2027'den Önce Claude Code'u Öğren: Sıfırdan Eksiksiz Rehber — Transcript
Full transcript
- 0:00Everyone's talking about Cloud Code.
- 0:01Because with it, people can do alone
- 0:03what used to be shared by an entire team
- 0:05. This means that if someone doing your
- 0:08job has learned it, they'll finish the
- 0:10same job much faster than you. If you're
- 0:12looking for work, that person is your
- 0:13potential candidate. If you're looking
- 0:15for clients, that person will be your
- 0:16competitor. So Cloud Code isn't
- 0:18something you'd rather learn; it's
- 0:20something you*must*learn. More and more
- 0:22people are learning it every month. The
- 0:24longer you wait, the more you'll be
- 0:25caught up from behind. You need to get
- 0:27this done before 2027. In this video,
- 0:29you'll learn Cloud Code from scratch,
- 0:31completely and thoroughly. We'll start
- 0:34with the setup.
- 0:35From Cloud MD to Skills, from MCPs to
- 0:37Sub Agents, you'll see everything on a
- 0:40real screen, explaining what it does. By
- 0:42the end of the video, you'll be someone
- 0:44who not only asks Cloud questions but
- 0:46also gets things done by it. Even if
- 0:48you've never used it before, that's okay
- 0:49. We're starting from the very beginning
- 0:51here. Let's clarify something first. You
- 0:53already know Cloud. You chat with it,
- 0:56you ask questions, and it answers you.
- 0:58Cloud Code is just like Cloud itself.
- 1:00The only difference is this: it runs
- 1:02inside a folder on your computer. Let me
- 1:04show you what that means with an example
- 1:05. When you want Cloud to do something in
- 1:08a chat, you upload a file. It writes you
- 1:10the answer. You then copy and paste that
- 1:12answer somewhere. Cloud Code doesn't
- 1:14have that copy-paste. For example, you
- 1:15say, "Gather all the notes in this
- 1:17folder into a single summary file."
- 1:19Cloud opens and reads those files,
- 1:21writes the summary, and puts the file in
- 1:23the folder itself. You can then look in
- 1:25the folder and find the file there.
- 1:26Don't think it's only for programmers
- 1:28just because it has "code" in its name.
- 1:32I use Cloud Code for most of the work on
- 1:34this channel's videos, from audio
- 1:36recording and cleaning to editing.
- 1:40I even use it for videos on my other
- 1:42social media channels and for my daily
- 1:44tasks.
- 1:46Everything done with folders and files
- 1:48is its job. You can use Cloud Code from
- 1:50different places. You know that black
- 1:52screen called Terminal? You can use it
- 1:54from there too. But there's no need to
- 1:56confuse yourself by using it when you're
- 1:57starting from scratch.
- 1:59Its own desktop is designed quite nicely
- 2:01and understandably.
- 2:03In this video, we'll be using the
- 2:05desktop application. Let's move on to
- 2:07installing the application. Installation
- 2:10takes only a few minutes, but before
- 2:12that, I need to mention something. Cloud
- 2:14Code isn't available in the free plan.
- 2:16You need to upgrade to a paid plan.
- 2:18There are three: Pro, Max, and Ultra.
- 2:20Pro is $ 20 per month. Max gives you
- 2:23five times the usage of Pro, and its
- 2:25price is also five times, so it's $ 100
- 2:27. Ultra is $ 200. This plan is for those
- 2:29who use the most and work very
- 2:31intensively. Prices may change over time
- 2:33. You can find the current prices on
- 2:35their official website. So, which one
- 2:37should you get? My recommendation is to
- 2:39start with Pro. Pro is perfectly
- 2:41sufficient for daily tasks. Antropin
- 2:43itself states on its website that Pro is
- 2:45for daily tasks and Max is for those who
- 2:47work all day. Test it with Pro first. Do
- 2:49some real work. If Pro's limit is enough
- 2:51for you, you can stay with Pro. You
- 2:53don't need to pay extra or upgrade to a
- 2:55higher plan.
- 2:57But if it's not enough, meaning you
- 2:58constantly hit the limit and have to
- 3:00wait while doing your work, then you can
- 3:02upgrade to Max.
- 3:04We'll see what the limit is and how it's
- 3:06reached in a later section.
- 3:08Even if you choose the wrong plan, you
- 3:10don't lose anything. When you upgrade to
- 3:11a higher plan, the remaining time from
- 3:13the old plan is deducted from the new
- 3:15one. You can cancel anytime. So, get a
- 3:17monthly plan for the first month, you
- 3:18can think about annual payments later.
- 3:20After you get your plan, you download
- 3:22the Cloud desktop application from the
- 3:24address on the screen, install it, and
- 3:26log in to your Cloud account. You don't
- 3:27need to install anything else. Cloud
- 3:29Code is already included in the
- 3:31application. When you open the
- 3:32application, you see two small icons in
- 3:34the upper left. The one on the left is
- 3:36the chat section, which is the Cloud you
- 3:38know, and the icon next to it with small
- 3:40separators is the code section. This is
- 3:41the Cloud that works with folders. So we
- 3:43click on it. On the left side, there are
- 3:45buttons like "new" and "search". Below
- 3:47that, there is a list. We call each
- 3:49conversation you have in Cloud Code a
- 3:51session. That list is a list of old
- 3:53sessions. Since you're just starting out
- 3:54, it will be empty. You don't need to
- 3:56look at it for now. I'll get there when
- 3:58you need to. As you can see, the setup
- 4:00is that simple.
- 4:02Now let's slowly dive into it.
- 4:04Let's go to the box at the bottom of the
- 4:06screen. That's where the real work is
- 4:07done.
- 4:08Now we've opened the code tab. We saw
- 4:10the box at the bottom of the screen.
- 4:12Inside, it says "Define a task" or "Ask
- 4:15a question." We call this the prompt box
- 4:17. It's like the place where you type
- 4:19messages in GPT, Gemini, or regular
- 4:22Cloud chat.
- 4:23You type what you want Cloud to do here
- 4:25and press Enter to send it.
- 4:27But before you type the first request,
- 4:29you need to specify two things: Where
- 4:31will Cloud run? Which folder will it run
- 4:34in? Both are in the line right below the
- 4:36box. There are other things in that line
- 4:37too. Let me explain what they are now so
- 4:40that nothing you see on the screen is
- 4:41unclear. The place on the left that says
- 4:44"This computer" is where you choose
- 4:46where Cloud will run. Since the files
- 4:47are on your computer, we'll choose that
- 4:49. When you open the menu, there's also a
- 4:52cloud below it.
- 4:53Cloud runs on Antropi's own computer. So
- 4:56it can't see the folders on your
- 4:57computer. It's not useful for us right
- 4:59now. The folder next to it is Cloud's
- 5:02workspace. It reads files there. It puts
- 5:04new files there. You select a folder by
- 5:07clicking "Open folder." You can also see
- 5:09the ones you've used before in the list
- 5:10. I want you to pay attention to
- 5:12something here. Cloud can modify the
- 5:14files inside the folder. So don't do the
- 5:16first tests on the original files.
- 5:18Create a new folder on your desktop
- 5:20called "Cloud Test". Copy the files you
- 5:22want to test into it. Even if Cloud
- 5:24misunderstands something, or does
- 5:26something wrong, your original files
- 5:28won't be affected. Now, leave the "
- 5:30Working Tree" checkbox next to the
- 5:32folder unchecked. This is for projects
- 5:34using Git. If you check it, Cloud will
- 5:36create a separate copy of the folder so
- 5:38it doesn't interfere with the original
- 5:39files. But our folder doesn't use Git.
- 5:41We don't need that for now. On the right
- 5:43side of the line, there are three things
- 5:45. Let's start with the model. Which
- 5:47version of Cloud will be running,
- 5:49meaning
- 5:50how much effort that version, or model,
- 5:52will consider for that task, and finally
- 5:54, the permissions mode. It currently
- 5:56says automatic. You can choose there
- 5:58whether Cloud should ask you when
- 6:00modifying a file or running something.
- 6:02Each of the three has its own section.
- 6:04We'll explain them in detail in this
- 6:06video.
- 6:07Permissions mode will come right after
- 6:09the first one. Because you can only
- 6:11better understand what it does after
- 6:12Cloud has done something. There's also a
- 6:15plus sign at the end of the box.
- 6:16Clicking it opens a menu.
- 6:18When you click the "Add File or Photo"
- 6:21button, you can add a PDF, image, or any
- 6:23other file to the Cloud from here.
- 6:26Dragging and dropping the file into the
- 6:27box does the same thing. Slash commands
- 6:30–even if you just type "slash" into the
- 6:32box, the same list opens. It's a way to
- 6:35give the Cloud ready-made commands.
- 6:37We'll use these commands when managing
- 6:39the context shortly. Skills you've
- 6:41created yourself can also be called from
- 6:43this list.
- 6:44I'll show them all individually when the
- 6:46time comes. There's also the MCP (
- 6:48Multiple Compartments) section. It
- 6:50connects the Cloud to other applications
- 6:52. "Add Plugin" installs ready-made
- 6:55packages. I'll cover both in detail in
- 6:57the video. For now, you can think of it
- 6:59as just an introduction to see what they
- 7:01are. The "Add Folder" button adds a
- 7:04second folder to the workspace. We'll be
- 7:06working with a single folder right now,
- 7:08so we don't need it. The "Import GitHub
- 7:10Issue" button is for developers to bring
- 7:12up their work files from GitHub. That's
- 7:14what interests us. Okay, now we can move
- 7:16on to writing the first request.
- 7:18We need a task to see what Cloud does. I
- 7:20deliberately left some things mixed up
- 7:22in the folder. There are a few invoices
- 7:24, a few YouTube covers, and a few notes
- 7:27files.
- 7:28When I look at the invoices, they're
- 7:31named in a jumbled way, like "invoice
- 7:34a93f-221," and so on. It's impossible to
- 7:37tell which invoice belongs to what just
- 7:39by the name. Sometimes, files we
- 7:40download in real life look like this too
- 7:43. Now I'm telling Cloud to examine each
- 7:45file in this folder individually,
- 7:46understand what each one is, give it a
- 7:48meaningful name, and organize it into
- 7:50folders according to its subject. I sent
- 7:52this prompt. Cloud doesn't just guess by
- 7:54name. It opens the files. It reads the
- 7:56text inside the PDF, looks at the images
- 7:58and understands what they are. And you
- 8:00don't have to wait while Cloud is
- 8:01working. The box stays open; it can't be
- 8:03closed. For example, if you have an idea
- 8:05, you can write "folder names should be
- 8:07without Turkish characters" and send it
- 8:09. Cloud sees it in the next step and
- 8:11incorporates it into its work. If you
- 8:12want to stop it completely, you press
- 8:14the stop button and it stops immediately
- 8:16. Now the job is done. When you looked
- 8:18at the folder, you saw an electricity
- 8:20bill inside the invoice file. Cloud
- 8:22changed the name accordingly. An
- 8:24internet bill starting with A 93 F
- 8:26appeared. The cover files went to a
- 8:28separate folder. The invoices went to
- 8:30another separate folder. Did you notice
- 8:31something? I'm very curious. Cloud
- 8:33didn't ask you anything when moving or
- 8:35renaming the files. Because the
- 8:37permissions mode was automatic.
- 8:38Everyone's files are different, and not
- 8:40everyone wants to work with the same
- 8:41level of trust. So, should Cloud have
- 8:43asked or not? That's determined by the
- 8:45permissions mode.
- 8:47Now, what we call permissions mode is
- 8:49how much Cloud can do without asking you
- 8:51. There are five here. I'll open the
- 8:53menu and explain them one by one. We'll
- 8:55go from the mode that asks the most to
- 8:57the one that asks the least. I'll leave
- 8:59the plan for last because it works a
- 9:01little differently.
- 9:02Let's start with manual. Manual Cloud is
- 9:04a mode where it asks you a question
- 9:06every time before modifying a file or
- 9:08running a command. It doesn't ask to
- 9:10read files, it asks to modify them. This
- 9:13is the most secure mode. But it's also
- 9:15the slowest. What it asks appears on the
- 9:17screen like this. Here you see what will
- 9:19change. You can either reject or allow.
- 9:21I want you to pay attention to something
- 9:23. The "Always Allow" button isn't just
- 9:25for this file; the text in the box
- 9:27itself says so. If you press it, that
- 9:29session switches to "Accept Edits" mode
- 9:31and it no longer asks for file edits. So
- 9:34, if you want to stay in manual mode,
- 9:36you have to press "Allow" every time.
- 9:38There's also an "Accept Edits" mode.
- 9:40Cloud handles file edits without asking
- 9:42. Simple file operations like opening
- 9:44folders, moving files, and renaming them
- 9:46are handled the same way. But if it's
- 9:48going to run another command, it will
- 9:49ask again. If you trust your files and
- 9:51want to proceed quickly, you can choose
- 9:53this mode.
- 9:54Now we come to the mode we use most
- 9:57often, the one you'll use constantly:
- 9:59automatic mode.
- 10:00Cloud doesn't ask every time. That's
- 10:02what you saw in the first instance. But
- 10:04for tasks like running commands or
- 10:06sending something to the internet, a
- 10:07separate control system kicks in in the
- 10:09background. This control checks two
- 10:11things: Does what's going to be done
- 10:13match your request? And it's done with
- 10:15an instruction hidden within a page that
- 10:16Cloud reads. It checks these things. If
- 10:19it sees it as risky, it blocks it.
- 10:21Reading and editing files inside a
- 10:22folder doesn't even come under this
- 10:24control. Commands like moving files do
- 10:26come under control, but it bypasses them
- 10:28because they match what you want. That's
- 10:30why it didn't ask anything on the first
- 10:32task. Since the default is auto mode,
- 10:33the application opens automatically.
- 10:35There's also a "bypass permissions" mode
- 10:38here. Think of it like giving Cloud a AK
- 10:40-47. It can do everything without asking
- 10:42. This mode is grayed out in the menu.
- 10:44It says "Enable in Cloud code settings.
- 10:46" So you have to enable it yourself from
- 10:48the settings. Antropik also says to only
- 10:50use this in an isolated environment,
- 10:52away from your computer. So they don't
- 10:54recommend enabling it if you're just
- 10:55starting out. I don't recommend it
- 10:57either.
- 10:58Auto mode does everything nicely. We
- 11:00also use a security mode because it runs
- 11:02in the background. Now, there's one
- 11:04thing that doesn't go without asking,
- 11:05regardless of which mode you're in: a
- 11:07command to delete the computer's main
- 11:09folder or the entire disk. It requires
- 11:10this in every mode, including the "
- 11:12bypass permissions" mode. The
- 11:14documentation says so, but I think it's
- 11:16absurd to completely trust it just
- 11:18because it says so.
- 11:20So if you're doing a very important job
- 11:22or you don't completely trust it, choose
- 11:24modes that suit you.
- 11:27Don't just trust the documentation or
- 11:29other people's methods and do things
- 11:31that way.
- 11:33So, which option do we choose here? If
- 11:35you're new to Cloud or working with a
- 11:37folder containing important files, you
- 11:39can start with manual mode. You can see
- 11:41what you're doing for a few tasks. Then
- 11:43, once you understand it better and feel
- 11:45confident, you can switch to automatic
- 11:47mode. Now, let's talk about the planning
- 11:49mode. Here, the fifth planning mode is
- 11:51slightly different from the others.
- 11:53Cloud reads the files. If necessary, it
- 11:55runs inspection commands and writes a
- 11:57plan for you. It doesn't change any
- 11:59files until you approve it. Once the
- 12:01plan is ready, Cloud asks how you want
- 12:02to proceed. There are three options. You
- 12:04can approve the plan and start in
- 12:06automatic mode. You can approve the plan
- 12:08and choose to approve each change
- 12:09individually, or you can choose not to
- 12:11approve and continue with the planning.
- 12:13You write what you want to change.
- 12:15Claude updates the plan accordingly.
- 12:17Once you approve, the plan closes, and
- 12:19Claude starts writing. For example,
- 12:21let's say, like in the first task, you
- 12:24need to rename and organize hundreds of
- 12:26files into folders, and reversing this
- 12:28task is difficult and costly. In the
- 12:30plan, Claude writes, "I will first open
- 12:32these folders, and I will name the files
- 12:34like this." You correct it by looking at
- 12:36it and separating the invoices month by
- 12:38month. Then he implements it. Something
- 12:40else happens in the plan. While Claude
- 12:42is reviewing the files, he has his
- 12:44ready-made assistant, a sub-agent, do
- 12:46this research. This way, everything
- 12:48that's read doesn't enter your
- 12:49conversation. We'll see what sub-agents
- 12:51are in detail in the sub-agents section
- 12:53. So, do I use the plan? No. I haven't
- 12:56used the plan for a long time. Because
- 12:58the models have developed so much that
- 13:00when you clearly state what you want, it
- 13:02can find the way itself. I made a
- 13:03separate video about this on the channel
- 13:05. It should have a name like "Claude's
- 13:07code was old." I explain that even the
- 13:10person who created the plan mode doesn't
- 13:12use it. I wanted to mention it here
- 13:13because it's in the menu and recommended
- 13:15for complex tasks in the Antropin
- 13:17documentation, so you should know about
- 13:19it. If you can't describe what you want
- 13:21, or if it's a difficult task to reverse
- 13:23and you want to see Claude's approach
- 13:25before starting, this might be useful
- 13:27for you. Otherwise, I don't think it's
- 13:29really necessary. Now we've done the
- 13:31first task. We selected the permission
- 13:33modes, but what if Claude's work isn't
- 13:35as good as you want?
- 13:37Then we'll look at the right end of the
- 13:39line at the bottom of the box.
- 13:42Now, looking at the right end of the
- 13:44line below the box, we see two settings
- 13:46: Model and Effort. Model refers to
- 13:48which version of Claude will perform the
- 13:50task, i.e., which model will do it.
- 13:52Effort, on the other hand, indicates how
- 13:54much thought that version will require
- 13:56at each step. We briefly saw this at the
- 13:58beginning. Now let's go into detail.
- 14:00First, I click on Model. Claude isn't a
- 14:02single model. Antropin has several
- 14:04models simultaneously. They're all
- 14:05listed one after the other in the menu.
- 14:07Next to Opus, it says Default. This is
- 14:09the default model in both Pro and Max.
- 14:11So, if you haven't changed anything,
- 14:13you're currently working with Opus as
- 14:14the default. But will we use Opus for
- 14:16every task? Anthropic's own
- 14:18documentation categorizes the models as
- 14:20follows: Sonnet is for everyday tasks,
- 14:22handles most coding tasks well, and is
- 14:24cheaper than Opus. Opus is for complex,
- 14:26multi-step, and thought-provoking tasks
- 14:29. Haiku is fast and light, sufficient
- 14:31for simple tasks. Fable is for the most
- 14:33difficult and longest-running tasks.
- 14:35Here, "cheap" and "expensive" refer to
- 14:37how much of your subscription cost. The
- 14:40large model is spending more than its
- 14:41limit for the same task. We'll see what
- 14:43that limit is in detail in the next
- 14:45section. Here, it's enough for you to
- 14:47know that. For example, I mostly use
- 14:49Opus for my video editing. Sometimes I
- 14:51use Sonnet as well. And I already
- 14:53explained Fable in detail in a separate
- 14:55video. You can find it on the channel.
- 14:57It should be the previous video. There's
- 14:59a line at the bottom of the menu called
- 15:01"Fast Mode." It makes the Cloud respond
- 15:03faster, but it requires extra credits.
- 15:05So it's a separate, paid service outside
- 15:07of your subscription. You don't need
- 15:09that when you start.
- 15:10Now let's talk about effort. Most people
- 15:12don't bring this up, but I think it's as
- 15:14important as choosing a model. The same
- 15:16model thinks quickly and works
- 15:17efficiently in low effort mode, while in
- 15:19high effort mode it thinks more at each
- 15:21step. The on-screen explanation also
- 15:23says this. Higher effort provides more
- 15:26comprehensive responses, but it takes
- 15:28longer and consumes its limits faster.
- 15:30There are five levels here. Let me tell
- 15:32you which one to use when, in order,
- 15:34according to Anthropic's own
- 15:35documentation. Low effort is for fast
- 15:37tasks where you control every result
- 15:39yourself. It says to use it for small
- 15:41tasks like drafting an idea or renaming
- 15:44a file. Medium effort is the default,
- 15:46and it says "recommended" next to it.
- 15:50Most everyday tasks where you know what
- 15:52you need can be done here. For example,
- 15:54in the first task, naming and organizing
- 15:56invoices into folders, the effort was on
- 15:58medium, and it was perfectly sufficient
- 15:59. Now, it says to use high effort for
- 16:01tasks where control is important and the
- 16:03probability of errors is high. For
- 16:05example, finding and fixing an error in
- 16:07a working system. X high, or extra high
- 16:10effort, requires more deep thinking and
- 16:12therefore more usage. Max, or maximum
- 16:15effort, is for tasks where you want the
- 16:16Cloud to solve a difficult problem from
- 16:18start to finish without you doing
- 16:20anything. There's also a warning next to
- 16:22it: 3.5 times or more usage. So it
- 16:24seriously uses up your limit. The
- 16:26document also adds this for maximum
- 16:28effort: the gain here can decrease. It
- 16:30says the Cloud can think more than
- 16:32necessary.
- 16:33By "the gain here can decrease," it
- 16:35means that you might not see much
- 16:37difference between a task done on High
- 16:39Effort and one done on Max.
- 16:41In other words, both can produce similar
- 16:43results.
- 16:45But here, because it's at the maximum,
- 16:47you're spending more tokens and thinking
- 16:49more.
- 16:51That's why your earnings can vary.
- 16:53Because you did a similar job in High
- 16:55Effort, and you did it at the maximum as
- 16:57well. But in one, you used a higher
- 17:00limit. So don't open every job to the
- 17:02maximum. Try a lower level first.
- 17:04Anthropic has another observation from
- 17:06its own tests that might help you when
- 17:08choosing. In High Effort, Cloud tries
- 17:10more possibilities. It checks your job
- 17:12more before answering and makes more
- 17:14decisions itself.
- 17:16In Low Effort, it gives you an earlier
- 17:18starting point. So, if you want to guide
- 17:20it by looking at each step, choose Low;
- 17:22if you want to say "finish this," choose
- 17:25High Effort. Sometimes you don't want to
- 17:27change the Effort. You just want it to
- 17:29think more deeply on a single question.
- 17:31Then write Ultra Think anywhere in your
- 17:32message. Look, for example, when you
- 17:34write it, the word lights up in color.
- 17:36This means Cloud Code recognized the
- 17:38word. It's thinking more deeply on that
- 17:39message. Your effort setting remains the
- 17:41same. You also need to pay attention to
- 17:43something. Other expressions like "
- 17:44Think" or "think more" don't work like
- 17:46this. They're going through as ordinary
- 17:48text. The word will be Ultra Think. Now
- 17:50, at the very bottom of the Effort menu
- 17:52, there's another section called Ultra
- 17:54Code. This isn't actually level 6 of
- 17:56Effort; it's a separate setting. When
- 17:58enabled, Cloud plans a workflow instead
- 18:00of working alone on a comprehensive
- 18:02project. It divides the work into
- 18:04multiple sub-agents and runs them
- 18:05together. We'll see what a sub-agent is
- 18:07in detail in its own section. For now,
- 18:09you need to know this: because multiple
- 18:11agents are working simultaneously, your
- 18:13limit might be reached faster. If you're
- 18:16just starting out or doing simple, small
- 18:18tasks, enabling this here would be
- 18:20unnecessary and pointless. So, keeping
- 18:22it off is better. Now, if you noticed,
- 18:25we saw the same warning for both large
- 18:27models and high-effort projects. We said
- 18:29it will consume your limit faster. So,
- 18:31how is this limit measured? The answer
- 18:33is how Cloud actually reads the text.
- 18:36Cloud doesn't read the text word for
- 18:38word like we do. It first breaks down
- 18:40every piece of text it receives into
- 18:41small pieces. We call these pieces
- 18:43tokens. A word is sometimes a single
- 18:45piece, sometimes two or three pieces.
- 18:47Cloud reads it using these pieces. When
- 18:49writing a response, it also writes it
- 18:51piece by piece. So, how much is a piece
- 18:53? According to Antropi's own dictionary
- 18:55, a token is approximately 3.5 English
- 18:57characters, roughly 3-4 letters. This is
- 19:00for the average English. You can count
- 19:01how many tokens a Turkish sentence would
- 19:03take, as shown on the screen. So why do
- 19:05we need to know this? Because when
- 19:07working with the Cloud, two important
- 19:09things are measured in tokens. First,
- 19:10how much information the Cloud can keep
- 19:12track of in a session, and second, how
- 19:14much you've spent from your subscription
- 19:16. We'll get to both of these in turn,
- 19:18but first we need to understand this:
- 19:20every message sent to the Cloud counts
- 19:22as a token. So we need to know what's
- 19:24going to the Cloud. Let's go back to the
- 19:26first task. I wrote two lines in the box
- 19:28. I said, "Look inside the invoices,
- 19:29name them, organize them into folders,"
- 19:31etc. The first thing that went to the
- 19:32Cloud was these two lines, but the Cloud
- 19:34couldn't learn what was written inside
- 19:35the invoice from my message. So it
- 19:37opened the invoices one by one. When the
- 19:39Cloud opens a file, the text inside the
- 19:41file goes to it. So each invoice entered
- 19:43the Cloud as a token. Then it ran a
- 19:46command to open the folders. The result
- 19:48of the command also came back to it as
- 19:50text. Finally, it wrote back to me
- 19:52saying, "You did these things." Actually
- 19:54, that answer is also a token. Now, when
- 19:56you look at it, the part I wrote is the
- 19:58smallest. The majority of the text sent
- 19:59to the Cloud during that process is the
- 20:01invoices themselves. The thing that
- 20:03consumes the most tokens is usually not
- 20:04what you write, but what the Cloud reads
- 20:06. So what happens to this text, where
- 20:08does it go? Now, everything that goes to
- 20:10the Cloud in a session accumulates. Your
- 20:12messages, the files it reads, command
- 20:14outputs, its own responses—it all piles
- 20:17up. When the Cloud responds to you, it
- 20:19looks at all of this accumulation
- 20:20together. For example, if I ask you
- 20:22which folder you put the electricity
- 20:24bill in at the end of the first task, it
- 20:25can answer. Because the moment it put
- 20:27that bill there, it's still in this
- 20:29accumulation. We call this accumulation
- 20:31context. That is, context. There is an
- 20:33upper limit to the context. That limit
- 20:35is called the context window. That is,
- 20:38the maximum amount of tokens that the
- 20:40Cloud can keep in mind at the same time
- 20:41in a session. In the model we are
- 20:43currently using, this limit is 1 million
- 20:45tokens. The context window is not
- 20:47completely empty before you write
- 20:48anything. Cloud Code puts some things
- 20:50into it at the beginning of each session
- 20:52. Cloud Code has its own instructions on
- 20:54how to operate, a list of tools it can
- 20:56use (like reading files, writing files,
- 20:59running commands, etc.), a persistent
- 21:01rule file in the folder, descriptions of
- 21:03skills, and the names of the tools
- 21:05you've connected. We'll see some of
- 21:07these individually in later sections.
- 21:10For now, it's enough to know that they
- 21:12take up space here. Let's see this in a
- 21:14real sense. I've opened a new session. I
- 21:16haven't typed anything yet. I type slash
- 21:18context in the box. As you can see, a
- 21:20card appeared in the chat. At the top it
- 21:22says total. That is, how many tokens the
- 21:24window has, what percentage of 1 million
- 21:26it is. Below, we can see line by line
- 21:28how much space each thing takes up. The
- 21:30names are in English. Let's look at them
- 21:32one by one. What's written here as
- 21:34system prompt are Cloud Code's own
- 21:35instructions. System tools are Cloud's
- 21:38own tools. MCP tools are the tools
- 21:40you've connected. Skills are
- 21:42descriptions of the skills the OS gave
- 21:44to Cloud. Messages are what we discussed
- 21:46in this session.
- 21:48At the bottom there's free space. There
- 21:50are a few more lines. Don't worry about
- 21:52those for now. Now look at the messages
- 21:54line. Almost zero. Because we haven't
- 21:56talked about anything yet. But part of
- 21:57the window is already full. That part is
- 21:59held by the lines above, the things
- 22:01Cloud Code put in at the beginning of
- 22:02the session. Now I'm running the same
- 22:04command in the session where the first
- 22:06task finished. As you can see, the lines
- 22:08above are almost the same. What's
- 22:09changed are the messages. Everything we
- 22:11talked about in the first task, every
- 22:13invoice Cloud read, is here. This is
- 22:15also where the tokens we just counted
- 22:17went. You don't need to open this dump
- 22:19with a command every time. When you send
- 22:21a message, the session details panel
- 22:23opens on the right side of the screen.
- 22:24The context card there shows the same
- 22:26lines. The only difference is that the
- 22:28card isn't live. It refreshes when
- 22:30Cloud's response is finished. Now the
- 22:32window has 1 million. As we can see,
- 22:34most of it is empty. Why should it be a
- 22:36problem if it's not full? How are we
- 22:37going to fill all this, and what happens
- 22:39when it is full? There are two reasons
- 22:40here. First, as the context grows, the
- 22:42quality of Cloud's use of it decreases.
- 22:45This is called context rot.
- 22:47Antropin explains it in his own article
- 22:49like this: As the number of tokens in
- 22:51the context window increases, the
- 22:52model's ability to correctly remember
- 22:54the information in that context
- 22:55decreases. The article says this is seen
- 22:57in all models.
- 22:59Some are slower. So the information is
- 23:01in the window, but Cloud can't use it as
- 23:03well as before. This starts before the
- 23:05window is full. Even if the window has 1
- 23:07million entries, it doesn't change that
- 23:09. Secondly, the window eventually fills
- 23:11in a long session. When it's close to
- 23:13full, Cloud doesn't close the code
- 23:15session. It automatically summarizes the
- 23:17conversation. It replaces the old
- 23:19conversation with the summary and
- 23:20continues that way. The Antropin
- 23:22document warns about this. The summary
- 23:24keeps track of your requests and
- 23:26important parts, but the detailed
- 23:27instructions at the beginning of the
- 23:29conversation can sometimes get lost. The
- 23:31result of both is the same. In a long
- 23:33session, something I say at the
- 23:35beginning might not be applied later.
- 23:37Let me show you with an example. In my
- 23:38first project, I interrupted Cloud and
- 23:40said that folder names should also be
- 23:42without Turkish characters, remember?
- 23:44Let's say I'm working for hours in the
- 23:46same session. I'm also editing other
- 23:48files. The conversation gets long and is
- 23:50summarized at some point. Now look. For
- 23:52example, it created a folder called "
- 23:54Contracts" with the letter "Ş". The rule
- 23:56I mentioned at the beginning is gone
- 23:57along with the summary. That's why we
- 23:59don't leave the context to Cloud code
- 24:01itself. We manage it ourselves. There
- 24:03are two commands and one setting for
- 24:04this. The first is compact. You initiate
- 24:07the summary yourself, without waiting
- 24:08for the automatic summary. This is
- 24:10called compression. You start the
- 24:11compression yourself. You type "slash
- 24:14compact" in the box and start it that
- 24:16way.
- 24:17The nice thing about this is that you
- 24:19can specify what you want it to keep.
- 24:21The automatic summary guesses what's
- 24:23important, but here you tell it.
- 24:26You tell it to keep file name rules and
- 24:28the last changes we made. The summaries
- 24:30preserve these.
- 24:32So when should you use this, when should
- 24:34you run it?
- 24:35The document says to run it between two
- 24:37jobs. Run this command after you've
- 24:39finished one job and before starting a
- 24:42new, longer job. If you do it in the
- 24:43middle of a job, the details of that job
- 24:45are compressed along with the summary.
- 24:47There's also an automatic compression
- 24:49threshold setting in the context card.
- 24:52You can set when the automatic summary (
- 24:54compact) will activate here.
- 24:57The explanation on the screen says: A
- 24:59lower value consumes less of the usage
- 25:02limit. Cloud summarizes older messages
- 25:04earlier. So if you bring the
- 25:06summarization forward, your limit is
- 25:08used less, but Cloud loses the details
- 25:10of older messages sooner. You can leave
- 25:12it on automatic at the start. Now,
- 25:14regarding the second command, it's `
- 25:17clear`. This actually works a little
- 25:19differently than the other one. There's
- 25:20no summary here. It completely clears
- 25:22the context. So you start from scratch.
- 25:25When do we use this? Let's say you've
- 25:27finished a task and you're moving on to
- 25:28something unrelated. For example, you
- 25:30worked on the invoice folder this
- 25:31morning, right? This afternoon, let's
- 25:33say you're going to create a Reels
- 25:35script. The conversation from this
- 25:37morning and the Reels script make no
- 25:39sense.
- 25:40They have no connection, but as long as
- 25:43you stay in that context, it takes up
- 25:45space and contributes to the context rod
- 25:47we saw earlier–the decay.
- 25:51So what do we do?
- 25:53If we're going to move on to a new task
- 25:55, a task without a connection, we run
- 25:58the ` clear` command and start again in
- 26:00a clean session. When you type ` clear`
- 26:02, a line appears in the chat saying the
- 26:04context has been cleared. Now, there's
- 26:06another way to do this on the desktop.
- 26:08You know that "new" button in the upper
- 26:10left? This opens a separate session, and
- 26:12the old session remains in the list on
- 26:13the left with its name. If you want to
- 26:15return to the invoice task tomorrow, you
- 26:17can click on it from the list. You can
- 26:19use this if you might return to the old
- 26:20conversation. Now let me explain how I
- 26:22do it in my own work. When the context
- 26:25reaches around 40%, I stop. I tell the
- 26:27Cloud to write everything we've done so
- 26:29far, what decisions we've made, and
- 26:30what's next to a file. Then I open a new
- 26:32session, read that file, and continue
- 26:34from where I left off. This way, nothing
- 26:37gets lost in the summary. The Cloud also
- 26:39doesn't start the new task with a
- 26:41cluttered context. So far, we've seen
- 26:43the first side of the token. That is,
- 26:45the part that the Cloud keeps track of.
- 26:46Now let's move on to the second side.
- 26:48How much did it spend from its
- 26:49subscription? There's something most
- 26:51people don't know here. The Cloud
- 26:52doesn't actually remember anything
- 26:54between two messages. That accumulated
- 26:56data I just described doesn't stay in
- 26:58the Cloud itself. It stays in the Cloud
- 26:59code. Every time you send a message, the
- 27:01Cloud code sends the entire context back
- 27:03to the Cloud from scratch. Instructions
- 27:05, tool list, all messages up to that
- 27:08point, files read, etc. And finally,
- 27:10your newly sent message is added. If
- 27:12this were processed from scratch every
- 27:14time, it would be very slow, and the
- 27:16limit would be reached very quickly.
- 27:18That's what caching is for. If the
- 27:19processed portion from the previous
- 27:21message remains unchanged, Cloud doesn't
- 27:23reprocess it. It reads it from the cache
- 27:25. It only reprocesses the newly added
- 27:27portion from the beginning. The portion
- 27:29read from the cache is both faster and
- 27:31consumes less than its limit.
- 27:33But consuming less doesn't mean it
- 27:35doesn't consume at all. As long as the
- 27:37morning's invoice conversation remains
- 27:39in the context, it's resent with every
- 27:41report message in the afternoon and
- 27:43counted each time.
- 27:44This is why Clear is important for the
- 27:46limit as well. Now, what determines that
- 27:48the cache remains unchanged? It
- 27:50determines this by comparing it to the
- 27:52beginning. If the beginning of the sent
- 27:54context is exactly the same as the
- 27:55previous one, it's read from the cache.
- 27:57If something changes at the beginning,
- 27:59everything after that point is
- 28:00reprocessed from the beginning. The
- 28:02document says this is a one-time thing.
- 28:04That message arrives slowly and consumes
- 28:06more. Then the new version enters the
- 28:08cache, it says. Then you continue, it
- 28:10says. But if the conversation is long,
- 28:12that single message can be quite costly
- 28:14. So you need to know what's causing the
- 28:16cache to malfunction. Now, the first
- 28:18thing is changing the model. Each model
- 28:20has its own cache. If you switch from
- 28:22Opus to Sonnet in the middle of a
- 28:24conversation, even if the content is the
- 28:26same, the entire conversation is
- 28:27processed from scratch without caching.
- 28:29So, remember that there's a cost to
- 28:31changing the mode in a long conversation
- 28:33.
- 28:34If you want to continue with another
- 28:36mode, starting in a new session is
- 28:37always better and cheaper. But effort is
- 28:40different. In most modes, changing
- 28:42effort also corrupts the cache, but the
- 28:45documentation says that in Opus 5.5,
- 28:47Sonnet 5.5, Haiku 5.5, and Fable,
- 28:50changing effort doesn't corrupt the
- 28:52cache. So, you can increase effort in
- 28:54the middle of a session while working in
- 28:56these modes. There's no problem there.
- 28:58Now, secondly, there's fast mode. We saw
- 29:00it in the mode menu. Fast mode also has
- 29:02a separate cache. If you open and close
- 29:04it in the middle of a conversation, the
- 29:05conversation up to that point is
- 29:06processed again from the beginning. It
- 29:08says to open a document at the beginning
- 29:09of the session. Thirdly, there's taking
- 29:11a long break. The cache doesn't last
- 29:13forever. If you don't send anything for
- 29:15a while, the cache is cleared. In your
- 29:17subscription plan, this period is 1 hour
- 29:20. Each message restarts the timer. So,
- 29:22it stays as long as you're working. When
- 29:24you return after a break longer than an
- 29:26hour, the entire context is reprocessed
- 29:27from scratch in your first message.
- 29:29Therefore, that message arrives slowly
- 29:31and consumes more effort. Subsequent
- 29:32messages are also processed from the
- 29:34cache. Fourthly, compact mode. Because
- 29:36it changes the conversation in summary,
- 29:38the cache is inevitably re-processed.
- 29:40This isn't expensive while the cache is
- 29:42still active, but if you use compact
- 29:44mode after a break longer than an hour,
- 29:45when the cache has been cleared, the
- 29:47entire conversation has to be
- 29:48reprocessed from the beginning. The
- 29:50document states that this is when
- 29:52compact mode is most expensive. So, if
- 29:54you return from a long break, don't use
- 29:57compact mode immediately. Use compact
- 29:59mode while working between two tasks.
- 30:01The document also mentions other points
- 30:03. For example, when you paste a large
- 30:05number of screenshots, they are pushed
- 30:07to older ones, and everything from that
- 30:08point onwards is reprocessed. But as a
- 30:11beginner, these four points are the ones
- 30:13you'll encounter most often.
- 30:16In short, select the model and fast mode
- 30:18at the beginning of the session.
- 30:21Know that there's a cost if you need to
- 30:22change it in the middle of a task. You
- 30:25can easily adjust the effort. You need
- 30:27to use compact mode between two tasks.
- 30:29If you take a long break, don't return
- 30:31to the same conversation. Start with a
- 30:33new message in a new session. Now, let's
- 30:35talk about the limit. Many people
- 30:37confuse this with context. Context
- 30:39refers to how full the session is, while
- 30:41limit refers to how much of your
- 30:42subscription you've used. Context
- 30:44empties during the session, but the
- 30:46limit doesn't. It resets when its time
- 30:48comes. There's a small circle at the far
- 30:51right of the line below the box.
- 30:52Clicking on it shows you how much of
- 30:54your plan you've used. There's a 5-hour
- 30:56limit and a weekly limit for all models
- 30:58. Next to each, it says when it will
- 31:01reset. It doesn't matter if you're using
- 31:03the cloud code on your desktop, in the
- 31:05terminal, or in any plugin. The limit is
- 31:08the same for all. If you want to see
- 31:10more details, type "usage" in the box. A
- 31:12usage card appears in the chat. At the
- 31:15top, there are the same bars. If you
- 31:17used Fab, there's a separate weekly bar
- 31:19for that. Below that is the session
- 31:21section. You see a cost figure there.
- 31:23Don't worry about that.
- 31:25The documentation clearly states that
- 31:27this figure is included in your usage
- 31:29plan, not billed to you.
- 31:31Below that is a breakdown. What I
- 31:33described in the cache, you can see here
- 31:35with the actual figure. In my test, the
- 31:38input was 8 tokens, while 199,700 tokens
- 31:41were read from the cache. So, almost all
- 31:43of the context sent to the Cloud in that
- 31:46message was read from the cache. Only 8
- 31:48tokens were processed initially. At the
- 31:50bottom, there's a section showing what's
- 31:52consuming your usage limits. It shows
- 31:54you the percentage of your limit
- 31:56consumed by what.
- 31:57From the menu next to it, you can view
- 32:00this session in the last 24 hours or the
- 32:02last 7 days.
- 32:03If your limit is suddenly used up so
- 32:05quickly, this should be the first place
- 32:07you check.
- 32:09To summarize, what's consuming your
- 32:11limit is the context resent with each
- 32:13message.
- 32:15The larger the context, the more tokens
- 32:17per message.
- 32:19The cache makes this cheaper, but it
- 32:20doesn't eliminate it.
- 32:22With a large model and high effort, it
- 32:24goes even faster.
- 32:26That's why compact and clear aren't just
- 32:29important for the Cloud not to lose what
- 32:31it initially said, but also for the
- 32:33limit itself. Now, something is missing
- 32:35. The Turkish character rule was lost in
- 32:37the summary. We can say that compact
- 32:39should keep this, but do we have to
- 32:41remind the Cloud every time? When you
- 32:42open a new session, Cloud won't know
- 32:44about that rule at all. The document's
- 32:46recommendation is clear here. Don't
- 32:48leave the rules that should be valid in
- 32:50every session in the chat; write them in
- 32:52a separate file. There's a special file
- 32:54in the folder specifically for this
- 32:55purpose.
- 32:57This special file is called cloud.md.
- 32:59It's a plain text file inside the folder
- 33:01. Cloud reads this file at the beginning
- 33:03of each session. So it knows what's
- 33:05written there without you having to tell
- 33:07it each time. Remember how the Turkish
- 33:08character rule was lost in the summary
- 33:10earlier? This doesn't happen with
- 33:12cloud.md. Because even after Cloud
- 33:14generates the summary, it reads the file
- 33:16from disk again and adds it to the
- 33:18session. Since the rule is in the file
- 33:20and not in the chat, it doesn't get lost
- 33:22along with the summary. So how do we
- 33:24create it? The easiest way is to tell
- 33:26Cloud. You write the rule and then tell
- 33:28it to add it to cloud.md. Cloud creates
- 33:30the file if it doesn't exist, or adds it
- 33:32if it does. There's also a ready-made
- 33:34command. You type ` init` in that box.
- 33:37If cloud.md doesn't exist in the folder
- 33:38, Cloud checks the folder and writes you
- 33:40one to start with. We already have one;
- 33:42we just created it. Then it doesn't
- 33:44overwrite it. It examines the folder,
- 33:46lists what it finds missing in the file
- 33:48, and finally asks if it wants to add it
- 33:50. If you say yes, it adds it; if you say
- 33:52leave it for now, it doesn't touch the
- 33:54file. So what are we going to do with
- 33:56this? Antropi's documentation has a good
- 33:58sentence. It says to think of the Cloud
- 34:01md as the place where you write things
- 34:03you'll explain repeatedly. It also tells
- 34:05you when to add it.
- 34:07If Cloud makes the same mistake a second
- 34:09time, if you're writing the correction
- 34:11you wrote in the previous session again
- 34:13in this session, or if someone else
- 34:15working in this folder will need the
- 34:17same information, then add it.
- 34:19When writing, pay attention to two
- 34:21things. First, write the rule in a
- 34:23verifiable way. If you write "name the
- 34:25files properly," what will Cloud
- 34:27understand from that? What does "
- 34:28properly" even mean? But if you write "
- 34:31don't use Turkish characters or spaces
- 34:32in file names, separate words with
- 34:34hyphens," then it's clear what it will
- 34:36do. You can check if it did or not. The
- 34:38documentation says exactly this. It says
- 34:41to write clearly enough to be verifiable
- 34:43. As for the second point, it recommends
- 34:44keeping it short. The documentation
- 34:46recommends under 200 lines. Why? Because
- 34:49this file is entered into the context
- 34:50from beginning to end in every session.
- 34:52The longer the file gets, the more space
- 34:54it takes up, and the less likely Cloud
- 34:56is to follow the rules. Also, if you've
- 34:58written two conflicting rules, Cloud can
- 35:00randomly choose one of them. You can
- 35:03occasionally open the file and delete or
- 35:05update the old or new rules. But you
- 35:07need to know this: Cloud MD isn't a list
- 35:10of prohibitions. Cloud reads it and
- 35:12tries to comply, but ultimately it's an
- 35:14instruction.
- 35:16If you want something to never happen,
- 35:18there's another way.
- 35:19We call that a hook.
- 35:21We'll see that at the end. Now, you can
- 35:23place Cloud MD in two places. If you put
- 35:26it inside the project folder, it's only
- 35:27readable when working in that folder. If
- 35:29you put it in Cloud's own settings
- 35:31folder on your computer, it's readable
- 35:33in all projects. If you have a rule that
- 35:35applies everywhere, like "always reply
- 35:37without emojis," you'll put it in the
- 35:39second place. Cloud also keeps its own
- 35:41notes.
- 35:42This is called auto-memory.
- 35:45When you correct Cloud or tell it a
- 35:47preference, if it thinks it will be
- 35:49useful in the future, it makes a note of
- 35:51it. When it does this, it writes in the
- 35:53chat that it has been saved and that
- 35:54there is a memory record. The difference
- 35:56here is...You write Cloud MD yourself.
- 35:59Cloud automatically writes the memory.
- 36:01Both are loaded at the beginning of each
- 36:02session. If you tell Cloud to remind you
- 36:04of something, it writes it to its memory
- 36:06. If you want a rule to be in Cloud MD,
- 36:08you can explicitly tell Cloud MD to add
- 36:11this rule. Now, after using Cloud MD for
- 36:13a while, the first thing you'll probably
- 36:15want to do is say, "This file is getting
- 36:17long, let's shorten it a bit." But
- 36:19sometimes Cloud deletes a rule you want
- 36:21to keep while shortening it. So what do
- 36:23you do then?
- 36:24For example, I told Cloud to shorten
- 36:26Cloud MD. It shortened it, but it also
- 36:28deleted the Turkish character rule. I
- 36:30didn't want that. There's no need to
- 36:32panic here. Cloud Cop automatically
- 36:34saves the current state of the files
- 36:36before Cloud touches anything, every
- 36:38time you send a message. They call this
- 36:40a checkpoint. Like in games, for example
- 36:43.
- 36:43You can think of it as a save point.
- 36:45Every message has a save point. When you
- 36:47want to undo something, you return to
- 36:49that point. So how do we undo it? There
- 36:51are three ways. All three lead to the
- 36:53same place. The first is that you type '
- 36:55slash rewind' into the box. A list of
- 36:57the messages you sent appears. Click on
- 37:00the message you want to return to.
- 37:02Secondly, when the box is empty, press
- 37:05ESC twice. The same list appears again.
- 37:08Thirdly, and the simplest way, hover
- 37:10your mouse over your own message. A
- 37:12rewind button appears below it.
- 37:14Pressing it takes you back to that
- 37:16message in the list. No matter which way
- 37:18you go, the same window appears. "
- 37:21Rewind to this message?" Your message is
- 37:23displayed here. Below it, you see how
- 37:25many files will be restored. In the
- 37:27restore section, you choose what will be
- 37:29restored. There are three options here.
- 37:31The first is code and chat. This is the
- 37:33default. Both the files and the
- 37:35conversation are restored to their state
- 37:37before that message. You can think of it
- 37:39as if you never sent that message.
- 37:41The second is chat only. Here, it
- 37:43doesn't touch the files. It only
- 37:44restores the conversation to that point
- 37:46. The third is code only. Here, "code"
- 37:48refers to your files, meaning the files
- 37:50are restored to their original state,
- 37:52but the conversation remains as it is.
- 37:54This means Cloud still knows what went
- 37:56wrong. I think the third option is more
- 37:58useful most of the time. Why? Your file
- 38:00is fixed, but Cloud remembers that you
- 38:02deleted the rule. Immediately after, you
- 38:04say "shorten it, but don't touch the
- 38:05Turkish character rule." It doesn't make
- 38:07the same mistake on the second attempt.
- 38:09Also, something nice happens when you
- 38:11rewind the chat. The message you
- 38:12selected returns to the prompt box. So
- 38:14you can correct your message and resend
- 38:16it. And if you rewind to the wrong
- 38:18message, you don't need to worry. A link
- 38:20appears saying "undo rewind" where you
- 38:23rewound. You can undo it from there. But
- 38:25you need to know the limitations of this
- 38:27feature. Because many people are
- 38:29mistaken here. Firstly, the save point
- 38:31only covers files that Cloud has
- 38:33modified using its own editing tool. If
- 38:35Cloud did something by running a command
- 38:37, it doesn't revert. It's written at the
- 38:39bottom of the window. In the first job,
- 38:41they moved files. They changed their
- 38:42names, you remember. Cloud does that
- 38:44with a command. So I don't want you to
- 38:46assume that those files will return to
- 38:48their original location when you rewind
- 38:50. Secondly, changes made by sub-agents
- 38:52often don't go into these save points.
- 38:55We'll see sub-agents in detail later.
- 38:58Thirdly, if you manually modified a file
- 39:01outside of Cloud, that's not tracked
- 39:03either. It's clearly stated in the
- 39:05documentation. This feature doesn't
- 39:07replace a permanent backup for quick
- 39:09rollback.
- 39:10They mention it so you have a backup if
- 39:12you're working with important files. By
- 39:14the way, you'll notice something when
- 39:16you undo and rewrite the message. To get
- 39:18good results from the cloud, you need to
- 39:20write the instructions long and clear.
- 39:22But if this is a task that repeats every
- 39:25week or every day, will you have to
- 39:27rewrite that long instruction from
- 39:29scratch every week, every day, every
- 39:31time you do the task?
- 39:33I have exactly that kind of task. When I
- 39:35finish a video, I give it the text and
- 39:37print the title, description, and tags.
- 39:39I was explaining the same things every
- 39:41time: how many titles I wanted, how long
- 39:43they should be, what the first line of
- 39:45the description should be, where the
- 39:47result should be saved, and so on.
- 39:49Skills are exactly for this. The
- 39:51Antropin documentation also says this.
- 39:53If you're pasting the same instruction,
- 39:55the same checklist, or a step-by-step
- 39:57task into the chat repeatedly, turn it
- 39:59into a skill. A skill is actually a
- 40:01folder. Inside, there's a file called
- 40:03skill.md. The folder name is the name of
- 40:05the skill. Let me show you, for example
- 40:07. As you can see, the file has two parts
- 40:09. At the top, between two lines, there's
- 40:12a short description. It explains what
- 40:14this skill does and when to use it.
- 40:16Below that are step-by-step instructions
- 40:18for Cloud. As you can see, it's all in
- 40:21Turkish. No code or anything. So how do
- 40:23we use this? There are two ways. First,
- 40:26you type "slash" in the box. The skill
- 40:29name appears in the list. You select it
- 40:31. You add the text and send it.
- 40:33Remember the "slash commands" line we
- 40:35saw in the plus menu earlier?
- 40:38Well, the skills are listed there. The
- 40:41second way is that you don't need to run
- 40:42any commands. You just say, "Write a
- 40:44title and description for this video
- 40:46text." Cloud checks your request. If it
- 40:48matches the description at the top, it
- 40:50unlocks the skill itself. That's why
- 40:52that description line is so important.
- 40:55You can write when to use it, using
- 40:57words people actually use.
- 40:59If Cloud isn't using the skill, that's
- 41:01the first place you should look. So how
- 41:03do we create this? The easiest way is to
- 41:05have Cloud write it. Let's say you've
- 41:07done a task from start to finish once in
- 41:09Cloud. And you're happy with the results
- 41:10. In the same conversation, you say, "
- 41:12Turn the work we just did into a skill.
- 41:14" Cloud writes the steps to a file. Then
- 41:16you open it and read it. If there are
- 41:18any missing parts, you correct them.
- 41:20Cloud notices the changes in the skill
- 41:22file without logging out.
- 41:24So you don't need to restart. There's
- 41:26also the application itself. In the left
- 41:28sidebar, you see a section called "
- 41:30Customization." When you open it, there
- 41:32are three tabs at the top: Skills, MCPs
- 41:35, and Plugins. For now, we're looking at
- 41:37skills. We'll see the other two when
- 41:39it's their turn.
- 41:41In the "Yours" section, the skills you
- 41:43created and those you unlocked from
- 41:45Antropik are listed. You can disable a
- 41:46skill you don't use without deleting it
- 41:48. Next to the disabled one, it says "
- 41:50Disabled." To create a new skill, press
- 41:52the "Add" button in the upper right.
- 41:54Four options appear. Three are for
- 41:56creating a skill, and one is for loading
- 41:58an existing skill. We'll get to loading
- 42:00in a moment. Let's look at creating
- 42:02first. The first is creating a skill. If
- 42:04you know what you want and want to write
- 42:06it yourself, you'll use this method. A
- 42:08form opens here. At the top, you write
- 42:10the name of the skill. Below, you write
- 42:12a description, meaning what it does and
- 42:14when to use it. This is the part we saw
- 42:16at the top of the skill MD earlier.
- 42:18Cloud opens the skill based on this
- 42:19sentence. You also write step-by-step
- 42:21instructions in the large area at the
- 42:23bottom. If you want to include an
- 42:25example file along with the instructions
- 42:27, there's an "add file" option in the
- 42:29upper right. You can select it from
- 42:30there. When you're finished, you click "
- 42:32create." Your skill is created. The
- 42:34second option is "create with Cloud." If
- 42:36you're not sure exactly what to write,
- 42:38you can choose this. When you click it,
- 42:40a new chat opens, and a pre-written
- 42:42message appears in the box.
- 42:45The message tells Cloud to create a
- 42:46skill with you, and first, to ask you
- 42:49what the skill should do.
- 42:51You send this. Cloud asks you a question
- 42:53. You answer as if you were having a
- 42:55normal conversation. Then, he already
- 42:57writes your skill. The third, and
- 42:59perhaps the most interesting, option is
- 43:01"record your screen." Here, you don't
- 43:03actually explain it, you demonstrate it
- 43:04. When you click it, a small window
- 43:06appears. It says that your screen clicks
- 43:08, typing, and voice will be recorded and
- 43:10then sent to Cloud to be converted into
- 43:12a usable skill. You select your
- 43:14microphone from the microphone button in
- 43:16the lower right. Since your voice also
- 43:18goes to Cloud, you can verbally explain
- 43:19what you're doing and why while you're
- 43:21doing the task. When you're ready, you
- 43:23press 'e' and do the task yourself from
- 43:26start to finish once. Once the recording
- 43:28is complete, Cloud converts it into a
- 43:30skill. You need to pay attention to
- 43:31something here. If you have any
- 43:33confidential information or private
- 43:35details, like entering a password,
- 43:36during recording, don't show them. The
- 43:38window already displays this warning.
- 43:40Now, no matter how you do it, the skill
- 43:42you add here is saved to your Cloud
- 43:44account, not to a folder on your
- 43:46computer. Therefore, no matter where you
- 43:48open Cloud from, it can work anywhere
- 43:49you log in with the same account. Now
- 43:51let's talk about the difference with
- 43:53Cloud MD. Because the two can sometimes
- 43:54get confused. Cloud MD loads completely
- 43:57in every session. Only the top
- 43:59description of the skill loads. The
- 44:01actual instructions only become
- 44:02available when the skill is used.
- 44:04Remember the line called "skills" in the
- 44:07context card? These are the descriptions
- 44:09there. So you can write a skill as long
- 44:11as you want. It takes up almost no space
- 44:13when you're not using it. The rule here
- 44:15is: the rules that apply in every
- 44:17session are the same for the skill side
- 44:19if you're doing a task step-by-step with
- 44:21Cloud MD. If a part of Cloud MD has
- 44:23turned into a recipe saying "do this
- 44:25before the rule, then do that," you can
- 44:27move that part to the skill file. Also,
- 44:29pay attention to something in a long
- 44:31session. When the context is full and a
- 44:33summary is extracted, the skills you
- 44:35used are added back, but only the
- 44:36beginning. Therefore, write the most
- 44:38important rule at the top of the skill
- 44:40file. You can also put the skill file in
- 44:42two places, like Cloud MD. If you put it
- 44:44in the project folder, it only works in
- 44:46that project; if you put it in the Cloud
- 44:48settings folder, it works in all
- 44:49projects on this computer. The ones you
- 44:51add from the customization section, as I
- 44:53said before, remain in your account. You
- 44:55also need to know this: you might not
- 44:57want some skills to automatically open
- 44:59on the cloud side. For example, if
- 45:01there's a skill that sends something or
- 45:03publishes something, you don't want the
- 45:05cloud to publish it or send a message
- 45:07for you by saying "the job seems ready.
- 45:09" To prevent this, you add a single line
- 45:12to the top of the skill file.
- 45:14From then on, only you can start that
- 45:16skill with slash commands.
- 45:18You don't always have to write the skill
- 45:20yourself. Skills are written according
- 45:22to an open standard. So they also work
- 45:25on other AI tools.
- 45:26People share the skills they create
- 45:28online. The easiest place to find
- 45:30ready-made skills is again the
- 45:31customization section. There, in the
- 45:33explore section, you'll find skills
- 45:35created by Antropin and its partners.
- 45:37They are categorized. The most uploaded
- 45:39ones are in a separate list. Each card
- 45:41already shows who published it and how
- 45:43many times it has been uploaded. You
- 45:45click the plus sign next to the one you
- 45:47like and you can easily add it. Now,
- 45:49let's say I have a skill I downloaded
- 45:51from the internet. What do we do to add
- 45:53it? Of course, we'll look at the fourth
- 45:56option. There you click on "upload skill
- 45:58". A page opens. You drag and drop the
- 46:00file to the area in the middle or click
- 46:02and select it from your computer. It
- 46:04already shows which files it accepts
- 46:06below. It will either be a single dot MD
- 46:09file with its name and description at
- 46:11the top, or a ZIP file containing "
- 46:13skill dot MD". Then you click "upload".
- 46:16The application also runs a security
- 46:18scan during installation. If you want,
- 46:20you can also put the downloaded folder
- 46:22as a file in the skill folder. You can
- 46:24also have this done by the cloud.
- 46:26Remember this: don't get complacent just
- 46:28because there's a scan. You need to be
- 46:30very careful here. Antropik's
- 46:32documentation is very clear on this.
- 46:34Only use skills from trusted sources.
- 46:37That is, from sources you've written
- 46:39yourself or from Antropik. Why? Because
- 46:41a skill isn't just instructions. It can
- 46:43also contain code that the cloud will
- 46:45execute. So a malicious skill could do
- 46:48something different than what's
- 46:49described in the instructions. For
- 46:51example, it could send your files
- 46:52somewhere, and it does this on your
- 46:54computer with your permissions.
- 46:56Therefore, if you download a skill from
- 46:58somewhere, open and read all the
- 47:00accompanying files before installing it
- 47:02. If you don't understand something, I
- 47:04don't recommend installing it here.
- 47:07Remember the tab next to the
- 47:09customization section?
- 47:12It teaches you how to do a task, but if
- 47:14the necessary materials aren't on your
- 47:16computer—for example, your invoices are
- 47:19in Gmail, your files are in Drive—it
- 47:21can't access them.
- 47:23Everything it's provided so far was in
- 47:25that folder on your computer. The cloud
- 47:27only sees that folder. The invoice in
- 47:29Gmail or the document in Drive doesn't
- 47:31exist for it; it can't see them. So what
- 47:33do you do in this situation? You go and
- 47:35download the invoice, put it in the
- 47:38folder, or open the email, copy the text
- 47:40, paste it, and so on. So you're the one
- 47:42transferring information from one device
- 47:44to another. Entropy's documentation
- 47:46describes this very moment. It says that
- 47:48the moment you realize you're copying
- 47:50information from another device to the
- 47:52chat, it's time to connect that device.
- 47:54So how do we connect here? There's
- 47:56something called ` connect`. This is an
- 47:59open standard, a common connection
- 48:01method that everyone can use. If an
- 48:03application has given support, it can
- 48:04connect to that application. After
- 48:06connecting, it reads the information
- 48:08there itself. If you allow it, it can
- 48:10even do work there. For example, in the
- 48:13desktop application, it's called an
- 48:15adapter. So it's the same thing here.
- 48:17The setup is done with a single click.
- 48:19Now you go to the customization section
- 48:21on the left side. At the top is the
- 48:23adapters, and next to it is ` discover`
- 48:26. You type the name of the application
- 48:28in the search box.
- 48:29The result shows what it is under each
- 48:31line.
- 48:33Is it a plugin?
- 48:35You click the plus sign next to it.
- 48:38Some lines have a community tag. Those
- 48:40were made by other companies, not the
- 48:42device's own company. In most cases, you
- 48:44just log in with that application's
- 48:46account and give permission, that's all
- 48:48. So the setup is very simple. If you
- 48:51want to see what you've connected later
- 48:53, you can go to the customization
- 48:54section on the left side. The status of
- 48:56each one is already written there. If
- 48:58it's connected, it has a checkmark; if
- 49:00it's not connected, it says "connect."
- 49:02If you've logged out, you need to click
- 49:04"reconnect."
- 49:05You can also remove it from here.
- 49:07There's also this: These MCPs are
- 49:09registered to your Cloud account. So, if
- 49:11you've connected something to the Cloud
- 49:13chat side, it appears here as well. You
- 49:15don't need to connect it again. Let's
- 49:17start with the first example. Let's
- 49:18start with Gmail. After connecting, I
- 49:20write this: List the emails that arrived
- 49:22this week and are awaiting a reply.
- 49:24Prepare a short draft reply for each,
- 49:26but don't send them. I want to emphasize
- 49:28that I said don't send them.
- 49:30Because the reply written by the Cloud
- 49:32goes under your name.
- 49:34So, initially, leave it in drafts.
- 49:37Read it there and check it.
- 49:39After checking it that way, send it.
- 49:41Sent emails cannot be undone. That is,
- 49:44the undo section's save points work for
- 49:46files in your folder. They don't work
- 49:48for emails that have already gone to
- 49:50Gmail. I'll just mention that warning
- 49:52here. Moving on to the second example,
- 49:54the Drive example. Let's say your
- 49:56invoices are in a folder in Drive. Here
- 49:58I write this...Find the invoices added
- 49:59this month in the invoices folder in
- 50:01Drive, and download them to the invoices
- 50:03folder on your computer. The first time
- 50:05, it asks for your permission to
- 50:07download files from Cloud Drive. Once
- 50:08you grant permission, it downloads them
- 50:10. There are no steps like downloading,
- 50:12dragging, or renaming. You are no longer
- 50:14the one carrying the information. MCPs
- 50:16aren't just for Google applications. Let
- 50:18me give two more examples. Both are
- 50:20tools that open the internet to the
- 50:21Cloud. The first is the File Crawl tool
- 50:23. The Cloud can already read a webpage.
- 50:26File Crawl takes this a step further. It
- 50:28crawls the entire site page by page and
- 50:30converts each page into plain text that
- 50:33the Cloud can easily read. Not just a
- 50:35single page, but the entire site. I use
- 50:37it when learning a new tool, for example
- 50:39. I say, "Pull all the pages from this
- 50:41tool's help site, save them to a folder
- 50:43, and then answer my questions from
- 50:45there," and so on. Now the Cloud
- 50:47responds not with what it remembers from
- 50:49the internet, but with the current state
- 50:51of the site.
- 50:52So, there are other uses for File Crawl
- 50:54, but I wanted to give one example here
- 50:57without making it too long. Looking at
- 50:59the second one, there's Zapier. Zapier
- 51:01has thousands of ready-made data
- 51:03collection tools created by others,
- 51:04which are called actors. For example,
- 51:06they can pull regularly generated data
- 51:08from places like Instagram, YouTube, and
- 51:10Google Maps. With MCP Cloud, you can
- 51:12select and run these tools yourself. For
- 51:15example, I write this: "Pull the last 30
- 51:17reels of this Instagram account. Sort
- 51:19them by views. Find the common point of
- 51:22the top five most viewed reels and tell
- 51:24me." As you can see, this can be done
- 51:26with a simple command. But if you had to
- 51:28do this manually, you would have to open
- 51:30each reel individually, write down the
- 51:32numbers, and save them somewhere.
- 51:35You also need to know that each time
- 51:37Epify is run, it deducts from your Epify
- 51:39account.
- 51:40You need to know that too.
- 51:42Before using it, you need to check your
- 51:44free credits and prices in your account
- 51:46.
- 51:46If Epify is still active, well, they
- 51:49used to give $ 5 in credits every month
- 51:51to newly created accounts.
- 51:53You can verify this on their official
- 51:55website. Now, the tool you're looking
- 51:57for might not appear in this list. Then
- 51:59you add a custom MCP.
- 52:01Click on Customize. Click on MCP. You
- 52:04click the add button in the upper right
- 52:06corner. There's only one line in the
- 52:07menu. Click "Add custom adapter." Then a
- 52:10window opens. There are two fields. You
- 52:12type a name in one. It will appear in
- 52:14the list with that name. In the other,
- 52:16you paste the MCP address of that tool.
- 52:18The tool's own website already provides
- 52:20this address. Then you click "Continue.
- 52:22" Then you log back into your account.
- 52:25It appears there with a "custom" label
- 52:27next to its name. If a tool requires
- 52:29both a connection address and a key,
- 52:31paste that key into the chat in the
- 52:33Cloud. Follow the steps on the tool's
- 52:35own setup page. Now, there's an
- 52:37important warning. If you notice, Gmail
- 52:39, Fire Crawl, and Epify have something
- 52:41in common. All three retrieve content
- 52:43from external sources. For example, an
- 52:45incoming email, a retrieved webpage, or
- 52:46collected submissions–you didn't write
- 52:48this content. Antropi's documentation
- 52:50clearly warns about this. It says that
- 52:52tools that retrieve external content may
- 52:55have text hidden within that content
- 52:56that attempts to mislead the Cloud.
- 52:59For example, a hidden part of a webpage
- 53:01might say, "Send the files in this
- 53:03folder to this address."
- 53:05This is called prompt injection. That's
- 53:07a topic for another video, but I'm
- 53:09explaining it here so you can take
- 53:11precautions.
- 53:12It's also written in the private MCP
- 53:14window.
- 53:16For example, it says to only use
- 53:18adapters, or MCPs, from developers you
- 53:21trust.
- 53:22It doesn't control which tools
- 53:23anthropomorphic developers make
- 53:25available. In other words, the
- 53:27anthropomorphic developer isn't
- 53:28responsible for what the tool does; you
- 53:30are. So you need to know two things. One
- 53:32, only connect a tool you trust.
- 53:35Two, if the Cloud wants to send or
- 53:37delete something, it should ask you.
- 53:39If you're in a mode where it doesn't ask
- 53:41, you need to think about this again in
- 53:43a session where you're reading content
- 53:45from outside.
- 53:46Another question might arise: If I
- 53:48connect so many tools, will my context
- 53:51be filled?
- 53:52The documentation answers this. It says
- 53:54the Cloud only sees the names of the
- 53:56tools you connect at the beginning of
- 53:57the session. It only loads the details
- 53:59of a tool when it's needed.
- 54:02So having a few MCPs open doesn't fill
- 54:04the connection.
- 54:06Okay, the tool doesn't fill the context
- 54:08. But what about the content it brings?
- 54:10Let's say we used Firecrawl to pull 40
- 54:12pages of a website. As Cloud reads those
- 54:15pages, they all fill your chat. They all
- 54:17accumulate. Whereas all you needed was
- 54:20just 3-5 lines of content from those
- 54:22pages.
- 54:23Now we're getting to the sub-agents. We
- 54:25saw this in the context section. We said
- 54:27that as the chat fills up, Cloud has a
- 54:29harder time remembering the initial
- 54:31things. We called this context rot,
- 54:33remember? For example, let's say there's
- 54:35a research project. This research
- 54:37project does exactly that. Cloud opens
- 54:39dozens of pages, reads them all. And all
- 54:41of it goes into your context. In the end
- 54:43, you get a summary, but everything
- 54:45that's read stays in the chat. That's
- 54:47why sub-agents exist.
- 54:50Well, not exactly for that reason, but
- 54:52they make things much easier here, and
- 54:54I'll explain sub-agents using this
- 54:56research example so you can understand
- 54:58better.
- 55:00Now, a sub-agent is a separate assistant
- 55:02that Cloud delegates a task to. They
- 55:05have their own context. They do all the
- 55:07reading on their side and only bring you
- 55:09the result.
- 55:10Those readings don't enter your chat.
- 55:12Actually, we've seen the sub-agent once
- 55:15before. We just haven't given it a name
- 55:16. For example, when Cloud code or files
- 55:18are being reviewed in plan mode, it uses
- 55:20its own pre-built sub-agent to perform
- 55:22this review. That's why the plan exits
- 55:24without filling up your chat. Cloud code
- 55:26has several such pre-built sub-agents.
- 55:28For example, a search agent that can
- 55:30only read and not modify anything. Cloud
- 55:32opens these itself when it needs to
- 55:34research something. You don't do
- 55:35anything.
- 55:37The nice thing about sub-agents is that
- 55:39you can have them perform independent
- 55:41tasks simultaneously.
- 55:43Let me give you an example. A very
- 55:45simple example. Let's say Cloud code
- 55:47researches the prices and usage limits
- 55:49of cursors and Codex. Use a separate
- 55:51sub-agent for each. Then I tell it to
- 55:54compare them all in a single table. As
- 55:55you can see, the three researches are
- 55:57done directly without waiting for each
- 55:59other. That is, they run simultaneously
- 56:02. They can work in parallel. Each one
- 56:04reads dozens of pages, but you will only
- 56:06get three short results and a table in
- 56:08your chat. In my test, when the three
- 56:10were finished and the table appeared,
- 56:12the context increased from 7%to 9%. If
- 56:14you want to see what they are doing
- 56:16while they are working, look at the
- 56:17panel on the right. When sub-agents are
- 56:19running, a section called "background
- 56:21tasks" appears. Each sub-agent is a row
- 56:23. Next to it is how long it has been
- 56:25running and a stop button.
- 56:27Clicking on a row shows what task Cloud
- 56:29has assigned to that sub-agent and which
- 56:31page it is currently reading. Now, if
- 56:34you frequently assign a task to a
- 56:36sub-agent, you can make it permanent.
- 56:38Again, the easiest way is to tell Cloud
- 56:40. For example, let's say you create a
- 56:43sub-agent that checks for spelling
- 56:45errors in your Twitter texts. It should
- 56:47only read, not modify the file. Let's
- 56:49say it uses the Haiku model. We saw the
- 56:51skill section in the Skill section. This
- 56:53is somewhat similar.
- 56:55The difference here is that the skill
- 56:57works within your chat. The sub-agent
- 56:59works in a separate context and brings
- 57:01you the result. Two lines in that file
- 57:03are important. One is that because you
- 57:05said the tools should only read, this
- 57:07sub-agent cannot touch your file. The
- 57:09second is the model. For a simple task
- 57:11like spell checking, high-q is
- 57:13sufficient and uses much less resources
- 57:15. That's why we choose high-q. The
- 57:17documentation lists this as one of the
- 57:19benefits of sub-agents. It talks about
- 57:21directing simple tasks to the model more
- 57:23quickly and cheaply. Cloud selects this
- 57:25sub-agent based on its description.
- 57:28You can also call it by name.
- 57:30For example, you say, "Check this text
- 57:32with the spell check agent," and it's
- 57:34called automatically. However, I should
- 57:36mention something. If you created the
- 57:38first sub-agent in this folder, you need
- 57:41to open a new session. I tried it in the
- 57:43same session, and Cloud couldn't find
- 57:44the agent. The documentation already
- 57:46states this. It says the first sub-agent
- 57:48in a project is loaded with a new
- 57:50session. Cloud also mentions that you
- 57:52may need to restart the session when
- 57:54creating it.
- 57:55Now, let's talk about the cost.
- 57:57Everything has a cost, and this has one
- 57:59too. What you need to know is this: Each
- 58:01sub-agent sends its own requests, and
- 58:03these deduct from your usage limit. The
- 58:06documentation warns about this. If you
- 58:08open many sub-agencies, and each one
- 58:09brings long results, you'll not only use
- 58:11up your limit but also fill up your chat
- 58:13with those results. So, it says don't
- 58:15open sub-agencies for every task.
- 58:17Sub-agencies aren't necessary. For
- 58:19example, if a side task is going to fill
- 58:21your chat with a lot of reading material
- 58:23, or if independent tasks are being done
- 58:25simultaneously, then use a sub-agent.
- 58:27How can we give an example here? Let's
- 58:28go with an example from the research.
- 58:30Let's say we're going to do research on
- 58:323-4 different topics. Calling in a
- 58:34sub-agent makes sense here. Why?
- 58:36Because the context of the main chat
- 58:38gets clogged while researching, and
- 58:41researching in 4-5 different categories
- 58:43takes a very long time.
- 58:45If we open 4-5 different accounts for
- 58:474-5 different categories, our context
- 58:50remains uncluttered. We get what we want
- 58:52cleanly, and we finish our work faster.
- 58:55But as the documentation says, these
- 58:57sub-agents consume our time limit while
- 58:59working. We need to take this into
- 59:01account when doing our work.
- 59:03Now, we've told the sub-agent to only
- 59:05read, and it can't touch the file.
- 59:08But how do we do this for the main one?
- 59:10How can we guarantee it will never do
- 59:12something?
- 59:14That's where hooks come in.
- 59:17Remember Cloud's D section? We wrote the
- 59:19rules there. Let's say you wrote "never
- 59:21touch files in the archive folder."
- 59:24Cloud usually complies with this. But
- 59:26the Antropin documentation clearly
- 59:28states that Cloud sees Cloud's D as
- 59:30context, not a mandatory rule. It's more
- 59:33like a request. In a long session with
- 59:35complex tasks, this request can
- 59:37sometimes be overlooked. But when it
- 59:39comes to hooks, a hook isn't a request.
- 59:41A hook is a small command that Cloud
- 59:43code automatically executes at specific
- 59:45times. For example, let's say it runs
- 59:48just before Cloud edits a file or when
- 59:50Cloud finishes its work. So the decision
- 59:53isn't Cloud's. When that moment comes,
- 59:55the hook runs automatically every time.
- 59:57The documentation explains it this way.
- 59:59It says if you want something to happen
- 1:00:01every time without the model having to
- 1:00:03choose, use hooks. Now, I'm writing this
- 1:00:05: Write a hook that prevents editing
- 1:00:08files in the archive folder. Cloud adds
- 1:00:10the hook to the configuration file. You
- 1:00:12don't need to write anything. When it's
- 1:00:15finished, it also says: This hook stops
- 1:00:17Cloud's file editing tool. But Cloud can
- 1:00:19still write to a file with a command.
- 1:00:21That path is still open, it says. It
- 1:00:23also says, "Tell me if you want me to
- 1:00:25close that." I want you to keep this in
- 1:00:26mind. We'll come back to this in a
- 1:00:28moment. Let's try this now. I'm opening
- 1:00:30a new session so that the hook is
- 1:00:32definitely loaded here. I'm asking Cloud
- 1:00:34to add a line to a file in the archive.
- 1:00:36As you can see, the hook stopped the
- 1:00:38editing before Cloud even finished. It
- 1:00:40also told Cloud why it was stopped.
- 1:00:41Cloud suggested two options, but you
- 1:00:43need to pay attention to one sentence.
- 1:00:45It says it didn't bypass the hook with a
- 1:00:47command. The reason is this: the hook
- 1:00:49looks at the tool Cloud uses. Our hook,
- 1:00:52however, captures Cloud's file editing
- 1:00:54tool. If Cloud tries to write to the
- 1:00:56same file with a command, this hook
- 1:00:58won't see it. The Antropin documentation
- 1:01:00explains it this way. The protected file
- 1:01:02instance is for the editing tool. So, if
- 1:01:04you want to check commands as well, the
- 1:01:06hook needs to look at the commands too.
- 1:01:09Cloud mentioned this during installation
- 1:01:11. So, when you install a hook, you need
- 1:01:13to know what it's capturing. Ours
- 1:01:15definitively stops editing.
- 1:01:18It works even if the rule in CloudMD is
- 1:01:20forgotten.
- 1:01:21So, what do you do if you want to see
- 1:01:23which hooks are installed? You ask Cloud
- 1:01:26. It looks at the configuration file and
- 1:01:28tells you. There's a command that lists
- 1:01:30hooks, but it only works in the terminal
- 1:01:32. If you type it on the desktop, a
- 1:01:34warning pops up saying this. Now, what
- 1:01:37else can be done with hooks?
- 1:01:39If you ask if hooks only prevent file
- 1:01:42editing, the answer is no.
- 1:01:44Hooks have two main functions. One
- 1:01:46approach is to prevent something from
- 1:01:48happening, like the one before, and the
- 1:01:49other is to ensure something happens
- 1:01:51every time. For example, Cloud could
- 1:01:53automatically format a file after
- 1:01:55editing it, or Cloud could send you a
- 1:01:57message when it finishes a task.
- 1:02:00I use this, for instance. Cloud
- 1:02:02automatically sends a notification when
- 1:02:04it's finished. But if you're just
- 1:02:06starting out, that protection example
- 1:02:08might be the most useful. Because a hook
- 1:02:10executes a real command, if it makes a
- 1:02:12mistake, it will make the same mistake
- 1:02:14every time. Therefore, you need to
- 1:02:16install few and clear hooks. In other
- 1:02:18words, you shouldn't overload the hooks
- 1:02:19. The rule here is simple: if you want
- 1:02:22it to happen generally, use CloudMD; if
- 1:02:24you want it to never happen or always
- 1:02:26happen, use a hook. So far, we've
- 1:02:29installed four things: Skill, MCP,
- 1:02:31sub-agent, and hook. Let's say someone
- 1:02:34has configured all of these for their
- 1:02:36own work, connected them, and it's
- 1:02:38working.
- 1:02:39You want the same thing. Do you have to
- 1:02:41copy them all one by one and put them in
- 1:02:43the right places? Now, that's where
- 1:02:45plugins come in. That's exactly what a
- 1:02:47plugin is.
- 1:02:49It's something that combines several
- 1:02:52skills, sub-agents, hooks, and MCPs into
- 1:02:54a single package. You install it, and
- 1:02:56everything comes with it at once. When
- 1:02:58the package creator updates it, you
- 1:03:00update too. But first, let me say this.
- 1:03:02It's clearly stated in the documentation
- 1:03:04. Skills, sub-agents, hooks, and MCPs
- 1:03:06work independently. It says you don't
- 1:03:09need a plugin.
- 1:03:10A plugin is only useful when you want
- 1:03:12several things together or when you want
- 1:03:14to use a pre-installed setup.
- 1:03:16Now, let's look at how to install it. We
- 1:03:18go to the customization section on the
- 1:03:20left side. We go to Plugins. From there
- 1:03:23, we click on Explore. At the top,
- 1:03:25there's a plugin featured by Antropi.
- 1:03:28Below that, new plugins are listed in
- 1:03:30cards. Each card shows who published it
- 1:03:32. For example, I'm opening Antropi's
- 1:03:34Design plugin. The plugin page has four
- 1:03:37tabs. The overview explains what it does
- 1:03:39.
- 1:03:40The Skills tab shows the skills included
- 1:03:42. Design, for example, has seven.
- 1:03:45You can call them all with a slash, or
- 1:03:48Cloud, which it already uses.
- 1:03:50The MCPs tab shows the tools it can
- 1:03:53connect to. For example, there are nine
- 1:03:55tools like Slack, Figma, Notion, and
- 1:03:58Gmail. There's something you need to
- 1:03:59know here. The plugin doesn't connect
- 1:04:01these tools for you. Just because you
- 1:04:03installed the plugin doesn't
- 1:04:04automatically connect your Slack account
- 1:04:06. Your Figma, Notion, and Gmail accounts
- 1:04:09don't connect automatically. You need to
- 1:04:11connect each one separately for Cloud to
- 1:04:13be able to use them, just like we do
- 1:04:16with MCPs.
- 1:04:17The content tab actually contains all
- 1:04:19the plugin files. For example, the
- 1:04:21skill.md files we see in the skills
- 1:04:23section are stored here in folders. You
- 1:04:25can open and read them. To install, you
- 1:04:27press the plus sign. In the corner, it
- 1:04:29says "Design plugin installed and ready
- 1:04:31to use." There's a switch in the upper
- 1:04:33right corner of the page. You can turn
- 1:04:34the plugin on and off this way. Like
- 1:04:36everything else, plugins have costs.
- 1:04:39First, context. The name and description
- 1:04:41of the skills and sub-agents within each
- 1:04:43open plugin are displayed in front of
- 1:04:45Cloud in every message. This is true
- 1:04:47even if you don't use them. We already
- 1:04:49saw this in the skills section. The
- 1:04:50skill description was always loaded.
- 1:04:52When done individually, it was less. The
- 1:04:54description of our spell check agent
- 1:04:57held 118 tokens in the context card. But
- 1:04:59when dozens are added with plugins, this
- 1:05:01accumulates.
- 1:05:03So you can disable the plugins you don't
- 1:05:05use. You don't need to uninstall them.
- 1:05:07Just disabling them is enough. Now,
- 1:05:09regarding the second point, trust. The
- 1:05:11trust part is very important. I want you
- 1:05:13to remember the warning in the Skill
- 1:05:15section. This is a slightly bigger
- 1:05:16version of that. Because the plugin can
- 1:05:18also contain Hooks and MCPs. This means
- 1:05:20that the moment you install it, it can
- 1:05:22run its own commands on your computer,
- 1:05:24with your permissions. The documentation
- 1:05:26clearly states this. It says it doesn't
- 1:05:28matter which market it comes from. So
- 1:05:29read what it contains before installing
- 1:05:31. The application also shows this. On
- 1:05:33its design page, it says above the MCPs
- 1:05:36that it can send data to other services
- 1:05:38. You can also check its files in the
- 1:05:40content tab. If you don't know who made
- 1:05:42it, I don't recommend installing it here
- 1:05:44.
- 1:05:45So, we've covered everything from the
- 1:05:47installation to the plugins. We've seen
- 1:05:49everything you need to open the cloud
- 1:05:51code and start using it in your own
- 1:05:53business. If you get stuck somewhere,
- 1:05:54you can go back to the relevant section
- 1:05:56and watch it again. The sections will be
- 1:05:58timestamped in the description. I will
- 1:06:00also add all the sources mentioned in
- 1:06:02the video to the description. They are
- 1:06:03all from Antropi's own official
- 1:06:05documentation. But you've seen each of
- 1:06:07them separately now.
- 1:06:09The main point is to use them all where
- 1:06:11they're needed within a project. That's
- 1:06:13exactly what we're doing in the next
- 1:06:14video. You know those videos you've been
- 1:06:16seeing everywhere on Instagram lately,
- 1:06:18the ones generated with Cloud Code and
- 1:06:20getting millions of views? These
- 1:06:21accounts are growing so fast. I've even
- 1:06:23seen one reach 300,000 followers in a
- 1:06:26month. I thought about opening such an
- 1:06:27account and I'll set everything up with
- 1:06:29Cloud Code. We'll start with an empty
- 1:06:31folder. I've already chosen the niche
- 1:06:33beforehand. I'll show it in the video.
- 1:06:35You can choose something else according
- 1:06:37to your interests. Then we'll write the
- 1:06:39video script. We'll have the voiceover
- 1:06:41and editing done, and we'll share it as
- 1:06:43a real reel.
- 1:06:45You'll see the Cloud MDI, skills, and
- 1:06:47sub-agents you see in this video,
- 1:06:49wherever they're needed, in a real
- 1:06:51project.
- 1:06:52You can subscribe to the channel so you
- 1:06:54don't miss the video when it comes out.
- 1:06:56If the video was helpful, you can like
- 1:06:58it. If you're watching from a mobile
- 1:07:00device, I'd be very happy if you hyped
- 1:07:02the video. And after the account is
- 1:07:04opened, I'll be sharing step-by-step how
- 1:07:06it's going, what happened and what
- 1:07:08didn't in the first few days, on Twitter
- 1:07:10. I'm also talking about tools that will
- 1:07:12be useful on Instagram. I'm showing you
- 1:07:14how to use them. The links to both will
- 1:07:17be in the description. Finally, if you'd
- 1:07:19like to see a separate, more detailed
- 1:07:20video on any of the topics in this video
- 1:07:22, you can write it in the comments. I'll
- 1:07:24start with the most requested one.
- 1:07:26See you in the next video.
About this transcript
This page contains the full transcript of 2027'den Önce Claude Code'u Öğren: Sıfırdan Eksiksiz Rehber by Ahmet Demirci | AI Builder, generated from the public captions YouTube serves with the video. The transcript has 13,725 words across 2,032 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.