YouTube2Text

【最新】Anthropic公式が推奨するClaude Codeの神活用法TOP12!初心者でもスグに使えるClaude Code活用法【AIで業務自動化】 — Transcript

by むっちょ【Claude Code攻略】 · 5,924 words · 900 segments · language en · Watch on YouTube

Full transcript

  1. 0:00When asking ClaudeCode for a difficult
  2. 0:02task, just look up the method, build it
  3. 0:04, and use this prompt; it increases the
  4. 0:06accuracy of Anthropic-level responses by
  5. 0:08over five times. For instructions to
  6. 0:10ClaudeCode, just keep these three things
  7. 0:12in mind: purpose, guardrails, and
  8. 0:13completion criteria, and you'll be fine
  9. 0:15. Seriously, just knowing this puts you
  10. 0:17way ahead of the curve. Hi everyone, I'm
  11. 0:19Muchi. Today, I’m going to talk about
  12. 0:21the top 10 god-tier ways to make
  13. 0:23ClaudeCode 100 times smarter, as
  14. 0:25announced by Anthropic. First, as a
  15. 0:27prerequisite, when learning about
  16. 0:29ClaudeCode, don't rely on random AI
  17. 0:31influencers; always use the official
  18. 0:33documentation from Anthropic. The reason
  19. 0:35is that the official documentation is
  20. 0:37the most accurate and up-to-date source
  21. 0:39of information. However, this official
  22. 0:41info is all in English and scattered
  23. 0:43everywhere, so frankly, it's quite—
  24. 0:45actually, very—tedious to research. So,
  25. 0:47this video is here to solve that hassle
  26. 0:49for you. Specifically, I’ve compiled the
  27. 0:52ClaudeCode usage methods recommended by
  28. 0:54Anthropic, and I'll be ranking 12 of
  29. 0:56them that I personally tried and found
  30. 0:58to be absolute must-dos.
  31. 1:00If you watch this video and put these
  32. 1:02tips into practice, beginners will start
  33. 1:04off using it correctly, and even
  34. 1:06advanced users will definitely find some
  35. 1:08"Aha!" moments they hadn't thought of,
  36. 1:10so please watch until the end. So,
  37. 1:12before we dive into the main content, I
  38. 1:14have a gift for you, which is right here
  39. 1:16. The Official God-Tier Usage Complete
  40. 1:18Kit. I've put all the prompts and
  41. 1:19detailed configuration settings used in
  42. 1:21this video into this kit. We’ll be going
  43. 1:23through this kit throughout the video,
  44. 1:25so I hope you’ll have it open and follow
  45. 1:27along as we go. To get it, just add my
  46. 1:28official LINE account via the link at
  47. 1:30the top of the description or this QR
  48. 1:32code, and send me the four kanji
  49. 1:33characters for "official usage." That
  50. 1:35will automatically register you. Also,
  51. 1:36for those who haven't finished the
  52. 1:38initial setup for ClaudeCode yet, I’ve
  53. 1:39created a ClaudeCode Initial Setup Guide
  54. 1:41. If you’d like to receive it, just send
  55. 1:42a message saying "initial setup" to the
  56. 1:44LINE account you just registered with.
  57. 1:45So, I hope everyone has successfully
  58. 1:47received it. Let’s get right into the
  59. 1:49first way to use it.
  60. 1:51The first way to use ClaudeCode, as
  61. 1:52recommended by the official team, is to
  62. 1:54tell it to research how professionals do
  63. 1:56it before you start building. What this
  64. 1:58means is, when you entrust ClaudeCode
  65. 1:59with a difficult task, instead of having
  66. 2:01it build immediately, you have it look
  67. 2:03up methods from professional engineers
  68. 2:05or official recommendations first, and
  69. 2:06then start the actual work. In fact,
  70. 2:08this approach is recommended by
  71. 2:10Anthropic employees, especially for
  72. 2:12tasks involving writing programming code
  73. 2:14, such as building a homepage or
  74. 2:15creating a membership site. However,
  75. 2:17just saying it might not make the
  76. 2:18difference clear, so I will show you how
  77. 2:20much the accuracy of the output changes
  78. 2:22depending on whether you say this one
  79. 2:23phrase or not. For this example, we’ll
  80. 2:25build a company homepage and compare the
  81. 2:27results. I’ve prepared a fictional
  82. 2:28renovation company and compiled the
  83. 2:30information to include on the site, like
  84. 2:32the company introduction, three services
  85. 2:34, key strengths, customer testimonials,
  86. 2:36and contact methods. So, from here, I’ll
  87. 2:38compare how the accuracy of the output
  88. 2:39changes based on whether or not you
  89. 2:41include the instruction to research
  90. 2:42professional methods first. Let’s start
  91. 2:44with the case where I don’t say it. "
  92. 2:46Based on the 'company_info.md' file in
  93. 2:48the folder, please create a homepage for
  94. 2:50this company." I’ll just tell it that.
  95. 2:52After waiting a little while, it has
  96. 2:54been created like this. Let’s take a
  97. 2:56look at it. Looking at it, you can see
  98. 2:57things like "28 years in business," past
  99. 2:59achievements, our services, reasons for
  100. 3:01being chosen, and customer testimonials
  101. 3:03—the minimum information I wrote in the
  102. 3:05notes is included. It does at least look
  103. 3:07like a proper homepage, doesn't it? But
  104. 3:09, honestly, looking at it like this,
  105. 3:11it’s just plain tacky, isn't it? I don’t
  106. 3:12think anyone would look at this website
  107. 3:14and decide to hire them for a renovation
  108. 3:15. So, what do we do from here? Next,
  109. 3:17we’ll use the method I mentioned earlier
  110. 3:18: having it research professional
  111. 3:19techniques first. I’ll say, "Please
  112. 3:21create a company website based on this
  113. 3:22company_info.md file.""At that time, I
  114. 3:24want you to research how professionals
  115. 3:25do it before you build it." That's how
  116. 3:27I’ll phrase it. Doing that, instead of
  117. 3:28just starting to build right away, it
  118. 3:30began with a research phase like this.
  119. 3:32So now, it’s searching the web for
  120. 3:33things like professional website design
  121. 3:35, how it looks on mobile, standard ways
  122. 3:37to improve page speed, how to structure
  123. 3:39the inquiry flow, and even official
  124. 3:40explanations from Google. Once it
  125. 3:42finishes the research, it moves straight
  126. 3:43to building, and the website was
  127. 3:45completed like this. Let’s take a look.
  128. 3:46This is really good. If you compare it
  129. 3:48to the two I showed earlier, this one is
  130. 3:50overwhelmingly more stylish. It has a
  131. 3:52calm color scheme, a large Mincho-style
  132. 3:53heading that says, "A house lasts longer
  133. 3:55than the time it takes to build," and on
  134. 3:56the right, it even drew an illustration
  135. 3:58of a house in the dappled sunlight. And
  136. 4:00this heading is actually the CEO's
  137. 4:01greeting. And below that...
  138. 4:12Just by adding that one phrase, "
  139. 4:14research how professionals do it," it
  140. 4:15changes this much. This method isn't
  141. 4:17just for complex tasks like building
  142. 4:19websites; you can try it on your
  143. 4:20everyday work by saying, "Research how a
  144. 4:22pro would do this before you start," and
  145. 4:23it will improve the quality of your
  146. 4:25output, so please try it starting today
  147. 4:26. So, that was the first officially
  148. 4:28recommended way to use it. It was to
  149. 4:29tell it to research how a professional
  150. 4:31would do it before creating the output.
  151. 4:33Next, the second officially recommended
  152. 4:35usage is to communicate three things:
  153. 4:37the goal, the guardrails, and the
  154. 4:38completion criteria. What this means is
  155. 4:40that when you ask Claude for something,
  156. 4:42you don't need to give detailed
  157. 4:43step-by-step instructions. In fact,
  158. 4:44Dario Amodei, the creator of Claude,
  159. 4:46says that once you explain these three
  160. 4:47things—the goal, the guardrails, and the
  161. 4:49completion criteria—it’s better to leave
  162. 4:51the specific steps to the model. So,
  163. 4:52whenever you give instructions to Claude
  164. 4:54, all you need to do is provide these
  165. 4:55three things. First is what you want it
  166. 4:57to do; this is the objective. Second is
  167. 4:59what it shouldn't do, which are the
  168. 5:00guardrails, and third is what
  169. 5:02constitutes completion, which is the
  170. 5:03finish condition. The key is not to
  171. 5:05specify more than these three things,
  172. 5:07which allows the cloud to think for
  173. 5:09itself and choose the best approach.
  174. 5:11Until about six months ago, we had to
  175. 5:13give detailed instructions, like "do
  176. 5:16step 1, then step 2, then verify with
  177. 5:18step 3," but AI is much smarter now;
  178. 5:20giving overly detailed instructions
  179. 5:22actually defeats the purpose because
  180. 5:25following the steps becomes the goal,
  181. 5:27even if a better way exists. After all,
  182. 5:30Cloud Code is smart enough to graduate
  183. 5:32top of its class from Harvard, so as
  184. 5:34long as you provide the objective,
  185. 5:36guardrails, and finish condition, it
  186. 5:38will handle the rest.
  187. 5:41Since explaining it might not fully
  188. 5:42convey the importance of these three
  189. 5:44things, I'll demonstrate by comparing
  190. 5:46the same task with and without providing
  191. 5:48those three criteria.
  192. 5:51Here, I have included a sample meeting
  193. 5:53notes memo and a price list like this
  194. 5:55one. These hearing notes are from a
  195. 5:57salesperson's meeting with a client,
  196. 5:58containing bits of info in the order
  197. 6:00they were discussed, like the struggle
  198. 6:02of retyping faxed orders every morning
  199. 6:04or a budget of around 40,000 yen. If you
  200. 6:06use AI meeting note tools, you probably
  201. 6:08end up with memos that look something
  202. 6:10like this. Down here, you'll find some
  203. 6:12internal notes that the salesperson
  204. 6:13wrote for themselves. And this price
  205. 6:15list includes three plans: Light,
  206. 6:17Standard, and Premium. So, I'm going to
  207. 6:19create a document based on this
  208. 6:20information, but first, I'll open a new
  209. 6:22conversation and ask for it without
  210. 6:24specifying the goal, guardrails, or
  211. 6:25completion criteria. Please create a
  212. 6:27document based on the hearing notes and
  213. 6:28price list in this folder. Well, that's
  214. 6:30how I'll ask. Then, a document is
  215. 6:32created like this. Looking inside, it's
  216. 6:34neatly organized by item, with sections
  217. 6:36like the client's overview, current
  218. 6:37business challenges, requests, and
  219. 6:39concerns. So, it seems that what was
  220. 6:41written in the notes and price list has
  221. 6:42been summarized to some extent. However
  222. 6:44, while it is organized, it's hard to
  223. 6:46tell what the main point is. For example
  224. 6:48, the price list for the three plans is
  225. 6:50just pasted in as-is, so it's not clear
  226. 6:51which one is recommended, and if you
  227. 6:53scroll down, you can see that the
  228. 6:55internal notes are still left in there.
  229. 6:57I mean, it would be a disaster if we
  230. 6:58handed this directly to a client, right
  231. 7:00? So, next, I will make the same request
  232. 7:01as before, but this time I'll include
  233. 7:03the goal, guardrails, and completion
  234. 7:04criteria. I'll use the same session for
  235. 7:06this. Please create the document again.
  236. 7:08I want to create a proposal to help
  237. 7:10President Hamasaki choose one of the
  238. 7:12three plans from Use Saisei. Please do
  239. 7:13not include any discounts or features
  240. 7:15not listed in the price sheet. Also, do
  241. 7:17not include any internal company matters
  242. 7:19or information about other clients.
  243. 7:21Please only use the figures found in the
  244. 7:22price list and hearing notes. The
  245. 7:24document should be about 5 to 7 pages
  246. 7:26long, covering the President's concerns
  247. 7:27, the recommended plan and the reason
  248. 7:29why, a comparison table of the three
  249. 7:31plans, costs, and next steps, and it
  250. 7:32should be finished when the President
  251. 7:34can make a decision after reading it
  252. 7:35once. This is how I'll ask. Also, since
  253. 7:37this is voice input, there's a typo, but
  254. 7:39the part about "President Hamasaki from
  255. 7:40Use Saisei" is the goal, and the part
  256. 7:42about "discounts or features not in the
  257. 7:43price list" represents the guardrails.
  258. 7:45So, for this document, the completion
  259. 7:47criteria is that it should be about 5 to
  260. 7:487 slides long. And by entering this, a 6
  261. 7:50-slide document was created like this.
  262. 7:52Oh, that's wonderful. Looking at the
  263. 7:54first page, it lists what needs to be
  264. 7:55decided in this document, and the second
  265. 7:56page shows the three concerns the
  266. 7:58president mentioned in the hearing memo
  267. 7:59. Then, on the third page, it clearly
  268. 8:00states that the recommendation is the "
  269. 8:01Standard" plan, and it uses arrows to
  270. 8:03connect the reasoning for that to each
  271. 8:04of the president's concerns. Looking at
  272. 8:05the fourth slide and beyond, it
  273. 8:07continues like this; clearly, this
  274. 8:08version is much better than the one
  275. 8:09before, isn't it? The reality is that
  276. 8:11even with the same AI and the same
  277. 8:12information, providing these three
  278. 8:14things—the objective, guardrails, and
  279. 8:15completion criteria—makes this much of a
  280. 8:17difference in the output. I think it's
  281. 8:19very easy to just tell an AI to "make it
  282. 8:21look good" when you're asking for
  283. 8:22something. But if you can at least
  284. 8:24articulate what "look good" means—which
  285. 8:26in this case is the objective—and
  286. 8:27communicate just that to the AI, I
  287. 8:29believe the accuracy of your output will
  288. 8:31continue to improve significantly. So,
  289. 8:33that concludes the second officially
  290. 8:35recommended usage: communicating the
  291. 8:37objective, guardrails, and completion
  292. 8:38criteria.
  293. 8:40Next, the third officially recommended
  294. 8:42usage is to have Claude ask you
  295. 8:43questions before it executes the task.
  296. 8:45What this means is that when you ask
  297. 8:47Claude to do something, you should have
  298. 8:48it ask you first if there's anything you
  299. 8:50forgot to mention or if anything is
  300. 8:51still unclear, and then proceed only
  301. 8:53after answering those questions. The
  302. 8:54reason is that if an AI encounters
  303. 8:56something it doesn't understand, it will
  304. 8:57fill in the gaps with its own
  305. 8:58assumptions and move forward. For
  306. 9:00instance, if you ask it to create a
  307. 9:01document based on this data, it wouldn't
  308. 9:03know whether the document is intended
  309. 9:04for your boss or for a client, right? In
  310. 9:06that case, Claude might just assume, "
  311. 9:08Given this content, it's probably for a
  312. 9:10boss," and move forward based on that
  313. 9:11assumption. And when that happens, a gap
  314. 9:13emerges between imagination and reality
  315. 9:15, ultimately lowering the accuracy of
  316. 9:16the output. So, the ideal state when
  317. 9:18delegating a task to Claude Code is to
  318. 9:19put everything into words so there's no
  319. 9:21room for Claude to fill in the blanks
  320. 9:23with imagination, but that's pretty
  321. 9:24difficult, isn't it? If we could do that
  322. 9:26, we wouldn't have any trouble. That's
  323. 9:27why you have Claude ask about the parts
  324. 9:29you haven't put into words first.
  325. 9:30Specific phrasing might be something
  326. 9:32like, "If you have any questions about
  327. 9:34this, please let me know, otherwise,
  328. 9:35please go ahead," or "In addition to
  329. 9:37this information, what else should I
  330. 9:38tell you to improve the accuracy of your
  331. 9:40output?" Honestly, just by saying this,
  332. 9:42Claude stops working based on
  333. 9:43assumptions, and you'll get the results
  334. 9:45you want without the output being
  335. 9:47off-target. In fact, Anthropic employees
  336. 9:49themselves have mentioned on their
  337. 9:50official blog that adding "Ask me
  338. 9:52everything you need to confirm before
  339. 9:53you start" is the most useful habit when
  340. 9:55using Claude Code. Really, just say, "
  341. 9:57If there is anything you don't
  342. 9:58understand, please ask me first; if you
  343. 10:00have no questions, please proceed." I
  344. 10:02personally think you could add this to
  345. 10:03every single request, so I really want
  346. 10:05you to adopt it as a standard habit. So
  347. 10:07, that was the third officially
  348. 10:09recommended usage: having Claude ask
  349. 10:10questions before execution.
  350. 10:13Next, the fourth officially recommended
  351. 10:15usage is to handle tasks without APIs
  352. 10:17through screen operation. Now, an API is
  353. 10:19like an entrance that allows you to
  354. 10:21control external services from Claude
  355. 10:22Code, and for services that have them,
  356. 10:24like X or YouTube, you can immediately
  357. 10:26integrate them with Claude Code, but
  358. 10:27there are services in the world that
  359. 10:29don't have this entrance. So, what do
  360. 10:31you do in those cases? That's where the
  361. 10:33method I mentioned earlier comes in:
  362. 10:34having Claude operate the screen
  363. 10:35directly. In fact, Claude Code can
  364. 10:37connect to Google Chrome and operate the
  365. 10:39browser itself—this is called "Claude in
  366. 10:40Chrome"—and the official documentation
  367. 10:42recommends this method when an API is
  368. 10:43unavailable. Even the creator of Claude
  369. 10:45, Anthropic, has stated that using this
  370. 10:47Claude screen control feature expands
  371. 10:48the range and quality of what you can
  372. 10:50create. So, I definitely want you to try
  373. 10:52it out, and for how to connect Claude to
  374. 10:53Chrome, check out this complete kit here
  375. 10:55. Just copy this and paste it into
  376. 10:57Claude Code, and it’ll be ready to use,
  377. 10:58so please follow these steps. That
  378. 11:00covers the official information, but
  379. 11:02I’ve taken it a step further by using a
  380. 11:04tool called Agent Desktop. Agent Desktop
  381. 11:06is released for free on GitHub by the
  382. 11:08web service company Vercel, and in short
  383. 11:10, it’s like an evolved version of the
  384. 11:12screen control I just showed you. The
  385. 11:14thing is, with the officially
  386. 11:15recommended method of having AI control
  387. 11:17your screen, it takes over your mouse
  388. 11:19movements, meaning you can't do your own
  389. 11:21work during that time. So, before I
  390. 11:23started using this Agent Desktop, I had
  391. 11:25to set screen control tasks to run when
  392. 11:27I wasn't using the PC, like right before
  393. 11:29bed or before heading to the gym.
  394. 11:31However, Agent Desktop doesn't take over
  395. 11:33your mouse, so you can continue your
  396. 11:34regular work while it handles
  397. 11:36screen-based tasks in the background.
  398. 11:38Also, it’s not limited to just one task
  399. 11:40; you can run two or more at once. While
  400. 11:42before you could only do one thing—like
  401. 11:45having it fetch info from LINE—with
  402. 11:47Agent Desktop, you can have it gather
  403. 11:49LINE info, research the latest AI news,
  404. 11:51and still do your own work
  405. 11:53simultaneously. I highly recommend this
  406. 11:55, so check out this complete kit right
  407. 11:56here. Installing Agent Desktop. If you
  408. 11:58copy this and paste it into Claude Code
  409. 12:00, it will handle everything from
  410. 12:01installation to the actual process. So
  411. 12:03that wraps up the fourth officially
  412. 12:05recommended use case: using screen
  413. 12:07control for tasks that lack an API.
  414. 12:09Next, the first recommended way to use
  415. 12:11it is to delegate heavy tasks to a
  416. 12:13separate AI and only receive the results
  417. 12:15. Heavy tasks here refer to jobs like
  418. 12:17reading a massive amount of documents,
  419. 12:19writing long articles, or generating
  420. 12:21many images. If you make the main Claude
  421. 12:23model handle all these high-processing
  422. 12:24tasks, the conversation becomes
  423. 12:26cluttered, and the accuracy of the
  424. 12:27output drops. So, the officially
  425. 12:29recommended method is to leave long
  426. 12:31research or large-scale output to
  427. 12:32another AI, or a "sub-agent," and have
  428. 12:34it return only the results to the main
  429. 12:36model. The way to delegate work to a
  430. 12:37sub-agent is incredibly simple: just
  431. 12:39tell it, "Delegate this task to a
  432. 12:40sub-agent and return only a short
  433. 12:41summary of the results to this chat."
  434. 12:43That is all you need to ask. I use these
  435. 12:45sub-agents quite a bit; for example,
  436. 12:46once I have a YouTube script, I set up
  437. 12:48multiple sub-agents to review it from
  438. 12:50various perspectives and have them
  439. 12:51report back if there are any points to
  440. 12:53address. Also, for this video, I was
  441. 12:55working on the script while
  442. 12:56simultaneously creating the complete kit
  443. 12:58offered as a bonus, so I used the
  444. 13:00sub-agent function to ensure there were
  445. 13:01no inconsistencies. However, since
  446. 13:03sub-agents simply double or triple your
  447. 13:05token usage, when I want to keep costs
  448. 13:07down, I don't split the work within
  449. 13:09Claude, but rather outsource it to other
  450. 13:11AIs. To explain what that means, I
  451. 13:13actually outsource two things: first,
  452. 13:14anything related to research, such as
  453. 13:16reading long documents, which I
  454. 13:18outsource to NotebookLM. Specifically,
  455. 13:20if I send a prompt like "Look up
  456. 13:21recommended washing machines with
  457. 13:23reviews" to Claude, it triggers
  458. 13:25NotebookLM via the cloud to conduct the
  459. 13:27research within NotebookLM itself. The
  460. 13:28reason I do this is basically a matter
  461. 13:30of role division; for instance,
  462. 13:32NotebookLM isn't as smart as Claude, but
  463. 13:34it is an AI that is exceptionally good
  464. 13:35at processing massive amounts of
  465. 13:37information. Plus, since it's free to
  466. 13:39use, you can summarize information while
  467. 13:41saving on Claude Code tokens. I've
  468. 13:43talked to many people, and I hear a lot
  469. 13:44of feedback that this integration with
  470. 13:46NotebookLM is excellent, so definitely
  471. 13:48go ahead and set it up while watching
  472. 13:50this video. As for how to connect it,
  473. 13:51you can use this from the complete kit.
  474. 13:53Just copy this to connect with
  475. 13:55NotebookLM and paste it into Claude Code
  476. 13:57, so please make use of it. I explain
  477. 13:59more effective ways to use this in this
  478. 14:00video here, so be sure to watch it after
  479. 14:02you're done with this one. So, that's
  480. 14:04the first bit of outsourcing; the second
  481. 14:06one is tasks related to creating images
  482. 14:08. For this, I leave it to DALL-E. DALL-E
  483. 14:10is an AI from OpenAI, the creators of
  484. 14:12ChatGPT, and I use it to create things
  485. 14:14like these diagrams and slide decks you
  486. 14:16see in my videos by giving instructions
  487. 14:18through Claude Code. The reason is
  488. 14:20simple: when it comes to creating images
  489. 14:22right now, DALL-E 3 is overwhelmingly
  490. 14:24higher in quality. By linking Claude
  491. 14:26Code and DALL-E like this, you can
  492. 14:28create stylish materials in an instant,
  493. 14:30so if you often have to prepare
  494. 14:32presentations, try linking these two.
  495. 14:33The linking method is the same as the
  496. 14:35NotebookLM one from earlier; it's this
  497. 14:37part of the complete kit. Just copy this
  498. 14:38and paste it into Claude Code to finish
  499. 14:40, so please give it a try. Also, I
  500. 14:42explain document creation in more detail
  501. 14:44in this video, so check that out after
  502. 14:46this one too. So, that covers the first
  503. 14:48officially recommended way to use it:
  504. 14:50offload heavy tasks to other AIs.
  505. 14:53Next, the sixth officially recommended
  506. 14:55way to use it is to adjust the effort
  507. 14:56based on the weight of the task. This "
  508. 14:58effort" is a setting that determines how
  509. 15:00deeply Claude Code thinks about a task.
  510. 15:02Most people probably ask for the same
  511. 15:03level of effort for every task, whether
  512. 15:05it's a minor job or one where mistakes
  513. 15:06would be a real problem. But as stated
  514. 15:08in Anthropic's official documentation, I
  515. 15:09highly recommend adjusting the effort
  516. 15:11level based on the weight of the task
  517. 15:12you're delegating. And when I talk about
  518. 15:14the weight of a task here, I don't mean
  519. 15:15the volume of work, but rather how much
  520. 15:17deep thinking is required. You can set
  521. 15:18this by clicking right here below the
  522. 15:20input field and selecting your desired
  523. 15:21effort level from there. So, how do I
  524. 15:23actually use this? For example, when
  525. 15:25writing YouTube scripts or creating
  526. 15:26presentations, I start with the initial
  527. 15:28brainstorming phase using the effort
  528. 15:30level recommended by the official
  529. 15:31guidelines. Once the direction is
  530. 15:32solidified to a certain extent, I move
  531. 15:34on to building out a more detailed
  532. 15:35structure—the skeleton, if you will.
  533. 15:37When I'm ready to create that, I set the
  534. 15:39effort to "Long-term" and then, once the
  535. 15:41structure is approved and I'm ready to
  536. 15:42write the script or build the slides, I
  537. 15:44set it to "Ultra" to finish it up. By
  538. 15:46the way, as a side note, it's generally
  539. 15:48advised not to switch between models
  540. 15:50like Opus or Sonnet within a single
  541. 15:52session, but since these effort levels
  542. 15:54are applied to each individual
  543. 15:56interaction, there's no problem at all
  544. 15:58with changing them within the same
  545. 15:59session. However, it might be hard to
  546. 16:01tell which effort level is suitable for
  547. 16:03your current task just from this video,
  548. 16:04so I’ve included an "Effort Setting
  549. 16:06Quick Reference Guide" in the complete
  550. 16:08kit. So, if you're ever in doubt, please
  551. 16:09take a look at that. That concludes the
  552. 16:11sixth officially recommended usage
  553. 16:13method: adjusting effort based on the
  554. 16:14weight of the task you're delegating.
  555. 16:16Moving on, the seventh officially
  556. 16:18recommended usage method is to give the
  557. 16:20AI the means to verify its own work.
  558. 16:22What this means is instead of just
  559. 16:23letting Claude build it and leaving it
  560. 16:25at that, you provide a checklist of
  561. 16:26requirements and make Claude revise its
  562. 16:28work until it passes every point. I mean
  563. 16:30, you should have it go as far as doing
  564. 16:31even that. Well, even if AI lets you
  565. 16:33generate a massive amount of output, if
  566. 16:34you have to check it with your own eyes
  567. 16:36every time, you’re just going to stay
  568. 16:37busy, right? That’s exactly why you
  569. 16:39should provide the criteria for checking
  570. 16:40to Claude. In fact, Boris Cherny
  571. 16:42mentioned that giving AI the means to
  572. 16:44verify its own work improves output
  573. 16:45accuracy by two to three times, and the
  574. 16:47official documentation also suggests
  575. 16:49providing a clear pass/fail checklist
  576. 16:50and continuing until it passes. I’ve
  577. 16:52actually adopted this method myself;
  578. 16:54when I write YouTube scripts, I don't
  579. 16:55look at them as soon as they're finished
  580. 16:57, but have them checked on Claude's end
  581. 16:59first. I’ve set it up so that it checks
  582. 17:01for things like logical consistency or
  583. 17:03monotonous sentence endings three times
  584. 17:05within Claude before presenting it to me
  585. 17:07. So, instead of checking everything
  586. 17:08myself every single time, only the items
  587. 17:10that pass the check are returned to me.
  588. 17:11The key to creating this checklist is to
  589. 17:13make it clear enough that anyone can
  590. 17:14tell if it's a pass or a fail, but then
  591. 17:16you might wonder how to actually build
  592. 17:18that kind of checklist for your own work
  593. 17:19. That’s where this complete kit comes
  594. 17:21in. "Create a Pass/Fail Checklist."
  595. 17:23Please copy this and paste it into
  596. 17:25Claude. By doing that, it will build a
  597. 17:27checklist tailored to your work, run a
  598. 17:29test grading, and then save it as a
  599. 17:30skill. So, if you’ve recently started
  600. 17:32using Claude and have been able to
  601. 17:33automate some of your tasks, but find
  602. 17:35that you still have to verify everything
  603. 17:36manually, so your workload hasn't
  604. 17:37actually decreased, then by all means,
  605. 17:39please try using the kit I just
  606. 17:40mentioned. So, that covers the seventh
  607. 17:42officially recommended usage. Which is
  608. 17:44giving AI the means to verify its own
  609. 17:46work.
  610. 17:47Next, the eighth officially recommended
  611. 17:49usage is to leave the final verification
  612. 17:51to a separate AI. This is because, with
  613. 17:53the seventh usage we just discussed,
  614. 17:54Claude is now able to determine whether
  615. 17:56its output passes or fails on its own.
  616. 17:57But this alone doesn't allow you to
  617. 17:59catch errors in areas not covered by the
  618. 18:01checklist. So, we need to have those
  619. 18:02checked too, but if you ask the same
  620. 18:04session that generated the output to
  621. 18:06also do the verification, it’s biased
  622. 18:08because it created the work itself. In
  623. 18:10short, even if there are things to fix,
  624. 18:12it might say there aren't any, kind of
  625. 18:14like when my grandma used to tell me I
  626. 18:16looked just like my uncle. It becomes
  627. 18:18extremely biased. That’s why for these
  628. 18:20final checks, you should have a
  629. 18:21different Claude check what the first
  630. 18:23one produced. This is what’s important.
  631. 18:25And by "a different Claude," I don't
  632. 18:26mean installing a new app, but rather
  633. 18:28starting a separate session. In fact,
  634. 18:30Anthropic's official documentation
  635. 18:31states that AIs tend to be lenient when
  636. 18:33grading their own work, so it's better
  637. 18:35to separate the creator role from the
  638. 18:36reviewer role. As for the prompt for the
  639. 18:38final check, here is the complete kit. "
  640. 18:40Final check of the completed work."
  641. 18:42Please try copying this, pasting it into
  642. 18:44Claude, and entering what you want
  643. 18:45verified below it. By using this prompt
  644. 18:47, you can, for example, have a business
  645. 18:49plan read by an investor, a customer,
  646. 18:51and a competitor all at once; this is a
  647. 18:53method the Claude team introduced on
  648. 18:55their blog. By automating the pass
  649. 18:56criteria and final verification using
  650. 18:58the seventh and eighth methods I
  651. 18:59discussed, your workload will decrease
  652. 19:01significantly and your business will
  653. 19:02accelerate, so please make use of the
  654. 19:04prompt I just showed you. So that
  655. 19:05concludes the eighth officially
  656. 19:06recommended way to use it. Which is to
  657. 19:08leave final verification to a different
  658. 19:10AI.
  659. 19:11Moving on, the ninth officially
  660. 19:13recommended way to use it is to shorten
  661. 19:14the "Claude System Prompt." This "
  662. 19:16Claude System Prompt" is like a set of
  663. 19:17instructions that Claude reads at the
  664. 19:19start of every conversation, where you
  665. 19:20write down things about yourself, work
  666. 19:22rules, and things you don't want it to
  667. 19:23do. Most people, when Claude doesn't
  668. 19:25listen to them, probably just keep
  669. 19:26adding more and more rules to their "
  670. 19:28claudemc" file. But actually, this is
  671. 19:29counterproductive; Anthropic's official
  672. 19:31documentation states that you should
  673. 19:32keep your "claudemc" file under 200
  674. 19:34lines. That's because the "claudemc"
  675. 19:35file is read every single time you have
  676. 19:37a conversation, so if it's too long, the
  677. 19:38important rules you really need followed
  678. 19:40just get buried. In short, writing more
  679. 19:41doesn't mean it will listen better;
  680. 19:42rather, keeping it short actually helps
  681. 19:44it follow your instructions properly. So
  682. 19:45, what should you actually write?
  683. 19:47According to the official documentation
  684. 19:48, just things Claude can't understand by
  685. 19:50reading the folder, or things you've had
  686. 19:51to warn it about twice. It says that
  687. 19:52these two things alone are enough for
  688. 19:54your "claudemc" file. Conversely, things
  689. 19:55Claude can understand by looking inside
  690. 19:57the folder can be read by Claude itself
  691. 19:58, so there's no need to write them down
  692. 19:59. However, those who use Claude Code a
  693. 20:01bit more might say, "No, even so, it
  694. 20:03doesn't listen, which is why I go out of
  695. 20:05my way to emphasize things with '
  696. 20:06IMPORTANT' or star marks." But that
  697. 20:08might actually be backfiring, and in
  698. 20:09fact, the official documentation states
  699. 20:11that you don't need to keep emphasizing
  700. 20:13things with the current model. This is
  701. 20:14exactly why you need to get your
  702. 20:16information from official documentation
  703. 20:18; AI is constantly evolving, so the "
  704. 20:20right answer" changes over time. For
  705. 20:22instance, writing detailed "claudemc"
  706. 20:23files or using star marks might have
  707. 20:25been the right approach back around Opus
  708. 20:274.6. But now we're at something like
  709. 20:29Opus 5.5, so the underlying assumptions
  710. 20:31have changed. The image of writing long
  711. 20:33, detailed "claudemc" files is like
  712. 20:35trying to make a middle schooler wear
  713. 20:36the clothes they wore in kindergarten
  714. 20:38just because they looked good back then
  715. 20:40. Let's keep up with its growth. I think
  716. 20:42this is actually quite common. People
  717. 20:43who started using Claude Code six months
  718. 20:45ago and are still using a "claudemc"
  719. 20:46file based on the specifications from
  720. 20:47back then. If that's the case, I hope
  721. 20:49this video encourages you to give your
  722. 20:51Claude.md file a makeover. As for how to
  723. 20:52do this, take a look at this part of the
  724. 20:54complete kit. Review and shorten your
  725. 20:56current Claude.md. If you copy this and
  726. 20:58paste it into Claude, it will review
  727. 21:00your Claude.md based on official
  728. 21:02standards. So, that covers the ninth
  729. 21:04officially recommended tip: shortening
  730. 21:06your Claude.md.
  731. 21:08Next is the tenth officially recommended
  732. 21:10tip: don't just create a project,
  733. 21:12nurture it. "Projects" are basically a
  734. 21:14system that lets you save step-by-step
  735. 21:15procedures like a manual for tasks you
  736. 21:17do repeatedly. By turning tasks like
  737. 21:18writing scripts or creating documents
  738. 21:20into projects, you can get consistent
  739. 21:22results every time with just a single
  740. 21:23request. The method recommended in the
  741. 21:25official documentation is to finish a
  742. 21:27task once without using a project, and
  743. 21:29then ask Claude at the end of the chat
  744. 21:31to turn that approach into a project.
  745. 21:32Because the successful conversation
  746. 21:34contains all your added explanations and
  747. 21:36the fixes for issues you encountered,
  748. 21:37you can incorporate all of that into the
  749. 21:39project. For the exact prompt to use,
  750. 21:41refer to this section of the complete
  751. 21:42kit. Turn a successful conversation into
  752. 21:44a project. I hope you'll copy and use
  753. 21:46this, so please make the most of it.
  754. 21:47However, I think many people just stop
  755. 21:49after they've created their project. But
  756. 21:51a project isn't finished once it's
  757. 21:53created; in many ways, it's more
  758. 21:54important to nurture it from there. In
  759. 21:56fact, the official Anthropic
  760. 21:58documentation introduces a way to
  761. 21:59nurture your projects by using them,
  762. 22:01seeing where you trip up, and then
  763. 22:02refining them. To help with this
  764. 22:04nurturing process, there's the "Project
  765. 22:06Creator," an official skill released by
  766. 22:07Anthropic for building your projects. If
  767. 22:09you add this, it will walk you through
  768. 22:11the process of testing and fixing the
  769. 22:12projects you've created. For this Skill
  770. 22:14Creator, you definitely want to be using
  771. 22:16Claude Code, so if you haven't set it up
  772. 22:17yet, check out this complete kit. I've
  773. 22:19included how to install the Skill
  774. 22:20Creator here, so please take advantage
  775. 22:22of it. That concludes the 10th
  776. 22:24officially recommended usage: don't just
  777. 22:26stop at creating a skill.
  778. 22:29Next, the 11th officially recommended
  779. 22:31usage is to set up hooks. A "hook" is a
  780. 22:33setting that forces a specific action to
  781. 22:35occur at certain times, like when Claude
  782. 22:38rewrites a file or finishes a task. As I
  783. 22:40mentioned with the 9th point, Claude
  784. 22:42will follow most rules if you set up the
  785. 22:44metadata correctly, but it isn't perfect
  786. 22:46. So, for things that absolutely must be
  787. 22:48followed, you need to set up a hook. In
  788. 22:50fact, Anthropic's official documentation
  789. 22:52states that for critical rules, you
  790. 22:54should use hooks rather than just
  791. 22:55metadata. Creating these hooks is super
  792. 22:57simple; you just describe them in your
  793. 22:59code. For example, you can tell it to
  794. 23:01generate a hook that always does "this"
  795. 23:02when "that" happens, and by just
  796. 23:04describing when and what to do, Claude
  797. 23:05Code will create the hook automatically
  798. 23:07. If you want to test if it works
  799. 23:09properly or learn how to remove a hook,
  800. 23:10check out this complete guide. Setting
  801. 23:12hooks for things you absolutely need to
  802. 23:14enforce. If you copy this and paste it
  803. 23:16into Claude Code, it will ask you
  804. 23:17questions one by one to create the hook
  805. 23:19and then have Claude test it out itself
  806. 23:21. So, if you find that Claude isn't
  807. 23:22following your instructions even after
  808. 23:24you've written the metadata correctly,
  809. 23:25please give this a try. That concludes
  810. 23:27the 11th officially recommended usage.
  811. 23:29That was setting up hooks.
  812. 23:32And finally, the 12th officially
  813. 23:33recommended usage is to verify five
  814. 23:35things before automating. So, what this
  815. 23:37means is that Claude Code has a "
  816. 23:39scheduled execution" feature that
  817. 23:40performs tasks automatically at set
  818. 23:42times, but you should check five things
  819. 23:44before entrusting it with a job. Because
  820. 23:45scheduled executions run while you're
  821. 23:47asleep or not watching, if Claude gets
  822. 23:49stuck along the way, it can't ask you
  823. 23:50for help. That is exactly why you need
  824. 23:52to decide on these five things
  825. 23:53beforehand and include them in your
  826. 23:54prompt so Claude doesn't get lost. These
  827. 23:56five points come from Anthropic's
  828. 23:58official documentation; the first is
  829. 23:59defining what "done" looks like. Second
  830. 24:01is a checklist to measure the criteria
  831. 24:02for success. Third is the scope of what
  832. 24:04it's allowed to touch. This essentially
  833. 24:06covers which files to read and what
  834. 24:07actions are prohibited. Fourth is when
  835. 24:09to stop; this includes deciding when to
  836. 24:10give up if a task won't finish, as well
  837. 24:12as setting a budget limit. And finally,
  838. 24:14the fifth point is where to send the
  839. 24:15results. These are the five pieces of
  840. 24:17information you should provide before
  841. 24:18setting up any scheduled execution or
  842. 24:19automation. I personally use scheduled
  843. 24:20executions for things like researching
  844. 24:22trending competitor videos, transcribing
  845. 24:24coaching or meeting recordings, and
  846. 24:25posting weekly group consulting archives
  847. 24:27to my membership site; I have various
  848. 24:28tasks running automatically like this.
  849. 24:30For every automated task I have here,
  850. 24:31I’ve specified all five: the completion
  851. 24:33criteria, the checklist, the prohibited
  852. 24:35actions, the limits, and the destination
  853. 24:37for the results. However, I know that
  854. 24:39deciding these five things every time
  855. 24:40can be a hassle, so for those moments,
  856. 24:42here is this complete kit. If you copy "
  857. 24:44Five Checks Before Automating" and paste
  858. 24:45it into Claude Code like this, it will
  859. 24:47fill in those five points and handle the
  860. 24:49scheduled execution setup for you. So,
  861. 24:51if you find yourself doing the same
  862. 24:52tasks every day but haven't automated
  863. 24:54them yet, pick one job to delegate and
  864. 24:56try using this prompt. So, that was the
  865. 24:5812th official recommendation: verify
  866. 25:00five things before automating. In this
  867. 25:02video, I’ve introduced all 12 ways to
  868. 25:04use Claude Code as recommended by
  869. 25:06Anthropic. If you’re going to start with
  870. 25:08just one thing today, please start by
  871. 25:09researching how professionals do it
  872. 25:11before you build, as I mentioned in the
  873. 25:12first tip. Try adding that one specific
  874. 25:14phrase I talked about. Truly, just
  875. 25:15adding that one phrase improves output
  876. 25:17accuracy more than you’d expect, and
  877. 25:18it’s something you can try right away,
  878. 25:20so I highly encourage you to use it. And
  879. 25:22once again, all the prompts and detailed
  880. 25:24settings used in this video are compiled
  881. 25:26in this "Official Ultimate Utilization
  882. 25:27Kit." So, if you haven’t received it yet
  883. 25:29, please add my official LINE account
  884. 25:30via the link at the top of the
  885. 25:32description or this QR code and send the
  886. 25:33message "Official Utilization." It’s set
  887. 25:35up to be sent to you automatically. Also
  888. 25:36, for those who want to learn more about
  889. 25:38the basics of Claude Code, I’m giving
  890. 25:40away a 20-chapter Claude Code
  891. 25:41Fundamentals Course for free in my 0-
  892. 25:42yen monthly online community, so please
  893. 25:44take advantage of that as well. You’ll
  894. 25:46receive information about the 0-yen
  895. 25:47community through the same LINE account
  896. 25:48you registered with earlier, so you can
  897. 25:50join from there. That wraps up this
  898. 25:52video. Thank you so much for watching
  899. 25:53until the end. I’ll see you all in the
  900. 25:55next video. Bye-bye.

About this transcript

This page contains the full transcript of 【最新】Anthropic公式が推奨するClaude Codeの神活用法TOP12!初心者でもスグに使えるClaude Code活用法【AIで業務自動化】 by むっちょ【Claude Code攻略】, generated from the public captions YouTube serves with the video. The transcript has 5,924 words across 900 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.