something called GitHub CLI. This is a mechanism that allows you to authorize in your GitHub from the terminal; GitHub is a web service for storing and writing your code and is located on the internet. GitHub CLI is installed through the terminal: you simply type the command to install GitHub. Since you are already experienced vibe coders, of course, you do not type it here but first in Codex. You write in Codex: “Write how I can install GitHub CLI,” and Codex will give you the exact installation steps. This is how everything works now.

Then, personally, I install FFMPEG codecs. Why? Because most automations I am currently making are based on voice interfaces: you say something or listen to something. For example, I have now trained myself to create Jira tasks through voice messages—I simply dictate for two minutes what I need to do in the Jira task, and then it immediately lands in Jira with all the checkboxes and fields. The voice interface is controlled by these codecs. There is another use as well: if you have Google Meet and need to make a transcription, it also works partly through this thing.

An important rule is that if, for example, you do not install FFMPEG codecs but try to work with audio content through vibe coding, you may end up in a situation—and with Claude you will definitely end up in one—where your AI starts trying to solve the lack of codecs by writing them from scratch itself. This leads to colossal token overspending. Always think about and check which packages, codecs, or libraries you need to solve a task before vibe-coding it.

Next, you install programming languages: Node.js and Python. During installation, Node.js will ask you to install all packages. Codex took a long time to decide whether we needed these packages; in the end, we decided to install it without packages. Install Python strictly as version 3.12 and strictly somewhere on drive C, because by default Python wants to install itself in the account folder. You do not need to memorize this: when you install Python, again ask Codex which Python to install and where. Codex will tell you everything. It is important that Python and Node.js are must-haves. Even if you are a Go programmer, Python is still the language of most automation scripts: all kinds of requests related to moving tasks into Jira, connecting to APIs, and so on—all of this will be in Python.

Next, we install VS Code. VS Code is a so-called IDE, a program in which programmers used to write code by hand. Since we are modern programmers, we will mainly use this thing to read files, navigate folders, and perform some other possible manual actions. You can often hear that the correct configuration for vibe coding is VS Code plus Claude Code or ChatGPT plugins, but in practice you will encounter a mass of limitations in VS Code that will waste your time. For example, the inability to drop screenshots and files in with one click, the lack of graphical interfaces for MCP servers, and so on—the interface is outdated right now.

Regarding a text editor for working with Markdown files, I can recommend Sublime—it is one of the best—or Notepad++. But right now, I myself work in a local version of Obsidian for Markdown files, because Obsidian has the super advantage of navigation between different Markdown files and folders, as well as the super advantage of simultaneously editing a file manually and editing it through AI tools.

Then comes the least obvious installation for Windows users: your desktop Codex accesses the internet, but by default its outgoing internet requests from your computer are blocked by the Firewall. This can lead to a whole bunch of problems, so do not forget to go into Defender Firewall and grant permission to Codex. Which permissions and where to find them—we ask Codex about all of that. But this is a real issue; it needs to be solved. Also, my vibe coding in Node.js required Windows Studio Build Tools; that also needs to be installed.

We have another step—setting up the so-called Path in Windows. These are configured paths for programs. Without correctly configured paths in Windows, traditionally, we will collect error messages during vibe coding.

Next comes selecting packages for programming languages. Packages are additions to programming languages that allow you to do something special. For example, if you want to work with voice messages in Python, you need to install voice packages. There are quite a lot of them, but what do we definitely need to install? Of course, a Python package for working with OpenAI.

So, you must have clean Windows in order not to encounter access problems, PowerShell, and programming languages. Second, you must have a properly installed environment, because if you click the wrong checkbox in Git, your Git will work with Codex not directly but in some unclear way. This will lead to you receiving messages from AI instead of the answers you need, such as: “Something is not working,” “This does not work,” “No access,” “You cannot do that,” “There is a sandbox here,” and so on. You need to install everything properly.

There is a super-lifehack in packages: if you do not have the OpenAI package and go on to call the OpenAI API, your coding tool, especially Claude Code, will start writing that package itself. This is where the super-secret of the insane Claude Code token consumption lies. It works worse, and the worse part is that if you give it a problem to solve, it begins solving it literally by trying to create the missing package, codecs, or library from scratch. And you will watch this for hours and days, without even suspecting that the problem can be solved in a couple of minutes. Codex is smarter and more honest: if you tell it to work with OpenAI, Codex will most likely suggest installing the missing package. You install the package and start working directly through it.

What does this lead to? You do not spend your time, do not spend AI tokens, and do not struggle with something that other people have already written a hundred times. This applies to packages, libraries, and other things—you need to understand and remember this clearly.

Compatibility of versions of different packages within your programming language version is also important, so before applying and vibe-coding, you always need to ask Codex: check the compatibility of the versions of the Python packages, for example, that we installed. Codex knows about compatibility, and if there are errors, it will fix them. You need to think about this and do it in advance.

A couple of weeks ago, I was researching how to accurately transcribe Google Meet and met some people who make Google Meet transcription, and they turned this transcription into a whole huge platform. They are proud of it. Although in fact, to make a transcription, we simply need to install FMPG codecs, install the necessary Python packages, write a Python script, connect it to Google Drive, download the file there, and transcribe it through our ready-made packages. Codex will tell you which other packages are needed. You install all these packages, and your transcription happens in one click.

And those people I spoke with today also use vibe coding, also use Claude Code, but they wrote the entire infrastructure effectively from scratch. They proudly showed me this application. I can see that they are proud, and I do not want to offend them, but I understand that this is implemented literally in several requests. I mean, really in a couple of requests. The question arises: how were they able to vibe-code this? The most likely answer is poorly conducted research and Claude Code's modest silence about properly setting up the environment.

Therefore, we always need to think about this in advance. When we connect to Jira, the first thing we need to think about is: what Python packages exist for working with Jira? What are we even planning to do there? What OpenAI packages are there for working with Jira? What Node.js packages are there for working with Jira? We look for these packages, find them, make sure they are made by official contributors, safe, and virus-free. We install all this into our development environment and only then launch our vibe coding and everything else. Otherwise, we will fix what does not need fixing and code infrastructure—which does not need to be coded.

By applying these simple rules for installing and configuring your vibe-coding environment on your desktop or laptop, you immediately make your work and token consumption dozens and hundreds of times more efficient. It is important that if you do not do this and do not think about it in advance, your Codex or Claude Code may never tell you about it during work; you will simply waste your time uselessly, burn a huge number of AI tokens, and become disappointed with the results of vibe coding.

AI intro 1 - Environment setup - AI builder diagram

Now the final thing. We have a wonderful diagram of how AI works, called AI architecture. Everyone will need to look at this diagram so that in the next session we can go through it substantively.

Overall, from the diagram you should understand why the whole world is going crazy over AI, including me.

The magic of AI is that it instantly connects several things that were previously completely impossible to connect.

First, it connects your computer, your laptop, with the folders on your computer—now AI works directly with your files.

Second, through the PowerShell terminal and Git, AI connects you with programming languages installed directly on your computer: Python, Node.js, any language you want can be installed.

Why does Codex need to be installed on a computer, why must it be desktop rather than a mobile application? Because your Codex specifically connects to your computer, to the folders and programs installed on it. In principle, you can connect to any program installed on your computer.

Third, you connect to your web services through special connectors, Plugins. In Codex, you connect your favorite GitHub, Figma, Jira, Gmail, Google Calendar, Google Drive—any web services that have a Plugin or API.

In fact, you connect to your entire web world, to all your web products: Notion, Google, Jira, Figma—anything you want.

Fourth—you connect to the OpenAI API, that is, to the brain, infinite computing power. Just as with the invention of the internet you connected to infinite information, with the invention of AI you connected to infinite computing power.

And all of this is managed through the Codex interface, essentially through a chat similar to ChatGPT.

And on the fifth side—your applications that you vibe-code. Any AI automations, services, and products that you can vibe-code yourself by installing all elements of this environment and connecting them through AI. This is how we assemble this puzzle of superpowers.

What will we do with project and product work? What is our supertask? We need to connect programming languages with corporate web services and then connect them with OpenAI AI models—voice, images, video, any models. And make AI automations so as to completely remove the executor's work from our processes; we should perform only the work of a process orchestrator. As soon as we encounter a manual task or something that requires execution—we add the necessary AI automation to our work. After AI integration, we now gain such a superpower.

Therefore, study this chart. On one side—your computer with programming languages, on another—connectors to services, on the third—infinite brains of AI models, on the fourth—our applications. These applications, mini apps, are what we will make.