Skill Picker
Discover and recommend the right skills at the right moment.
What this skill does with npx: It runs npx skills find [query] to search the skills registry and display results. It does NOT run npx skills add on its own — install commands are always shown to the user for manual execution or explicit confirmation.
Core Philosophy
Don't match keywords. Understand intent.
The difference:
- - Keyword match: User says "GitHub" → search "github"
- Intent match: User says "我的PR没人review" → understand they need async code review automation → recommend
gh-issues + coding-agent as a combination
Always ask: What outcome does the user actually want? Then find the shortest path to that outcome using available skills.
Security principle: npx skills find is used for searching only. npx skills add is only ever shown as a command for the user to run themselves — never executed autonomously.
When to Activate
Explicit triggers (user asks directly)
- - "find a skill for X"
- "is there a skill that can..."
- "how do I automate X"
- "can you do X" (where X is specialized)
- "我想自动化X" / "有没有能做X的skill"
- "帮我找一个可以..."
Implicit triggers (proactive — suggest only, never auto-install)
Suggest when you detect:
- - Repetition: User manually does the same thing 2+ times in a session → suggest a skill that could help
- Friction: User copy-pastes between tools → suggest an integration skill
- Capability gap: User asks for something you can't do well alone → find a skill that fills the gap
- Workflow opportunity: User completes step 1 of a multi-step process → suggest skills for steps 2-3
Example proactive suggestion:
User manually summarizes a PDF, then pastes it somewhere else.
→ "I notice you're doing this manually. The summarize skill could automate this flow. Want me to show you how to install it?"
Important: Proactive suggestions are informational only. Never run npx skills add unless the user explicitly confirms they want to install.
Intent Analysis Framework
Before searching, extract three things:
CODEBLOCK0
Intent Translation Examples
| User says | Outcome | Obstacle | Best skill search |
|---|
| "help me review this PR" | Merged, high-quality code | Manual review takes time | INLINECODE9 + INLINECODE10 |
| "I need to send updates to my team" |
Team is informed | Manually composing messages |
gog (Gmail) or
feishu-doc |
| "summarize this YouTube video" | Get key points fast | Can't watch 2-hour video |
summarize |
| "我想监控竞品动态" | Know when competitors ship | No alert system |
blogwatcher +
summarize |
| "帮我管理邮件" | Zero inbox | Too many emails |
himalaya +
gog |
| "我的日历太乱了" | Clear, organized schedule | Manual scheduling |
gog calendar + proactive-agent |
Search Process
Step 1: Clarify if ambiguous (max 1 question)
If the intent is unclear, ask ONE focused question:
"Are you looking to automate this completely, or just get help doing it faster?"
Never ask more than one clarifying question. Make a reasonable assumption and proceed.
Step 2: Search with intent-based queries
CODEBLOCK1
Use outcome words, not task words:
- - ❌
npx skills find email (too broad) - ✅
npx skills find email automation inbox zero (outcome-focused) - ❌ INLINECODE21
- ✅
npx skills find pull request review automation notify (outcome-focused)
Try 2-3 different query angles before concluding nothing exists.
Step 3: Evaluate results with quality signals
When presenting skills, always include:
- - Download count (from skills.sh) — proxy for reliability
- Combination potential — does it work better with another skill?
- Fit score — how well does this match the user's specific outcome?
Step 4: Recommend combinations, not just single skills
Most outcomes need more than one skill. Think in workflows:
CODEBLOCK2
Response Format
When skills are found
CODEBLOCK3
⚠️ Always present npx skills add commands for the user to run manually — never execute them on behalf of the user without explicit confirmation.
When no skills are found
Don't just give up. Three-part response:
CODEBLOCK4
Common Skill Categories & Best Searches
Productivity & Communication
| Need | Search query | Top skills |
|---|
| Email management | INLINECODE24 | INLINECODE25 , INLINECODE26 |
| Calendar scheduling |
calendar scheduling meeting |
gog |
| Team updates |
team notification summary |
gog,
feishu-doc |
| Document creation |
document write generate |
feishu-doc,
prose |
Development & Code
| Need | Search query | Top skills |
|---|
| PR review | INLINECODE35 | INLINECODE36 , INLINECODE37 |
| Issue management |
github issues automate |
github,
gh-issues |
| Code changes |
diff review changes |
diffs |
| Multi-agent coding |
agent code build |
coding-agent,
prose |
Knowledge & Research
| Need | Search query | Top skills |
|---|
| Summarize content | INLINECODE46 | INLINECODE47 |
| Monitor updates |
monitor rss blog watch |
blogwatcher |
| Search places |
places search location |
gsplaces |
| Agent memory |
memory knowledge graph |
ontology |
System & Automation
| Need | Search query | Top skills |
|---|
| Self-improvement | INLINECODE54 | INLINECODE55 |
| Find more skills |
discover install skills |
skill-picker |
| Security check |
security audit skill vetting |
skill-vetter |
| Smart routing |
route model ai select |
acp-router |
Hardware & Home
| Need | Search query | Top skills |
|---|
| Music control | INLINECODE62 | INLINECODE63 |
| Sleep tracking |
sleep pod temperature |
eightctl |
| Camera capture |
camera rtsp snapshot |
camsnap |
Proactive Skill Suggestions (Mid-conversation)
If during any task you notice a skill would significantly improve the outcome, mention it naturally:
"There's a skill called [name] that would help here. Want me to show you the install command?"
Then wait for the user to confirm before providing the command. Do not run any install command proactively.
Threshold for suggesting: skill would save >2 manual steps OR significantly improve output quality.
Don't suggest more than one new skill per conversation unless the user is explicitly building a workflow.
Commands Reference
CODEBLOCK5
ℹ️ npx skills add is intentionally shown without -g or -y flags. Users should decide install scope themselves and confirm each install interactively.
Quality Signals to Always Check
Before recommending, verify at skills.sh:
CODEBLOCK6
Language Support
This skill operates in the user's language.
- - If user writes in Chinese → respond in Chinese, search in English, present results in Chinese
- If user writes in English → respond in English throughout
Chinese intent → English search translation examples:
- - "帮我自动回复邮件" → INLINECODE72
- "监控竞品网站更新" → INLINECODE73
- "整理我的GitHub Issues" → INLINECODE74
- "把会议内容发给团队" → INLINECODE75
Skill Picker
在合适的时机发现并推荐合适的技能。
该技能通过 npx 执行的操作:运行 npx skills find [query] 搜索技能注册表并显示结果。它不会自行运行 npx skills add——安装命令始终展示给用户,供其手动执行或明确确认。
核心理念
不要匹配关键词。要理解意图。
区别如下:
- - 关键词匹配:用户说GitHub → 搜索 github
- 意图匹配:用户说我的PR没人review → 理解他们需要异步代码审查自动化 → 推荐 gh-issues + coding-agent 组合
始终要问:用户真正想要的结果是什么? 然后利用可用技能找到达成该结果的最短路径。
安全原则:npx skills find 仅用于搜索。npx skills add 始终仅作为供用户自行运行的命令展示——绝不自主执行。
何时激活
显式触发(用户直接询问)
- - 找一个能做X的技能
- 有没有技能可以……
- 我该如何自动化X
- 你能做X吗(其中X是专业领域)
- 我想自动化X / 有没有能做X的skill
- 帮我找一个可以……
隐式触发(主动建议——仅建议,绝不自动安装)
当检测到以下情况时提出建议:
- - 重复操作:用户在会话中手动执行相同操作2次以上 → 建议一个能提供帮助的技能
- 摩擦点:用户在工具之间复制粘贴 → 建议一个集成技能
- 能力缺口:用户要求你做你单独无法很好完成的事情 → 寻找能填补缺口的技能
- 工作流机会:用户完成多步骤流程的第一步 → 为第2-3步建议技能
主动建议示例:
用户手动总结一个PDF,然后粘贴到其他地方。
→ 我注意到你在手动操作。summarize 技能可以自动化这个流程。需要我告诉你怎么安装吗?
重要:主动建议仅为信息性。除非用户明确确认想要安装,否则绝不运行 npx skills add。
意图分析框架
在搜索之前,提取三个要素:
- 1. 结果 — 用户想要拥有/感受/达成什么?
- 障碍 — 什么阻碍了他们达成该结果?
- 上下文 — 他们是谁?他们已经在使用什么工具?
意图转换示例
| 用户说 | 结果 | 障碍 | 最佳技能搜索 |
|---|
| 帮我review这个PR | 合并高质量代码 | 手动审查耗时 | pr review + gh-issues |
| 我需要给团队发更新 |
团队知情 | 手动撰写消息 | gog (Gmail) 或 feishu-doc |
| 总结这个YouTube视频 | 快速获取要点 | 无法观看2小时视频 | summarize |
| 我想监控竞品动态 | 了解竞品发布 | 没有提醒系统 | blogwatcher + summarize |
| 帮我管理邮件 | 收件箱归零 | 邮件太多 | himalaya + gog |
| 我的日历太乱了 | 清晰有序的日程 | 手动安排 | gog 日历 + proactive-agent |
搜索流程
第一步:如有歧义则澄清(最多1个问题)
如果意图不明确,问一个聚焦的问题:
你是想完全自动化这个操作,还是只是想更快地完成它?
绝不多问一个以上的澄清问题。做出合理的假设并继续。
第二步:使用基于意图的查询进行搜索
bash
这是该技能自主执行的唯一 npx 命令
npx skills find [基于意图的查询]
使用结果导向的词,而非任务导向的词:
- - ❌ npx skills find email(过于宽泛)
- ✅ npx skills find email automation inbox zero(聚焦结果)
- ❌ npx skills find github
- ✅ npx skills find pull request review automation notify(聚焦结果)
在断定没有结果之前,尝试2-3个不同的查询角度。
第三步:通过质量信号评估结果
展示技能时,始终包含:
- - 下载量(来自 skills.sh)—— 可靠性的代理指标
- 组合潜力—— 是否与另一个技能配合更好?
- 匹配度评分—— 与用户特定结果的匹配程度如何?
第四步:推荐组合,而非单一技能
大多数结果需要不止一个技能。从工作流角度思考:
用户结果:让我的团队自动获取更新
单一技能回答(较弱):
→ gog (Gmail)
组合回答(更强):
→ gog(发送更新)+ summarize(浓缩内容)+ proactive-agent(自动触发)
→ 这三个技能共同处理完整工作流:检测 → 总结 → 发送
回复格式
找到技能时
我找到了 [N] 个与你想要做的事情相匹配的技能。
最佳匹配:[技能名称](★ [星级] · [下载量] 次安装)
[一句话:该技能能实现什么结果,而非技术上的功能]
要安装,请自行运行以下命令:
npx skills add [所有者/仓库@技能]
了解更多:https://skills.sh/[路径]
[如果推荐组合:]
与 [技能名称2] 配合效果更佳:
两者结合可处理 [完整工作流描述]。
要安装两者,请运行:
npx skills add [技能1]
npx skills add [技能2]
在安装之前,如果你需要更多详情,请告诉我。
⚠️ 始终将 npx skills add 命令展示给用户手动运行——未经用户明确确认,绝不代表用户执行。
未找到技能时
不要就此放弃。采用三部分回复:
- 1. 确认:我搜索了 [X] 和 [Y],但没有找到完全匹配的。
- 2. 最接近:最接近的可用技能是 [名称]——它涵盖了 [重叠部分],
但无法处理 [缺口部分]。
- 3. 选项:
a) 我可以用内置能力直接处理这个——需要我继续吗?
b) 这看起来是一个足够普遍的需求,应该存在相应的技能。
你可以创建一个:npx skills init [建议名称]
c) 直接查看 skills.sh——生态系统更新频繁。
常见技能类别与最佳搜索
生产力与沟通
| 需求 | 搜索查询 | 顶级技能 |
|---|
| 邮件管理 | email inbox automation | himalaya, gog |
| 日历安排 |
calendar scheduling meeting | gog |
| 团队更新 | team notification summary | gog, feishu-doc |
| 文档创建 | document write generate | feishu-doc, prose |
开发与代码
| 需求 | 搜索查询 | 顶级技能 |
|---|
| PR审查 | pull request review code | gh-issues, coding-agent |
| Issue管理 |
github issues automate | github, gh-issues |
| 代码变更 | diff review changes | diffs |
| 多智能体编码 | agent code build | coding-agent, prose |
知识与研究
| 需求 | 搜索查询 | 顶级技能 |
|---|
| 内容总结 | summarize url pdf video | summarize |
| 监控更新 |
monitor rss blog watch | blogwatcher |
| 地点搜索 | places search location | gsplaces |
| 智能体记忆 | memory knowledge graph | ontology |
系统与自动化
| 需求 | 搜索查询 | 顶级技能 |
|---|
| 自我改进 | agent learn improve | self-improving-agent |
| 查找更多技能 |
discover install skills | skill-picker |
| 安全检查 | security audit skill vetting | skill-vetter |
| 智能路由 | route model ai select | acp-router |
硬件与家居
| 需求 | 搜索查询 | 顶级技能 |
|---|
| 音乐控制 | sonos speaker audio | sonoscli |
| 睡眠追踪 |
sleep pod temperature | eightctl |
| 摄像头捕捉 | camera rtsp snapshot | camsnap |
主动技能建议(对话中)
如果在任何任务过程中,你注意到某个技能能显著改善结果,自然地提及:
有一个叫 [名称] 的技能在这里会很有帮助。需要我展示安装命令吗?
然后等待用户确认后再提供命令。不要主动运行任何安装命令。
建议阈值:技能能节省 >2 个手动步骤或显著提高输出质量。
除非