Claude и Claude Code: полный гайд от установки до сложных задач — Transcript
Full transcript
- 0:00This video will be a comprehensive guide
- 0:03on setting up Claude and Claude Code.
- 0:06In this video, guys, there will be
- 0:08everything you need to know to use these
- 0:11tools as simply and effectively as
- 0:13possible.
- 0:14And this video will consist of two parts
- 0:17. The first is for absolute beginners:
- 0:19where to download it, how the interface
- 0:21is structured, the difference between
- 0:23various operating modes, what everything
- 0:25is for, and so on.
- 0:27The second part will be about how to set
- 0:29up Claude and Claude Code for advanced
- 0:31work.
- 0:33We will create a working folder, connect
- 0:35the necessary services, and add a
- 0:37selection of skills that I recommend and
- 0:39use daily.
- 0:40And I will show you three ways to work
- 0:42with Claude that will unlock its
- 0:44potential.
- 0:46There will be timestamps in the video
- 0:48description, so you can jump straight to
- 0:50the section you need or find your way
- 0:52back later to review something.
- 0:54And additionally, guys, I have also
- 0:57prepared a guide for you. It will be
- 0:59available on my Telegram channel. I will
- 1:01explain what is in this guide and how to
- 1:03use it throughout the video. Guys, this
- 1:05will be an incredibly useful video. You
- 1:07simply won't need another video. Let's
- 1:09go. So, let's start with the very basics
- 1:11. Where to use it and how to download it
- 1:13. Claude has a web version that you can
- 1:16use right in your browser, and there is
- 1:19a desktop application. Spoiler alert: it
- 1:22is the desktop application that unlocks
- 1:25all the possibilities of this tool. You
- 1:27can download it from the official
- 1:29website. From this one right here. I
- 1:31will leave the link in the description.
- 1:32We select our operating system, download
- 1:35the file, install it, and log into the
- 1:37account. You can also authorize using
- 1:40Google. Okay. Now, we will need the
- 1:42downloaded application a little later,
- 1:44but we will start our acquaintance with
- 1:46the web version so you can understand
- 1:48the difference. And at the same time, I
- 1:50will tell you about Claude's models and
- 1:51the differences in subscriptions. Okay.
- 1:54We open our Claude in the browser, and
- 1:56this is what the start screen looks like
- 1:57.
- 1:59In the center is the field where we
- 2:01write, and if you click the plus sign,
- 2:03you can attach files, attach documents,
- 2:05or select a connected tool.
- 2:07Next to it, we also have a microphone to
- 2:09dictate tasks by voice. This is an
- 2:12absolute must-have when working with
- 2:13agents, because naturally, it is easier
- 2:15to speak a lot more details than to
- 2:17write a prompt. Speaking for myself, I
- 2:20would train myself to write more using
- 2:22text. Why? Because when using our voice
- 2:25, we tend to say many, many things with
- 2:27very little substance.
- 2:29With text, however, we approach the task
- 2:31more thoughtfully.
- 2:32Therefore, if we are talking about
- 2:34architecture, if we are talking about
- 2:37planning your project or your results in
- 2:39general, I would practice writing it out
- 2:41in text and only use voice for minor
- 2:44additions.
- 2:45That is exactly how I structure my work
- 2:47, and it is exactly what I recommend you
- 2:48do. Because with the arrival of AI, the
- 2:51most valuable skill is defining the task
- 2:53correctly and communicating exactly what
- 2:56you want to do. Let's move on. Click on
- 2:58the model name, and a list drops down.
- 3:01The menu contains four families. Let me
- 3:04explain. There is Haiku, Sonnet, Opus,
- 3:06and Fable.
- 3:08They differ in speed, capabilities, and
- 3:10, naturally, token limit consumption.
- 3:13After all, tokens are the currency of
- 3:14the future. I have a separate video on
- 3:16that, you can watch it later. Briefly
- 3:18about the models. Haiku is our fastest
- 3:21and most economical model. We choose it
- 3:24for short answers, for summarizing text
- 3:26, or for extracting, for example,
- 3:27specific data from a document. In other
- 3:30words, even the simplest model handles
- 3:32simple tasks well and with high quality
- 3:34. That does not mean it is bad. No.
- 3:37Sonnet is our universal model, more for
- 3:39everyday work.
- 3:41For texts, code, analysis, and
- 3:42multi-step tasks, Sonnet is a good fit.
- 3:44Use it, don't be afraid, it is good. Now
- 3:47for Opus.
- 3:48Opus is suitable for complex analysis
- 3:50and tasks that require deep thought.
- 3:52For example, comparing several studies,
- 3:55checking conclusions, fact-checking data
- 3:58, and synthesizing it all. We use Opus.
- 4:01And the fourth model is Fable. A model
- 4:03for the most complex and long-term
- 4:05projects with many steps. This model
- 4:08takes longer to think and, naturally,
- 4:10consumes the most limits.
- 4:12Therefore, for example, I use it only
- 4:14when I need to think through a truly
- 4:16architectural solution or when I need to
- 4:19compare and see the big picture.
- 4:21An example regarding my health.
- 4:23I have test results, doctor's
- 4:24recommendations, I need to compare
- 4:26studies I found, and synthesize all of
- 4:28this into a complete picture.
- 4:30I can use Fable for that.
- 4:32Or I want to create a custom system for
- 4:34myself, a dashboard with a lot of data.
- 4:37I need it to be architecturally
- 4:39well-thought-out. I can use Fable for
- 4:41that as well. You understand, right, for
- 4:42what kind of tasks this is needed.
- 4:43Therefore, for a fast task we choose a
- 4:45fast model, for a complex one we assign
- 4:47a more complex model. Personally, I use
- 4:49Opus most often, but I use Opus with
- 4:52sub-agents.
- 4:53Meaning, I have a main Opus, and then it
- 4:55has sub-agents that already use Sonnet
- 4:56and Haiku.
- 4:58We will talk about all this, and you
- 4:59will understand. And in the second part
- 5:01, I will show you how to combine all of
- 5:03this in one task. What I just mentioned
- 5:05. For example, how a more expensive
- 5:07model directs cheaper ones. This is
- 5:09called orchestration. Because right now
- 5:11, it is absolutely essential to know how
- 5:13to build these kinds of links so you
- 5:15don't burn through all your limits in a
- 5:17day.
- 5:18Tokens are the currency of the future,
- 5:19so we treat them exactly the same way as
- 5:21your money. We approach it wisely. Okay
- 5:23, let's get back to the interface. Which
- 5:25models are available? Your menu will
- 5:27depend on your subscription. Let me
- 5:29explain. On the free plan, if you happen
- 5:31to use it, you only have access to Haiku
- 5:33and Sonnet, so you won't see other
- 5:35models. On Pro, Opus is added to them.
- 5:38And Fable can be used for an additional
- 5:40fee within Opus.
- 5:42On Max and Ultra plans, all four models
- 5:44are available, and Fable is already
- 5:46included in the subscription.
- 5:48But remember, guys, you can spend up to
- 5:50half of your weekly limit on it. Now,
- 5:53next to the model selector, we have
- 5:56another toggle. And this toggle is where
- 5:59you can set permissions. Meaning when
- 6:01Claude can act on its own and when it
- 6:04needs to ask you. For example, in the
- 6:06future, you might want it to check your
- 6:08email. Naturally, you can grant it
- 6:10permission here. Now let's look at the
- 6:12menu on the left. Let's break down what
- 6:14each section is for. We have chats.
- 6:16Chats, I think it's clear, are our
- 6:17conversation history in Claude. Here you
- 6:20can find a past dialogue you need,
- 6:21continue where you left off, or simply
- 6:23open the dialogues you need that are
- 6:25tailored to specific topics.
- 6:27We shouldn't have everything mixed up in
- 6:29one chat. A separate chat for a separate
- 6:31topic. Now the next one is Projects.
- 6:33Projects are separate workspaces for us
- 6:35. Within projects, we can group multiple
- 6:38chats, collect documents, and general
- 6:40instructions on one specific project
- 6:42topic. And Claude will use these
- 6:45materials as context. So, to start
- 6:48working, it won't need to be reminded or
- 6:49have things uploaded again; it will have
- 6:51all of this by default. For example, you
- 6:53could create a project where you write
- 6:55posts for your social media. Accordingly
- 6:57, it would already contain examples of
- 6:59your posts, rules, and the topics you
- 7:01write about.
- 7:03And when you start working within that
- 7:05project, you won't need to re-explain
- 7:07what you're doing or how you write. I
- 7:09think that's clear. But we will analyze
- 7:11projects in detail in the second part of
- 7:12the video. Next, we have the Artifacts
- 7:14tab.
- 7:16Artifacts will store the materials you
- 7:18created in Claude, so you can easily
- 7:20find them instead of digging through
- 7:22chats to look for them.
- 7:25For instance, if you created HTML pages
- 7:27in some chat, or built small apps,
- 7:28visualizations, or documents.
- 7:31Basically, all of that will be there,
- 7:32and you'll be able to find it easily.
- 7:33The next tab is Scheduled Tasks. These
- 7:36are our scheduled tasks.
- 7:39For example, you can set your agent to
- 7:42send you news or a summary once a week
- 7:44or every day at a specific time.
- 7:48If you forget or want to check what
- 7:50tasks are scheduled, you can go to this
- 7:52tab and view the schedule to keep track
- 7:54of these runs.
- 7:56The next tab we have is Design.
- 7:58This is a section for visual work.
- 8:01Here you can start a presentation,
- 8:04design a page, create an interface, or
- 8:06make an interactive prototype.
- 8:09It’s a dedicated tab with a UI/UX design
- 8:11environment where you can change
- 8:12absolutely everything. It is amazing
- 8:14what they have done.
- 8:16And again, what you create here will
- 8:18then be stored in artifacts.
- 8:20You can easily return to them as well.
- 8:21And another very important tab is
- 8:23Personalization. This is where our
- 8:25skills are located. Skills are
- 8:27instructions for specific step-by-step
- 8:29tasks. Connectors are also found here.
- 8:32Connectors are how Claude links to your
- 8:34other services so you can work within
- 8:37them.
- 8:38For instance, if you work with your CRM
- 8:40, Telegram, or an application like
- 8:42Eleven Labs for voice generation.
- 8:46This is where it will all be stored, so
- 8:48you can easily go in, see your
- 8:49connectors, and link new apps to your
- 8:51Claude so it can interact with them.
- 8:54This will be located here.
- 8:55Again, in the second part of the video,
- 8:57we will return here, install the
- 8:58recommended skills, and connect the
- 9:00necessary services, so we will work with
- 9:02it more.
- 9:03Next, we move on to the important part:
- 9:04Chat and Cowork. Another thing important
- 9:06for me to tell you regarding the
- 9:07interface.
- 9:08We already released a guide on Claude,
- 9:10and back then there was a distinction
- 9:12between three modes: normal chat, Cowork
- 9:15mode for longer office tasks, and Code
- 9:17mode.
- 9:18Well, it's also called Claude Code,
- 9:20which we need for web coding to build
- 9:21tools, websites, and various complex
- 9:23artifacts.
- 9:24So, in the updated version, since they
- 9:26are constantly updating the app, the
- 9:28first two modes have been merged: the
- 9:29normal chat and Cowork mode.
- 9:31Now, in one conversation, you can
- 9:34discuss an idea and then ask it to find
- 9:36materials, compare them, and assemble an
- 9:39artifact. This means Claude now
- 9:40understands for itself when to just
- 9:42discuss things and when to activate
- 9:44Cowork mode. This now happens
- 9:45automatically under the hood. They
- 9:47simplified it. But the Code mode has
- 9:49remained separate.
- 9:51And here we have a toggle, and that is
- 9:53where many possibilities are hidden. And
- 9:56, by the way, the word "code" is
- 9:58important. There is no need to be scared
- 10:00. This mode, it seems to me, actually
- 10:02provides even more opportunities for
- 10:04those who don’t code at all, but who
- 10:07have their own project ideas, product
- 10:09ideas, and creativity. Meaning it helps
- 10:12by coding for you. The main thing is for
- 10:15you to describe it well.
- 10:17And in the browser, the code works with
- 10:19cloud projects, but next I will show you
- 10:21another option—Cloud Code with a folder
- 10:23on your computer.
- 10:24Accordingly, we are now moving from the
- 10:26browser to the installed Cloud Code
- 10:28application on the computer. What is
- 10:30important to say? Everything I have just
- 10:32told you about the interface up to this
- 10:34point is absolutely applicable to the
- 10:36computer application as well.
- 10:39There are all the same tabs, all the
- 10:41same buttons for you to understand.
- 10:43Then you might ask, what is the
- 10:45difference?
- 10:46The main difference is that when Cloud
- 10:48Code runs specifically on your computer
- 10:50, it gains access to your computer's
- 10:52file system.
- 10:55It can read documents, create new ones
- 10:57on your computer, edit them, and
- 10:58organize everything into folders.
- 11:02And the files and all the results will
- 11:04be stored on your disk, and you can
- 11:05return to them in future tasks.
- 11:08Furthermore, it can use the capabilities
- 11:11of your computer, you see? It can write
- 11:13and execute code on your computer, it
- 11:16can run installed programs and tools,
- 11:19and even use the browser itself.
- 11:21And for all this, your computer's
- 11:23processor, RAM, and, for suitable tasks
- 11:25, the graphics card are utilized.
- 11:28For example, for audio transcription, it
- 11:31can download and run a separate local
- 11:33model directly on your computer to use
- 11:36it.
- 11:37And then the entire recording will be
- 11:39processed on your device, and the
- 11:40finished text will be saved to a file,
- 11:42you see? Naturally, you will need to
- 11:44grant permission for everything it does
- 11:45on your computer, so be attentive to
- 11:47this. It is incredibly convenient.
- 11:49I have many agents working, and they
- 11:51periodically open my calendar on their
- 11:53own, add various tasks, and sort
- 11:55everything out themselves.
- 11:57The only thing is, of course, that you
- 11:58need to monitor this periodically.
- 12:00Meaning, do not trust AI completely;
- 12:02periodically check these folders, verify
- 12:04all the permissions it has, and see what
- 12:06it changes or adds. We are looking out
- 12:09for security. Trust, but verify. And an
- 12:11important note—Cloud itself continues to
- 12:14run in the cloud, but the commands and
- 12:16programs it manages are executed on our
- 12:19computer. So, in essence, you are giving
- 12:22Claude hands inside your computer,
- 12:24access to files and tools it can use to
- 12:26perform its work. Basically, that is the
- 12:29value of this mode. Next, I will explain
- 12:31in detail how to customize this workflow
- 12:32for yourself. But for now, let's move on
- 12:34to the initial settings. We have covered
- 12:36the basics, right? Now we will configure
- 12:38Claude itself.
- 12:40Open the Claude app and, in the
- 12:42environment selection where you will be
- 12:44working with Claude, specify this
- 12:45computer.
- 12:47In other words, by doing this, we
- 12:49confirm that Claude is allowed to run on
- 12:50this computer. We are pointing it out to
- 12:53him. Now, let's give Claude a working
- 12:55folder.
- 12:56This will be the place where he creates
- 12:59files, stores all instructions, and
- 13:01keeps everything updated.
- 13:03In short, this will be his workspace.
- 13:06Create a separate folder on your desktop
- 13:08called "Claude Projects". You can do
- 13:11this yourself via Finder or File
- 13:13Explorer, or you can even ask Claude to
- 13:15create the folder.
- 13:17Just write to him: "Create a Claude
- 13:18Projects folder on the desktop." That's
- 13:20it, he will create it for you. He will
- 13:22ask for permission, you grant it, and he
- 13:23will create it.
- 13:24Once the folder is ready, click "Select
- 13:26Folder" and open the specific folder we
- 13:29just created.
- 13:30This is how we give the agent access to
- 13:32our workspace.
- 13:34And from now on, this will be the main
- 13:36folder for working with Claude.
- 13:39Next, ask him directly in the chat to
- 13:41save all materials inside it and create
- 13:43a separate subfolder for each project.
- 13:47This way, documents, spreadsheets, code
- 13:50, and finished results will always be
- 13:52organized, and you can easily find them
- 13:55.
- 13:55Next, we need to set up your
- 13:57personalization. I will share a little
- 14:00life hack with you. This comes from me.
- 14:03You can do it on your own if you know
- 14:04how, but I will give you a different
- 14:06method that is simpler and even more
- 14:07effective.
- 14:09To set up personalization and configure
- 14:10Claude for your needs, we won't rack our
- 14:12brains thinking about what to tell him;
- 14:14we will ask him to conduct a short
- 14:16interview with us. Just turn on
- 14:18dictation and say: "I just installed
- 14:20Claude.
- 14:21Help me configure it for my tasks. Ask
- 14:25me what I do, what projects I have, what
- 14:27services I work with, and in what format
- 14:29I want to receive the results." And now
- 14:31he will ask you questions, and you will
- 14:34answer them one by one. You can tell him
- 14:36that you have several projects so that
- 14:38he creates different subfolders for you
- 14:39right away. Here you can share the full
- 14:42picture of your work, and he will take
- 14:44it into account. For instance, I explain
- 14:47that I have several projects and in one
- 14:48of them, I prepare videos and texts. I
- 14:50need scripts. In research, for example,
- 14:52primary sources are important to me. If
- 14:54there isn't enough information, please
- 14:56ask me to clarify.
- 14:58Basically, by answering these questions
- 15:00, you're syncing up with the AI agent so
- 15:02it understands exactly what you want and
- 15:04how you intend to work with it.
- 15:07It’s just like getting to know a new
- 15:09person. Now, here is a very important
- 15:11question: where will it store this
- 15:13information?
- 15:14Well, in case you missed it, I mentioned
- 15:16earlier that it will store it in
- 15:18specific files locally on your computer
- 15:20.
- 15:21So, we’re going to walk through that
- 15:23right now. This is extremely, extremely
- 15:25important. So, memory and instructions.
- 15:28After the interview, we ask Claude: "
- 15:30Save my general rules to the user
- 15:32Claude.md file and show me where it is.
- 15:34"
- 15:36To be clear, this will be a very
- 15:37important place that you—guys—need to
- 15:39visit and periodically review to ensure
- 15:42everything is correct, and see what it
- 15:44changes or adds. We open this file.
- 15:47Here is recorded what Claude should know
- 15:49about us and how we want to work with it
- 15:51.
- 15:52Basically, this will be a standard text
- 15:55file with permanent instructions.
- 15:58You can open it, read it, and edit it if
- 16:01the AI misunderstood something.
- 16:04You can add important details here
- 16:05yourself so you don't have to keep
- 16:07telling it to redo or add things; just
- 16:08go in and add them manually. It will be
- 16:10much easier that way. And this Claude.md
- 16:12is simultaneously your agent's identity
- 16:15, its knowledge about you, and the rules
- 16:17for working with you. Once again.
- 16:20Your agent will maintain this main user
- 16:23file by default.
- 16:26I’m just saying that you can go in there
- 16:27and change things yourself. Next to this
- 16:29file, there may be auxiliary files, for
- 16:33example, called Memory. Memory is a more
- 16:36detailed record about us. In most cases
- 16:38, we don't need to mess with that. This
- 16:41file will simply sit in your program's
- 16:43folder and get accessed by the script
- 16:45often.
- 16:46That is, the agent itself will fill it,
- 16:48update it, and record details it deems
- 16:50important for working with you.
- 16:53But this is not the only instruction
- 16:57file, Claude.md.
- 16:59A very important detail is that these
- 17:02kinds of instructions can be stored at
- 17:04three different levels simultaneously
- 17:06and used for different tasks.
- 17:09So, I’m going to break everything down
- 17:11for you in detail right now.
- 17:14Your agent can have several files that
- 17:16are all named Claude.md.
- 17:19And only the location of the file
- 17:21determines which tasks its rules are
- 17:23meant for. Let me explain. The first
- 17:25level is the general rules used for
- 17:27working with Cloud in general.
- 17:30This is the custom cloud.md in the
- 17:33hidden Cloud folder located inside your
- 17:35user folder on the computer.
- 17:38That’s exactly the one we just looked at
- 17:40.
- 17:41Here, I repeat, the language of
- 17:42communication, your preferences for
- 17:44responses, and general work requirements
- 17:46are stored.
- 17:48The second level, where I recommend
- 17:50setting instructions and creating an
- 17:52additional cloud.md, will be our main
- 17:54Projects Cloud folder that we created on
- 17:56the desktop.
- 17:58In other words, we ask Cloud to create
- 18:01another cloud.md inside it.
- 18:04And here we save more specific and
- 18:06detailed knowledge about what we do,
- 18:09what our work areas are, such as, in my
- 18:12case, separate tracks for online
- 18:15education and content production.
- 18:19This is where we outline the materials
- 18:21and where to save the results.
- 18:24These rules will be general for all
- 18:26projects within this folder.
- 18:28And the third level is the Cloud.md
- 18:31located in a specific project folder.
- 18:34Here, there will be narrower rules,
- 18:36materials, and requirements related to a
- 18:39specific task.
- 18:40You see, right? You may have many
- 18:42different tasks that you do with agents
- 18:44. You likely won't have just one agent.
- 18:46Different departments. We understand. I
- 18:48will show you with an example now.
- 18:49Content project. For example, we've done
- 18:52the user instructions and the general
- 18:54project rules.
- 18:55Now, inside the Projects Cloud folder,
- 18:57let's create a Content folder. So, this
- 18:59will be a separate project for us. In
- 19:01this folder, we will prepare posts and
- 19:03materials about artificial intelligence
- 19:04, so let's write it down. Well, you,
- 19:06accordingly, adapt all of this to
- 19:08yourself.
- 19:09Inside the Projects Cloud folder, create
- 19:11a Content folder.
- 19:12Posts about AI will be written here.
- 19:15Create a cloud.md inside and write
- 19:17instructions for this specific project
- 19:19in it. Topics, material formats, text
- 19:22style, and the fact-checking process.
- 19:25Save the finished results in a "Results
- 19:28" subfolder.
- 19:30And ask me everything needed to create
- 19:33this project.
- 19:34We understand.
- 19:36He won't just make it all up from his
- 19:37head; he will ask us now, and we must
- 19:39answer every single question.
- 19:41And then, after he analyzes them, he
- 19:43will create everything properly. There,
- 19:45we’ve stated our requirements, for
- 19:47example. Now we open this Content folder
- 19:49and see the cloud.md inside. MD. And
- 19:52here, as we can see, the work on content
- 19:54is already described: who we are writing
- 19:55for, what texts we are preparing, what
- 19:57we check, and where we put the finished
- 19:59materials. So, this is a special
- 20:01instruction specifically for this
- 20:03project. I think you’ve got it. Now, how
- 20:05will all these three levels work
- 20:07together?
- 20:08We create a new chat and ask it to start
- 20:11working on a new Telegram post.
- 20:13What is happening inside Claude?
- 20:16It is currently reading and considering
- 20:18the general user rules from the
- 20:20top-level file.
- 20:22You see, right? Where, for example, it
- 20:24says to answer in Russian, to keep
- 20:25answers brief if possible—basically,
- 20:27general instructions for all my projects
- 20:29.
- 20:30Then, it reads the instructions from the
- 20:33main "Claude Projects" folder and
- 20:35understands that I have an online
- 20:37education stream and a content stream,
- 20:40and that they are connected.
- 20:42So, it understands my work context and
- 20:44realizes it needs to go into the content
- 20:46folder.
- 20:48And after that, it reads the rules of
- 20:49the content project itself, which
- 20:51details exactly how I write posts and
- 20:53where the result should be saved.
- 20:55You see, right? So, all these files are
- 20:57three levels, and they complement each
- 21:00other. This reading of the instructions
- 21:02takes mere fractions of a second.
- 21:05We don't notice it, but when you have
- 21:07all these levels, the agent has enough
- 21:10context to perform tasks for us as
- 21:12efficiently and personalized as possible
- 21:15.
- 21:16For instance, it won't emphasize points
- 21:18in posts that would contradict the
- 21:19principles we promote in our online
- 21:21school, for example.
- 21:23Even though that's a completely
- 21:24different project, right? Online
- 21:25education is a completely different
- 21:27project. Right now, we are writing a
- 21:28project for posts. But it takes all of
- 21:29that into account because everything is
- 21:31interconnected. My life is
- 21:33interconnected with this.
- 21:35I also have separate projects for my
- 21:36health, and I have separate projects for
- 21:38my other ventures.
- 21:40These are all important details because
- 21:42, despite them being different
- 21:43businesses and different directions,
- 21:45they are all linked to me, which means
- 21:47they all affect me.
- 21:49And overall, my approach to each project
- 21:51defines me as a person, and it will
- 21:52define you too.
- 21:54But there is another scenario.
- 21:56Another scenario could be that you
- 21:57create a separate Claude MD folder that
- 21:59won't take your work context into
- 22:01account at all, but will be focused, for
- 22:03example, on analyzing your test results
- 22:05, if you don't want it to be linked.
- 22:08And this project will work in isolation
- 22:10. How do you make it work in isolation?
- 22:13You need to start a chat from within
- 22:15that specific new "lab results analysis
- 22:17" folder.
- 22:18It is done like this.
- 22:20When we create a chat and start a
- 22:21dialogue within this specific narrow
- 22:23project, Claude will only consider the
- 22:25general user instructions and the
- 22:26instructions of that specific project,
- 22:28separate from everything else.
- 22:31And you can have as many such divisions
- 22:33and sub-systems as you like. Again,
- 22:35don't worry if something is unclear
- 22:37right now. In the process of working,
- 22:39you will build all the necessary levels
- 22:41and directions for yourself. Okay, let's
- 22:43move on.
- 22:45Now we need to give our Claude more
- 22:47hands and more tools so that it can be
- 22:49more useful to you.
- 22:51I will tell you how and which services
- 22:54you should connect first. First and
- 22:56foremost, of course, you need to connect
- 22:59Google, specifically Google Drive, for
- 23:01example, Gmail, and Google Calendar.
- 23:04Through these, Claude will gain access
- 23:06to our documents, emails, and meetings—
- 23:09essentially our plans—and will be able
- 23:11to add events to them and remind us
- 23:13about them.
- 23:15To do this, we go to personalization,
- 23:18select connectors, and find the service
- 23:20we need.
- 23:21For example, here is Google Drive. Click
- 23:24connect. Then we log into our account,
- 23:26and after authorization, the service
- 23:28appears among those connected. That's it
- 23:30; now it can access it, and they can
- 23:32exchange information. And we add mail
- 23:34and calendar using the same principle.
- 23:37If, for example, you also want to create
- 23:39documents, edit them, or make
- 23:41presentations, you can additionally
- 23:43connect Google Docs, Sheets, and Slides
- 23:45.
- 23:45Convenient.
- 23:47And now, through these connectors, your
- 23:49agent will be able to create and edit
- 23:50your documents, spreadsheets, and slides
- 23:52.
- 23:53In short, everything you need. Schedule
- 23:55meetings for you and set reminders in
- 23:57the calendar.
- 23:58And, of course, besides Google, you
- 24:00understand there are dozens of other
- 24:01platforms and services you use, and you
- 24:03can connect them here using the same
- 24:05principle.
- 24:06For example, if you use Dropbox, connect
- 24:08it. If you use Figma, connect it. If you
- 24:11use Notion, connect it. There is a huge
- 24:14list here. How do you work with these
- 24:15connectors? Let me explain. Just write
- 24:17in the chat.
- 24:19For example, if you want to work with
- 24:20the calendar, write: "Find the next work
- 24:22call in the calendar, look at the
- 24:24created document, and the latest emails
- 24:25on this topic."
- 24:27"Gather a short summary of what to
- 24:28discuss, what questions are open, and
- 24:30what needs to be solved." And we see
- 24:32right in the chat how Claude begins to
- 24:34use different tools one by one.
- 24:36Everything is ready; it checks the
- 24:38calendar and completes the task for us.
- 24:39Everything is clear. Now, an important
- 24:41question that might cross your mind.
- 24:44What if I didn't find the service I want
- 24:46my agent to work with in the list?
- 24:49For this, you will now be introduced to
- 24:51a new concept: MCP.
- 24:53MCP is a special adapter that helps an
- 24:56agent work with almost any service.
- 24:59I have a separate video about MCP, so I
- 25:02won't go into technical details; I'll
- 25:04just show you how to implement it. If
- 25:06you find you lack the knowledge, just
- 25:07watch that video and you will understand
- 25:09everything. For example, let's connect
- 25:11Telegram via MCP. We'll use it, for
- 25:13example, to retrieve agreements,
- 25:15commands, messages, and files needed for
- 25:17our tasks. So, we just type exactly that
- 25:19into the chat. Help me set up an MCP
- 25:21connection to Telegram. I want to read
- 25:24messages in my work chat, download files
- 25:27, and track agreements. Basically, we
- 25:29write down everything we need from this
- 25:31service. Do you understand? This is
- 25:32important. It’s not just about
- 25:33connecting Telegram and letting the
- 25:35agent set everything up for you. Yes, it
- 25:37really is just that simple. The only
- 25:39thing, guys, is that I must mention,
- 25:41since we have a large media presence, we
- 25:43must naturally keep security and privacy
- 25:46in mind.
- 25:47That’s why I always create a separate
- 25:49email, a separate Google account, and a
- 25:51separate Telegram for Claude, not my
- 25:53main Telegram, not my main Google email
- 25:56, and so on.
- 25:57We need to be mindful of this. You know
- 25:59what I mean? There is also an option to
- 26:01use a Telegram bot, which you would need
- 26:04to add to all the groups you want to
- 26:06read messages from later. But there are
- 26:08limitations; the bot won't be able to
- 26:10download files larger than 20 MB.
- 26:13For instance, that’s critical for me,
- 26:15because our team often sends video files
- 26:17, so my preferred option is to connect a
- 26:20real account.
- 26:21But I will use a separate work number, a
- 26:23separate work account, not my personal
- 26:25one.
- 26:26Now, to connect to a Telegram account,
- 26:28you will need Telegram API data.
- 26:31The agent will go to my.telegram.org
- 26:33itself, open API Development Tools, and
- 26:36create an application. Two parameters
- 26:39will appear there: API ID and API Hash.
- 26:42The agent will take them and save them
- 26:44to the local connection settings.
- 26:46Alternatively, it might ask you to visit
- 26:48those pages yourself to get these
- 26:50parameters and pass them over securely.
- 26:53And naturally, we confirm the Telegram
- 26:56login ourselves.
- 26:58And guys, this is also very important—
- 27:00never paste any APIs, passwords, or
- 27:02authorization codes into your chat with
- 27:05the model.
- 27:06In principle, they have already
- 27:08addressed this, and the security system
- 27:10is set up so that it will guide you on
- 27:12where and how to save them.
- 27:15If it doesn't ask, then you ask it: "
- 27:17Create a separate file where I should
- 27:19load or paste these API passwords, and
- 27:21walk me through it—tell me where to put
- 27:24what."
- 27:25It will create instructions for you. We
- 27:27never send anything important like that
- 27:28into chats or the cloud. Keep that in
- 27:30mind.
- 27:31Also, I’ve left a special file in the
- 27:33Telegram channel that contains
- 27:34everything needed to connect to Telegram
- 27:36.
- 27:37So, if anything, just give it to your
- 27:39agent, and it will set everything up for
- 27:41you. So, that’s basically it—nothing
- 27:43complicated.
- 27:44After setup, we open the list of MCP
- 27:47connections here in this tab and check
- 27:49the status of Telegram.
- 27:52Then we type: read the messages in the
- 27:54work chat from today and tell me what
- 27:56decisions we made and what questions
- 27:58remain open.
- 27:59This is a check. After all connections
- 28:01are made, we must verify our agent to
- 28:03see if it understood everything and set
- 28:05it up correctly. Important. Accordingly
- 28:07, our agent goes to Telegram, reads the
- 28:09messages, and brings us the answers.
- 28:12And again, in exactly the same way, we
- 28:14can set up almost any other service that
- 28:16doesn't have an official connector in
- 28:18the app.
- 28:20What else can be said here?
- 28:23If you don't know, or you want some tips
- 28:25on what else you can connect, I'll just
- 28:27show you my basic list of what needs to
- 28:30be connected.
- 28:32It is very simple.
- 28:33I would recommend connecting all Google
- 28:35services if you work in Google, Telegram
- 28:37or the messenger you use for
- 28:39correspondence, and my favorite search
- 28:40engine.
- 28:42The search engine enhances your agent's
- 28:44capabilities. It is a special search
- 28:46engine for agents. It is called Exa.
- 28:50In short, it is a special tool so that
- 28:52agents can search the internet
- 28:53semantically, that is, by meaning.
- 28:56Understood. Let's move on. Now, skills.
- 28:59I will tell you everything you need to
- 29:01know about skills. What are they,
- 29:02generally speaking? A skill is a saved
- 29:05instruction for a specific task. What
- 29:07does it look like?
- 29:09It looks like a regular text file, but
- 29:11it contains all the details, meaning a
- 29:13specific workflow of yours, step by step
- 29:15.
- 29:16For example, how to prepare a brief, how
- 29:18to format a page correctly, how to edit
- 29:21text.
- 29:22Inside a skill, there can be rules,
- 29:24examples, links to supporting files, and
- 29:26references.
- 29:27By the way, I have a separate detailed
- 29:29video specifically about how skills are
- 29:31structured internally, so I won't go
- 29:33into technical details now; you can
- 29:34watch that. I will only say the main
- 29:37thing: a skill is essentially digitized
- 29:39expertise. And right now, 1,000 skills
- 29:41are publicly available. Designers, for
- 29:43example, digitize their expertise on how
- 29:46pages should look, post their expertise
- 29:49publicly, and you can simply take this
- 29:51expertise and use it in your project.
- 29:54There are paid skills, there are free
- 29:56skills, and there are huge marketplaces
- 29:57for these skills. Where are they located
- 29:59? We have skills within Claude itself.
- 30:02We also have a separate website. How do
- 30:04you find the skills that are in Claude?
- 30:05They are located in the personalization
- 30:07section.
- 30:08We open personalization, then Skills,
- 30:10and here we see all the pre-installed
- 30:12skills that Claude has already set up
- 30:14for us.
- 30:15We can also add skills ourselves, that
- 30:17is, upload them to our agent. Skills can
- 30:19be uploaded here directly as a file; you
- 30:21can add a text file here, one you
- 30:23downloaded somewhere or someone gave you
- 30:25, and upload it here. Claude will read
- 30:27it, Claude will add it, and now it can
- 30:29use it for a specific task. But there is
- 30:31a simpler way to install a skill.
- 30:34You can actually just give Claude a link
- 30:36to a skill you find on the internet.
- 30:38Usually our skills, or rather most of
- 30:40them, the most capabilities are found on
- 30:43GitHub. It is a kind of social network
- 30:45for programmers, but it is used by
- 30:47non-programmers now too.
- 30:49And we can just provide a link to the
- 30:51skill we need, the one we like, and ask
- 30:53your agent to install it by itself. How
- 30:56do you search for these skills on GitHub
- 30:57? There are entire collections of skills
- 30:59. That is, these are separate
- 31:01repositories with individual skills.
- 31:03And if you want, you don't even have to
- 31:05go to GitHub yourself and search for
- 31:06anything there.
- 31:08We can simply ask Claude itself to find
- 31:10the necessary skills, for example.
- 31:13For instance, you want to create a
- 31:14specific website right now.
- 31:15You just ask it: find suitable skills on
- 31:18GitHub and explain what each one does.
- 31:20Even if you don't understand anything at
- 31:22all, this is the simplest way, right?
- 31:24That's it, it opens, it searches, it
- 31:26suggests, it sends you a list, and you
- 31:28read and study it.
- 31:30If you like a certain skill, you must
- 31:32check it for safety. That means you ask
- 31:35it to check for safety, to see if
- 31:37everything is okay, if there are any
- 31:39pitfalls, or if there are any hidden
- 31:40instructions. Actually, tell me, tell me
- 31:42, because it is very important.
- 31:44Absolutely anyone can post any skill on
- 31:46GitHub. And there are skills that are
- 31:48not verified, and there are skills that
- 31:50could potentially hack you. Therefore,
- 31:52even if someone doesn't mention it, we
- 31:54always check everything.
- 31:56And if it has checked the skill and
- 31:58everything is great and it is safe, only
- 32:00then do we ask it to install it.
- 32:02And immediately, as soon as you have
- 32:04installed a specific skill, you ask it
- 32:06to work with it. Because it is one thing
- 32:07that someone created it for themselves,
- 32:09but it is another thing how you can use
- 32:10someone else's skill for your own tasks
- 32:11. If you like it, you keep it in your
- 32:13arsenal. If you don't like it, you just
- 32:15delete the skill saying, I didn't like
- 32:16it. Or, if you liked it but something
- 32:19about it doesn't suit you, you also have
- 32:21the option to redo, supplement, or
- 32:23improve it. For example, you like
- 32:25everything about it, but you don't like
- 32:26certain details. You just write: listen
- 32:28, I don't like these specific details,
- 32:29change them to fit our needs. That's it
- 32:31, it will take a ready-made skill; it
- 32:33won't be doing half the work from
- 32:34scratch, just adapting and improving it
- 32:36for you. And you will have your own
- 32:38improved, your very own skill.
- 32:40If you ask me now which skills I would
- 32:42recommend, I will tell you about four
- 32:45skills that I suggest starting with if
- 32:47you are starting from scratch. Each of
- 32:49them is useful for its own specific task
- 32:51.
- 32:52The first skill is one I would call a
- 32:54meta-skill.
- 32:56It helps us think through the task
- 32:57itself from different angles and how we
- 32:59are going to solve it.
- 33:01The workflow is defined within this
- 33:03skill.
- 33:05That is, first Claude studies the
- 33:06context, then it asks clarifying
- 33:08questions, then proposes several
- 33:09approaches and discusses with us which
- 33:11one to choose.
- 33:13The key thing here is that you might
- 33:15simply not think about a certain side of
- 33:18your issue, and this skill helps you
- 33:20look at it and suggests things you
- 33:22aren't noticing.
- 33:23For example, you want to figure out who
- 33:25your audience is, who your main client
- 33:27is.
- 33:27You export everything you need from your
- 33:30side, from your Telegram channel, and
- 33:32ask it to analyze it.
- 33:34And with the help of this skill, the
- 33:36agent can suggest that you also look at
- 33:38, for example, your region's
- 33:39demographics to get a broader view of
- 33:41your issue.
- 33:42You see, right? So it looks for a few
- 33:45more angles on your task. It is
- 33:47incredibly useful. I personally start
- 33:49every new task with a discussion like
- 33:51this.
- 33:52I answer various questions, and we sort
- 33:54of brainstorm the idea together, compare
- 33:56options, and agree on an approach.
- 33:58And only after that do we move on to
- 34:00execution.
- 34:01The second skill I recommend is one I
- 34:04suggest for web coding, when you'll be
- 34:06creating something with an agent,
- 34:08meaning where you need a visual
- 34:10interface, like making websites or apps
- 34:12, and you want them to look good.
- 34:16This skill contains a collection of
- 34:18visual styles, palettes, and font
- 34:20pairings, and Claude will rely on them
- 34:22to, for example, choose the design for
- 34:25your projects, headings, colors, spacing
- 34:27, buttons, and other interface elements
- 34:29.
- 34:30For example, we ask Claude to build a
- 34:32landing page that describes our channel
- 34:34and our principles.
- 34:36And we immediately add: use and select
- 34:38the style, colors, and fonts for this
- 34:40project, and make a selection for me so
- 34:42I can choose.
- 34:43That’s it, we look at the various
- 34:45options it immediately suggests and ask
- 34:47it to apply one to the page. If you
- 34:49managed to get something. Today we ended
- 34:50up with a page like this. I also use
- 34:52this skill myself. I already posted a
- 34:54selection of visual skills on Telegram,
- 34:56and this one is there, by the way. And
- 34:59there are other important skills and
- 35:01tools for this field too, so if you are
- 35:03interested in visuals, go to Telegram
- 35:05and look for that very useful post.
- 35:08Next, our third skill is called "It
- 35:11scans the text for things that
- 35:13immediately give away an AI."
- 35:16These are repetitive introductory
- 35:18phrases, unnecessary contrasts,
- 35:20antitheses, pompous wording, too many
- 35:22dashes, and artificially choppy
- 35:24sentences.
- 35:25Well, I think everything is clear here.
- 35:27This is just for anyone writing with AI
- 35:29.
- 35:30And the fourth skill is called "It
- 35:32contains both a video creation tool and
- 35:34a skill that explains to the agent how
- 35:36to work with it."
- 35:38The agent, for example, writes the code
- 35:40, and it then turns it into a video.
- 35:43I think there are a lot of tasks related
- 35:45to video content right now, but even if
- 35:48you don't post anything, you can just
- 35:50ask it to explain a complex topic with a
- 35:52short, captioned, animated video, right
- 35:54?
- 35:55It’s not necessarily for creating videos
- 35:56for social media. It can just be a video
- 35:58created for your colleagues, for your
- 36:00child, or for yourself. It's incredibly
- 36:02interesting. This way of working with AI
- 36:04is trending right now. That is, the most
- 36:06advanced users are using it for such
- 36:08tasks.
- 36:09And, of course, you can also make Reels
- 36:11, subtitles, intros, and animated
- 36:13graphics with it.
- 36:15Well, guys, up to this point, we have
- 36:17prepared for work.
- 36:19We created folders, saved instructions,
- 36:22connected services, and added skills.
- 36:25Now let's figure out how exactly to work
- 36:27with it. So everything before this was
- 36:30just the setup. These are the tools, the
- 36:33instructions, everything needed for the
- 36:35work. But how exactly to work is
- 36:37different. So I will show you three ways
- 36:40to work effectively, three AI modes. In
- 36:42Claude, there are essentially three ways
- 36:45of working. We have the plan mode, the
- 36:47goal mode, and the orchestration mode.
- 36:49One by one. You need the plan mode when
- 36:52the specific result you want to get is
- 36:54not yet clear.
- 36:56First, we build and agree on the steps
- 36:58together with Claude, and only then,
- 37:00after you have agreed, does it proceed
- 37:02to execute the plan. And you see what
- 37:05result you arrived at. I'll show you an
- 37:07example. Here is a table with our
- 37:08channel's analytics. There are 57
- 37:10episodes, views, subscriptions, and
- 37:12other metrics for different periods. And
- 37:15our task is to choose the next four
- 37:17episodes, based on the data. For example
- 37:20, I don't really understand what the
- 37:21episodes could be, what specific topics
- 37:23should be in this plan.
- 37:25But I do know that I need to check which
- 37:27videos are getting views, which are
- 37:29bringing in subscribers, and where
- 37:30success is repeated across several
- 37:32episodes.
- 37:33To do this, we turn on plan mode and
- 37:35write. Look at the channel analytics.
- 37:38We need to choose four topics that will
- 37:40help get views and new subscribers.
- 37:42Let's leave one episode for an
- 37:43experiment. First, propose an analysis
- 37:45plan and clear it with me.
- 37:47Then, accordingly, in this mode, Claude
- 37:50now suggests steps to complete this task
- 37:52.
- 37:53We check, answer questions, and correct
- 37:55the approach if we need to adjust it.
- 37:58When you are satisfied with everything,
- 38:00we say the plan works, go ahead. That's
- 38:02it. And now Claude will carry out
- 38:04everything specifically, following the
- 38:06steps you agreed upon. We understand why
- 38:09we need it. Accordingly, we open the
- 38:11result, four topics and a justification
- 38:13for each. That is what plan mode is for
- 38:15. Now we have goal mode. In goal mode,
- 38:18you know what you want to get in the end
- 38:20. You already know the result, of course
- 38:23, but in this case, you just don't know
- 38:25which steps and tools to use to get
- 38:27there. Accordingly, here we set the goal
- 38:29and the criteria for completion. But the
- 38:32path to this goal, to the result, Claude
- 38:34will choose for himself. For example,
- 38:36let's take the same spreadsheet. But
- 38:37here we will set a different task.
- 38:39Choose an episode and a specific clip
- 38:41that is worth turning into a short
- 38:43vertical video and making a finished
- 38:45Reel.
- 38:46To do this, we select goal mode and
- 38:48write: Look at the channel analytics,
- 38:50choose an episode for a short clip, find
- 38:53the necessary tools and skills yourself
- 38:55; I need a finished, edited Reel. And
- 38:57it's better to clarify with him right
- 39:00away what "finished" means. And if
- 39:02necessary, you can add to this chat, for
- 39:04example, a vertical video of about a
- 39:06minute with a clear message, readable
- 39:08subtitles, and good sound. Well, so that
- 39:10he understands what "finished" means to
- 39:12you. What, accordingly, happens next in
- 39:14this mode? Claude studies the analytics
- 39:16, he chooses the most promising video in
- 39:19his opinion, and then he goes into our
- 39:21archive, takes the files from that
- 39:23episode, connects a transcriber, matches
- 39:26frames with audio, then cuts out a clip
- 39:28, adds graphics, and sends you the
- 39:30finished video. That is, the agent opens
- 39:33the file himself, checks himself if it
- 39:34meets the "finished video" requirements
- 39:36, if the subtitles are readable, if the
- 39:38voice is clearly audible, if the thought
- 39:40is understandable to someone who hasn't
- 39:41seen the full episode. And if something
- 39:44needs to be corrected, the agent himself
- 39:46goes and redoes it. And it will keep
- 39:48doing this until the goal is achieved.
- 39:51So, in your goal, be sure to explicitly
- 39:53write down what is important to you and
- 39:55what the exact final outcome should be,
- 39:56you understand? Then it will keep
- 39:58redoing the task until it reaches that
- 40:00point. That is, the goal itself helps
- 40:02your agent continue working until the
- 40:05result meets the criteria you set.
- 40:09If it lacks a source or tool access for
- 40:11the next step, Claude will first try to
- 40:13solve the problem itself and only then
- 40:15ask you for help.
- 40:18And now, the third mode is the most
- 40:20complex, but also the most advanced
- 40:23workflow for working with Claude.
- 40:26This is orchestration and sub-agents.
- 40:29Let me explain.
- 40:30Orchestration is a way to divide the
- 40:32work within a single task between
- 40:34different models.
- 40:36Because the main Claude, using an
- 40:38expensive and smart model, coordinates
- 40:41the process, while separate, cheaper
- 40:43assistants receive their parts of the
- 40:45work. For example, let's return to our
- 40:48table and again pick the topic of four
- 40:50future episodes.
- 40:52But now we instruct Claude to involve
- 40:53three sub-agents and assign them
- 40:55different roles.
- 40:56The first, for instance, will search for
- 40:58topics with steady reach.
- 41:00The second will look at view depth,
- 41:02subscriptions, or shares.
- 41:04The third will check their calculations
- 41:06and conclusions. That is, it will look
- 41:08for errors, comparisons, and
- 41:09counterexamples. We run the first two in
- 41:11parallel, while the checker first
- 41:13studies the data and then their results
- 41:15.
- 41:16And we ask the main Claude to now
- 41:18compare these arguments and make the
- 41:20final editorial decision.
- 41:23So, in this way, we distribute the load
- 41:25and, consequently, the token costs. This
- 41:28is insanely important, because in the
- 41:30future, tokens are literally your money
- 41:32, and orchestration helps distribute
- 41:35them wisely. That is, orchestration
- 41:37helps allocate tokens more rationally.
- 41:39Complex, high-volume tasks that don't
- 41:41require much brainpower use fewer of
- 41:43your limits.
- 41:44Complex, high-volume tasks that don't
- 41:46require much brainpower use fewer of
- 41:48your limits because these tasks are
- 41:50performed by cheaper models. And the
- 41:53smartest, most expensive model analyzes
- 41:55the results and provides you with
- 41:57ready-made solutions. If you want to
- 41:59know my recommendation, I personally use
- 42:02Opus as my default main model. And for
- 42:04super architectural tasks, I use Haiku,
- 42:06but that is extremely rare. But Sonnet,
- 42:08for example, I use as a sub-agent, an
- 42:10assistant that performs all the routine
- 42:11work.
- 42:13Phew, I think I've told you absolutely
- 42:15everything in this video that you need
- 42:17to get started.
- 42:19And that is the guide we’ve put together
- 42:21. I think everything I’ve shared will be
- 42:23enough for you. If you find you’re
- 42:25missing something, you can watch my
- 42:26additional videos.
- 42:27And don't worry if there's something you
- 42:29don't quite understand while watching
- 42:30the video. Everything comes with
- 42:32practice.
- 42:33Start with this guide, start working,
- 42:36and you'll figure everything out as you
- 42:38go.
- 42:39I've left all the instructions, settings
- 42:40, prompts, and links to skills and tools
- 42:42in my Telegram channel; go there, grab
- 42:44them, try them out, and experiment.
- 42:46Please leave a like, subscribe, comment
- 42:47, and share with your friends, because
- 42:49this was an incredibly useful video.
- 42:50That's it, see you in the next video.
- 42:52Bye.
About this transcript
This page contains the full transcript of Claude и Claude Code: полный гайд от установки до сложных задач by Web3nity, generated from the public captions YouTube serves with the video. The transcript has 7,818 words across 1,258 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.