agent-skills-zh

中文 | English

Agent Skills for Claude Code, Codex and 20+ coding agents — built for Chinese-speaking developers. Translate English docs into Chinese, archive WeChat / Xiaohongshu / X content as Markdown, keep multi-repo Git hygiene, direct GPT-image prompts, and run autonomous improvement loops — one SKILL.md per capability, install only what you need.

面向中文开发者的 Agent Skills 集合。 英文技术文档精准翻译、公众号/小红书/X 内容存档、多仓库 Git 归位、GPT 生图提示词导演、AI 自治改进循环——每个能力一个 SKILL.md,按需安装,Claude Code 与 Codex 通用。

GitHub stars License: MIT Skills Claude Code Codex skills.sh

agent-skills-zh skill map

Quick start · 快速开始

# Works with Claude Code, Codex, Cursor, Copilot, Windsurf, Gemini CLI and more
npx skills add hanzhangzzz/agent-skills-zh            # pick skills interactively
npx skills add hanzhangzzz/agent-skills-zh -s doc-reader -g   # install one skill globally

Then just talk to your agent — every skill lists its own trigger phrases (English and Chinese) in its SKILL.md, e.g. /doc-reader <url>, 排版, 归位, 做点什么.

Skills · 技能一览

Skill What it does · 做什么 Use when · 何时用 Trigger · 触发
doc-reader Translate an English article/PDF into Chinese section by section, keep 100% of images, build a 3-column preview (original · translation · AI slides). 英文技术文档/PDF 章节级精准翻译,图片全保留,三栏本地预览 Reading long English blogs, papers, docs and wanting a faithful side-by-side Chinese version /doc-reader <URL or PDF> · 翻译这篇文章
wechat-article-md-local Save a WeChat Official Account article as local Markdown with images. 公众号文章下载为本地 Markdown,图片本地化 You receive an mp.weixin.qq.com link and want to archive, quote or analyze it auto on mp.weixin.qq.com links · 下载公众号文章
xiaohongshu-downloader Download a Xiaohongshu (RedNote) video and transcribe the voice-over with Whisper into Markdown. 小红书视频下载 + 口播逐字稿 You receive a Xiaohongshu video link and need the spoken content as text auto on xiaohongshu.com / xhslink.com links · 小红书视频转文字
x-article-download Download a tweet, a long-form X article or a whole account to Markdown. X/Twitter 单条或整账号批量下载 You receive an x.com link and want to archive or analyze it auto on x.com links · 下载推文
hkr-render WeChat publishing pipeline: Markdown → styled HTML (7 themes) → cover image → draft box, multi-article supported. 公众号排版 → 封面 → 推送草稿箱 Publishing WeChat articles; dark mode and mobile font sizes tuned 排版 · 微信排版 · /format
md2view Re-encode Markdown into a traceable two-column reading view (diagrams, flows, matrices) with every element anchored to the source. Markdown 重编码为可溯源的双栏阅读视图 Retrospectives, reports, specs and long docs that people need to absorb and share /md2view · 信息重组
gpt-image2-prompt-director Turn a weak idea into a production-grade GPT image2 brief, with a built-in benchmark. GPT image2 提示词导演 + 评测门禁 Avatars, sticker packs, infographics, covers, posters, product shots, or repairing a drifting prompt $gpt-image2-prompt-director · 生图提示词
repo-tidy Git tidy-up + parallel-task base: back to latest master, prune merged branches/worktrees, --new opens a task branch (parallel worktree when busy); SessionStart hook injects repo status. 仓库归位与并行任务底座 Multi-repo + multi-task AI work where task branches pile up 归位 · 开新任务 · repo tidy
repo-map Local repository map: self-healing index of all local git repos; a hook injects path + read/write role whenever a repo name is mentioned. 本地仓库地图,提到仓库名自动注入路径 Cross-repo references where the AI keeps asking for paths 仓库地图 · repo-map
harness Minimal Harness Engineering: Inspector → Worker → Reviewer loop driven by a shared TODO.md. 三角色 AI 自治改进循环 You want an agent to continuously inspect, fix and review a repo, on demand or on a schedule /harness · 启动 harness
do-something Answer feedback, choose work backed by real value evidence, finish and verify a durable outcome, or return NO-OP. do/* auto-merge requires separate value and execution verdicts. 自主推进项目:有价值且真正完成才交付,无事可做就停 An idle project where the agent may work unattended without manufacturing tasks to keep the loop busy /do-something · 做点什么 · 自己看着办
ci-review Every PR/MR gets an execution verdict; configured bot branches also need a separate value verdict before deterministic auto-merge. Skill Markdown is reviewed as behavior code and failed verdicts are red checks. CI 里的验证型代码审查机器人,机器人分支价值与执行双门禁 You want machine-verified execution, plus a value gate before unattended bot branches enter main /ci-review · 装一个 CI 代码审查
git-push-guard Hook-only plugin: intercepts direct pushes to master/main, asks for confirmation, per-repo allowlist. 纯 hook:直推默认分支拦截 You let an agent commit and want shared-branch discipline enforced auto on git push to master/main (plugin install only)
hook-test-kit Scaffold behavior-matrix tests for Claude Code hooks: scratch fixtures, stdin JSON, EMPTY/! assertion protocol, mutation-experiment finish. 给 hook 脚本补行为测试 写了/改了 hook 脚本要测试 · /hook-test-kit

效果一览 · Gallery

每张图都是真实产出或 skill 自带的真实文案/看板,不是示意图。

Install · 安装

1. npx skills (any agent · 推荐)

npx skills add hanzhangzzz/agent-skills-zh --list      # see what's inside
npx skills add hanzhangzzz/agent-skills-zh              # choose interactively
npx skills add hanzhangzzz/agent-skills-zh -s doc-reader -s md2view -g   # specific skills, user-level
npx skills update                                      # keep them fresh

npx skills detects Claude Code, Codex, Cursor, Copilot, Windsurf, Gemini CLI, Cline and others, and installs to the right directories.

2. Claude Code plugin marketplace (hooks auto-enabled · hook 自动生效)

/plugin marketplace add hanzhangzzz/agent-skills-zh

/plugin install repo-tidy@agent-skills-zh        # + SessionStart repo-status hook
/plugin install repo-map@agent-skills-zh         # + UserPromptSubmit repo-map hook
/plugin install git-push-guard@agent-skills-zh   # hook-only: guard master/main
/plugin install doc-reader@agent-skills-zh
/plugin install md2view@agent-skills-zh
/plugin install hkr-render@agent-skills-zh       # needs config.json (WeChat AppID/Secret), see its SKILL.md
/plugin install do-something@agent-skills-zh
/plugin install harness@agent-skills-zh
/plugin install gpt-image2-prompt-director@agent-skills-zh
/plugin install wechat-article-md-local@agent-skills-zh
/plugin install x-article-download@agent-skills-zh
/plugin install xiaohongshu-downloader@agent-skills-zh

One plugin = one skill, so you can install, update (/plugin update <name>@agent-skills-zh) and uninstall each independently. Plugins with hooks (repo-tidy, repo-map, git-push-guard) work immediately — no settings.json edits.

3. Copy manually · 手动复制

git clone https://github.com/hanzhangzzz/agent-skills-zh.git
cd agent-skills-zh
cp -r doc-reader ~/.claude/skills/     # Claude Code
cp -r doc-reader ~/.codex/skills/      # Codex
# same for any other skill directory

Notes · 说明:

  • repo-map auto-injection and repo-tidy session status rely on Claude Code hooks; on Codex the commands still work, the automatic injection does not.
  • hkr-render needs a config.json with your WeChat AppID/Secret before first use (see its SKILL.md).
  • doc-reader slide generation uses the local Codex CLI's built-in imagegen: codex must be on PATH and logged in. Use --no-ppt to skip slides.

4. Let your agent install it · 让 Agent 帮你装

Paste this to your agent:

Install the skill "doc-reader" from https://github.com/hanzhangzzz/agent-skills-zh
(English technical article/PDF → accurate Chinese, images kept, 3-column preview HTML)
安装 skill:hkr-render
描述:公众号完整管线:Markdown 排版(7 个主题)→ 封面 → 推送草稿箱(支持多图文)
安装源:https://github.com/hanzhangzzz/agent-skills-zh

Why these skills · 为什么是这些

  • Content that lives on Chinese platforms (WeChat, Xiaohongshu, plus X) is hard for agents to reach; the three downloaders turn links into local Markdown the agent can actually read, quote and translate.
  • Reading English docs in Chinese should not lose images, structure or terminology — doc-reader keeps all three and adds AI slides on top.
  • Multi-repo, multi-task agent work rots Git state fast; repo-tidy and repo-map keep every task on a clean branch with the right paths.
  • Agents that improve things while you are awayharness and do-something are two levels of autonomy, from a reviewed three-role loop to "just pick the best thing and ship it".

doc-reader

一条命令,把一篇英文技术文章变成可对照阅读的中文版:左栏原文、中栏译文(章节级翻译,术语一致,图片 100% 保留)、右栏每个章节一张 AI 知识卡幻灯片。下面是真实产出,输入是 Anthropic 的一篇客户案例:

doc-reader 三栏预览:原文 · 译文 · AI 幻灯片

幻灯片由本机 Codex CLI 内置 imagegen 生成,不需要图片 API Key。提示词经过单变量实验校准:画面可见汉字控制在 250 字以内,字形错误从平均 2.5 处/图降到 0(实验记录见 doc-reader/instructions.md)。

doc-reader 生成的 AI 幻灯片示例

/doc-reader https://claude.com/blog/how-warp-builds-self-improving-agents-on-claude
/doc-reader ./paper.pdf --no-ppt      # 只翻译,不生成幻灯片

hkr-render

Markdown 一键变成公众号可直接发布的排版,7 个主题,深色模式与手机端字号已调优;可选生成封面并推送到草稿箱,支持多篇合一条图文:

hkr-render 主题总览

排版这篇文章 ./article.md    # 打开主题画廊,用你的真实文章预览 7 个主题,选中即复制微信兼容 HTML
生成封面                     # 可选:封面亮度检查,浅色底自动加渐变暗化
推送到公众号草稿箱           # 可选:需要 config.json 里的 AppID/Secret

GPT Image2 Prompt Director

GPT Image2 Prompt Director overview

gpt-image2-prompt-director 不是普通"风格词生成器"。它更像一个会先做判断的图像创意总监:先识别你真正要交付的视觉资产,再把弱输入改写成能直接喂给 GPT image2 的完整 brief,并用内置 benchmark 检查 prompt 是否只是看起来完整、实际却导向错误。

真实出图示例

下面这些图片来自同一批 GPT 真实生成结果,源文件均为 regenerated-* 开头;为了 README 加载速度,这里嵌入的是 900px 版本。

字体海报 文物信息图
GPT image2 typography poster example GPT image2 relic infographic example
小红书卡片 日记漫画
GPT image2 Xiaohongshu card example GPT image2 diary comic example

它解决什么问题

很多生图失败不是模型不行,而是 prompt 从一开始就把任务定义错了:

  • 想要平台头像,却写成了电影感肖像。
  • 想要小红书卡片,却只堆了一串高级形容词。
  • 想要知识图鉴,却没有事实核验、信息层级和阅读路径。
  • 想要表情包,却没有一致角色、独立单格和情绪清晰度。
  • 想修 prompt,却不知道它到底缺角色、内容策划、视觉系统还是反失败约束。

这个 skill 会把"帮我做张高级图"拆成更可靠的结构:角色定位、反定义、输入契约、内部策划、内容策划、视觉系统、反失败约束、输出规格和自检。

能力文案板

适合人群

  • 内容创作者:公众号、小红书、X、视频封面和知识卡片。
  • 个人品牌作者:长期头像、IP 形象、表达系统。
  • 设计协作者:需要把模糊需求翻译成可执行视觉 brief。
  • Agent/Skill 作者:想用 benchmark 给生图 prompt 做回归评测。

核心能力

  • 弱点子增强:从一句主题、标题、对象或场景,扩展成完整生图 brief。
  • 无点子模式:用户只有身份、平台或目标时,先生成 10-20 个候选玩法,再选最强方向展开。
  • 头像/IP 专项:默认优先 3 x 3 头像探索,强调圆形裁切、80px 可读性、脸部优先和符号层级。
  • 表情包专项:默认 4 x 4 表情包系统,强调角色一致性、情绪清晰度和单格可用性。
  • artifact-first:先确定图像容器,比如信息图、图鉴、海报、卡片、产品图、头像资产,而不是先堆风格词。
  • 评测门禁:内置 40 个 benchmark cases、结构评分、能力覆盖评分和 hard gates。

一句话触发

$gpt-image2-prompt-director 帮我把"AI 时代的个人知识库"做成一张适合公众号头图的高级图片 prompt。

修复触发

$gpt-image2-prompt-director 下面这个头像 prompt 太像真人写真了,帮我修成可长期使用的平台头像资产,并解释问题在哪里。

效果试跑

仓库内置的 evaluator 可以验证 prompt 是否覆盖了必要结构和专项 hard gates。

cd gpt-image2-prompt-director
node scripts/eval_prompt_director.mjs \
  --case-id 40 \
  --prompt-file examples/readme-avatar-demo.md \
  --report /tmp/gpt-image2-prompt-director-readme-demo.md \
  --fail-under 80 \
  --strict

本 README 的头像案例已跑通:

Case 输入 结果
40 "公众号个人头像,有个性,有辨识度,略男性化,AI 先锋,独立作者,酷酷的;我不想要写实真人头像。" 总分 100,structure 100,capability 100,hard gates 通过

你也可以跑完整基准集,观察当前 prompt 框架和 archived expert prompts 的差距:

cd gpt-image2-prompt-director
node scripts/eval_prompt_director.mjs --gold --report /tmp/gpt-image2-prompt-director-gold-report.md

说明:顶部流程图是示意图;"真实出图示例"来自本地同批 regenerated-* GPT 输出。实际生成质量仍取决于最终模型、输入素材和提示词执行环境。

Harness

harness 是一个轻量的 Agent 自治循环:Inspector 负责巡检和拆任务,Worker 负责领取并修改,Reviewer 负责复审和验证。三者通过项目里的 TODO.md 共享状态,避免 Agent 在复杂仓库里"想到哪改到哪"。

能力文案板

适合人群

  • 想让 Agent 定期巡检代码质量和未完成工作的项目维护者。
  • 想把"发现问题 - 修复问题 - 复审验证"变成固定流程的工程团队。
  • 想让多个 Agent 通过 TODO.md 看板协作,而不是靠聊天上下文接力。

核心能力

  • Inspector:读取项目文档和当前状态,发现可执行任务,写入 TODO.md
  • Worker:领取 [待领取] 任务,修改代码,运行相关验证,再标记 [待审查]
  • Reviewer:检查 diff、验证结果和项目约束,通过后记录 Done Log,必要时提交。
  • Loop 模式:持续消化任务,直到没有可安全推进的项。
  • Cron 模式:在 Claude Code session 内定时触发 inspect/work/loop/full cycle。

一句话触发

/harness 这轮重点关注静默数据丢失、环境变量错误和缺少验证的代码路径。

常用命令

cd <project> && bash ~/.codex/skills/harness/inspector.sh
cd <project> && bash ~/.codex/skills/harness/worker-reviewer.sh --loop
HARNESS_PROJECT_DIR=/abs/path/to/project bash ~/.codex/skills/harness/inspector.sh

md2view

md2view 双栏同步阅读器

md2view 把一份 Markdown 重新编码成人类读得进去的视图——不是渲染加样式,而是抽出信息结构换一种编码(架构图 / 流程链 / 对比矩阵 / 标注卡),且每个视图元素都能一键回到原文出处。它是 md2html 的继任者:旧的转格式,新的转视图。

核心理念:md 是给 AI 和 git 的权威源,HTML 是给人的消费投影;人类该读的不是文字墙,是按信息类型选最优编码的信息设计;有损压缩 + 可回溯 = 无损(每个元素溯源回原文,左原文右重组双栏锚定同步)。v4 起,表达归模型自由创作(组件词汇 + 设计系统),核证归机器(data-sources 锚点 + 词法锚点 + 数字逐字 + 100% 来源覆盖),生产者必须截图自审闭环。

它和市面 md→html 工具正交:那些是美化器(优化第一眼好看),md2view 是可信重编码器(优化敢拿去做决策、敢溯源微调)。

一句话触发

/md2view 把这份复盘 md 变成双栏视图,左边原文右边重组

详见 md2view/README.md

Contributing · 贡献

Each skill is a directory with a SKILL.md (YAML frontmatter: name = directory name, description = what + when to use, ≤1024 chars) plus optional scripts/, references/, assets/. No hard-coded paths. See CLAUDE.md for the full checklist, then open a PR — issues and skill requests are welcome.

If a skill saved you time, a ⭐ helps other people find it.

Local development · 本地开发

# list available skills
find . -maxdepth 2 -name SKILL.md

# validate the Claude Code plugin manifest
claude plugin validate .

# run the GPT image2 prompt director sample evaluation
cd gpt-image2-prompt-director
node scripts/eval_prompt_director.mjs --case-id 40 --prompt-file examples/readme-avatar-demo.md --fail-under 80 --strict

Layout · 目录

.
├── .claude-plugin/marketplace.json   # Claude Code plugin marketplace (one plugin per skill)
├── MARKETPLACE.md                    # machine-readable registry for agents
├── doc-reader/                       # SKILL.md + instructions.md + scripts/
├── wechat-article-md-local/
├── xiaohongshu-downloader/
├── x-article-download/
├── hkr-render/
├── md2view/
├── gpt-image2-prompt-director/       # SKILL.md + examples/ + references/ + evals/ + scripts/
├── repo-tidy/
├── repo-map/
├── harness/                          # SKILL.md + prompts/ + inspector.sh + worker-reviewer.sh
├── do-something/
├── ci-review/
├── git-push-guard/                   # hook-only plugin
└── assets/readme/

License

MIT