claude-code-skills-zh: A Chinese-Language Skill Library for Claude Code

The gap this fills for Chinese-speaking Claude Code users

Claude Code skills are reusable instruction modules. The README states the mechanism plainly: a single SKILL.md file is enough to give Claude a specialised role such as code reviewer, security auditor or documentation writer. The barrier is not the format, which is small, but writing the instructions. A skill that produces useful output needs a defined input scope, a stated output format and enough domain detail to steer the model away from generic advice. Most developers will not write that from scratch, and the material available in English assumes an English-language workflow.

This repository targets that second group. Its stated audience is developers who want to add professional capabilities to Claude Code quickly, individual developers and teams looking for skills that are readable in Chinese and installable by copying, and people who want ready-made skills for recurring tasks: code review, README writing, API testing, performance analysis and refactoring suggestions. The repository is MIT licensed and the primary language listed is Python, which suggests some skills ship supporting scripts rather than only Markdown. The README does not enumerate those scripts, so treat that as something to check per skill directory.

What actually ships: five original skills versus the curated tables

The repository has two halves and they are not equivalent. The first half is five original skills the README highlights for first-time visitors: zh-code-reviewer, which the README says outputs a Chinese review report; zh-readme, which the README describes as analysing a project before writing a Chinese README; api-tester, which the README says generates and verifies API tests against a real contract after confirming environment and authorisation; refactor-advisor, which the README says finds code smells and gives actionable refactoring suggestions; and perf-profiler, which the README says locates performance bottlenecks using a reproducible baseline and profiler evidence.

The second half is a long table of third-party projects with one-line descriptions and star counts. That table is a reading list, not a component of the library. Nothing in the supplied material indicates the entries were installed, reviewed or version-pinned by the maintainer. The star figures printed beside them (values in the hundreds of thousands for projects like superpowers and openclaw) are inconsistent with what those repositories plausibly have, which is a reason to distrust the table as a source of fact rather than a reason to distrust the five original skills. The description text also carries emoji and marketing phrasing throughout. Judge the repository on the skill directories, because those are the part with a defined interface.

How a skill is structured and how Claude Code picks it up

The README gives one architectural fact and no more: a skill is a directory containing SKILL.md, and installing means placing that directory under ~/.claude/skills/. Everything downstream follows from that. Claude Code reads the skill's front matter to decide when the skill is relevant, then loads the instruction body when it activates. The repository does not document the front matter fields in the material provided, so the authoritative description of the format is the skill files themselves.

This is the point where the README stops being useful and the directories start. A reader deciding whether to adopt should open skills/zh-code-reviewer/SKILL.md and read the front matter and the instruction body. Two things to look for: whether the name and description fields are specific enough that the skill activates on the right requests rather than on every code-related prompt, and whether the body defines an output contract. For a review skill, an output contract means a fixed report shape (findings grouped by severity, each with a file and line reference) rather than a request to "review this code". The README's descriptions of the five skills suggest this kind of contract exists, particularly for api-tester, which the README says confirms environment and authorisation before executing anything. That ordering matters: a test-generating skill that runs requests without a confirmation step is a liability on a production endpoint.

Installing from GitHub, or through the mirror when that fails

The README gives three installation routes and all three are copy operations. The first is a clone followed by a recursive copy of every skill directory into the user-level skills folder:

git clone https://github.com/laolaoshiren/claude-code-skills-zh.git cp -r claude-code-skills-zh/skills/* ~/.claude/skills/

The second is the same clone routed through a proxy host for readers who cannot reach GitHub directly:

git clone https://gh-proxy.com/https://github.com/laolaoshiren/claude-code-skills-zh.git cp -r claude-code-skills-zh/skills/* ~/.claude/skills/

The third avoids git entirely by downloading the archive, either from https://github.com/laolaoshiren/claude-code-skills-zh/archive/main.zip or the mirrored equivalent at https://gh-proxy.com/https://github.com/laolaoshiren/claude-code-skills-zh/archive/main.zip.

Two practical notes the README does not spell out. The recursive copy puts every skill in the repository into your skills directory at once, including any you do not want active; copying individual directories gives you control over what Claude Code can invoke. Second, the proxy route hands your clone request to a third-party host, so the archive you receive is only as trustworthy as that host. The README presents the mirror as a convenience for network-restricted users, which is a fair description, but it is a supply-chain decision and should be made deliberately.

Where this repository is the wrong tool

The curated half ages badly and there is no mechanism in the material to keep it honest. The README claims continuous maintenance and the badge shows an update date, but the third-party table carries no per-entry timestamp, no version pin and no note about what was checked. A skill listed a year ago may have been renamed, archived or rewritten. If your reason for opening this repository is to find installable third-party skills, you are better served by the upstream directories the table itself points to, such as awesome-claude-code or the official Anthropic skills repository, where the entry and the project live in the same place.

There is a second, sharper limitation. The original skills are written for Chinese output. zh-code-reviewer produces a Chinese review report and zh-readme produces a Chinese README. That is the point of the repository, but it means these skills are a poor fit for a team whose review comments, documentation and commit conventions are in English. Mixing them in produces bilingual artefacts that nobody wants to maintain. The api-tester and perf-profiler skills are less language-bound in their output, but their instructions are still written in Chinese, so a reviewer who cannot read the skill body cannot audit what the agent has been told to do. For a regulated codebase, an unauditable instruction file is a blocker regardless of how good the skill is.

Compared with a general awesome-list

The obvious alternative is a pure index such as awesome-claude-code, which the README itself lists. The difference in approach is real. An awesome-list is a pointer collection: it tells you a project exists and links to it, and the project's own repository is the source of truth for installation and behaviour. This repository inverts that for its five original skills. The skill and its documentation ship together, so the install command, the directory name and the instructions are consistent by construction.

The cost of that inversion is scope. A general index can list hundreds of projects because it does no maintenance on any of them; this repository can only vouch for what it writes. The README's curated table tries to have both, and that is where the quality drops. If you want discovery, use a general index. If you want a working skill for Chinese code review, the original directory is the reason to be here, and the table is noise. A second alternative worth naming is writing your own SKILL.md. Given that the format is one file, a team with specific review conventions may get more from a 200-line skill written in-house than from any general-purpose reviewer skill, because the in-house version can encode the conventions the team actually enforces.

Licence, upkeep and what to check before you commit

The repository is MIT licensed, which permits commercial use, modification and redistribution provided the licence and copyright notice are retained. That is the whole of the licence implication here; a skill directory copied into ~/.claude/skills/ is a copy, and if you redistribute the repository or a modified version of it you carry the notice with it. This is not legal advice, and the curated table links to third-party projects under their own licences, which this repository's MIT licence does not cover. If you install a third-party skill from the table, check that project's licence separately.

Upkeep is the open question. The README states the repository is updated regularly with new skills, README changes and category restructuring, and the badge shows a recent date. What is not visible in the supplied material is a release history: no releases were retrieved, so there is no versioned artefact to pin. That means upgrading is a re-clone and re-copy, and any local edits you made to a skill file will be overwritten by the recursive copy command. Keep your modifications in a separate directory or track them in your own repository before you run cp -r again. If you only need one skill, clone once and copy that directory, which sidesteps the overwrite problem entirely.

Editorial conclusion

Adopt it if you write code with Claude Code and want Chinese-language skill definitions for review, README drafting, API testing, refactoring and profiling, installed by copying directories into ~/.claude/skills/. Skip it if you need English-language skills or a maintained index of third-party projects, since the curated tables are copied descriptions with no verification dates. Before installing, open skills/zh-code-reviewer/SKILL.md and confirm the front matter and instruction body match your Claude Code version, because that file, not the README, is what the tool actually reads.

Official sources

1. Issues
2. laolaoshiren/claude-code-skills-zh on GitHub
3. License: MIT
4. Project website
5. README