YouTube2Text

2027'den Önce Claude Code'u Öğren: Sıfırdan Eksiksiz Rehber — Transcript

by Ahmet Demirci | AI Builder · 13,725 words · 2,032 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Everyone's talking about Cloud Code.
  2. 0:01Because with it, people can do alone
  3. 0:03what used to be shared by an entire team
  4. 0:05. This means that if someone doing your
  5. 0:08job has learned it, they'll finish the
  6. 0:10same job much faster than you. If you're
  7. 0:12looking for work, that person is your
  8. 0:13potential candidate. If you're looking
  9. 0:15for clients, that person will be your
  10. 0:16competitor. So Cloud Code isn't
  11. 0:18something you'd rather learn; it's
  12. 0:20something you*must*learn. More and more
  13. 0:22people are learning it every month. The
  14. 0:24longer you wait, the more you'll be
  15. 0:25caught up from behind. You need to get
  16. 0:27this done before 2027. In this video,
  17. 0:29you'll learn Cloud Code from scratch,
  18. 0:31completely and thoroughly. We'll start
  19. 0:34with the setup.
  20. 0:35From Cloud MD to Skills, from MCPs to
  21. 0:37Sub Agents, you'll see everything on a
  22. 0:40real screen, explaining what it does. By
  23. 0:42the end of the video, you'll be someone
  24. 0:44who not only asks Cloud questions but
  25. 0:46also gets things done by it. Even if
  26. 0:48you've never used it before, that's okay
  27. 0:49. We're starting from the very beginning
  28. 0:51here. Let's clarify something first. You
  29. 0:53already know Cloud. You chat with it,
  30. 0:56you ask questions, and it answers you.
  31. 0:58Cloud Code is just like Cloud itself.
  32. 1:00The only difference is this: it runs
  33. 1:02inside a folder on your computer. Let me
  34. 1:04show you what that means with an example
  35. 1:05. When you want Cloud to do something in
  36. 1:08a chat, you upload a file. It writes you
  37. 1:10the answer. You then copy and paste that
  38. 1:12answer somewhere. Cloud Code doesn't
  39. 1:14have that copy-paste. For example, you
  40. 1:15say, "Gather all the notes in this
  41. 1:17folder into a single summary file."
  42. 1:19Cloud opens and reads those files,
  43. 1:21writes the summary, and puts the file in
  44. 1:23the folder itself. You can then look in
  45. 1:25the folder and find the file there.
  46. 1:26Don't think it's only for programmers
  47. 1:28just because it has "code" in its name.
  48. 1:32I use Cloud Code for most of the work on
  49. 1:34this channel's videos, from audio
  50. 1:36recording and cleaning to editing.
  51. 1:40I even use it for videos on my other
  52. 1:42social media channels and for my daily
  53. 1:44tasks.
  54. 1:46Everything done with folders and files
  55. 1:48is its job. You can use Cloud Code from
  56. 1:50different places. You know that black
  57. 1:52screen called Terminal? You can use it
  58. 1:54from there too. But there's no need to
  59. 1:56confuse yourself by using it when you're
  60. 1:57starting from scratch.
  61. 1:59Its own desktop is designed quite nicely
  62. 2:01and understandably.
  63. 2:03In this video, we'll be using the
  64. 2:05desktop application. Let's move on to
  65. 2:07installing the application. Installation
  66. 2:10takes only a few minutes, but before
  67. 2:12that, I need to mention something. Cloud
  68. 2:14Code isn't available in the free plan.
  69. 2:16You need to upgrade to a paid plan.
  70. 2:18There are three: Pro, Max, and Ultra.
  71. 2:20Pro is $ 20 per month. Max gives you
  72. 2:23five times the usage of Pro, and its
  73. 2:25price is also five times, so it's $ 100
  74. 2:27. Ultra is $ 200. This plan is for those
  75. 2:29who use the most and work very
  76. 2:31intensively. Prices may change over time
  77. 2:33. You can find the current prices on
  78. 2:35their official website. So, which one
  79. 2:37should you get? My recommendation is to
  80. 2:39start with Pro. Pro is perfectly
  81. 2:41sufficient for daily tasks. Antropin
  82. 2:43itself states on its website that Pro is
  83. 2:45for daily tasks and Max is for those who
  84. 2:47work all day. Test it with Pro first. Do
  85. 2:49some real work. If Pro's limit is enough
  86. 2:51for you, you can stay with Pro. You
  87. 2:53don't need to pay extra or upgrade to a
  88. 2:55higher plan.
  89. 2:57But if it's not enough, meaning you
  90. 2:58constantly hit the limit and have to
  91. 3:00wait while doing your work, then you can
  92. 3:02upgrade to Max.
  93. 3:04We'll see what the limit is and how it's
  94. 3:06reached in a later section.
  95. 3:08Even if you choose the wrong plan, you
  96. 3:10don't lose anything. When you upgrade to
  97. 3:11a higher plan, the remaining time from
  98. 3:13the old plan is deducted from the new
  99. 3:15one. You can cancel anytime. So, get a
  100. 3:17monthly plan for the first month, you
  101. 3:18can think about annual payments later.
  102. 3:20After you get your plan, you download
  103. 3:22the Cloud desktop application from the
  104. 3:24address on the screen, install it, and
  105. 3:26log in to your Cloud account. You don't
  106. 3:27need to install anything else. Cloud
  107. 3:29Code is already included in the
  108. 3:31application. When you open the
  109. 3:32application, you see two small icons in
  110. 3:34the upper left. The one on the left is
  111. 3:36the chat section, which is the Cloud you
  112. 3:38know, and the icon next to it with small
  113. 3:40separators is the code section. This is
  114. 3:41the Cloud that works with folders. So we
  115. 3:43click on it. On the left side, there are
  116. 3:45buttons like "new" and "search". Below
  117. 3:47that, there is a list. We call each
  118. 3:49conversation you have in Cloud Code a
  119. 3:51session. That list is a list of old
  120. 3:53sessions. Since you're just starting out
  121. 3:54, it will be empty. You don't need to
  122. 3:56look at it for now. I'll get there when
  123. 3:58you need to. As you can see, the setup
  124. 4:00is that simple.
  125. 4:02Now let's slowly dive into it.
  126. 4:04Let's go to the box at the bottom of the
  127. 4:06screen. That's where the real work is
  128. 4:07done.
  129. 4:08Now we've opened the code tab. We saw
  130. 4:10the box at the bottom of the screen.
  131. 4:12Inside, it says "Define a task" or "Ask
  132. 4:15a question." We call this the prompt box
  133. 4:17. It's like the place where you type
  134. 4:19messages in GPT, Gemini, or regular
  135. 4:22Cloud chat.
  136. 4:23You type what you want Cloud to do here
  137. 4:25and press Enter to send it.
  138. 4:27But before you type the first request,
  139. 4:29you need to specify two things: Where
  140. 4:31will Cloud run? Which folder will it run
  141. 4:34in? Both are in the line right below the
  142. 4:36box. There are other things in that line
  143. 4:37too. Let me explain what they are now so
  144. 4:40that nothing you see on the screen is
  145. 4:41unclear. The place on the left that says
  146. 4:44"This computer" is where you choose
  147. 4:46where Cloud will run. Since the files
  148. 4:47are on your computer, we'll choose that
  149. 4:49. When you open the menu, there's also a
  150. 4:52cloud below it.
  151. 4:53Cloud runs on Antropi's own computer. So
  152. 4:56it can't see the folders on your
  153. 4:57computer. It's not useful for us right
  154. 4:59now. The folder next to it is Cloud's
  155. 5:02workspace. It reads files there. It puts
  156. 5:04new files there. You select a folder by
  157. 5:07clicking "Open folder." You can also see
  158. 5:09the ones you've used before in the list
  159. 5:10. I want you to pay attention to
  160. 5:12something here. Cloud can modify the
  161. 5:14files inside the folder. So don't do the
  162. 5:16first tests on the original files.
  163. 5:18Create a new folder on your desktop
  164. 5:20called "Cloud Test". Copy the files you
  165. 5:22want to test into it. Even if Cloud
  166. 5:24misunderstands something, or does
  167. 5:26something wrong, your original files
  168. 5:28won't be affected. Now, leave the "
  169. 5:30Working Tree" checkbox next to the
  170. 5:32folder unchecked. This is for projects
  171. 5:34using Git. If you check it, Cloud will
  172. 5:36create a separate copy of the folder so
  173. 5:38it doesn't interfere with the original
  174. 5:39files. But our folder doesn't use Git.
  175. 5:41We don't need that for now. On the right
  176. 5:43side of the line, there are three things
  177. 5:45. Let's start with the model. Which
  178. 5:47version of Cloud will be running,
  179. 5:49meaning
  180. 5:50how much effort that version, or model,
  181. 5:52will consider for that task, and finally
  182. 5:54, the permissions mode. It currently
  183. 5:56says automatic. You can choose there
  184. 5:58whether Cloud should ask you when
  185. 6:00modifying a file or running something.
  186. 6:02Each of the three has its own section.
  187. 6:04We'll explain them in detail in this
  188. 6:06video.
  189. 6:07Permissions mode will come right after
  190. 6:09the first one. Because you can only
  191. 6:11better understand what it does after
  192. 6:12Cloud has done something. There's also a
  193. 6:15plus sign at the end of the box.
  194. 6:16Clicking it opens a menu.
  195. 6:18When you click the "Add File or Photo"
  196. 6:21button, you can add a PDF, image, or any
  197. 6:23other file to the Cloud from here.
  198. 6:26Dragging and dropping the file into the
  199. 6:27box does the same thing. Slash commands
  200. 6:30–even if you just type "slash" into the
  201. 6:32box, the same list opens. It's a way to
  202. 6:35give the Cloud ready-made commands.
  203. 6:37We'll use these commands when managing
  204. 6:39the context shortly. Skills you've
  205. 6:41created yourself can also be called from
  206. 6:43this list.
  207. 6:44I'll show them all individually when the
  208. 6:46time comes. There's also the MCP (
  209. 6:48Multiple Compartments) section. It
  210. 6:50connects the Cloud to other applications
  211. 6:52. "Add Plugin" installs ready-made
  212. 6:55packages. I'll cover both in detail in
  213. 6:57the video. For now, you can think of it
  214. 6:59as just an introduction to see what they
  215. 7:01are. The "Add Folder" button adds a
  216. 7:04second folder to the workspace. We'll be
  217. 7:06working with a single folder right now,
  218. 7:08so we don't need it. The "Import GitHub
  219. 7:10Issue" button is for developers to bring
  220. 7:12up their work files from GitHub. That's
  221. 7:14what interests us. Okay, now we can move
  222. 7:16on to writing the first request.
  223. 7:18We need a task to see what Cloud does. I
  224. 7:20deliberately left some things mixed up
  225. 7:22in the folder. There are a few invoices
  226. 7:24, a few YouTube covers, and a few notes
  227. 7:27files.
  228. 7:28When I look at the invoices, they're
  229. 7:31named in a jumbled way, like "invoice
  230. 7:34a93f-221," and so on. It's impossible to
  231. 7:37tell which invoice belongs to what just
  232. 7:39by the name. Sometimes, files we
  233. 7:40download in real life look like this too
  234. 7:43. Now I'm telling Cloud to examine each
  235. 7:45file in this folder individually,
  236. 7:46understand what each one is, give it a
  237. 7:48meaningful name, and organize it into
  238. 7:50folders according to its subject. I sent
  239. 7:52this prompt. Cloud doesn't just guess by
  240. 7:54name. It opens the files. It reads the
  241. 7:56text inside the PDF, looks at the images
  242. 7:58and understands what they are. And you
  243. 8:00don't have to wait while Cloud is
  244. 8:01working. The box stays open; it can't be
  245. 8:03closed. For example, if you have an idea
  246. 8:05, you can write "folder names should be
  247. 8:07without Turkish characters" and send it
  248. 8:09. Cloud sees it in the next step and
  249. 8:11incorporates it into its work. If you
  250. 8:12want to stop it completely, you press
  251. 8:14the stop button and it stops immediately
  252. 8:16. Now the job is done. When you looked
  253. 8:18at the folder, you saw an electricity
  254. 8:20bill inside the invoice file. Cloud
  255. 8:22changed the name accordingly. An
  256. 8:24internet bill starting with A 93 F
  257. 8:26appeared. The cover files went to a
  258. 8:28separate folder. The invoices went to
  259. 8:30another separate folder. Did you notice
  260. 8:31something? I'm very curious. Cloud
  261. 8:33didn't ask you anything when moving or
  262. 8:35renaming the files. Because the
  263. 8:37permissions mode was automatic.
  264. 8:38Everyone's files are different, and not
  265. 8:40everyone wants to work with the same
  266. 8:41level of trust. So, should Cloud have
  267. 8:43asked or not? That's determined by the
  268. 8:45permissions mode.
  269. 8:47Now, what we call permissions mode is
  270. 8:49how much Cloud can do without asking you
  271. 8:51. There are five here. I'll open the
  272. 8:53menu and explain them one by one. We'll
  273. 8:55go from the mode that asks the most to
  274. 8:57the one that asks the least. I'll leave
  275. 8:59the plan for last because it works a
  276. 9:01little differently.
  277. 9:02Let's start with manual. Manual Cloud is
  278. 9:04a mode where it asks you a question
  279. 9:06every time before modifying a file or
  280. 9:08running a command. It doesn't ask to
  281. 9:10read files, it asks to modify them. This
  282. 9:13is the most secure mode. But it's also
  283. 9:15the slowest. What it asks appears on the
  284. 9:17screen like this. Here you see what will
  285. 9:19change. You can either reject or allow.
  286. 9:21I want you to pay attention to something
  287. 9:23. The "Always Allow" button isn't just
  288. 9:25for this file; the text in the box
  289. 9:27itself says so. If you press it, that
  290. 9:29session switches to "Accept Edits" mode
  291. 9:31and it no longer asks for file edits. So
  292. 9:34, if you want to stay in manual mode,
  293. 9:36you have to press "Allow" every time.
  294. 9:38There's also an "Accept Edits" mode.
  295. 9:40Cloud handles file edits without asking
  296. 9:42. Simple file operations like opening
  297. 9:44folders, moving files, and renaming them
  298. 9:46are handled the same way. But if it's
  299. 9:48going to run another command, it will
  300. 9:49ask again. If you trust your files and
  301. 9:51want to proceed quickly, you can choose
  302. 9:53this mode.
  303. 9:54Now we come to the mode we use most
  304. 9:57often, the one you'll use constantly:
  305. 9:59automatic mode.
  306. 10:00Cloud doesn't ask every time. That's
  307. 10:02what you saw in the first instance. But
  308. 10:04for tasks like running commands or
  309. 10:06sending something to the internet, a
  310. 10:07separate control system kicks in in the
  311. 10:09background. This control checks two
  312. 10:11things: Does what's going to be done
  313. 10:13match your request? And it's done with
  314. 10:15an instruction hidden within a page that
  315. 10:16Cloud reads. It checks these things. If
  316. 10:19it sees it as risky, it blocks it.
  317. 10:21Reading and editing files inside a
  318. 10:22folder doesn't even come under this
  319. 10:24control. Commands like moving files do
  320. 10:26come under control, but it bypasses them
  321. 10:28because they match what you want. That's
  322. 10:30why it didn't ask anything on the first
  323. 10:32task. Since the default is auto mode,
  324. 10:33the application opens automatically.
  325. 10:35There's also a "bypass permissions" mode
  326. 10:38here. Think of it like giving Cloud a AK
  327. 10:40-47. It can do everything without asking
  328. 10:42. This mode is grayed out in the menu.
  329. 10:44It says "Enable in Cloud code settings.
  330. 10:46" So you have to enable it yourself from
  331. 10:48the settings. Antropik also says to only
  332. 10:50use this in an isolated environment,
  333. 10:52away from your computer. So they don't
  334. 10:54recommend enabling it if you're just
  335. 10:55starting out. I don't recommend it
  336. 10:57either.
  337. 10:58Auto mode does everything nicely. We
  338. 11:00also use a security mode because it runs
  339. 11:02in the background. Now, there's one
  340. 11:04thing that doesn't go without asking,
  341. 11:05regardless of which mode you're in: a
  342. 11:07command to delete the computer's main
  343. 11:09folder or the entire disk. It requires
  344. 11:10this in every mode, including the "
  345. 11:12bypass permissions" mode. The
  346. 11:14documentation says so, but I think it's
  347. 11:16absurd to completely trust it just
  348. 11:18because it says so.
  349. 11:20So if you're doing a very important job
  350. 11:22or you don't completely trust it, choose
  351. 11:24modes that suit you.
  352. 11:27Don't just trust the documentation or
  353. 11:29other people's methods and do things
  354. 11:31that way.
  355. 11:33So, which option do we choose here? If
  356. 11:35you're new to Cloud or working with a
  357. 11:37folder containing important files, you
  358. 11:39can start with manual mode. You can see
  359. 11:41what you're doing for a few tasks. Then
  360. 11:43, once you understand it better and feel
  361. 11:45confident, you can switch to automatic
  362. 11:47mode. Now, let's talk about the planning
  363. 11:49mode. Here, the fifth planning mode is
  364. 11:51slightly different from the others.
  365. 11:53Cloud reads the files. If necessary, it
  366. 11:55runs inspection commands and writes a
  367. 11:57plan for you. It doesn't change any
  368. 11:59files until you approve it. Once the
  369. 12:01plan is ready, Cloud asks how you want
  370. 12:02to proceed. There are three options. You
  371. 12:04can approve the plan and start in
  372. 12:06automatic mode. You can approve the plan
  373. 12:08and choose to approve each change
  374. 12:09individually, or you can choose not to
  375. 12:11approve and continue with the planning.
  376. 12:13You write what you want to change.
  377. 12:15Claude updates the plan accordingly.
  378. 12:17Once you approve, the plan closes, and
  379. 12:19Claude starts writing. For example,
  380. 12:21let's say, like in the first task, you
  381. 12:24need to rename and organize hundreds of
  382. 12:26files into folders, and reversing this
  383. 12:28task is difficult and costly. In the
  384. 12:30plan, Claude writes, "I will first open
  385. 12:32these folders, and I will name the files
  386. 12:34like this." You correct it by looking at
  387. 12:36it and separating the invoices month by
  388. 12:38month. Then he implements it. Something
  389. 12:40else happens in the plan. While Claude
  390. 12:42is reviewing the files, he has his
  391. 12:44ready-made assistant, a sub-agent, do
  392. 12:46this research. This way, everything
  393. 12:48that's read doesn't enter your
  394. 12:49conversation. We'll see what sub-agents
  395. 12:51are in detail in the sub-agents section
  396. 12:53. So, do I use the plan? No. I haven't
  397. 12:56used the plan for a long time. Because
  398. 12:58the models have developed so much that
  399. 13:00when you clearly state what you want, it
  400. 13:02can find the way itself. I made a
  401. 13:03separate video about this on the channel
  402. 13:05. It should have a name like "Claude's
  403. 13:07code was old." I explain that even the
  404. 13:10person who created the plan mode doesn't
  405. 13:12use it. I wanted to mention it here
  406. 13:13because it's in the menu and recommended
  407. 13:15for complex tasks in the Antropin
  408. 13:17documentation, so you should know about
  409. 13:19it. If you can't describe what you want
  410. 13:21, or if it's a difficult task to reverse
  411. 13:23and you want to see Claude's approach
  412. 13:25before starting, this might be useful
  413. 13:27for you. Otherwise, I don't think it's
  414. 13:29really necessary. Now we've done the
  415. 13:31first task. We selected the permission
  416. 13:33modes, but what if Claude's work isn't
  417. 13:35as good as you want?
  418. 13:37Then we'll look at the right end of the
  419. 13:39line at the bottom of the box.
  420. 13:42Now, looking at the right end of the
  421. 13:44line below the box, we see two settings
  422. 13:46: Model and Effort. Model refers to
  423. 13:48which version of Claude will perform the
  424. 13:50task, i.e., which model will do it.
  425. 13:52Effort, on the other hand, indicates how
  426. 13:54much thought that version will require
  427. 13:56at each step. We briefly saw this at the
  428. 13:58beginning. Now let's go into detail.
  429. 14:00First, I click on Model. Claude isn't a
  430. 14:02single model. Antropin has several
  431. 14:04models simultaneously. They're all
  432. 14:05listed one after the other in the menu.
  433. 14:07Next to Opus, it says Default. This is
  434. 14:09the default model in both Pro and Max.
  435. 14:11So, if you haven't changed anything,
  436. 14:13you're currently working with Opus as
  437. 14:14the default. But will we use Opus for
  438. 14:16every task? Anthropic's own
  439. 14:18documentation categorizes the models as
  440. 14:20follows: Sonnet is for everyday tasks,
  441. 14:22handles most coding tasks well, and is
  442. 14:24cheaper than Opus. Opus is for complex,
  443. 14:26multi-step, and thought-provoking tasks
  444. 14:29. Haiku is fast and light, sufficient
  445. 14:31for simple tasks. Fable is for the most
  446. 14:33difficult and longest-running tasks.
  447. 14:35Here, "cheap" and "expensive" refer to
  448. 14:37how much of your subscription cost. The
  449. 14:40large model is spending more than its
  450. 14:41limit for the same task. We'll see what
  451. 14:43that limit is in detail in the next
  452. 14:45section. Here, it's enough for you to
  453. 14:47know that. For example, I mostly use
  454. 14:49Opus for my video editing. Sometimes I
  455. 14:51use Sonnet as well. And I already
  456. 14:53explained Fable in detail in a separate
  457. 14:55video. You can find it on the channel.
  458. 14:57It should be the previous video. There's
  459. 14:59a line at the bottom of the menu called
  460. 15:01"Fast Mode." It makes the Cloud respond
  461. 15:03faster, but it requires extra credits.
  462. 15:05So it's a separate, paid service outside
  463. 15:07of your subscription. You don't need
  464. 15:09that when you start.
  465. 15:10Now let's talk about effort. Most people
  466. 15:12don't bring this up, but I think it's as
  467. 15:14important as choosing a model. The same
  468. 15:16model thinks quickly and works
  469. 15:17efficiently in low effort mode, while in
  470. 15:19high effort mode it thinks more at each
  471. 15:21step. The on-screen explanation also
  472. 15:23says this. Higher effort provides more
  473. 15:26comprehensive responses, but it takes
  474. 15:28longer and consumes its limits faster.
  475. 15:30There are five levels here. Let me tell
  476. 15:32you which one to use when, in order,
  477. 15:34according to Anthropic's own
  478. 15:35documentation. Low effort is for fast
  479. 15:37tasks where you control every result
  480. 15:39yourself. It says to use it for small
  481. 15:41tasks like drafting an idea or renaming
  482. 15:44a file. Medium effort is the default,
  483. 15:46and it says "recommended" next to it.
  484. 15:50Most everyday tasks where you know what
  485. 15:52you need can be done here. For example,
  486. 15:54in the first task, naming and organizing
  487. 15:56invoices into folders, the effort was on
  488. 15:58medium, and it was perfectly sufficient
  489. 15:59. Now, it says to use high effort for
  490. 16:01tasks where control is important and the
  491. 16:03probability of errors is high. For
  492. 16:05example, finding and fixing an error in
  493. 16:07a working system. X high, or extra high
  494. 16:10effort, requires more deep thinking and
  495. 16:12therefore more usage. Max, or maximum
  496. 16:15effort, is for tasks where you want the
  497. 16:16Cloud to solve a difficult problem from
  498. 16:18start to finish without you doing
  499. 16:20anything. There's also a warning next to
  500. 16:22it: 3.5 times or more usage. So it
  501. 16:24seriously uses up your limit. The
  502. 16:26document also adds this for maximum
  503. 16:28effort: the gain here can decrease. It
  504. 16:30says the Cloud can think more than
  505. 16:32necessary.
  506. 16:33By "the gain here can decrease," it
  507. 16:35means that you might not see much
  508. 16:37difference between a task done on High
  509. 16:39Effort and one done on Max.
  510. 16:41In other words, both can produce similar
  511. 16:43results.
  512. 16:45But here, because it's at the maximum,
  513. 16:47you're spending more tokens and thinking
  514. 16:49more.
  515. 16:51That's why your earnings can vary.
  516. 16:53Because you did a similar job in High
  517. 16:55Effort, and you did it at the maximum as
  518. 16:57well. But in one, you used a higher
  519. 17:00limit. So don't open every job to the
  520. 17:02maximum. Try a lower level first.
  521. 17:04Anthropic has another observation from
  522. 17:06its own tests that might help you when
  523. 17:08choosing. In High Effort, Cloud tries
  524. 17:10more possibilities. It checks your job
  525. 17:12more before answering and makes more
  526. 17:14decisions itself.
  527. 17:16In Low Effort, it gives you an earlier
  528. 17:18starting point. So, if you want to guide
  529. 17:20it by looking at each step, choose Low;
  530. 17:22if you want to say "finish this," choose
  531. 17:25High Effort. Sometimes you don't want to
  532. 17:27change the Effort. You just want it to
  533. 17:29think more deeply on a single question.
  534. 17:31Then write Ultra Think anywhere in your
  535. 17:32message. Look, for example, when you
  536. 17:34write it, the word lights up in color.
  537. 17:36This means Cloud Code recognized the
  538. 17:38word. It's thinking more deeply on that
  539. 17:39message. Your effort setting remains the
  540. 17:41same. You also need to pay attention to
  541. 17:43something. Other expressions like "
  542. 17:44Think" or "think more" don't work like
  543. 17:46this. They're going through as ordinary
  544. 17:48text. The word will be Ultra Think. Now
  545. 17:50, at the very bottom of the Effort menu
  546. 17:52, there's another section called Ultra
  547. 17:54Code. This isn't actually level 6 of
  548. 17:56Effort; it's a separate setting. When
  549. 17:58enabled, Cloud plans a workflow instead
  550. 18:00of working alone on a comprehensive
  551. 18:02project. It divides the work into
  552. 18:04multiple sub-agents and runs them
  553. 18:05together. We'll see what a sub-agent is
  554. 18:07in detail in its own section. For now,
  555. 18:09you need to know this: because multiple
  556. 18:11agents are working simultaneously, your
  557. 18:13limit might be reached faster. If you're
  558. 18:16just starting out or doing simple, small
  559. 18:18tasks, enabling this here would be
  560. 18:20unnecessary and pointless. So, keeping
  561. 18:22it off is better. Now, if you noticed,
  562. 18:25we saw the same warning for both large
  563. 18:27models and high-effort projects. We said
  564. 18:29it will consume your limit faster. So,
  565. 18:31how is this limit measured? The answer
  566. 18:33is how Cloud actually reads the text.
  567. 18:36Cloud doesn't read the text word for
  568. 18:38word like we do. It first breaks down
  569. 18:40every piece of text it receives into
  570. 18:41small pieces. We call these pieces
  571. 18:43tokens. A word is sometimes a single
  572. 18:45piece, sometimes two or three pieces.
  573. 18:47Cloud reads it using these pieces. When
  574. 18:49writing a response, it also writes it
  575. 18:51piece by piece. So, how much is a piece
  576. 18:53? According to Antropi's own dictionary
  577. 18:55, a token is approximately 3.5 English
  578. 18:57characters, roughly 3-4 letters. This is
  579. 19:00for the average English. You can count
  580. 19:01how many tokens a Turkish sentence would
  581. 19:03take, as shown on the screen. So why do
  582. 19:05we need to know this? Because when
  583. 19:07working with the Cloud, two important
  584. 19:09things are measured in tokens. First,
  585. 19:10how much information the Cloud can keep
  586. 19:12track of in a session, and second, how
  587. 19:14much you've spent from your subscription
  588. 19:16. We'll get to both of these in turn,
  589. 19:18but first we need to understand this:
  590. 19:20every message sent to the Cloud counts
  591. 19:22as a token. So we need to know what's
  592. 19:24going to the Cloud. Let's go back to the
  593. 19:26first task. I wrote two lines in the box
  594. 19:28. I said, "Look inside the invoices,
  595. 19:29name them, organize them into folders,"
  596. 19:31etc. The first thing that went to the
  597. 19:32Cloud was these two lines, but the Cloud
  598. 19:34couldn't learn what was written inside
  599. 19:35the invoice from my message. So it
  600. 19:37opened the invoices one by one. When the
  601. 19:39Cloud opens a file, the text inside the
  602. 19:41file goes to it. So each invoice entered
  603. 19:43the Cloud as a token. Then it ran a
  604. 19:46command to open the folders. The result
  605. 19:48of the command also came back to it as
  606. 19:50text. Finally, it wrote back to me
  607. 19:52saying, "You did these things." Actually
  608. 19:54, that answer is also a token. Now, when
  609. 19:56you look at it, the part I wrote is the
  610. 19:58smallest. The majority of the text sent
  611. 19:59to the Cloud during that process is the
  612. 20:01invoices themselves. The thing that
  613. 20:03consumes the most tokens is usually not
  614. 20:04what you write, but what the Cloud reads
  615. 20:06. So what happens to this text, where
  616. 20:08does it go? Now, everything that goes to
  617. 20:10the Cloud in a session accumulates. Your
  618. 20:12messages, the files it reads, command
  619. 20:14outputs, its own responses—it all piles
  620. 20:17up. When the Cloud responds to you, it
  621. 20:19looks at all of this accumulation
  622. 20:20together. For example, if I ask you
  623. 20:22which folder you put the electricity
  624. 20:24bill in at the end of the first task, it
  625. 20:25can answer. Because the moment it put
  626. 20:27that bill there, it's still in this
  627. 20:29accumulation. We call this accumulation
  628. 20:31context. That is, context. There is an
  629. 20:33upper limit to the context. That limit
  630. 20:35is called the context window. That is,
  631. 20:38the maximum amount of tokens that the
  632. 20:40Cloud can keep in mind at the same time
  633. 20:41in a session. In the model we are
  634. 20:43currently using, this limit is 1 million
  635. 20:45tokens. The context window is not
  636. 20:47completely empty before you write
  637. 20:48anything. Cloud Code puts some things
  638. 20:50into it at the beginning of each session
  639. 20:52. Cloud Code has its own instructions on
  640. 20:54how to operate, a list of tools it can
  641. 20:56use (like reading files, writing files,
  642. 20:59running commands, etc.), a persistent
  643. 21:01rule file in the folder, descriptions of
  644. 21:03skills, and the names of the tools
  645. 21:05you've connected. We'll see some of
  646. 21:07these individually in later sections.
  647. 21:10For now, it's enough to know that they
  648. 21:12take up space here. Let's see this in a
  649. 21:14real sense. I've opened a new session. I
  650. 21:16haven't typed anything yet. I type slash
  651. 21:18context in the box. As you can see, a
  652. 21:20card appeared in the chat. At the top it
  653. 21:22says total. That is, how many tokens the
  654. 21:24window has, what percentage of 1 million
  655. 21:26it is. Below, we can see line by line
  656. 21:28how much space each thing takes up. The
  657. 21:30names are in English. Let's look at them
  658. 21:32one by one. What's written here as
  659. 21:34system prompt are Cloud Code's own
  660. 21:35instructions. System tools are Cloud's
  661. 21:38own tools. MCP tools are the tools
  662. 21:40you've connected. Skills are
  663. 21:42descriptions of the skills the OS gave
  664. 21:44to Cloud. Messages are what we discussed
  665. 21:46in this session.
  666. 21:48At the bottom there's free space. There
  667. 21:50are a few more lines. Don't worry about
  668. 21:52those for now. Now look at the messages
  669. 21:54line. Almost zero. Because we haven't
  670. 21:56talked about anything yet. But part of
  671. 21:57the window is already full. That part is
  672. 21:59held by the lines above, the things
  673. 22:01Cloud Code put in at the beginning of
  674. 22:02the session. Now I'm running the same
  675. 22:04command in the session where the first
  676. 22:06task finished. As you can see, the lines
  677. 22:08above are almost the same. What's
  678. 22:09changed are the messages. Everything we
  679. 22:11talked about in the first task, every
  680. 22:13invoice Cloud read, is here. This is
  681. 22:15also where the tokens we just counted
  682. 22:17went. You don't need to open this dump
  683. 22:19with a command every time. When you send
  684. 22:21a message, the session details panel
  685. 22:23opens on the right side of the screen.
  686. 22:24The context card there shows the same
  687. 22:26lines. The only difference is that the
  688. 22:28card isn't live. It refreshes when
  689. 22:30Cloud's response is finished. Now the
  690. 22:32window has 1 million. As we can see,
  691. 22:34most of it is empty. Why should it be a
  692. 22:36problem if it's not full? How are we
  693. 22:37going to fill all this, and what happens
  694. 22:39when it is full? There are two reasons
  695. 22:40here. First, as the context grows, the
  696. 22:42quality of Cloud's use of it decreases.
  697. 22:45This is called context rot.
  698. 22:47Antropin explains it in his own article
  699. 22:49like this: As the number of tokens in
  700. 22:51the context window increases, the
  701. 22:52model's ability to correctly remember
  702. 22:54the information in that context
  703. 22:55decreases. The article says this is seen
  704. 22:57in all models.
  705. 22:59Some are slower. So the information is
  706. 23:01in the window, but Cloud can't use it as
  707. 23:03well as before. This starts before the
  708. 23:05window is full. Even if the window has 1
  709. 23:07million entries, it doesn't change that
  710. 23:09. Secondly, the window eventually fills
  711. 23:11in a long session. When it's close to
  712. 23:13full, Cloud doesn't close the code
  713. 23:15session. It automatically summarizes the
  714. 23:17conversation. It replaces the old
  715. 23:19conversation with the summary and
  716. 23:20continues that way. The Antropin
  717. 23:22document warns about this. The summary
  718. 23:24keeps track of your requests and
  719. 23:26important parts, but the detailed
  720. 23:27instructions at the beginning of the
  721. 23:29conversation can sometimes get lost. The
  722. 23:31result of both is the same. In a long
  723. 23:33session, something I say at the
  724. 23:35beginning might not be applied later.
  725. 23:37Let me show you with an example. In my
  726. 23:38first project, I interrupted Cloud and
  727. 23:40said that folder names should also be
  728. 23:42without Turkish characters, remember?
  729. 23:44Let's say I'm working for hours in the
  730. 23:46same session. I'm also editing other
  731. 23:48files. The conversation gets long and is
  732. 23:50summarized at some point. Now look. For
  733. 23:52example, it created a folder called "
  734. 23:54Contracts" with the letter "Ş". The rule
  735. 23:56I mentioned at the beginning is gone
  736. 23:57along with the summary. That's why we
  737. 23:59don't leave the context to Cloud code
  738. 24:01itself. We manage it ourselves. There
  739. 24:03are two commands and one setting for
  740. 24:04this. The first is compact. You initiate
  741. 24:07the summary yourself, without waiting
  742. 24:08for the automatic summary. This is
  743. 24:10called compression. You start the
  744. 24:11compression yourself. You type "slash
  745. 24:14compact" in the box and start it that
  746. 24:16way.
  747. 24:17The nice thing about this is that you
  748. 24:19can specify what you want it to keep.
  749. 24:21The automatic summary guesses what's
  750. 24:23important, but here you tell it.
  751. 24:26You tell it to keep file name rules and
  752. 24:28the last changes we made. The summaries
  753. 24:30preserve these.
  754. 24:32So when should you use this, when should
  755. 24:34you run it?
  756. 24:35The document says to run it between two
  757. 24:37jobs. Run this command after you've
  758. 24:39finished one job and before starting a
  759. 24:42new, longer job. If you do it in the
  760. 24:43middle of a job, the details of that job
  761. 24:45are compressed along with the summary.
  762. 24:47There's also an automatic compression
  763. 24:49threshold setting in the context card.
  764. 24:52You can set when the automatic summary (
  765. 24:54compact) will activate here.
  766. 24:57The explanation on the screen says: A
  767. 24:59lower value consumes less of the usage
  768. 25:02limit. Cloud summarizes older messages
  769. 25:04earlier. So if you bring the
  770. 25:06summarization forward, your limit is
  771. 25:08used less, but Cloud loses the details
  772. 25:10of older messages sooner. You can leave
  773. 25:12it on automatic at the start. Now,
  774. 25:14regarding the second command, it's `
  775. 25:17clear`. This actually works a little
  776. 25:19differently than the other one. There's
  777. 25:20no summary here. It completely clears
  778. 25:22the context. So you start from scratch.
  779. 25:25When do we use this? Let's say you've
  780. 25:27finished a task and you're moving on to
  781. 25:28something unrelated. For example, you
  782. 25:30worked on the invoice folder this
  783. 25:31morning, right? This afternoon, let's
  784. 25:33say you're going to create a Reels
  785. 25:35script. The conversation from this
  786. 25:37morning and the Reels script make no
  787. 25:39sense.
  788. 25:40They have no connection, but as long as
  789. 25:43you stay in that context, it takes up
  790. 25:45space and contributes to the context rod
  791. 25:47we saw earlier–the decay.
  792. 25:51So what do we do?
  793. 25:53If we're going to move on to a new task
  794. 25:55, a task without a connection, we run
  795. 25:58the ` clear` command and start again in
  796. 26:00a clean session. When you type ` clear`
  797. 26:02, a line appears in the chat saying the
  798. 26:04context has been cleared. Now, there's
  799. 26:06another way to do this on the desktop.
  800. 26:08You know that "new" button in the upper
  801. 26:10left? This opens a separate session, and
  802. 26:12the old session remains in the list on
  803. 26:13the left with its name. If you want to
  804. 26:15return to the invoice task tomorrow, you
  805. 26:17can click on it from the list. You can
  806. 26:19use this if you might return to the old
  807. 26:20conversation. Now let me explain how I
  808. 26:22do it in my own work. When the context
  809. 26:25reaches around 40%, I stop. I tell the
  810. 26:27Cloud to write everything we've done so
  811. 26:29far, what decisions we've made, and
  812. 26:30what's next to a file. Then I open a new
  813. 26:32session, read that file, and continue
  814. 26:34from where I left off. This way, nothing
  815. 26:37gets lost in the summary. The Cloud also
  816. 26:39doesn't start the new task with a
  817. 26:41cluttered context. So far, we've seen
  818. 26:43the first side of the token. That is,
  819. 26:45the part that the Cloud keeps track of.
  820. 26:46Now let's move on to the second side.
  821. 26:48How much did it spend from its
  822. 26:49subscription? There's something most
  823. 26:51people don't know here. The Cloud
  824. 26:52doesn't actually remember anything
  825. 26:54between two messages. That accumulated
  826. 26:56data I just described doesn't stay in
  827. 26:58the Cloud itself. It stays in the Cloud
  828. 26:59code. Every time you send a message, the
  829. 27:01Cloud code sends the entire context back
  830. 27:03to the Cloud from scratch. Instructions
  831. 27:05, tool list, all messages up to that
  832. 27:08point, files read, etc. And finally,
  833. 27:10your newly sent message is added. If
  834. 27:12this were processed from scratch every
  835. 27:14time, it would be very slow, and the
  836. 27:16limit would be reached very quickly.
  837. 27:18That's what caching is for. If the
  838. 27:19processed portion from the previous
  839. 27:21message remains unchanged, Cloud doesn't
  840. 27:23reprocess it. It reads it from the cache
  841. 27:25. It only reprocesses the newly added
  842. 27:27portion from the beginning. The portion
  843. 27:29read from the cache is both faster and
  844. 27:31consumes less than its limit.
  845. 27:33But consuming less doesn't mean it
  846. 27:35doesn't consume at all. As long as the
  847. 27:37morning's invoice conversation remains
  848. 27:39in the context, it's resent with every
  849. 27:41report message in the afternoon and
  850. 27:43counted each time.
  851. 27:44This is why Clear is important for the
  852. 27:46limit as well. Now, what determines that
  853. 27:48the cache remains unchanged? It
  854. 27:50determines this by comparing it to the
  855. 27:52beginning. If the beginning of the sent
  856. 27:54context is exactly the same as the
  857. 27:55previous one, it's read from the cache.
  858. 27:57If something changes at the beginning,
  859. 27:59everything after that point is
  860. 28:00reprocessed from the beginning. The
  861. 28:02document says this is a one-time thing.
  862. 28:04That message arrives slowly and consumes
  863. 28:06more. Then the new version enters the
  864. 28:08cache, it says. Then you continue, it
  865. 28:10says. But if the conversation is long,
  866. 28:12that single message can be quite costly
  867. 28:14. So you need to know what's causing the
  868. 28:16cache to malfunction. Now, the first
  869. 28:18thing is changing the model. Each model
  870. 28:20has its own cache. If you switch from
  871. 28:22Opus to Sonnet in the middle of a
  872. 28:24conversation, even if the content is the
  873. 28:26same, the entire conversation is
  874. 28:27processed from scratch without caching.
  875. 28:29So, remember that there's a cost to
  876. 28:31changing the mode in a long conversation
  877. 28:33.
  878. 28:34If you want to continue with another
  879. 28:36mode, starting in a new session is
  880. 28:37always better and cheaper. But effort is
  881. 28:40different. In most modes, changing
  882. 28:42effort also corrupts the cache, but the
  883. 28:45documentation says that in Opus 5.5,
  884. 28:47Sonnet 5.5, Haiku 5.5, and Fable,
  885. 28:50changing effort doesn't corrupt the
  886. 28:52cache. So, you can increase effort in
  887. 28:54the middle of a session while working in
  888. 28:56these modes. There's no problem there.
  889. 28:58Now, secondly, there's fast mode. We saw
  890. 29:00it in the mode menu. Fast mode also has
  891. 29:02a separate cache. If you open and close
  892. 29:04it in the middle of a conversation, the
  893. 29:05conversation up to that point is
  894. 29:06processed again from the beginning. It
  895. 29:08says to open a document at the beginning
  896. 29:09of the session. Thirdly, there's taking
  897. 29:11a long break. The cache doesn't last
  898. 29:13forever. If you don't send anything for
  899. 29:15a while, the cache is cleared. In your
  900. 29:17subscription plan, this period is 1 hour
  901. 29:20. Each message restarts the timer. So,
  902. 29:22it stays as long as you're working. When
  903. 29:24you return after a break longer than an
  904. 29:26hour, the entire context is reprocessed
  905. 29:27from scratch in your first message.
  906. 29:29Therefore, that message arrives slowly
  907. 29:31and consumes more effort. Subsequent
  908. 29:32messages are also processed from the
  909. 29:34cache. Fourthly, compact mode. Because
  910. 29:36it changes the conversation in summary,
  911. 29:38the cache is inevitably re-processed.
  912. 29:40This isn't expensive while the cache is
  913. 29:42still active, but if you use compact
  914. 29:44mode after a break longer than an hour,
  915. 29:45when the cache has been cleared, the
  916. 29:47entire conversation has to be
  917. 29:48reprocessed from the beginning. The
  918. 29:50document states that this is when
  919. 29:52compact mode is most expensive. So, if
  920. 29:54you return from a long break, don't use
  921. 29:57compact mode immediately. Use compact
  922. 29:59mode while working between two tasks.
  923. 30:01The document also mentions other points
  924. 30:03. For example, when you paste a large
  925. 30:05number of screenshots, they are pushed
  926. 30:07to older ones, and everything from that
  927. 30:08point onwards is reprocessed. But as a
  928. 30:11beginner, these four points are the ones
  929. 30:13you'll encounter most often.
  930. 30:16In short, select the model and fast mode
  931. 30:18at the beginning of the session.
  932. 30:21Know that there's a cost if you need to
  933. 30:22change it in the middle of a task. You
  934. 30:25can easily adjust the effort. You need
  935. 30:27to use compact mode between two tasks.
  936. 30:29If you take a long break, don't return
  937. 30:31to the same conversation. Start with a
  938. 30:33new message in a new session. Now, let's
  939. 30:35talk about the limit. Many people
  940. 30:37confuse this with context. Context
  941. 30:39refers to how full the session is, while
  942. 30:41limit refers to how much of your
  943. 30:42subscription you've used. Context
  944. 30:44empties during the session, but the
  945. 30:46limit doesn't. It resets when its time
  946. 30:48comes. There's a small circle at the far
  947. 30:51right of the line below the box.
  948. 30:52Clicking on it shows you how much of
  949. 30:54your plan you've used. There's a 5-hour
  950. 30:56limit and a weekly limit for all models
  951. 30:58. Next to each, it says when it will
  952. 31:01reset. It doesn't matter if you're using
  953. 31:03the cloud code on your desktop, in the
  954. 31:05terminal, or in any plugin. The limit is
  955. 31:08the same for all. If you want to see
  956. 31:10more details, type "usage" in the box. A
  957. 31:12usage card appears in the chat. At the
  958. 31:15top, there are the same bars. If you
  959. 31:17used Fab, there's a separate weekly bar
  960. 31:19for that. Below that is the session
  961. 31:21section. You see a cost figure there.
  962. 31:23Don't worry about that.
  963. 31:25The documentation clearly states that
  964. 31:27this figure is included in your usage
  965. 31:29plan, not billed to you.
  966. 31:31Below that is a breakdown. What I
  967. 31:33described in the cache, you can see here
  968. 31:35with the actual figure. In my test, the
  969. 31:38input was 8 tokens, while 199,700 tokens
  970. 31:41were read from the cache. So, almost all
  971. 31:43of the context sent to the Cloud in that
  972. 31:46message was read from the cache. Only 8
  973. 31:48tokens were processed initially. At the
  974. 31:50bottom, there's a section showing what's
  975. 31:52consuming your usage limits. It shows
  976. 31:54you the percentage of your limit
  977. 31:56consumed by what.
  978. 31:57From the menu next to it, you can view
  979. 32:00this session in the last 24 hours or the
  980. 32:02last 7 days.
  981. 32:03If your limit is suddenly used up so
  982. 32:05quickly, this should be the first place
  983. 32:07you check.
  984. 32:09To summarize, what's consuming your
  985. 32:11limit is the context resent with each
  986. 32:13message.
  987. 32:15The larger the context, the more tokens
  988. 32:17per message.
  989. 32:19The cache makes this cheaper, but it
  990. 32:20doesn't eliminate it.
  991. 32:22With a large model and high effort, it
  992. 32:24goes even faster.
  993. 32:26That's why compact and clear aren't just
  994. 32:29important for the Cloud not to lose what
  995. 32:31it initially said, but also for the
  996. 32:33limit itself. Now, something is missing
  997. 32:35. The Turkish character rule was lost in
  998. 32:37the summary. We can say that compact
  999. 32:39should keep this, but do we have to
  1000. 32:41remind the Cloud every time? When you
  1001. 32:42open a new session, Cloud won't know
  1002. 32:44about that rule at all. The document's
  1003. 32:46recommendation is clear here. Don't
  1004. 32:48leave the rules that should be valid in
  1005. 32:50every session in the chat; write them in
  1006. 32:52a separate file. There's a special file
  1007. 32:54in the folder specifically for this
  1008. 32:55purpose.
  1009. 32:57This special file is called cloud.md.
  1010. 32:59It's a plain text file inside the folder
  1011. 33:01. Cloud reads this file at the beginning
  1012. 33:03of each session. So it knows what's
  1013. 33:05written there without you having to tell
  1014. 33:07it each time. Remember how the Turkish
  1015. 33:08character rule was lost in the summary
  1016. 33:10earlier? This doesn't happen with
  1017. 33:12cloud.md. Because even after Cloud
  1018. 33:14generates the summary, it reads the file
  1019. 33:16from disk again and adds it to the
  1020. 33:18session. Since the rule is in the file
  1021. 33:20and not in the chat, it doesn't get lost
  1022. 33:22along with the summary. So how do we
  1023. 33:24create it? The easiest way is to tell
  1024. 33:26Cloud. You write the rule and then tell
  1025. 33:28it to add it to cloud.md. Cloud creates
  1026. 33:30the file if it doesn't exist, or adds it
  1027. 33:32if it does. There's also a ready-made
  1028. 33:34command. You type ` init` in that box.
  1029. 33:37If cloud.md doesn't exist in the folder
  1030. 33:38, Cloud checks the folder and writes you
  1031. 33:40one to start with. We already have one;
  1032. 33:42we just created it. Then it doesn't
  1033. 33:44overwrite it. It examines the folder,
  1034. 33:46lists what it finds missing in the file
  1035. 33:48, and finally asks if it wants to add it
  1036. 33:50. If you say yes, it adds it; if you say
  1037. 33:52leave it for now, it doesn't touch the
  1038. 33:54file. So what are we going to do with
  1039. 33:56this? Antropi's documentation has a good
  1040. 33:58sentence. It says to think of the Cloud
  1041. 34:01md as the place where you write things
  1042. 34:03you'll explain repeatedly. It also tells
  1043. 34:05you when to add it.
  1044. 34:07If Cloud makes the same mistake a second
  1045. 34:09time, if you're writing the correction
  1046. 34:11you wrote in the previous session again
  1047. 34:13in this session, or if someone else
  1048. 34:15working in this folder will need the
  1049. 34:17same information, then add it.
  1050. 34:19When writing, pay attention to two
  1051. 34:21things. First, write the rule in a
  1052. 34:23verifiable way. If you write "name the
  1053. 34:25files properly," what will Cloud
  1054. 34:27understand from that? What does "
  1055. 34:28properly" even mean? But if you write "
  1056. 34:31don't use Turkish characters or spaces
  1057. 34:32in file names, separate words with
  1058. 34:34hyphens," then it's clear what it will
  1059. 34:36do. You can check if it did or not. The
  1060. 34:38documentation says exactly this. It says
  1061. 34:41to write clearly enough to be verifiable
  1062. 34:43. As for the second point, it recommends
  1063. 34:44keeping it short. The documentation
  1064. 34:46recommends under 200 lines. Why? Because
  1065. 34:49this file is entered into the context
  1066. 34:50from beginning to end in every session.
  1067. 34:52The longer the file gets, the more space
  1068. 34:54it takes up, and the less likely Cloud
  1069. 34:56is to follow the rules. Also, if you've
  1070. 34:58written two conflicting rules, Cloud can
  1071. 35:00randomly choose one of them. You can
  1072. 35:03occasionally open the file and delete or
  1073. 35:05update the old or new rules. But you
  1074. 35:07need to know this: Cloud MD isn't a list
  1075. 35:10of prohibitions. Cloud reads it and
  1076. 35:12tries to comply, but ultimately it's an
  1077. 35:14instruction.
  1078. 35:16If you want something to never happen,
  1079. 35:18there's another way.
  1080. 35:19We call that a hook.
  1081. 35:21We'll see that at the end. Now, you can
  1082. 35:23place Cloud MD in two places. If you put
  1083. 35:26it inside the project folder, it's only
  1084. 35:27readable when working in that folder. If
  1085. 35:29you put it in Cloud's own settings
  1086. 35:31folder on your computer, it's readable
  1087. 35:33in all projects. If you have a rule that
  1088. 35:35applies everywhere, like "always reply
  1089. 35:37without emojis," you'll put it in the
  1090. 35:39second place. Cloud also keeps its own
  1091. 35:41notes.
  1092. 35:42This is called auto-memory.
  1093. 35:45When you correct Cloud or tell it a
  1094. 35:47preference, if it thinks it will be
  1095. 35:49useful in the future, it makes a note of
  1096. 35:51it. When it does this, it writes in the
  1097. 35:53chat that it has been saved and that
  1098. 35:54there is a memory record. The difference
  1099. 35:56here is...You write Cloud MD yourself.
  1100. 35:59Cloud automatically writes the memory.
  1101. 36:01Both are loaded at the beginning of each
  1102. 36:02session. If you tell Cloud to remind you
  1103. 36:04of something, it writes it to its memory
  1104. 36:06. If you want a rule to be in Cloud MD,
  1105. 36:08you can explicitly tell Cloud MD to add
  1106. 36:11this rule. Now, after using Cloud MD for
  1107. 36:13a while, the first thing you'll probably
  1108. 36:15want to do is say, "This file is getting
  1109. 36:17long, let's shorten it a bit." But
  1110. 36:19sometimes Cloud deletes a rule you want
  1111. 36:21to keep while shortening it. So what do
  1112. 36:23you do then?
  1113. 36:24For example, I told Cloud to shorten
  1114. 36:26Cloud MD. It shortened it, but it also
  1115. 36:28deleted the Turkish character rule. I
  1116. 36:30didn't want that. There's no need to
  1117. 36:32panic here. Cloud Cop automatically
  1118. 36:34saves the current state of the files
  1119. 36:36before Cloud touches anything, every
  1120. 36:38time you send a message. They call this
  1121. 36:40a checkpoint. Like in games, for example
  1122. 36:43.
  1123. 36:43You can think of it as a save point.
  1124. 36:45Every message has a save point. When you
  1125. 36:47want to undo something, you return to
  1126. 36:49that point. So how do we undo it? There
  1127. 36:51are three ways. All three lead to the
  1128. 36:53same place. The first is that you type '
  1129. 36:55slash rewind' into the box. A list of
  1130. 36:57the messages you sent appears. Click on
  1131. 37:00the message you want to return to.
  1132. 37:02Secondly, when the box is empty, press
  1133. 37:05ESC twice. The same list appears again.
  1134. 37:08Thirdly, and the simplest way, hover
  1135. 37:10your mouse over your own message. A
  1136. 37:12rewind button appears below it.
  1137. 37:14Pressing it takes you back to that
  1138. 37:16message in the list. No matter which way
  1139. 37:18you go, the same window appears. "
  1140. 37:21Rewind to this message?" Your message is
  1141. 37:23displayed here. Below it, you see how
  1142. 37:25many files will be restored. In the
  1143. 37:27restore section, you choose what will be
  1144. 37:29restored. There are three options here.
  1145. 37:31The first is code and chat. This is the
  1146. 37:33default. Both the files and the
  1147. 37:35conversation are restored to their state
  1148. 37:37before that message. You can think of it
  1149. 37:39as if you never sent that message.
  1150. 37:41The second is chat only. Here, it
  1151. 37:43doesn't touch the files. It only
  1152. 37:44restores the conversation to that point
  1153. 37:46. The third is code only. Here, "code"
  1154. 37:48refers to your files, meaning the files
  1155. 37:50are restored to their original state,
  1156. 37:52but the conversation remains as it is.
  1157. 37:54This means Cloud still knows what went
  1158. 37:56wrong. I think the third option is more
  1159. 37:58useful most of the time. Why? Your file
  1160. 38:00is fixed, but Cloud remembers that you
  1161. 38:02deleted the rule. Immediately after, you
  1162. 38:04say "shorten it, but don't touch the
  1163. 38:05Turkish character rule." It doesn't make
  1164. 38:07the same mistake on the second attempt.
  1165. 38:09Also, something nice happens when you
  1166. 38:11rewind the chat. The message you
  1167. 38:12selected returns to the prompt box. So
  1168. 38:14you can correct your message and resend
  1169. 38:16it. And if you rewind to the wrong
  1170. 38:18message, you don't need to worry. A link
  1171. 38:20appears saying "undo rewind" where you
  1172. 38:23rewound. You can undo it from there. But
  1173. 38:25you need to know the limitations of this
  1174. 38:27feature. Because many people are
  1175. 38:29mistaken here. Firstly, the save point
  1176. 38:31only covers files that Cloud has
  1177. 38:33modified using its own editing tool. If
  1178. 38:35Cloud did something by running a command
  1179. 38:37, it doesn't revert. It's written at the
  1180. 38:39bottom of the window. In the first job,
  1181. 38:41they moved files. They changed their
  1182. 38:42names, you remember. Cloud does that
  1183. 38:44with a command. So I don't want you to
  1184. 38:46assume that those files will return to
  1185. 38:48their original location when you rewind
  1186. 38:50. Secondly, changes made by sub-agents
  1187. 38:52often don't go into these save points.
  1188. 38:55We'll see sub-agents in detail later.
  1189. 38:58Thirdly, if you manually modified a file
  1190. 39:01outside of Cloud, that's not tracked
  1191. 39:03either. It's clearly stated in the
  1192. 39:05documentation. This feature doesn't
  1193. 39:07replace a permanent backup for quick
  1194. 39:09rollback.
  1195. 39:10They mention it so you have a backup if
  1196. 39:12you're working with important files. By
  1197. 39:14the way, you'll notice something when
  1198. 39:16you undo and rewrite the message. To get
  1199. 39:18good results from the cloud, you need to
  1200. 39:20write the instructions long and clear.
  1201. 39:22But if this is a task that repeats every
  1202. 39:25week or every day, will you have to
  1203. 39:27rewrite that long instruction from
  1204. 39:29scratch every week, every day, every
  1205. 39:31time you do the task?
  1206. 39:33I have exactly that kind of task. When I
  1207. 39:35finish a video, I give it the text and
  1208. 39:37print the title, description, and tags.
  1209. 39:39I was explaining the same things every
  1210. 39:41time: how many titles I wanted, how long
  1211. 39:43they should be, what the first line of
  1212. 39:45the description should be, where the
  1213. 39:47result should be saved, and so on.
  1214. 39:49Skills are exactly for this. The
  1215. 39:51Antropin documentation also says this.
  1216. 39:53If you're pasting the same instruction,
  1217. 39:55the same checklist, or a step-by-step
  1218. 39:57task into the chat repeatedly, turn it
  1219. 39:59into a skill. A skill is actually a
  1220. 40:01folder. Inside, there's a file called
  1221. 40:03skill.md. The folder name is the name of
  1222. 40:05the skill. Let me show you, for example
  1223. 40:07. As you can see, the file has two parts
  1224. 40:09. At the top, between two lines, there's
  1225. 40:12a short description. It explains what
  1226. 40:14this skill does and when to use it.
  1227. 40:16Below that are step-by-step instructions
  1228. 40:18for Cloud. As you can see, it's all in
  1229. 40:21Turkish. No code or anything. So how do
  1230. 40:23we use this? There are two ways. First,
  1231. 40:26you type "slash" in the box. The skill
  1232. 40:29name appears in the list. You select it
  1233. 40:31. You add the text and send it.
  1234. 40:33Remember the "slash commands" line we
  1235. 40:35saw in the plus menu earlier?
  1236. 40:38Well, the skills are listed there. The
  1237. 40:41second way is that you don't need to run
  1238. 40:42any commands. You just say, "Write a
  1239. 40:44title and description for this video
  1240. 40:46text." Cloud checks your request. If it
  1241. 40:48matches the description at the top, it
  1242. 40:50unlocks the skill itself. That's why
  1243. 40:52that description line is so important.
  1244. 40:55You can write when to use it, using
  1245. 40:57words people actually use.
  1246. 40:59If Cloud isn't using the skill, that's
  1247. 41:01the first place you should look. So how
  1248. 41:03do we create this? The easiest way is to
  1249. 41:05have Cloud write it. Let's say you've
  1250. 41:07done a task from start to finish once in
  1251. 41:09Cloud. And you're happy with the results
  1252. 41:10. In the same conversation, you say, "
  1253. 41:12Turn the work we just did into a skill.
  1254. 41:14" Cloud writes the steps to a file. Then
  1255. 41:16you open it and read it. If there are
  1256. 41:18any missing parts, you correct them.
  1257. 41:20Cloud notices the changes in the skill
  1258. 41:22file without logging out.
  1259. 41:24So you don't need to restart. There's
  1260. 41:26also the application itself. In the left
  1261. 41:28sidebar, you see a section called "
  1262. 41:30Customization." When you open it, there
  1263. 41:32are three tabs at the top: Skills, MCPs
  1264. 41:35, and Plugins. For now, we're looking at
  1265. 41:37skills. We'll see the other two when
  1266. 41:39it's their turn.
  1267. 41:41In the "Yours" section, the skills you
  1268. 41:43created and those you unlocked from
  1269. 41:45Antropik are listed. You can disable a
  1270. 41:46skill you don't use without deleting it
  1271. 41:48. Next to the disabled one, it says "
  1272. 41:50Disabled." To create a new skill, press
  1273. 41:52the "Add" button in the upper right.
  1274. 41:54Four options appear. Three are for
  1275. 41:56creating a skill, and one is for loading
  1276. 41:58an existing skill. We'll get to loading
  1277. 42:00in a moment. Let's look at creating
  1278. 42:02first. The first is creating a skill. If
  1279. 42:04you know what you want and want to write
  1280. 42:06it yourself, you'll use this method. A
  1281. 42:08form opens here. At the top, you write
  1282. 42:10the name of the skill. Below, you write
  1283. 42:12a description, meaning what it does and
  1284. 42:14when to use it. This is the part we saw
  1285. 42:16at the top of the skill MD earlier.
  1286. 42:18Cloud opens the skill based on this
  1287. 42:19sentence. You also write step-by-step
  1288. 42:21instructions in the large area at the
  1289. 42:23bottom. If you want to include an
  1290. 42:25example file along with the instructions
  1291. 42:27, there's an "add file" option in the
  1292. 42:29upper right. You can select it from
  1293. 42:30there. When you're finished, you click "
  1294. 42:32create." Your skill is created. The
  1295. 42:34second option is "create with Cloud." If
  1296. 42:36you're not sure exactly what to write,
  1297. 42:38you can choose this. When you click it,
  1298. 42:40a new chat opens, and a pre-written
  1299. 42:42message appears in the box.
  1300. 42:45The message tells Cloud to create a
  1301. 42:46skill with you, and first, to ask you
  1302. 42:49what the skill should do.
  1303. 42:51You send this. Cloud asks you a question
  1304. 42:53. You answer as if you were having a
  1305. 42:55normal conversation. Then, he already
  1306. 42:57writes your skill. The third, and
  1307. 42:59perhaps the most interesting, option is
  1308. 43:01"record your screen." Here, you don't
  1309. 43:03actually explain it, you demonstrate it
  1310. 43:04. When you click it, a small window
  1311. 43:06appears. It says that your screen clicks
  1312. 43:08, typing, and voice will be recorded and
  1313. 43:10then sent to Cloud to be converted into
  1314. 43:12a usable skill. You select your
  1315. 43:14microphone from the microphone button in
  1316. 43:16the lower right. Since your voice also
  1317. 43:18goes to Cloud, you can verbally explain
  1318. 43:19what you're doing and why while you're
  1319. 43:21doing the task. When you're ready, you
  1320. 43:23press 'e' and do the task yourself from
  1321. 43:26start to finish once. Once the recording
  1322. 43:28is complete, Cloud converts it into a
  1323. 43:30skill. You need to pay attention to
  1324. 43:31something here. If you have any
  1325. 43:33confidential information or private
  1326. 43:35details, like entering a password,
  1327. 43:36during recording, don't show them. The
  1328. 43:38window already displays this warning.
  1329. 43:40Now, no matter how you do it, the skill
  1330. 43:42you add here is saved to your Cloud
  1331. 43:44account, not to a folder on your
  1332. 43:46computer. Therefore, no matter where you
  1333. 43:48open Cloud from, it can work anywhere
  1334. 43:49you log in with the same account. Now
  1335. 43:51let's talk about the difference with
  1336. 43:53Cloud MD. Because the two can sometimes
  1337. 43:54get confused. Cloud MD loads completely
  1338. 43:57in every session. Only the top
  1339. 43:59description of the skill loads. The
  1340. 44:01actual instructions only become
  1341. 44:02available when the skill is used.
  1342. 44:04Remember the line called "skills" in the
  1343. 44:07context card? These are the descriptions
  1344. 44:09there. So you can write a skill as long
  1345. 44:11as you want. It takes up almost no space
  1346. 44:13when you're not using it. The rule here
  1347. 44:15is: the rules that apply in every
  1348. 44:17session are the same for the skill side
  1349. 44:19if you're doing a task step-by-step with
  1350. 44:21Cloud MD. If a part of Cloud MD has
  1351. 44:23turned into a recipe saying "do this
  1352. 44:25before the rule, then do that," you can
  1353. 44:27move that part to the skill file. Also,
  1354. 44:29pay attention to something in a long
  1355. 44:31session. When the context is full and a
  1356. 44:33summary is extracted, the skills you
  1357. 44:35used are added back, but only the
  1358. 44:36beginning. Therefore, write the most
  1359. 44:38important rule at the top of the skill
  1360. 44:40file. You can also put the skill file in
  1361. 44:42two places, like Cloud MD. If you put it
  1362. 44:44in the project folder, it only works in
  1363. 44:46that project; if you put it in the Cloud
  1364. 44:48settings folder, it works in all
  1365. 44:49projects on this computer. The ones you
  1366. 44:51add from the customization section, as I
  1367. 44:53said before, remain in your account. You
  1368. 44:55also need to know this: you might not
  1369. 44:57want some skills to automatically open
  1370. 44:59on the cloud side. For example, if
  1371. 45:01there's a skill that sends something or
  1372. 45:03publishes something, you don't want the
  1373. 45:05cloud to publish it or send a message
  1374. 45:07for you by saying "the job seems ready.
  1375. 45:09" To prevent this, you add a single line
  1376. 45:12to the top of the skill file.
  1377. 45:14From then on, only you can start that
  1378. 45:16skill with slash commands.
  1379. 45:18You don't always have to write the skill
  1380. 45:20yourself. Skills are written according
  1381. 45:22to an open standard. So they also work
  1382. 45:25on other AI tools.
  1383. 45:26People share the skills they create
  1384. 45:28online. The easiest place to find
  1385. 45:30ready-made skills is again the
  1386. 45:31customization section. There, in the
  1387. 45:33explore section, you'll find skills
  1388. 45:35created by Antropin and its partners.
  1389. 45:37They are categorized. The most uploaded
  1390. 45:39ones are in a separate list. Each card
  1391. 45:41already shows who published it and how
  1392. 45:43many times it has been uploaded. You
  1393. 45:45click the plus sign next to the one you
  1394. 45:47like and you can easily add it. Now,
  1395. 45:49let's say I have a skill I downloaded
  1396. 45:51from the internet. What do we do to add
  1397. 45:53it? Of course, we'll look at the fourth
  1398. 45:56option. There you click on "upload skill
  1399. 45:58". A page opens. You drag and drop the
  1400. 46:00file to the area in the middle or click
  1401. 46:02and select it from your computer. It
  1402. 46:04already shows which files it accepts
  1403. 46:06below. It will either be a single dot MD
  1404. 46:09file with its name and description at
  1405. 46:11the top, or a ZIP file containing "
  1406. 46:13skill dot MD". Then you click "upload".
  1407. 46:16The application also runs a security
  1408. 46:18scan during installation. If you want,
  1409. 46:20you can also put the downloaded folder
  1410. 46:22as a file in the skill folder. You can
  1411. 46:24also have this done by the cloud.
  1412. 46:26Remember this: don't get complacent just
  1413. 46:28because there's a scan. You need to be
  1414. 46:30very careful here. Antropik's
  1415. 46:32documentation is very clear on this.
  1416. 46:34Only use skills from trusted sources.
  1417. 46:37That is, from sources you've written
  1418. 46:39yourself or from Antropik. Why? Because
  1419. 46:41a skill isn't just instructions. It can
  1420. 46:43also contain code that the cloud will
  1421. 46:45execute. So a malicious skill could do
  1422. 46:48something different than what's
  1423. 46:49described in the instructions. For
  1424. 46:51example, it could send your files
  1425. 46:52somewhere, and it does this on your
  1426. 46:54computer with your permissions.
  1427. 46:56Therefore, if you download a skill from
  1428. 46:58somewhere, open and read all the
  1429. 47:00accompanying files before installing it
  1430. 47:02. If you don't understand something, I
  1431. 47:04don't recommend installing it here.
  1432. 47:07Remember the tab next to the
  1433. 47:09customization section?
  1434. 47:12It teaches you how to do a task, but if
  1435. 47:14the necessary materials aren't on your
  1436. 47:16computer—for example, your invoices are
  1437. 47:19in Gmail, your files are in Drive—it
  1438. 47:21can't access them.
  1439. 47:23Everything it's provided so far was in
  1440. 47:25that folder on your computer. The cloud
  1441. 47:27only sees that folder. The invoice in
  1442. 47:29Gmail or the document in Drive doesn't
  1443. 47:31exist for it; it can't see them. So what
  1444. 47:33do you do in this situation? You go and
  1445. 47:35download the invoice, put it in the
  1446. 47:38folder, or open the email, copy the text
  1447. 47:40, paste it, and so on. So you're the one
  1448. 47:42transferring information from one device
  1449. 47:44to another. Entropy's documentation
  1450. 47:46describes this very moment. It says that
  1451. 47:48the moment you realize you're copying
  1452. 47:50information from another device to the
  1453. 47:52chat, it's time to connect that device.
  1454. 47:54So how do we connect here? There's
  1455. 47:56something called ` connect`. This is an
  1456. 47:59open standard, a common connection
  1457. 48:01method that everyone can use. If an
  1458. 48:03application has given support, it can
  1459. 48:04connect to that application. After
  1460. 48:06connecting, it reads the information
  1461. 48:08there itself. If you allow it, it can
  1462. 48:10even do work there. For example, in the
  1463. 48:13desktop application, it's called an
  1464. 48:15adapter. So it's the same thing here.
  1465. 48:17The setup is done with a single click.
  1466. 48:19Now you go to the customization section
  1467. 48:21on the left side. At the top is the
  1468. 48:23adapters, and next to it is ` discover`
  1469. 48:26. You type the name of the application
  1470. 48:28in the search box.
  1471. 48:29The result shows what it is under each
  1472. 48:31line.
  1473. 48:33Is it a plugin?
  1474. 48:35You click the plus sign next to it.
  1475. 48:38Some lines have a community tag. Those
  1476. 48:40were made by other companies, not the
  1477. 48:42device's own company. In most cases, you
  1478. 48:44just log in with that application's
  1479. 48:46account and give permission, that's all
  1480. 48:48. So the setup is very simple. If you
  1481. 48:51want to see what you've connected later
  1482. 48:53, you can go to the customization
  1483. 48:54section on the left side. The status of
  1484. 48:56each one is already written there. If
  1485. 48:58it's connected, it has a checkmark; if
  1486. 49:00it's not connected, it says "connect."
  1487. 49:02If you've logged out, you need to click
  1488. 49:04"reconnect."
  1489. 49:05You can also remove it from here.
  1490. 49:07There's also this: These MCPs are
  1491. 49:09registered to your Cloud account. So, if
  1492. 49:11you've connected something to the Cloud
  1493. 49:13chat side, it appears here as well. You
  1494. 49:15don't need to connect it again. Let's
  1495. 49:17start with the first example. Let's
  1496. 49:18start with Gmail. After connecting, I
  1497. 49:20write this: List the emails that arrived
  1498. 49:22this week and are awaiting a reply.
  1499. 49:24Prepare a short draft reply for each,
  1500. 49:26but don't send them. I want to emphasize
  1501. 49:28that I said don't send them.
  1502. 49:30Because the reply written by the Cloud
  1503. 49:32goes under your name.
  1504. 49:34So, initially, leave it in drafts.
  1505. 49:37Read it there and check it.
  1506. 49:39After checking it that way, send it.
  1507. 49:41Sent emails cannot be undone. That is,
  1508. 49:44the undo section's save points work for
  1509. 49:46files in your folder. They don't work
  1510. 49:48for emails that have already gone to
  1511. 49:50Gmail. I'll just mention that warning
  1512. 49:52here. Moving on to the second example,
  1513. 49:54the Drive example. Let's say your
  1514. 49:56invoices are in a folder in Drive. Here
  1515. 49:58I write this...Find the invoices added
  1516. 49:59this month in the invoices folder in
  1517. 50:01Drive, and download them to the invoices
  1518. 50:03folder on your computer. The first time
  1519. 50:05, it asks for your permission to
  1520. 50:07download files from Cloud Drive. Once
  1521. 50:08you grant permission, it downloads them
  1522. 50:10. There are no steps like downloading,
  1523. 50:12dragging, or renaming. You are no longer
  1524. 50:14the one carrying the information. MCPs
  1525. 50:16aren't just for Google applications. Let
  1526. 50:18me give two more examples. Both are
  1527. 50:20tools that open the internet to the
  1528. 50:21Cloud. The first is the File Crawl tool
  1529. 50:23. The Cloud can already read a webpage.
  1530. 50:26File Crawl takes this a step further. It
  1531. 50:28crawls the entire site page by page and
  1532. 50:30converts each page into plain text that
  1533. 50:33the Cloud can easily read. Not just a
  1534. 50:35single page, but the entire site. I use
  1535. 50:37it when learning a new tool, for example
  1536. 50:39. I say, "Pull all the pages from this
  1537. 50:41tool's help site, save them to a folder
  1538. 50:43, and then answer my questions from
  1539. 50:45there," and so on. Now the Cloud
  1540. 50:47responds not with what it remembers from
  1541. 50:49the internet, but with the current state
  1542. 50:51of the site.
  1543. 50:52So, there are other uses for File Crawl
  1544. 50:54, but I wanted to give one example here
  1545. 50:57without making it too long. Looking at
  1546. 50:59the second one, there's Zapier. Zapier
  1547. 51:01has thousands of ready-made data
  1548. 51:03collection tools created by others,
  1549. 51:04which are called actors. For example,
  1550. 51:06they can pull regularly generated data
  1551. 51:08from places like Instagram, YouTube, and
  1552. 51:10Google Maps. With MCP Cloud, you can
  1553. 51:12select and run these tools yourself. For
  1554. 51:15example, I write this: "Pull the last 30
  1555. 51:17reels of this Instagram account. Sort
  1556. 51:19them by views. Find the common point of
  1557. 51:22the top five most viewed reels and tell
  1558. 51:24me." As you can see, this can be done
  1559. 51:26with a simple command. But if you had to
  1560. 51:28do this manually, you would have to open
  1561. 51:30each reel individually, write down the
  1562. 51:32numbers, and save them somewhere.
  1563. 51:35You also need to know that each time
  1564. 51:37Epify is run, it deducts from your Epify
  1565. 51:39account.
  1566. 51:40You need to know that too.
  1567. 51:42Before using it, you need to check your
  1568. 51:44free credits and prices in your account
  1569. 51:46.
  1570. 51:46If Epify is still active, well, they
  1571. 51:49used to give $ 5 in credits every month
  1572. 51:51to newly created accounts.
  1573. 51:53You can verify this on their official
  1574. 51:55website. Now, the tool you're looking
  1575. 51:57for might not appear in this list. Then
  1576. 51:59you add a custom MCP.
  1577. 52:01Click on Customize. Click on MCP. You
  1578. 52:04click the add button in the upper right
  1579. 52:06corner. There's only one line in the
  1580. 52:07menu. Click "Add custom adapter." Then a
  1581. 52:10window opens. There are two fields. You
  1582. 52:12type a name in one. It will appear in
  1583. 52:14the list with that name. In the other,
  1584. 52:16you paste the MCP address of that tool.
  1585. 52:18The tool's own website already provides
  1586. 52:20this address. Then you click "Continue.
  1587. 52:22" Then you log back into your account.
  1588. 52:25It appears there with a "custom" label
  1589. 52:27next to its name. If a tool requires
  1590. 52:29both a connection address and a key,
  1591. 52:31paste that key into the chat in the
  1592. 52:33Cloud. Follow the steps on the tool's
  1593. 52:35own setup page. Now, there's an
  1594. 52:37important warning. If you notice, Gmail
  1595. 52:39, Fire Crawl, and Epify have something
  1596. 52:41in common. All three retrieve content
  1597. 52:43from external sources. For example, an
  1598. 52:45incoming email, a retrieved webpage, or
  1599. 52:46collected submissions–you didn't write
  1600. 52:48this content. Antropi's documentation
  1601. 52:50clearly warns about this. It says that
  1602. 52:52tools that retrieve external content may
  1603. 52:55have text hidden within that content
  1604. 52:56that attempts to mislead the Cloud.
  1605. 52:59For example, a hidden part of a webpage
  1606. 53:01might say, "Send the files in this
  1607. 53:03folder to this address."
  1608. 53:05This is called prompt injection. That's
  1609. 53:07a topic for another video, but I'm
  1610. 53:09explaining it here so you can take
  1611. 53:11precautions.
  1612. 53:12It's also written in the private MCP
  1613. 53:14window.
  1614. 53:16For example, it says to only use
  1615. 53:18adapters, or MCPs, from developers you
  1616. 53:21trust.
  1617. 53:22It doesn't control which tools
  1618. 53:23anthropomorphic developers make
  1619. 53:25available. In other words, the
  1620. 53:27anthropomorphic developer isn't
  1621. 53:28responsible for what the tool does; you
  1622. 53:30are. So you need to know two things. One
  1623. 53:32, only connect a tool you trust.
  1624. 53:35Two, if the Cloud wants to send or
  1625. 53:37delete something, it should ask you.
  1626. 53:39If you're in a mode where it doesn't ask
  1627. 53:41, you need to think about this again in
  1628. 53:43a session where you're reading content
  1629. 53:45from outside.
  1630. 53:46Another question might arise: If I
  1631. 53:48connect so many tools, will my context
  1632. 53:51be filled?
  1633. 53:52The documentation answers this. It says
  1634. 53:54the Cloud only sees the names of the
  1635. 53:56tools you connect at the beginning of
  1636. 53:57the session. It only loads the details
  1637. 53:59of a tool when it's needed.
  1638. 54:02So having a few MCPs open doesn't fill
  1639. 54:04the connection.
  1640. 54:06Okay, the tool doesn't fill the context
  1641. 54:08. But what about the content it brings?
  1642. 54:10Let's say we used Firecrawl to pull 40
  1643. 54:12pages of a website. As Cloud reads those
  1644. 54:15pages, they all fill your chat. They all
  1645. 54:17accumulate. Whereas all you needed was
  1646. 54:20just 3-5 lines of content from those
  1647. 54:22pages.
  1648. 54:23Now we're getting to the sub-agents. We
  1649. 54:25saw this in the context section. We said
  1650. 54:27that as the chat fills up, Cloud has a
  1651. 54:29harder time remembering the initial
  1652. 54:31things. We called this context rot,
  1653. 54:33remember? For example, let's say there's
  1654. 54:35a research project. This research
  1655. 54:37project does exactly that. Cloud opens
  1656. 54:39dozens of pages, reads them all. And all
  1657. 54:41of it goes into your context. In the end
  1658. 54:43, you get a summary, but everything
  1659. 54:45that's read stays in the chat. That's
  1660. 54:47why sub-agents exist.
  1661. 54:50Well, not exactly for that reason, but
  1662. 54:52they make things much easier here, and
  1663. 54:54I'll explain sub-agents using this
  1664. 54:56research example so you can understand
  1665. 54:58better.
  1666. 55:00Now, a sub-agent is a separate assistant
  1667. 55:02that Cloud delegates a task to. They
  1668. 55:05have their own context. They do all the
  1669. 55:07reading on their side and only bring you
  1670. 55:09the result.
  1671. 55:10Those readings don't enter your chat.
  1672. 55:12Actually, we've seen the sub-agent once
  1673. 55:15before. We just haven't given it a name
  1674. 55:16. For example, when Cloud code or files
  1675. 55:18are being reviewed in plan mode, it uses
  1676. 55:20its own pre-built sub-agent to perform
  1677. 55:22this review. That's why the plan exits
  1678. 55:24without filling up your chat. Cloud code
  1679. 55:26has several such pre-built sub-agents.
  1680. 55:28For example, a search agent that can
  1681. 55:30only read and not modify anything. Cloud
  1682. 55:32opens these itself when it needs to
  1683. 55:34research something. You don't do
  1684. 55:35anything.
  1685. 55:37The nice thing about sub-agents is that
  1686. 55:39you can have them perform independent
  1687. 55:41tasks simultaneously.
  1688. 55:43Let me give you an example. A very
  1689. 55:45simple example. Let's say Cloud code
  1690. 55:47researches the prices and usage limits
  1691. 55:49of cursors and Codex. Use a separate
  1692. 55:51sub-agent for each. Then I tell it to
  1693. 55:54compare them all in a single table. As
  1694. 55:55you can see, the three researches are
  1695. 55:57done directly without waiting for each
  1696. 55:59other. That is, they run simultaneously
  1697. 56:02. They can work in parallel. Each one
  1698. 56:04reads dozens of pages, but you will only
  1699. 56:06get three short results and a table in
  1700. 56:08your chat. In my test, when the three
  1701. 56:10were finished and the table appeared,
  1702. 56:12the context increased from 7%to 9%. If
  1703. 56:14you want to see what they are doing
  1704. 56:16while they are working, look at the
  1705. 56:17panel on the right. When sub-agents are
  1706. 56:19running, a section called "background
  1707. 56:21tasks" appears. Each sub-agent is a row
  1708. 56:23. Next to it is how long it has been
  1709. 56:25running and a stop button.
  1710. 56:27Clicking on a row shows what task Cloud
  1711. 56:29has assigned to that sub-agent and which
  1712. 56:31page it is currently reading. Now, if
  1713. 56:34you frequently assign a task to a
  1714. 56:36sub-agent, you can make it permanent.
  1715. 56:38Again, the easiest way is to tell Cloud
  1716. 56:40. For example, let's say you create a
  1717. 56:43sub-agent that checks for spelling
  1718. 56:45errors in your Twitter texts. It should
  1719. 56:47only read, not modify the file. Let's
  1720. 56:49say it uses the Haiku model. We saw the
  1721. 56:51skill section in the Skill section. This
  1722. 56:53is somewhat similar.
  1723. 56:55The difference here is that the skill
  1724. 56:57works within your chat. The sub-agent
  1725. 56:59works in a separate context and brings
  1726. 57:01you the result. Two lines in that file
  1727. 57:03are important. One is that because you
  1728. 57:05said the tools should only read, this
  1729. 57:07sub-agent cannot touch your file. The
  1730. 57:09second is the model. For a simple task
  1731. 57:11like spell checking, high-q is
  1732. 57:13sufficient and uses much less resources
  1733. 57:15. That's why we choose high-q. The
  1734. 57:17documentation lists this as one of the
  1735. 57:19benefits of sub-agents. It talks about
  1736. 57:21directing simple tasks to the model more
  1737. 57:23quickly and cheaply. Cloud selects this
  1738. 57:25sub-agent based on its description.
  1739. 57:28You can also call it by name.
  1740. 57:30For example, you say, "Check this text
  1741. 57:32with the spell check agent," and it's
  1742. 57:34called automatically. However, I should
  1743. 57:36mention something. If you created the
  1744. 57:38first sub-agent in this folder, you need
  1745. 57:41to open a new session. I tried it in the
  1746. 57:43same session, and Cloud couldn't find
  1747. 57:44the agent. The documentation already
  1748. 57:46states this. It says the first sub-agent
  1749. 57:48in a project is loaded with a new
  1750. 57:50session. Cloud also mentions that you
  1751. 57:52may need to restart the session when
  1752. 57:54creating it.
  1753. 57:55Now, let's talk about the cost.
  1754. 57:57Everything has a cost, and this has one
  1755. 57:59too. What you need to know is this: Each
  1756. 58:01sub-agent sends its own requests, and
  1757. 58:03these deduct from your usage limit. The
  1758. 58:06documentation warns about this. If you
  1759. 58:08open many sub-agencies, and each one
  1760. 58:09brings long results, you'll not only use
  1761. 58:11up your limit but also fill up your chat
  1762. 58:13with those results. So, it says don't
  1763. 58:15open sub-agencies for every task.
  1764. 58:17Sub-agencies aren't necessary. For
  1765. 58:19example, if a side task is going to fill
  1766. 58:21your chat with a lot of reading material
  1767. 58:23, or if independent tasks are being done
  1768. 58:25simultaneously, then use a sub-agent.
  1769. 58:27How can we give an example here? Let's
  1770. 58:28go with an example from the research.
  1771. 58:30Let's say we're going to do research on
  1772. 58:323-4 different topics. Calling in a
  1773. 58:34sub-agent makes sense here. Why?
  1774. 58:36Because the context of the main chat
  1775. 58:38gets clogged while researching, and
  1776. 58:41researching in 4-5 different categories
  1777. 58:43takes a very long time.
  1778. 58:45If we open 4-5 different accounts for
  1779. 58:474-5 different categories, our context
  1780. 58:50remains uncluttered. We get what we want
  1781. 58:52cleanly, and we finish our work faster.
  1782. 58:55But as the documentation says, these
  1783. 58:57sub-agents consume our time limit while
  1784. 58:59working. We need to take this into
  1785. 59:01account when doing our work.
  1786. 59:03Now, we've told the sub-agent to only
  1787. 59:05read, and it can't touch the file.
  1788. 59:08But how do we do this for the main one?
  1789. 59:10How can we guarantee it will never do
  1790. 59:12something?
  1791. 59:14That's where hooks come in.
  1792. 59:17Remember Cloud's D section? We wrote the
  1793. 59:19rules there. Let's say you wrote "never
  1794. 59:21touch files in the archive folder."
  1795. 59:24Cloud usually complies with this. But
  1796. 59:26the Antropin documentation clearly
  1797. 59:28states that Cloud sees Cloud's D as
  1798. 59:30context, not a mandatory rule. It's more
  1799. 59:33like a request. In a long session with
  1800. 59:35complex tasks, this request can
  1801. 59:37sometimes be overlooked. But when it
  1802. 59:39comes to hooks, a hook isn't a request.
  1803. 59:41A hook is a small command that Cloud
  1804. 59:43code automatically executes at specific
  1805. 59:45times. For example, let's say it runs
  1806. 59:48just before Cloud edits a file or when
  1807. 59:50Cloud finishes its work. So the decision
  1808. 59:53isn't Cloud's. When that moment comes,
  1809. 59:55the hook runs automatically every time.
  1810. 59:57The documentation explains it this way.
  1811. 59:59It says if you want something to happen
  1812. 1:00:01every time without the model having to
  1813. 1:00:03choose, use hooks. Now, I'm writing this
  1814. 1:00:05: Write a hook that prevents editing
  1815. 1:00:08files in the archive folder. Cloud adds
  1816. 1:00:10the hook to the configuration file. You
  1817. 1:00:12don't need to write anything. When it's
  1818. 1:00:15finished, it also says: This hook stops
  1819. 1:00:17Cloud's file editing tool. But Cloud can
  1820. 1:00:19still write to a file with a command.
  1821. 1:00:21That path is still open, it says. It
  1822. 1:00:23also says, "Tell me if you want me to
  1823. 1:00:25close that." I want you to keep this in
  1824. 1:00:26mind. We'll come back to this in a
  1825. 1:00:28moment. Let's try this now. I'm opening
  1826. 1:00:30a new session so that the hook is
  1827. 1:00:32definitely loaded here. I'm asking Cloud
  1828. 1:00:34to add a line to a file in the archive.
  1829. 1:00:36As you can see, the hook stopped the
  1830. 1:00:38editing before Cloud even finished. It
  1831. 1:00:40also told Cloud why it was stopped.
  1832. 1:00:41Cloud suggested two options, but you
  1833. 1:00:43need to pay attention to one sentence.
  1834. 1:00:45It says it didn't bypass the hook with a
  1835. 1:00:47command. The reason is this: the hook
  1836. 1:00:49looks at the tool Cloud uses. Our hook,
  1837. 1:00:52however, captures Cloud's file editing
  1838. 1:00:54tool. If Cloud tries to write to the
  1839. 1:00:56same file with a command, this hook
  1840. 1:00:58won't see it. The Antropin documentation
  1841. 1:01:00explains it this way. The protected file
  1842. 1:01:02instance is for the editing tool. So, if
  1843. 1:01:04you want to check commands as well, the
  1844. 1:01:06hook needs to look at the commands too.
  1845. 1:01:09Cloud mentioned this during installation
  1846. 1:01:11. So, when you install a hook, you need
  1847. 1:01:13to know what it's capturing. Ours
  1848. 1:01:15definitively stops editing.
  1849. 1:01:18It works even if the rule in CloudMD is
  1850. 1:01:20forgotten.
  1851. 1:01:21So, what do you do if you want to see
  1852. 1:01:23which hooks are installed? You ask Cloud
  1853. 1:01:26. It looks at the configuration file and
  1854. 1:01:28tells you. There's a command that lists
  1855. 1:01:30hooks, but it only works in the terminal
  1856. 1:01:32. If you type it on the desktop, a
  1857. 1:01:34warning pops up saying this. Now, what
  1858. 1:01:37else can be done with hooks?
  1859. 1:01:39If you ask if hooks only prevent file
  1860. 1:01:42editing, the answer is no.
  1861. 1:01:44Hooks have two main functions. One
  1862. 1:01:46approach is to prevent something from
  1863. 1:01:48happening, like the one before, and the
  1864. 1:01:49other is to ensure something happens
  1865. 1:01:51every time. For example, Cloud could
  1866. 1:01:53automatically format a file after
  1867. 1:01:55editing it, or Cloud could send you a
  1868. 1:01:57message when it finishes a task.
  1869. 1:02:00I use this, for instance. Cloud
  1870. 1:02:02automatically sends a notification when
  1871. 1:02:04it's finished. But if you're just
  1872. 1:02:06starting out, that protection example
  1873. 1:02:08might be the most useful. Because a hook
  1874. 1:02:10executes a real command, if it makes a
  1875. 1:02:12mistake, it will make the same mistake
  1876. 1:02:14every time. Therefore, you need to
  1877. 1:02:16install few and clear hooks. In other
  1878. 1:02:18words, you shouldn't overload the hooks
  1879. 1:02:19. The rule here is simple: if you want
  1880. 1:02:22it to happen generally, use CloudMD; if
  1881. 1:02:24you want it to never happen or always
  1882. 1:02:26happen, use a hook. So far, we've
  1883. 1:02:29installed four things: Skill, MCP,
  1884. 1:02:31sub-agent, and hook. Let's say someone
  1885. 1:02:34has configured all of these for their
  1886. 1:02:36own work, connected them, and it's
  1887. 1:02:38working.
  1888. 1:02:39You want the same thing. Do you have to
  1889. 1:02:41copy them all one by one and put them in
  1890. 1:02:43the right places? Now, that's where
  1891. 1:02:45plugins come in. That's exactly what a
  1892. 1:02:47plugin is.
  1893. 1:02:49It's something that combines several
  1894. 1:02:52skills, sub-agents, hooks, and MCPs into
  1895. 1:02:54a single package. You install it, and
  1896. 1:02:56everything comes with it at once. When
  1897. 1:02:58the package creator updates it, you
  1898. 1:03:00update too. But first, let me say this.
  1899. 1:03:02It's clearly stated in the documentation
  1900. 1:03:04. Skills, sub-agents, hooks, and MCPs
  1901. 1:03:06work independently. It says you don't
  1902. 1:03:09need a plugin.
  1903. 1:03:10A plugin is only useful when you want
  1904. 1:03:12several things together or when you want
  1905. 1:03:14to use a pre-installed setup.
  1906. 1:03:16Now, let's look at how to install it. We
  1907. 1:03:18go to the customization section on the
  1908. 1:03:20left side. We go to Plugins. From there
  1909. 1:03:23, we click on Explore. At the top,
  1910. 1:03:25there's a plugin featured by Antropi.
  1911. 1:03:28Below that, new plugins are listed in
  1912. 1:03:30cards. Each card shows who published it
  1913. 1:03:32. For example, I'm opening Antropi's
  1914. 1:03:34Design plugin. The plugin page has four
  1915. 1:03:37tabs. The overview explains what it does
  1916. 1:03:39.
  1917. 1:03:40The Skills tab shows the skills included
  1918. 1:03:42. Design, for example, has seven.
  1919. 1:03:45You can call them all with a slash, or
  1920. 1:03:48Cloud, which it already uses.
  1921. 1:03:50The MCPs tab shows the tools it can
  1922. 1:03:53connect to. For example, there are nine
  1923. 1:03:55tools like Slack, Figma, Notion, and
  1924. 1:03:58Gmail. There's something you need to
  1925. 1:03:59know here. The plugin doesn't connect
  1926. 1:04:01these tools for you. Just because you
  1927. 1:04:03installed the plugin doesn't
  1928. 1:04:04automatically connect your Slack account
  1929. 1:04:06. Your Figma, Notion, and Gmail accounts
  1930. 1:04:09don't connect automatically. You need to
  1931. 1:04:11connect each one separately for Cloud to
  1932. 1:04:13be able to use them, just like we do
  1933. 1:04:16with MCPs.
  1934. 1:04:17The content tab actually contains all
  1935. 1:04:19the plugin files. For example, the
  1936. 1:04:21skill.md files we see in the skills
  1937. 1:04:23section are stored here in folders. You
  1938. 1:04:25can open and read them. To install, you
  1939. 1:04:27press the plus sign. In the corner, it
  1940. 1:04:29says "Design plugin installed and ready
  1941. 1:04:31to use." There's a switch in the upper
  1942. 1:04:33right corner of the page. You can turn
  1943. 1:04:34the plugin on and off this way. Like
  1944. 1:04:36everything else, plugins have costs.
  1945. 1:04:39First, context. The name and description
  1946. 1:04:41of the skills and sub-agents within each
  1947. 1:04:43open plugin are displayed in front of
  1948. 1:04:45Cloud in every message. This is true
  1949. 1:04:47even if you don't use them. We already
  1950. 1:04:49saw this in the skills section. The
  1951. 1:04:50skill description was always loaded.
  1952. 1:04:52When done individually, it was less. The
  1953. 1:04:54description of our spell check agent
  1954. 1:04:57held 118 tokens in the context card. But
  1955. 1:04:59when dozens are added with plugins, this
  1956. 1:05:01accumulates.
  1957. 1:05:03So you can disable the plugins you don't
  1958. 1:05:05use. You don't need to uninstall them.
  1959. 1:05:07Just disabling them is enough. Now,
  1960. 1:05:09regarding the second point, trust. The
  1961. 1:05:11trust part is very important. I want you
  1962. 1:05:13to remember the warning in the Skill
  1963. 1:05:15section. This is a slightly bigger
  1964. 1:05:16version of that. Because the plugin can
  1965. 1:05:18also contain Hooks and MCPs. This means
  1966. 1:05:20that the moment you install it, it can
  1967. 1:05:22run its own commands on your computer,
  1968. 1:05:24with your permissions. The documentation
  1969. 1:05:26clearly states this. It says it doesn't
  1970. 1:05:28matter which market it comes from. So
  1971. 1:05:29read what it contains before installing
  1972. 1:05:31. The application also shows this. On
  1973. 1:05:33its design page, it says above the MCPs
  1974. 1:05:36that it can send data to other services
  1975. 1:05:38. You can also check its files in the
  1976. 1:05:40content tab. If you don't know who made
  1977. 1:05:42it, I don't recommend installing it here
  1978. 1:05:44.
  1979. 1:05:45So, we've covered everything from the
  1980. 1:05:47installation to the plugins. We've seen
  1981. 1:05:49everything you need to open the cloud
  1982. 1:05:51code and start using it in your own
  1983. 1:05:53business. If you get stuck somewhere,
  1984. 1:05:54you can go back to the relevant section
  1985. 1:05:56and watch it again. The sections will be
  1986. 1:05:58timestamped in the description. I will
  1987. 1:06:00also add all the sources mentioned in
  1988. 1:06:02the video to the description. They are
  1989. 1:06:03all from Antropi's own official
  1990. 1:06:05documentation. But you've seen each of
  1991. 1:06:07them separately now.
  1992. 1:06:09The main point is to use them all where
  1993. 1:06:11they're needed within a project. That's
  1994. 1:06:13exactly what we're doing in the next
  1995. 1:06:14video. You know those videos you've been
  1996. 1:06:16seeing everywhere on Instagram lately,
  1997. 1:06:18the ones generated with Cloud Code and
  1998. 1:06:20getting millions of views? These
  1999. 1:06:21accounts are growing so fast. I've even
  2000. 1:06:23seen one reach 300,000 followers in a
  2001. 1:06:26month. I thought about opening such an
  2002. 1:06:27account and I'll set everything up with
  2003. 1:06:29Cloud Code. We'll start with an empty
  2004. 1:06:31folder. I've already chosen the niche
  2005. 1:06:33beforehand. I'll show it in the video.
  2006. 1:06:35You can choose something else according
  2007. 1:06:37to your interests. Then we'll write the
  2008. 1:06:39video script. We'll have the voiceover
  2009. 1:06:41and editing done, and we'll share it as
  2010. 1:06:43a real reel.
  2011. 1:06:45You'll see the Cloud MDI, skills, and
  2012. 1:06:47sub-agents you see in this video,
  2013. 1:06:49wherever they're needed, in a real
  2014. 1:06:51project.
  2015. 1:06:52You can subscribe to the channel so you
  2016. 1:06:54don't miss the video when it comes out.
  2017. 1:06:56If the video was helpful, you can like
  2018. 1:06:58it. If you're watching from a mobile
  2019. 1:07:00device, I'd be very happy if you hyped
  2020. 1:07:02the video. And after the account is
  2021. 1:07:04opened, I'll be sharing step-by-step how
  2022. 1:07:06it's going, what happened and what
  2023. 1:07:08didn't in the first few days, on Twitter
  2024. 1:07:10. I'm also talking about tools that will
  2025. 1:07:12be useful on Instagram. I'm showing you
  2026. 1:07:14how to use them. The links to both will
  2027. 1:07:17be in the description. Finally, if you'd
  2028. 1:07:19like to see a separate, more detailed
  2029. 1:07:20video on any of the topics in this video
  2030. 1:07:22, you can write it in the comments. I'll
  2031. 1:07:24start with the most requested one.
  2032. 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.