find-skills

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Find Skills

查找技能

Find the right agent skill for a task by searching every major registry at once and presenting each board on its own native metric — instead of trusting a single site's leaderboard.

通过同时搜索所有主流注册表，并按各平台原生指标展示结果，为任务找到合适的Agent技能——而非仅依赖单个网站的排行榜。

When to use this skill

何时使用本技能

The user asks "how do I do X" and X is a common task that likely has a published skill
"find a skill for X" / "is there a skill for X" / "what should I install for X"
The user wants to extend agent capabilities (testing, design, deployment, a specific model/API, a domain workflow)
You're about to build something from scratch that a battle-tested skill might already cover

用户询问“我该如何做X”，且X是常见任务，很可能已有发布的技能
“找一个用于X的技能” / “有没有用于X的技能” / “我该为X安装什么”
用户想要扩展Agent能力（测试、设计、部署、特定模型/API、领域工作流）
你正打算从零构建某功能，但已有经过实战检验的技能可覆盖该需求

Why multi-source matters

多源搜索的重要性

Each registry shows only its own slice, with different signals:

Registry	Search	Signals it exposes	Blind spot
skills.sh	`GET /api/search?q=`	lifetime installs, source repo	no stars, no summary, install count lags ~2.5h
clawhub.ai	`/api/search` + `/api/skill?slug=`	summary, installs, stars, versions	smaller corpus than skills.sh
GitHub	`gh search repos --topic {claude-skills,agent-skills,claude-code-skills}`	repo stars, description, maintenance	repo-level, not skill-level; only topic-tagged repos

A skill can rank #1 on one site with 50 installs while a 1,300-install equivalent sits unranked on another. Searching only one registry gives a biased answer. This skill queries all three in parallel, ranks each board by its own native metric, flags skills that appear on both registries (matched on the normalized name, both directions — so

face-swap

↔

faceswap

line up), and shows the top of every board so you see the whole ecosystem — not one site's leaderboard.

每个注册表仅展示自身的内容切片，且信号维度不同：

注册表	搜索方式	暴露的信号	盲区
skills.sh	`GET /api/search?q=`	累计安装量、源码仓库	无星标数、无摘要、安装量延迟约2.5小时
clawhub.ai	`/api/search` + `/api/skill?slug=`	摘要、安装量、星标数、版本	内容量少于skills.sh
GitHub	`gh search repos --topic {claude-skills,agent-skills,claude-code-skills}`	仓库星标数、描述、维护状态	基于仓库维度而非技能维度；仅包含带主题标签的仓库

某个技能可能在一个网站以50次安装量排名第一，而另一个拥有1300次安装量的同类技能在其他网站却未上榜。仅搜索一个注册表会得到有偏差的结果。本技能会并行查询这三个平台，按各平台原生指标对结果排序，标记在多个注册表中出现的技能（通过标准化名称匹配，双向匹配——例如

face-swap

与

faceswap

会被识别为同一技能），并展示各平台的榜首结果，让你了解整个生态，而非单个网站的排行榜。

How to run it

运行方式

Run the bundled aggregator with the user's need as the query:

bash

bash scripts/find.sh "<what the user needs>"

Examples:

bash

bash scripts/find.sh "react performance"
bash scripts/find.sh "pdf form filling" --limit 8
bash scripts/find.sh "video generation" --scan 5
bash scripts/find.sh "deploy to vercel" --no-scan --json

Flags:

```
--limit N
```
— results per source (default 10; non-numeric falls back to 10)
```
--scan K
```
— security-scan the top K installable candidates on each registry (default 2, max 5)
```
--no-scan
```
— skip the security scan (faster; the scan adds a few seconds because it fetches each candidate's real SKILL.md)
```
--json
```
— emit machine-readable JSON instead of the formatted report
```
-h
```
/
```
--help
```
— print usage and exit

The script needs

curl

and

jq

. It uses

gh

for the GitHub section and for fetching skills.sh skill bodies, and

unzip

for clawhub skill bodies during the scan — all optional and degraded gracefully if absent. No API keys are required.

使用捆绑的聚合器，以用户需求作为查询词运行：

bash

bash scripts/find.sh "<用户需求>"

示例：

bash

bash scripts/find.sh "react performance"
bash scripts/find.sh "pdf form filling" --limit 8
bash scripts/find.sh "video generation" --scan 5
bash scripts/find.sh "deploy to vercel" --no-scan --json

参数：

```
--limit N
```
— 每个源的结果数量（默认10；非数值参数会 fallback 到10）
```
--scan K
```
— 对每个注册表中排名前K的可安装候选技能进行安全扫描（默认2，最大5）
```
--no-scan
```
— 跳过安全扫描（速度更快；扫描会增加几秒耗时，因为需要获取每个候选技能的真实SKILL.md文件）
```
--json
```
— 输出机器可读的JSON格式，而非格式化报告
```
-h
```
/
```
--help
```
— 打印使用说明并退出

脚本需要

curl

和

jq

。GitHub部分以及获取skills.sh技能内容会用到

gh

，扫描clawhub技能内容会用到

unzip

——这些工具都是可选的，如果缺失会自动降级处理。无需API密钥。

Run adjacent-term queries — one search is not enough

运行相邻术语查询——单次搜索不够

The registries index by a skill's name, not its meaning. A single query will miss great skills filed under a sibling term. For any non-trivial need, run 2–3 searches across the adjacent vocabulary before you conclude, then pool the results.

This is not optional polish — it routinely changes the answer. Example: searching

"ui ux design"

tops out around 19k installs, but

"frontend design"

surfaces

anthropics/skills/frontend-design

at 443k installs — the single best skill for the same need, completely invisible to the first query.

Pick adjacent terms by domain, e.g.:

UI/UX →

frontend design

design system

web design

tailwind shadcn

dashboard ui

testing →
```
e2e
```
,
```
playwright
```
,
```
unit tests
```
,
```
test automation
```
deploy →
```
deployment
```
,
```
ci cd
```
,
```
docker
```
,
```
vercel
```
/
```
kubernetes
```
docs →
```
documentation
```
,
```
readme
```
,
```
api docs
```
,
```
changelog
```

Treat the union of these runs as your candidate pool, then apply the rubric below to the whole pool — not to one query's results.

注册表按技能的名称索引，而非语义。单次查询可能会错过归类在同类术语下的优质技能。对于任何非简单需求，在得出结论前运行2-3次相邻词汇的搜索，然后汇总结果。

这不是可选的优化步骤——它通常会改变最终答案。例如：搜索

"ui ux design"

的最高安装量约为19k，但搜索

"frontend design"

会找到

anthropics/skills/frontend-design

，其安装量高达443k——这是同一需求下的最佳技能，但在第一次查询中完全不可见。

根据领域选择相邻术语，例如：

UI/UX →

frontend design

design system

web design

tailwind shadcn

dashboard ui

测试 →
```
e2e
```
,
```
playwright
```
,
```
unit tests
```
,
```
test automation
```
部署 →
```
deployment
```
,
```
ci cd
```
,
```
docker
```
,
```
vercel
```
/
```
kubernetes
```
文档 →
```
documentation
```
,
```
readme
```
,
```
api docs
```
,
```
changelog
```

将多次搜索的结果合并作为候选池，然后对整个候选池应用以下评估标准——而非单次查询的结果。

How to read the output

如何解读输出

Design principle: no invented "quality score." There is no composite number you have to trust. Each board is ranked by that board's own native popularity metric, and the only computed value shown is

match %

— the share of the user's query words that appear in a skill's name/summary. That's a transparent relevance hint, not a verdict.

The report, top-down:

Banner + telemetry — the query, a source health line (
```
skills.sh ✓ 5·1.2s   clawhub ✓ 0·2.7s   github ✓ 8·1.1s
```
) showing which boards responded / hit count / latency, and how many skills are installed locally. A
```
✗ failed
```
/
```
– skipped
```
source means partial results — say so.
▶ TOP OF EACH BOARD — the #1 entry on each board by that board's native metric (skills.sh installs / clawhub installs·stars / GitHub match%·stars). This is keyword + popularity, NOT a vetted pick — it's a starting point, and it can be wrong (a 100%-keyword-match repo may be off-topic once you read it). Always confirm by reading before you forward it. If a board has no match it says so.
✓ ALREADY ON YOUR MACHINE — appears only when a result is already installed. A positive nudge: tell the user they already own a good skill and needn't install anything. (Heuristic: matches the result name against folder names under
```
~/.agents/skills
```
/
```
~/.claude/skills
```
; a skill installed under a renamed folder can be missed.)
📦 skills.sh — ranked by installs. 1-click installable. Each row shows its source repo (so same-named skills from different repos are distinguishable).
```
(also on clawhub)
```
marks a normalized-name match on the other registry. Top
```
K
```
carry a risk badge.
🪝 clawhub — ranked by installs · stars. 1-click installable.
```
(also on skills.sh)
```
marks the reverse cross-post. Top
```
K
```
carry a risk badge.
🐙 github — repos across the
```
claude-skills
```
/
```
agent-skills
```
/
```
claude-code-skills
```
topics (merged, de-duped), ranked by match% then stars, so a high-star but off-topic repo (e.g. a 14k-star tool that merely mentions "UI") sinks below a 100%-match repo. Only topic-tagged repos appear — untagged skill repos are invisible here. Not 1-click installs — review before use.
📚 curated lists — human-vetted awesome-lists for a final sanity check.

设计原则：不使用虚构的“质量分数”。无需信任任何合成数值。每个平台的结果都按该平台自身的原生流行度指标排序，唯一展示的计算值是

匹配度%

——即用户查询词出现在技能名称/摘要中的比例。这是一个透明的相关性提示，而非最终结论。

报告从上到下包括：

标题栏 + 遥测信息 — 查询词、源健康状态行（例如
```
skills.sh ✓ 5·1.2s   clawhub ✓ 0·2.7s   github ✓ 8·1.1s
```
），展示各平台是否响应/结果数量/延迟，以及本地已安装的技能数量。
```
✗ failed
```
/
```
– skipped
```
表示该源返回部分结果——需告知用户。
▶ 各平台榜首结果 — 每个平台按原生指标排名第一的条目（skills.sh按安装量 / clawhub按安装量·星标数 / GitHub按匹配度%·星标数）。这是关键词+流行度的结果，并非经过审核的推荐——只是一个起点，可能存在错误（例如100%关键词匹配的仓库实际内容可能偏离主题）。在推荐前务必阅读内容确认。如果某个平台无匹配结果会明确说明。
✓ 本地已安装 — 仅当结果已安装时显示。这是一个积极提示：告知用户他们已拥有合适的技能，无需再安装。（判断逻辑：将结果名称与
```
~/.agents/skills
```
/
```
~/.claude/skills
```
下的文件夹名匹配；如果技能安装时重命名了文件夹，可能会被遗漏。）
📦 skills.sh — 按安装量排序。支持一键安装。每行显示源码仓库（以便区分同名但不同仓库的技能）。
```
(also on clawhub)
```
表示该技能在另一个注册表中有标准化名称匹配的条目。排名前K的条目带有风险标识。
🪝 clawhub — 按安装量·星标数排序。支持一键安装。
```
(also on skills.sh)
```
表示反向跨平台发布的条目。排名前K的条目带有风险标识。
🐙 github — 来自
```
claude-skills
```
/
```
agent-skills
```
/
```
claude-code-skills
```
主题的仓库（合并、去重），按匹配度%然后星标数排序，因此高星标但偏离主题的仓库（例如14k星标的工具仅提及“UI”）会排在100%匹配的仓库之后。仅包含带主题标签的仓库——未标记的技能仓库不会显示在此处。不支持一键安装——使用前需审核。
📚 精选列表 — 人工审核的优质列表，用于最终的合理性检查。

How to choose the right skill (decision rubric)

如何选择合适的技能（决策标准）

The script gathers evidence; you make the call. Don't dump the table and ask the user to decide — that's not helpful. Form a clear, defensible recommendation. The professional move is to be opinionated and show your reasoning grounded in real signals.

Step 1 — Read before you judge. Never recommend from metadata alone. Open the actual SKILL.md / README of the top 2–3 candidates (the script already fetches bodies for the scanned ones; for GitHub repos fetch the README). Ask: does it actually do the user's specific task, or just share keywords? A 10k-install skill that's off-topic loses to a 200-install one that nails it.

Step 2 — Weigh the signals, in this order:

Fit — does the skill's documented behaviour match the real need? (read the trigger/scope, not just
```
match %
```
). This dominates everything else.
Popularity as social proof — among genuinely-fitting skills, higher installs/stars means more people validated it. Use the board-native numbers.
Depth & maintenance — does the body cover the user's specific sub-need? Is there a recent version, real examples, a clear scope? Thin or abandoned skills lose.
Safety — a
```
⛔ RISKY
```
scan result disqualifies unless the user accepts the risk knowingly;
```
⚠ caution
```
needs a heads-up.
Cross-posting — present on both registries is a mild positive (independently published/maintained).

Step 3 — Break ties by preferring the one you actually read and can vouch for, with the narrowest clear scope and fewest surprising dependencies.

Step 4 — Deliver a verdict, not a menu:

One primary recommendation with a one-sentence why citing concrete signals ("12k installs, and its SKILL.md covers form-filling specifically, which is your case").
1–2 alternatives framed by need ("pick this instead if you want X").
If the best option is a GitHub repo rather than a 1-click registry skill, say so and still recommend it — note it needs manual review/install.
Only say "nothing fits" when you've read the top candidates and they genuinely don't. Then offer to do the task directly or scaffold a new skill.

State native facts, never an invented score. A professional answer reads like: "
ui-components
(529 installs) is my pick — I read it; it covers shadcn + Radix + design tokens + forms, exactly your case, and its tools are read-only. The 363★
ai-design-components
repo looks tempting but it's a 76-skill full-stack grab-bag, not UI-focused — skip it unless you want everything." "Here are the numbers, you choose" is not.

脚本负责收集证据；由你做出决策。不要直接输出表格让用户自行选择——这毫无帮助。要给出清晰、有依据的推荐。专业的做法是给出明确观点，并基于真实信号说明理由。

步骤1 — 先阅读再判断。永远不要仅根据元数据推荐。打开排名前2-3的候选技能的真实SKILL.md / README（脚本已获取扫描条目的内容；GitHub仓库需自行获取README）。问自己：它是否真正满足用户的具体任务，还是仅包含关键词？ 10k安装量但偏离主题的技能，不如200安装量但完全匹配需求的技能。

步骤2 — 按以下优先级权衡信号：

匹配度 — 技能的文档描述是否符合真实需求？（阅读触发条件/适用范围，而非仅看
```
匹配度%
```
）。这是最重要的因素。
流行度作为社交证明 — 在真正匹配的技能中，更高的安装量/星标数意味着更多人验证过它。使用平台原生数据。
深度与维护状态 — 内容是否覆盖用户的具体子需求？是否有最新版本、真实示例、明确的适用范围？内容单薄或已废弃的技能应排除。
安全性 —
```
⛔ RISKY
```
扫描结果会取消推荐资格，除非用户明确接受风险；
```
⚠ caution
```
需要提前告知用户。
跨平台发布 — 在多个注册表中出现是一个轻微的积极信号（独立发布/维护）。

步骤3 — 打破平局的方法是优先选择你实际阅读过、可以担保的技能，且该技能的适用范围最窄、依赖项最少。

步骤4 — 给出结论，而非选项列表：

一个主要推荐，并用一句话说明理由，引用具体信号（例如“12k安装量，且其SKILL.md明确涵盖表单填充，完全符合你的需求”）。
1-2个按需求划分的替代选项（例如“如果你需要X功能，可以选择这个”）。
如果最佳选项是GitHub仓库而非一键安装的注册表技能，明确说明并仍然推荐它——注意需要手动审核/安装。
仅当你阅读了顶级候选技能后发现确实没有匹配的，才说“没有合适的技能”。然后主动提出直接完成任务或帮助搭建新技能。

陈述客观事实，而非虚构分数。专业的回答应该是：“我推荐
ui-components
（529次安装）——我已阅读其内容；它涵盖shadcn + Radix + 设计令牌 + 表单，完全符合你的需求，且其工具为只读模式。363★的
ai-design-components
仓库看起来很吸引人，但它是包含76个技能的全栈合集，并非专注于UI——除非你需要全套功能，否则不推荐。” 而不是“给你这些数据，你自己选”。

The security scan

安全扫描

For the top

results the script fetches the actual SKILL.md (skills.sh → GitHub raw via the source repo; clawhub → its download zip) and greps for risk signals:

Flag	Level	Meaning
`curl-pipe-install`	⛔ risky	pipes a remote script straight into `sh` / `bash` — the #1 audit-failing pattern
`eval-remote` / `base64-pipe-exec`	⛔ risky	executes fetched or obfuscated code at runtime
`broad-tool-grant`	⚠ caution	`allowed-tools` grants `Bash(*)` or unrestricted tool access
`reads-secrets`	⚠ caution	references `~/.ssh` , `~/.aws` , `.env` , private keys
`solicits-credentials`	⚠ caution	asks the user to paste an API key / token / password

✓ clean

= none found;

? unscanned

= body couldn't be retrieved (rank beyond

, private repo, or no body endpoint). A badge is a heuristic prompt to read the skill yourself before recommending, not a guarantee either way. Never recommend a

⛔ RISKY

skill without explicitly warning the user what it does.

对于排名前K的结果，脚本会获取真实的SKILL.md文件（skills.sh → 通过源码仓库获取GitHub原始文件；clawhub → 通过下载zip文件），并排查风险信号：

标识	级别	含义
`curl-pipe-install`	⛔ 高风险	将远程脚本直接通过管道传入 `sh` / `bash` 执行——这是排名第一的审计失败模式
`eval-remote` / `base64-pipe-exec`	⛔ 高风险	在运行时执行获取的或混淆的代码
`broad-tool-grant`	⚠ 需注意	`allowed-tools` 授予 `Bash(*)` 或无限制的工具访问权限
`reads-secrets`	⚠ 需注意	引用 `~/.ssh` , `~/.aws` , `.env` , 私钥等敏感信息
`solicits-credentials`	⚠ 需注意	要求用户粘贴API密钥/令牌/密码

✓ clean

= 未发现风险；

? unscanned

= 无法获取内容（排名超出K、私有仓库或无内容端点）。标识是一个启发式提示，提醒你在推荐前自行阅读技能内容，而非绝对的安全或危险保证。永远不要在未明确告知用户风险的情况下推荐

⛔ RISKY

技能。

Recommending and installing

When nothing good turns up

无合适结果时的处理

If no result clears a reasonable bar, say so plainly, offer to do the task directly with general capabilities, and mention the user can scaffold their own skill (

npx skills init <name>

). Don't oversell a weak match.

如果没有结果达到合理标准，直接告知用户，主动提出使用通用能力直接完成任务，并提及用户可以搭建自己的技能（

npx skills init <name>

）。不要过度推销匹配度低的技能。

Security & privacy

安全与隐私

The script only issues read-only HTTP GETs to public APIs (
```
skills.sh
```
,
```
clawhub.ai
```
,
```
raw.githubusercontent.com
```
) and read-only
```
gh
```
queries. It sends nothing but the query string and writes only temp files (a result JSON and, during the scan, candidate SKILL.md bodies it discards on exit).
It requires no API keys or tokens.
The built-in security scan is itself a defense: it inspects candidates before you trust them. But it is a heuristic — a clean badge is not a security audit.
Installed skills run with full agent permissions. Treat any discovered skill as untrusted third-party code: review its
```
SKILL.md
```
and any bundled scripts before running it, and be wary of low-install skills from unknown authors that request broad tool access or fetch remote code at runtime.
This skill never auto-installs anything; installation is always an explicit, user-confirmed step.

脚本仅向公共API（
```
skills.sh
```
,
```
clawhub.ai
```
,
```
raw.githubusercontent.com
```
）发送只读HTTP GET请求，并执行只读
```
gh
```
查询。仅发送查询字符串，且仅写入临时文件（结果JSON以及扫描期间的候选SKILL.md内容，退出时会自动删除）。
无需API密钥或令牌。
内置的安全扫描本身就是一种防御机制：在信任候选技能前先检查其内容。但这只是启发式检查——
```
clean
```
标识并不等同于安全审计。
已安装的技能拥有Agent的全部权限。将任何发现的技能视为不受信任的第三方代码：运行前审核其
```
SKILL.md
```
和所有捆绑脚本，警惕来自未知作者、安装量低且请求广泛工具访问权限或在运行时获取远程代码的技能。
本技能永远不会自动安装任何内容；安装始终是明确的、需用户确认的步骤。

find-skills

Original

Translation

Find Skills

查找技能

When to use this skill

何时使用本技能

Why multi-source matters

多源搜索的重要性

How to run it

运行方式

Run adjacent-term queries — one search is not enough

运行相邻术语查询——单次搜索不够

How to read the output

如何解读输出

How to choose the right skill (decision rubric)

如何选择合适的技能（决策标准）

The security scan

安全扫描

Recommending and installing

推荐与安装

When nothing good turns up

无合适结果时的处理

Security & privacy

安全与隐私

See also

相关链接