Introduction to Agent Skills | Searching in Japanese with cccskills
機械翻訳 / Machine-translated
A
AI Friends
2日前・17分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever thought, "I heard you can add skills to Claude Code, but where do I find them?" Endlessly navigating English GitHub repositories is exhausting. This article introduces how to use "cccskills," a site you can search in Japanese, and how to try out popular skills.
Agent Skills is a system for teaching AI how to perform specific tasks. All it contains is a single folder with a file called SKILL.md inside.
According to the official Claude Code documentation, you write a name and description at the top of the file (in a settings section called YAML front matter), followed by the steps written in Markdown below. The location is ~/.claude/skills/ if you want to use it across your entire computer, or .claude/skills/ if it's just for a specific project.
What's interesting is how skills are loaded. Normally, only the "description" of each skill is held in Claude's memory, and the full content is only loaded when it becomes necessary. This means that even if you install many skills, performance is unlikely to slow down.
There are two ways to invoke a skill: typing /skill-name to call it directly, or having Claude use it automatically when the content of a request matches the skill's description.
This format was originally created by Anthropic and is now an open standard published on the official Agent Skills website. The same skills can also be used with other compatible tools such as Cursor, Gemini CLI, and GitHub Copilot.
Skills are convenient once you know how to use them. However, there was a problem: there were simply too many to search through. That's where cccskills comes in.
Its tagline is "Search for what you want to do, in Japanese." It aggregates skills published on GitHub and adds Japanese descriptions to them.
As of the top page in October 2026, approximately 91,000 skills are listed, of which 2,845 have Japanese descriptions. While that's still only a portion of the total, highly rated skills tend to have Japanese descriptions.
Searching is simple: just type Japanese words like "議事録" (meeting minutes), "小説" (novel), or "SEO" into the search box. Popular search terms include code review, PowerPoint, Excel, design, and testing.
Categories include 14 classifications such as "Development," "Writing & Creativity," "Business," and "Data & AI." Each skill's page includes a Japanese overview and suggested use cases.
On each skill's page on cccskills, the installation command is displayed. Just copy it and paste it into your terminal. For example, it looks something like this:
npx skills add https://github.com/coji/natural-japanese --skill natural-japanese
Node.js must be installed to use this command.
If commands aren't your thing, you can also ask Claude to "install this skill." You provide Claude with the skill's GitHub path, have it check the contents, and then ask it to place the files in ~/.claude/skills/natural-japanese/.
Here's an important note: because skills are "instruction sets for AI," there's no guarantee that malicious instructions haven't been mixed in. The cccskills pages themselves state that they do not guarantee the quality or safety of skills.
The license for publicly available skills follows each individual repository. Before installing, check the SKILL.md and any bundled scripts on the original GitHub with your own eyes.
Now for a concrete example. Let's look at the natural-japanese page listed under "Writing & Creativity" on cccskills.
The author is coji, and at the time of writing, the GitHub repository had approximately 1,900 stars (1,899 on the page), 53 forks, and an MIT license.
It's a skill for writing and rewriting professional Japanese text to make it more readable. Target use cases include meeting minutes, research reports, internal manuals, proposals, emails, and slide outlines. It also aims to suppress the "AI-ish" quality that tends to creep into AI-generated text.
Imagine a newcomer tasked with taking meeting minutes. They transcribe a recording, have AI summarize it into minutes, but the result feels somehow unnatural—not quite something they'd feel comfortable showing their manager.
In a situation like that, asking to "make this more readable" invokes the skill. According to the repository description, it is automatically triggered by requests such as "make this more readable," "write with the conclusion first," or "fix unnatural expressions."
There are two modes: "Quick" for everyday documents, and "Full" for important documents or longer texts. Since Full takes more time, it tells you the estimated duration before it begins.
There's also a score usage mode that only evaluates and scores the text without rewriting it—giving you a 0–100 naturalness rating for something you wrote yourself.
The process follows the steps of planning, writing, automated checking, and repeated revision. A script called lint.py is used for checking, identifying formulaic phrasing and translationese. Whether to revise or retain each instance is left to the AI's contextual judgment.
Finally, the text is scored on six criteria. The passing standard is 90 or more on each criterion, with an average of 92 or higher. The bundled package contains 26 files in total.
Running it requires a Python tool called uv. Also, semantic.py, which performs semantic detection, is an experimental feature that downloads approximately 1 GB of model data on first use. If you just want to try it out casually, you can skip this feature.
cccskills isn't the only way to obtain skills. Here's a summary of the differences:
A reasonable approach is to start by installing popular ones from cccskills, then create your own once you're comfortable.
Work involving Japanese documents—meeting minutes, reports, emails—is endless. Since many overseas skills are designed for English documents, skills that include procedures tailored for Japanese are valuable. Skills like natural-japanese, which even include scripts for checking Japanese phrasing, are prime examples of this.
cccskills also has creator-facing features: free registration with just an email address, the ability to publish skills by uploading a zip file, bundled "skill sets" of multiple skills, and support for paid sales.
It can serve as a place to package and share know-how you've refined on the job in Japan. However, at the time of writing, the skill set count displayed on the page was 0—it appears to still be in a growing phase.
A. natural-japanese is published under the MIT license. cccskills itself also states that publicly available GitHub skills follow the license of each respective repository. That said, some creators sell skills for a fee, so check what's displayed on the page.
A. Agent Skills is an open standard, and many tools support it, including Cursor, Gemini CLI, and GitHub Copilot. However, depending on the skill, some may be written with a specific tool in mind.
A. Yes. To install, you can ask Claude to "install this skill," and after installation, you simply make requests in ordinary Japanese. However, some skills—like natural-japanese—require tools such as uv to run.
A. Read the SKILL.md and any bundled scripts before installing. Look for any inexplicable external communications or file deletion commands. cccskills does not guarantee the quality or safety of skills.
As a next step, try searching "議事録" (meeting minutes) on cccskills and open the page for a skill that catches your interest.
This article is a cross-post from AI Friends.
