【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
Full transcript
- 0:00When asking ClaudeCode for a difficult
- 0:02task, just look up the method, build it
- 0:04, and use this prompt; it increases the
- 0:06accuracy of Anthropic-level responses by
- 0:08over five times. For instructions to
- 0:10ClaudeCode, just keep these three things
- 0:12in mind: purpose, guardrails, and
- 0:13completion criteria, and you'll be fine
- 0:15. Seriously, just knowing this puts you
- 0:17way ahead of the curve. Hi everyone, I'm
- 0:19Muchi. Today, I’m going to talk about
- 0:21the top 10 god-tier ways to make
- 0:23ClaudeCode 100 times smarter, as
- 0:25announced by Anthropic. First, as a
- 0:27prerequisite, when learning about
- 0:29ClaudeCode, don't rely on random AI
- 0:31influencers; always use the official
- 0:33documentation from Anthropic. The reason
- 0:35is that the official documentation is
- 0:37the most accurate and up-to-date source
- 0:39of information. However, this official
- 0:41info is all in English and scattered
- 0:43everywhere, so frankly, it's quite—
- 0:45actually, very—tedious to research. So,
- 0:47this video is here to solve that hassle
- 0:49for you. Specifically, I’ve compiled the
- 0:52ClaudeCode usage methods recommended by
- 0:54Anthropic, and I'll be ranking 12 of
- 0:56them that I personally tried and found
- 0:58to be absolute must-dos.
- 1:00If you watch this video and put these
- 1:02tips into practice, beginners will start
- 1:04off using it correctly, and even
- 1:06advanced users will definitely find some
- 1:08"Aha!" moments they hadn't thought of,
- 1:10so please watch until the end. So,
- 1:12before we dive into the main content, I
- 1:14have a gift for you, which is right here
- 1:16. The Official God-Tier Usage Complete
- 1:18Kit. I've put all the prompts and
- 1:19detailed configuration settings used in
- 1:21this video into this kit. We’ll be going
- 1:23through this kit throughout the video,
- 1:25so I hope you’ll have it open and follow
- 1:27along as we go. To get it, just add my
- 1:28official LINE account via the link at
- 1:30the top of the description or this QR
- 1:32code, and send me the four kanji
- 1:33characters for "official usage." That
- 1:35will automatically register you. Also,
- 1:36for those who haven't finished the
- 1:38initial setup for ClaudeCode yet, I’ve
- 1:39created a ClaudeCode Initial Setup Guide
- 1:41. If you’d like to receive it, just send
- 1:42a message saying "initial setup" to the
- 1:44LINE account you just registered with.
- 1:45So, I hope everyone has successfully
- 1:47received it. Let’s get right into the
- 1:49first way to use it.
- 1:51The first way to use ClaudeCode, as
- 1:52recommended by the official team, is to
- 1:54tell it to research how professionals do
- 1:56it before you start building. What this
- 1:58means is, when you entrust ClaudeCode
- 1:59with a difficult task, instead of having
- 2:01it build immediately, you have it look
- 2:03up methods from professional engineers
- 2:05or official recommendations first, and
- 2:06then start the actual work. In fact,
- 2:08this approach is recommended by
- 2:10Anthropic employees, especially for
- 2:12tasks involving writing programming code
- 2:14, such as building a homepage or
- 2:15creating a membership site. However,
- 2:17just saying it might not make the
- 2:18difference clear, so I will show you how
- 2:20much the accuracy of the output changes
- 2:22depending on whether you say this one
- 2:23phrase or not. For this example, we’ll
- 2:25build a company homepage and compare the
- 2:27results. I’ve prepared a fictional
- 2:28renovation company and compiled the
- 2:30information to include on the site, like
- 2:32the company introduction, three services
- 2:34, key strengths, customer testimonials,
- 2:36and contact methods. So, from here, I’ll
- 2:38compare how the accuracy of the output
- 2:39changes based on whether or not you
- 2:41include the instruction to research
- 2:42professional methods first. Let’s start
- 2:44with the case where I don’t say it. "
- 2:46Based on the 'company_info.md' file in
- 2:48the folder, please create a homepage for
- 2:50this company." I’ll just tell it that.
- 2:52After waiting a little while, it has
- 2:54been created like this. Let’s take a
- 2:56look at it. Looking at it, you can see
- 2:57things like "28 years in business," past
- 2:59achievements, our services, reasons for
- 3:01being chosen, and customer testimonials
- 3:03—the minimum information I wrote in the
- 3:05notes is included. It does at least look
- 3:07like a proper homepage, doesn't it? But
- 3:09, honestly, looking at it like this,
- 3:11it’s just plain tacky, isn't it? I don’t
- 3:12think anyone would look at this website
- 3:14and decide to hire them for a renovation
- 3:15. So, what do we do from here? Next,
- 3:17we’ll use the method I mentioned earlier
- 3:18: having it research professional
- 3:19techniques first. I’ll say, "Please
- 3:21create a company website based on this
- 3:22company_info.md file.""At that time, I
- 3:24want you to research how professionals
- 3:25do it before you build it." That's how
- 3:27I’ll phrase it. Doing that, instead of
- 3:28just starting to build right away, it
- 3:30began with a research phase like this.
- 3:32So now, it’s searching the web for
- 3:33things like professional website design
- 3:35, how it looks on mobile, standard ways
- 3:37to improve page speed, how to structure
- 3:39the inquiry flow, and even official
- 3:40explanations from Google. Once it
- 3:42finishes the research, it moves straight
- 3:43to building, and the website was
- 3:45completed like this. Let’s take a look.
- 3:46This is really good. If you compare it
- 3:48to the two I showed earlier, this one is
- 3:50overwhelmingly more stylish. It has a
- 3:52calm color scheme, a large Mincho-style
- 3:53heading that says, "A house lasts longer
- 3:55than the time it takes to build," and on
- 3:56the right, it even drew an illustration
- 3:58of a house in the dappled sunlight. And
- 4:00this heading is actually the CEO's
- 4:01greeting. And below that...
- 4:12Just by adding that one phrase, "
- 4:14research how professionals do it," it
- 4:15changes this much. This method isn't
- 4:17just for complex tasks like building
- 4:19websites; you can try it on your
- 4:20everyday work by saying, "Research how a
- 4:22pro would do this before you start," and
- 4:23it will improve the quality of your
- 4:25output, so please try it starting today
- 4:26. So, that was the first officially
- 4:28recommended way to use it. It was to
- 4:29tell it to research how a professional
- 4:31would do it before creating the output.
- 4:33Next, the second officially recommended
- 4:35usage is to communicate three things:
- 4:37the goal, the guardrails, and the
- 4:38completion criteria. What this means is
- 4:40that when you ask Claude for something,
- 4:42you don't need to give detailed
- 4:43step-by-step instructions. In fact,
- 4:44Dario Amodei, the creator of Claude,
- 4:46says that once you explain these three
- 4:47things—the goal, the guardrails, and the
- 4:49completion criteria—it’s better to leave
- 4:51the specific steps to the model. So,
- 4:52whenever you give instructions to Claude
- 4:54, all you need to do is provide these
- 4:55three things. First is what you want it
- 4:57to do; this is the objective. Second is
- 4:59what it shouldn't do, which are the
- 5:00guardrails, and third is what
- 5:02constitutes completion, which is the
- 5:03finish condition. The key is not to
- 5:05specify more than these three things,
- 5:07which allows the cloud to think for
- 5:09itself and choose the best approach.
- 5:11Until about six months ago, we had to
- 5:13give detailed instructions, like "do
- 5:16step 1, then step 2, then verify with
- 5:18step 3," but AI is much smarter now;
- 5:20giving overly detailed instructions
- 5:22actually defeats the purpose because
- 5:25following the steps becomes the goal,
- 5:27even if a better way exists. After all,
- 5:30Cloud Code is smart enough to graduate
- 5:32top of its class from Harvard, so as
- 5:34long as you provide the objective,
- 5:36guardrails, and finish condition, it
- 5:38will handle the rest.
- 5:41Since explaining it might not fully
- 5:42convey the importance of these three
- 5:44things, I'll demonstrate by comparing
- 5:46the same task with and without providing
- 5:48those three criteria.
- 5:51Here, I have included a sample meeting
- 5:53notes memo and a price list like this
- 5:55one. These hearing notes are from a
- 5:57salesperson's meeting with a client,
- 5:58containing bits of info in the order
- 6:00they were discussed, like the struggle
- 6:02of retyping faxed orders every morning
- 6:04or a budget of around 40,000 yen. If you
- 6:06use AI meeting note tools, you probably
- 6:08end up with memos that look something
- 6:10like this. Down here, you'll find some
- 6:12internal notes that the salesperson
- 6:13wrote for themselves. And this price
- 6:15list includes three plans: Light,
- 6:17Standard, and Premium. So, I'm going to
- 6:19create a document based on this
- 6:20information, but first, I'll open a new
- 6:22conversation and ask for it without
- 6:24specifying the goal, guardrails, or
- 6:25completion criteria. Please create a
- 6:27document based on the hearing notes and
- 6:28price list in this folder. Well, that's
- 6:30how I'll ask. Then, a document is
- 6:32created like this. Looking inside, it's
- 6:34neatly organized by item, with sections
- 6:36like the client's overview, current
- 6:37business challenges, requests, and
- 6:39concerns. So, it seems that what was
- 6:41written in the notes and price list has
- 6:42been summarized to some extent. However
- 6:44, while it is organized, it's hard to
- 6:46tell what the main point is. For example
- 6:48, the price list for the three plans is
- 6:50just pasted in as-is, so it's not clear
- 6:51which one is recommended, and if you
- 6:53scroll down, you can see that the
- 6:55internal notes are still left in there.
- 6:57I mean, it would be a disaster if we
- 6:58handed this directly to a client, right
- 7:00? So, next, I will make the same request
- 7:01as before, but this time I'll include
- 7:03the goal, guardrails, and completion
- 7:04criteria. I'll use the same session for
- 7:06this. Please create the document again.
- 7:08I want to create a proposal to help
- 7:10President Hamasaki choose one of the
- 7:12three plans from Use Saisei. Please do
- 7:13not include any discounts or features
- 7:15not listed in the price sheet. Also, do
- 7:17not include any internal company matters
- 7:19or information about other clients.
- 7:21Please only use the figures found in the
- 7:22price list and hearing notes. The
- 7:24document should be about 5 to 7 pages
- 7:26long, covering the President's concerns
- 7:27, the recommended plan and the reason
- 7:29why, a comparison table of the three
- 7:31plans, costs, and next steps, and it
- 7:32should be finished when the President
- 7:34can make a decision after reading it
- 7:35once. This is how I'll ask. Also, since
- 7:37this is voice input, there's a typo, but
- 7:39the part about "President Hamasaki from
- 7:40Use Saisei" is the goal, and the part
- 7:42about "discounts or features not in the
- 7:43price list" represents the guardrails.
- 7:45So, for this document, the completion
- 7:47criteria is that it should be about 5 to
- 7:487 slides long. And by entering this, a 6
- 7:50-slide document was created like this.
- 7:52Oh, that's wonderful. Looking at the
- 7:54first page, it lists what needs to be
- 7:55decided in this document, and the second
- 7:56page shows the three concerns the
- 7:58president mentioned in the hearing memo
- 7:59. Then, on the third page, it clearly
- 8:00states that the recommendation is the "
- 8:01Standard" plan, and it uses arrows to
- 8:03connect the reasoning for that to each
- 8:04of the president's concerns. Looking at
- 8:05the fourth slide and beyond, it
- 8:07continues like this; clearly, this
- 8:08version is much better than the one
- 8:09before, isn't it? The reality is that
- 8:11even with the same AI and the same
- 8:12information, providing these three
- 8:14things—the objective, guardrails, and
- 8:15completion criteria—makes this much of a
- 8:17difference in the output. I think it's
- 8:19very easy to just tell an AI to "make it
- 8:21look good" when you're asking for
- 8:22something. But if you can at least
- 8:24articulate what "look good" means—which
- 8:26in this case is the objective—and
- 8:27communicate just that to the AI, I
- 8:29believe the accuracy of your output will
- 8:31continue to improve significantly. So,
- 8:33that concludes the second officially
- 8:35recommended usage: communicating the
- 8:37objective, guardrails, and completion
- 8:38criteria.
- 8:40Next, the third officially recommended
- 8:42usage is to have Claude ask you
- 8:43questions before it executes the task.
- 8:45What this means is that when you ask
- 8:47Claude to do something, you should have
- 8:48it ask you first if there's anything you
- 8:50forgot to mention or if anything is
- 8:51still unclear, and then proceed only
- 8:53after answering those questions. The
- 8:54reason is that if an AI encounters
- 8:56something it doesn't understand, it will
- 8:57fill in the gaps with its own
- 8:58assumptions and move forward. For
- 9:00instance, if you ask it to create a
- 9:01document based on this data, it wouldn't
- 9:03know whether the document is intended
- 9:04for your boss or for a client, right? In
- 9:06that case, Claude might just assume, "
- 9:08Given this content, it's probably for a
- 9:10boss," and move forward based on that
- 9:11assumption. And when that happens, a gap
- 9:13emerges between imagination and reality
- 9:15, ultimately lowering the accuracy of
- 9:16the output. So, the ideal state when
- 9:18delegating a task to Claude Code is to
- 9:19put everything into words so there's no
- 9:21room for Claude to fill in the blanks
- 9:23with imagination, but that's pretty
- 9:24difficult, isn't it? If we could do that
- 9:26, we wouldn't have any trouble. That's
- 9:27why you have Claude ask about the parts
- 9:29you haven't put into words first.
- 9:30Specific phrasing might be something
- 9:32like, "If you have any questions about
- 9:34this, please let me know, otherwise,
- 9:35please go ahead," or "In addition to
- 9:37this information, what else should I
- 9:38tell you to improve the accuracy of your
- 9:40output?" Honestly, just by saying this,
- 9:42Claude stops working based on
- 9:43assumptions, and you'll get the results
- 9:45you want without the output being
- 9:47off-target. In fact, Anthropic employees
- 9:49themselves have mentioned on their
- 9:50official blog that adding "Ask me
- 9:52everything you need to confirm before
- 9:53you start" is the most useful habit when
- 9:55using Claude Code. Really, just say, "
- 9:57If there is anything you don't
- 9:58understand, please ask me first; if you
- 10:00have no questions, please proceed." I
- 10:02personally think you could add this to
- 10:03every single request, so I really want
- 10:05you to adopt it as a standard habit. So
- 10:07, that was the third officially
- 10:09recommended usage: having Claude ask
- 10:10questions before execution.
- 10:13Next, the fourth officially recommended
- 10:15usage is to handle tasks without APIs
- 10:17through screen operation. Now, an API is
- 10:19like an entrance that allows you to
- 10:21control external services from Claude
- 10:22Code, and for services that have them,
- 10:24like X or YouTube, you can immediately
- 10:26integrate them with Claude Code, but
- 10:27there are services in the world that
- 10:29don't have this entrance. So, what do
- 10:31you do in those cases? That's where the
- 10:33method I mentioned earlier comes in:
- 10:34having Claude operate the screen
- 10:35directly. In fact, Claude Code can
- 10:37connect to Google Chrome and operate the
- 10:39browser itself—this is called "Claude in
- 10:40Chrome"—and the official documentation
- 10:42recommends this method when an API is
- 10:43unavailable. Even the creator of Claude
- 10:45, Anthropic, has stated that using this
- 10:47Claude screen control feature expands
- 10:48the range and quality of what you can
- 10:50create. So, I definitely want you to try
- 10:52it out, and for how to connect Claude to
- 10:53Chrome, check out this complete kit here
- 10:55. Just copy this and paste it into
- 10:57Claude Code, and it’ll be ready to use,
- 10:58so please follow these steps. That
- 11:00covers the official information, but
- 11:02I’ve taken it a step further by using a
- 11:04tool called Agent Desktop. Agent Desktop
- 11:06is released for free on GitHub by the
- 11:08web service company Vercel, and in short
- 11:10, it’s like an evolved version of the
- 11:12screen control I just showed you. The
- 11:14thing is, with the officially
- 11:15recommended method of having AI control
- 11:17your screen, it takes over your mouse
- 11:19movements, meaning you can't do your own
- 11:21work during that time. So, before I
- 11:23started using this Agent Desktop, I had
- 11:25to set screen control tasks to run when
- 11:27I wasn't using the PC, like right before
- 11:29bed or before heading to the gym.
- 11:31However, Agent Desktop doesn't take over
- 11:33your mouse, so you can continue your
- 11:34regular work while it handles
- 11:36screen-based tasks in the background.
- 11:38Also, it’s not limited to just one task
- 11:40; you can run two or more at once. While
- 11:42before you could only do one thing—like
- 11:45having it fetch info from LINE—with
- 11:47Agent Desktop, you can have it gather
- 11:49LINE info, research the latest AI news,
- 11:51and still do your own work
- 11:53simultaneously. I highly recommend this
- 11:55, so check out this complete kit right
- 11:56here. Installing Agent Desktop. If you
- 11:58copy this and paste it into Claude Code
- 12:00, it will handle everything from
- 12:01installation to the actual process. So
- 12:03that wraps up the fourth officially
- 12:05recommended use case: using screen
- 12:07control for tasks that lack an API.
- 12:09Next, the first recommended way to use
- 12:11it is to delegate heavy tasks to a
- 12:13separate AI and only receive the results
- 12:15. Heavy tasks here refer to jobs like
- 12:17reading a massive amount of documents,
- 12:19writing long articles, or generating
- 12:21many images. If you make the main Claude
- 12:23model handle all these high-processing
- 12:24tasks, the conversation becomes
- 12:26cluttered, and the accuracy of the
- 12:27output drops. So, the officially
- 12:29recommended method is to leave long
- 12:31research or large-scale output to
- 12:32another AI, or a "sub-agent," and have
- 12:34it return only the results to the main
- 12:36model. The way to delegate work to a
- 12:37sub-agent is incredibly simple: just
- 12:39tell it, "Delegate this task to a
- 12:40sub-agent and return only a short
- 12:41summary of the results to this chat."
- 12:43That is all you need to ask. I use these
- 12:45sub-agents quite a bit; for example,
- 12:46once I have a YouTube script, I set up
- 12:48multiple sub-agents to review it from
- 12:50various perspectives and have them
- 12:51report back if there are any points to
- 12:53address. Also, for this video, I was
- 12:55working on the script while
- 12:56simultaneously creating the complete kit
- 12:58offered as a bonus, so I used the
- 13:00sub-agent function to ensure there were
- 13:01no inconsistencies. However, since
- 13:03sub-agents simply double or triple your
- 13:05token usage, when I want to keep costs
- 13:07down, I don't split the work within
- 13:09Claude, but rather outsource it to other
- 13:11AIs. To explain what that means, I
- 13:13actually outsource two things: first,
- 13:14anything related to research, such as
- 13:16reading long documents, which I
- 13:18outsource to NotebookLM. Specifically,
- 13:20if I send a prompt like "Look up
- 13:21recommended washing machines with
- 13:23reviews" to Claude, it triggers
- 13:25NotebookLM via the cloud to conduct the
- 13:27research within NotebookLM itself. The
- 13:28reason I do this is basically a matter
- 13:30of role division; for instance,
- 13:32NotebookLM isn't as smart as Claude, but
- 13:34it is an AI that is exceptionally good
- 13:35at processing massive amounts of
- 13:37information. Plus, since it's free to
- 13:39use, you can summarize information while
- 13:41saving on Claude Code tokens. I've
- 13:43talked to many people, and I hear a lot
- 13:44of feedback that this integration with
- 13:46NotebookLM is excellent, so definitely
- 13:48go ahead and set it up while watching
- 13:50this video. As for how to connect it,
- 13:51you can use this from the complete kit.
- 13:53Just copy this to connect with
- 13:55NotebookLM and paste it into Claude Code
- 13:57, so please make use of it. I explain
- 13:59more effective ways to use this in this
- 14:00video here, so be sure to watch it after
- 14:02you're done with this one. So, that's
- 14:04the first bit of outsourcing; the second
- 14:06one is tasks related to creating images
- 14:08. For this, I leave it to DALL-E. DALL-E
- 14:10is an AI from OpenAI, the creators of
- 14:12ChatGPT, and I use it to create things
- 14:14like these diagrams and slide decks you
- 14:16see in my videos by giving instructions
- 14:18through Claude Code. The reason is
- 14:20simple: when it comes to creating images
- 14:22right now, DALL-E 3 is overwhelmingly
- 14:24higher in quality. By linking Claude
- 14:26Code and DALL-E like this, you can
- 14:28create stylish materials in an instant,
- 14:30so if you often have to prepare
- 14:32presentations, try linking these two.
- 14:33The linking method is the same as the
- 14:35NotebookLM one from earlier; it's this
- 14:37part of the complete kit. Just copy this
- 14:38and paste it into Claude Code to finish
- 14:40, so please give it a try. Also, I
- 14:42explain document creation in more detail
- 14:44in this video, so check that out after
- 14:46this one too. So, that covers the first
- 14:48officially recommended way to use it:
- 14:50offload heavy tasks to other AIs.
- 14:53Next, the sixth officially recommended
- 14:55way to use it is to adjust the effort
- 14:56based on the weight of the task. This "
- 14:58effort" is a setting that determines how
- 15:00deeply Claude Code thinks about a task.
- 15:02Most people probably ask for the same
- 15:03level of effort for every task, whether
- 15:05it's a minor job or one where mistakes
- 15:06would be a real problem. But as stated
- 15:08in Anthropic's official documentation, I
- 15:09highly recommend adjusting the effort
- 15:11level based on the weight of the task
- 15:12you're delegating. And when I talk about
- 15:14the weight of a task here, I don't mean
- 15:15the volume of work, but rather how much
- 15:17deep thinking is required. You can set
- 15:18this by clicking right here below the
- 15:20input field and selecting your desired
- 15:21effort level from there. So, how do I
- 15:23actually use this? For example, when
- 15:25writing YouTube scripts or creating
- 15:26presentations, I start with the initial
- 15:28brainstorming phase using the effort
- 15:30level recommended by the official
- 15:31guidelines. Once the direction is
- 15:32solidified to a certain extent, I move
- 15:34on to building out a more detailed
- 15:35structure—the skeleton, if you will.
- 15:37When I'm ready to create that, I set the
- 15:39effort to "Long-term" and then, once the
- 15:41structure is approved and I'm ready to
- 15:42write the script or build the slides, I
- 15:44set it to "Ultra" to finish it up. By
- 15:46the way, as a side note, it's generally
- 15:48advised not to switch between models
- 15:50like Opus or Sonnet within a single
- 15:52session, but since these effort levels
- 15:54are applied to each individual
- 15:56interaction, there's no problem at all
- 15:58with changing them within the same
- 15:59session. However, it might be hard to
- 16:01tell which effort level is suitable for
- 16:03your current task just from this video,
- 16:04so I’ve included an "Effort Setting
- 16:06Quick Reference Guide" in the complete
- 16:08kit. So, if you're ever in doubt, please
- 16:09take a look at that. That concludes the
- 16:11sixth officially recommended usage
- 16:13method: adjusting effort based on the
- 16:14weight of the task you're delegating.
- 16:16Moving on, the seventh officially
- 16:18recommended usage method is to give the
- 16:20AI the means to verify its own work.
- 16:22What this means is instead of just
- 16:23letting Claude build it and leaving it
- 16:25at that, you provide a checklist of
- 16:26requirements and make Claude revise its
- 16:28work until it passes every point. I mean
- 16:30, you should have it go as far as doing
- 16:31even that. Well, even if AI lets you
- 16:33generate a massive amount of output, if
- 16:34you have to check it with your own eyes
- 16:36every time, you’re just going to stay
- 16:37busy, right? That’s exactly why you
- 16:39should provide the criteria for checking
- 16:40to Claude. In fact, Boris Cherny
- 16:42mentioned that giving AI the means to
- 16:44verify its own work improves output
- 16:45accuracy by two to three times, and the
- 16:47official documentation also suggests
- 16:49providing a clear pass/fail checklist
- 16:50and continuing until it passes. I’ve
- 16:52actually adopted this method myself;
- 16:54when I write YouTube scripts, I don't
- 16:55look at them as soon as they're finished
- 16:57, but have them checked on Claude's end
- 16:59first. I’ve set it up so that it checks
- 17:01for things like logical consistency or
- 17:03monotonous sentence endings three times
- 17:05within Claude before presenting it to me
- 17:07. So, instead of checking everything
- 17:08myself every single time, only the items
- 17:10that pass the check are returned to me.
- 17:11The key to creating this checklist is to
- 17:13make it clear enough that anyone can
- 17:14tell if it's a pass or a fail, but then
- 17:16you might wonder how to actually build
- 17:18that kind of checklist for your own work
- 17:19. That’s where this complete kit comes
- 17:21in. "Create a Pass/Fail Checklist."
- 17:23Please copy this and paste it into
- 17:25Claude. By doing that, it will build a
- 17:27checklist tailored to your work, run a
- 17:29test grading, and then save it as a
- 17:30skill. So, if you’ve recently started
- 17:32using Claude and have been able to
- 17:33automate some of your tasks, but find
- 17:35that you still have to verify everything
- 17:36manually, so your workload hasn't
- 17:37actually decreased, then by all means,
- 17:39please try using the kit I just
- 17:40mentioned. So, that covers the seventh
- 17:42officially recommended usage. Which is
- 17:44giving AI the means to verify its own
- 17:46work.
- 17:47Next, the eighth officially recommended
- 17:49usage is to leave the final verification
- 17:51to a separate AI. This is because, with
- 17:53the seventh usage we just discussed,
- 17:54Claude is now able to determine whether
- 17:56its output passes or fails on its own.
- 17:57But this alone doesn't allow you to
- 17:59catch errors in areas not covered by the
- 18:01checklist. So, we need to have those
- 18:02checked too, but if you ask the same
- 18:04session that generated the output to
- 18:06also do the verification, it’s biased
- 18:08because it created the work itself. In
- 18:10short, even if there are things to fix,
- 18:12it might say there aren't any, kind of
- 18:14like when my grandma used to tell me I
- 18:16looked just like my uncle. It becomes
- 18:18extremely biased. That’s why for these
- 18:20final checks, you should have a
- 18:21different Claude check what the first
- 18:23one produced. This is what’s important.
- 18:25And by "a different Claude," I don't
- 18:26mean installing a new app, but rather
- 18:28starting a separate session. In fact,
- 18:30Anthropic's official documentation
- 18:31states that AIs tend to be lenient when
- 18:33grading their own work, so it's better
- 18:35to separate the creator role from the
- 18:36reviewer role. As for the prompt for the
- 18:38final check, here is the complete kit. "
- 18:40Final check of the completed work."
- 18:42Please try copying this, pasting it into
- 18:44Claude, and entering what you want
- 18:45verified below it. By using this prompt
- 18:47, you can, for example, have a business
- 18:49plan read by an investor, a customer,
- 18:51and a competitor all at once; this is a
- 18:53method the Claude team introduced on
- 18:55their blog. By automating the pass
- 18:56criteria and final verification using
- 18:58the seventh and eighth methods I
- 18:59discussed, your workload will decrease
- 19:01significantly and your business will
- 19:02accelerate, so please make use of the
- 19:04prompt I just showed you. So that
- 19:05concludes the eighth officially
- 19:06recommended way to use it. Which is to
- 19:08leave final verification to a different
- 19:10AI.
- 19:11Moving on, the ninth officially
- 19:13recommended way to use it is to shorten
- 19:14the "Claude System Prompt." This "
- 19:16Claude System Prompt" is like a set of
- 19:17instructions that Claude reads at the
- 19:19start of every conversation, where you
- 19:20write down things about yourself, work
- 19:22rules, and things you don't want it to
- 19:23do. Most people, when Claude doesn't
- 19:25listen to them, probably just keep
- 19:26adding more and more rules to their "
- 19:28claudemc" file. But actually, this is
- 19:29counterproductive; Anthropic's official
- 19:31documentation states that you should
- 19:32keep your "claudemc" file under 200
- 19:34lines. That's because the "claudemc"
- 19:35file is read every single time you have
- 19:37a conversation, so if it's too long, the
- 19:38important rules you really need followed
- 19:40just get buried. In short, writing more
- 19:41doesn't mean it will listen better;
- 19:42rather, keeping it short actually helps
- 19:44it follow your instructions properly. So
- 19:45, what should you actually write?
- 19:47According to the official documentation
- 19:48, just things Claude can't understand by
- 19:50reading the folder, or things you've had
- 19:51to warn it about twice. It says that
- 19:52these two things alone are enough for
- 19:54your "claudemc" file. Conversely, things
- 19:55Claude can understand by looking inside
- 19:57the folder can be read by Claude itself
- 19:58, so there's no need to write them down
- 19:59. However, those who use Claude Code a
- 20:01bit more might say, "No, even so, it
- 20:03doesn't listen, which is why I go out of
- 20:05my way to emphasize things with '
- 20:06IMPORTANT' or star marks." But that
- 20:08might actually be backfiring, and in
- 20:09fact, the official documentation states
- 20:11that you don't need to keep emphasizing
- 20:13things with the current model. This is
- 20:14exactly why you need to get your
- 20:16information from official documentation
- 20:18; AI is constantly evolving, so the "
- 20:20right answer" changes over time. For
- 20:22instance, writing detailed "claudemc"
- 20:23files or using star marks might have
- 20:25been the right approach back around Opus
- 20:274.6. But now we're at something like
- 20:29Opus 5.5, so the underlying assumptions
- 20:31have changed. The image of writing long
- 20:33, detailed "claudemc" files is like
- 20:35trying to make a middle schooler wear
- 20:36the clothes they wore in kindergarten
- 20:38just because they looked good back then
- 20:40. Let's keep up with its growth. I think
- 20:42this is actually quite common. People
- 20:43who started using Claude Code six months
- 20:45ago and are still using a "claudemc"
- 20:46file based on the specifications from
- 20:47back then. If that's the case, I hope
- 20:49this video encourages you to give your
- 20:51Claude.md file a makeover. As for how to
- 20:52do this, take a look at this part of the
- 20:54complete kit. Review and shorten your
- 20:56current Claude.md. If you copy this and
- 20:58paste it into Claude, it will review
- 21:00your Claude.md based on official
- 21:02standards. So, that covers the ninth
- 21:04officially recommended tip: shortening
- 21:06your Claude.md.
- 21:08Next is the tenth officially recommended
- 21:10tip: don't just create a project,
- 21:12nurture it. "Projects" are basically a
- 21:14system that lets you save step-by-step
- 21:15procedures like a manual for tasks you
- 21:17do repeatedly. By turning tasks like
- 21:18writing scripts or creating documents
- 21:20into projects, you can get consistent
- 21:22results every time with just a single
- 21:23request. The method recommended in the
- 21:25official documentation is to finish a
- 21:27task once without using a project, and
- 21:29then ask Claude at the end of the chat
- 21:31to turn that approach into a project.
- 21:32Because the successful conversation
- 21:34contains all your added explanations and
- 21:36the fixes for issues you encountered,
- 21:37you can incorporate all of that into the
- 21:39project. For the exact prompt to use,
- 21:41refer to this section of the complete
- 21:42kit. Turn a successful conversation into
- 21:44a project. I hope you'll copy and use
- 21:46this, so please make the most of it.
- 21:47However, I think many people just stop
- 21:49after they've created their project. But
- 21:51a project isn't finished once it's
- 21:53created; in many ways, it's more
- 21:54important to nurture it from there. In
- 21:56fact, the official Anthropic
- 21:58documentation introduces a way to
- 21:59nurture your projects by using them,
- 22:01seeing where you trip up, and then
- 22:02refining them. To help with this
- 22:04nurturing process, there's the "Project
- 22:06Creator," an official skill released by
- 22:07Anthropic for building your projects. If
- 22:09you add this, it will walk you through
- 22:11the process of testing and fixing the
- 22:12projects you've created. For this Skill
- 22:14Creator, you definitely want to be using
- 22:16Claude Code, so if you haven't set it up
- 22:17yet, check out this complete kit. I've
- 22:19included how to install the Skill
- 22:20Creator here, so please take advantage
- 22:22of it. That concludes the 10th
- 22:24officially recommended usage: don't just
- 22:26stop at creating a skill.
- 22:29Next, the 11th officially recommended
- 22:31usage is to set up hooks. A "hook" is a
- 22:33setting that forces a specific action to
- 22:35occur at certain times, like when Claude
- 22:38rewrites a file or finishes a task. As I
- 22:40mentioned with the 9th point, Claude
- 22:42will follow most rules if you set up the
- 22:44metadata correctly, but it isn't perfect
- 22:46. So, for things that absolutely must be
- 22:48followed, you need to set up a hook. In
- 22:50fact, Anthropic's official documentation
- 22:52states that for critical rules, you
- 22:54should use hooks rather than just
- 22:55metadata. Creating these hooks is super
- 22:57simple; you just describe them in your
- 22:59code. For example, you can tell it to
- 23:01generate a hook that always does "this"
- 23:02when "that" happens, and by just
- 23:04describing when and what to do, Claude
- 23:05Code will create the hook automatically
- 23:07. If you want to test if it works
- 23:09properly or learn how to remove a hook,
- 23:10check out this complete guide. Setting
- 23:12hooks for things you absolutely need to
- 23:14enforce. If you copy this and paste it
- 23:16into Claude Code, it will ask you
- 23:17questions one by one to create the hook
- 23:19and then have Claude test it out itself
- 23:21. So, if you find that Claude isn't
- 23:22following your instructions even after
- 23:24you've written the metadata correctly,
- 23:25please give this a try. That concludes
- 23:27the 11th officially recommended usage.
- 23:29That was setting up hooks.
- 23:32And finally, the 12th officially
- 23:33recommended usage is to verify five
- 23:35things before automating. So, what this
- 23:37means is that Claude Code has a "
- 23:39scheduled execution" feature that
- 23:40performs tasks automatically at set
- 23:42times, but you should check five things
- 23:44before entrusting it with a job. Because
- 23:45scheduled executions run while you're
- 23:47asleep or not watching, if Claude gets
- 23:49stuck along the way, it can't ask you
- 23:50for help. That is exactly why you need
- 23:52to decide on these five things
- 23:53beforehand and include them in your
- 23:54prompt so Claude doesn't get lost. These
- 23:56five points come from Anthropic's
- 23:58official documentation; the first is
- 23:59defining what "done" looks like. Second
- 24:01is a checklist to measure the criteria
- 24:02for success. Third is the scope of what
- 24:04it's allowed to touch. This essentially
- 24:06covers which files to read and what
- 24:07actions are prohibited. Fourth is when
- 24:09to stop; this includes deciding when to
- 24:10give up if a task won't finish, as well
- 24:12as setting a budget limit. And finally,
- 24:14the fifth point is where to send the
- 24:15results. These are the five pieces of
- 24:17information you should provide before
- 24:18setting up any scheduled execution or
- 24:19automation. I personally use scheduled
- 24:20executions for things like researching
- 24:22trending competitor videos, transcribing
- 24:24coaching or meeting recordings, and
- 24:25posting weekly group consulting archives
- 24:27to my membership site; I have various
- 24:28tasks running automatically like this.
- 24:30For every automated task I have here,
- 24:31I’ve specified all five: the completion
- 24:33criteria, the checklist, the prohibited
- 24:35actions, the limits, and the destination
- 24:37for the results. However, I know that
- 24:39deciding these five things every time
- 24:40can be a hassle, so for those moments,
- 24:42here is this complete kit. If you copy "
- 24:44Five Checks Before Automating" and paste
- 24:45it into Claude Code like this, it will
- 24:47fill in those five points and handle the
- 24:49scheduled execution setup for you. So,
- 24:51if you find yourself doing the same
- 24:52tasks every day but haven't automated
- 24:54them yet, pick one job to delegate and
- 24:56try using this prompt. So, that was the
- 24:5812th official recommendation: verify
- 25:00five things before automating. In this
- 25:02video, I’ve introduced all 12 ways to
- 25:04use Claude Code as recommended by
- 25:06Anthropic. If you’re going to start with
- 25:08just one thing today, please start by
- 25:09researching how professionals do it
- 25:11before you build, as I mentioned in the
- 25:12first tip. Try adding that one specific
- 25:14phrase I talked about. Truly, just
- 25:15adding that one phrase improves output
- 25:17accuracy more than you’d expect, and
- 25:18it’s something you can try right away,
- 25:20so I highly encourage you to use it. And
- 25:22once again, all the prompts and detailed
- 25:24settings used in this video are compiled
- 25:26in this "Official Ultimate Utilization
- 25:27Kit." So, if you haven’t received it yet
- 25:29, please add my official LINE account
- 25:30via the link at the top of the
- 25:32description or this QR code and send the
- 25:33message "Official Utilization." It’s set
- 25:35up to be sent to you automatically. Also
- 25:36, for those who want to learn more about
- 25:38the basics of Claude Code, I’m giving
- 25:40away a 20-chapter Claude Code
- 25:41Fundamentals Course for free in my 0-
- 25:42yen monthly online community, so please
- 25:44take advantage of that as well. You’ll
- 25:46receive information about the 0-yen
- 25:47community through the same LINE account
- 25:48you registered with earlier, so you can
- 25:50join from there. That wraps up this
- 25:52video. Thank you so much for watching
- 25:53until the end. I’ll see you all in the
- 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.