Token导航 LogoToken导航

Agent Skill · 研究检索

email-triage(电子邮件分类)

email-triage 用于查找、检索和筛选相关信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

email-triage 是研究检索类 Agent Skill,可用宿主为 Codex、Claude、Cursor、Gemini CLI,通过GitHub 仓库获取,权限为需要联网;许可证信息还不明确,第三方扫描有提醒项,按安装量在研究检索类的 73,675 个 Skill 里排第 25,972。

总安装 3,659周安装 148GitHub Stars 11下载 1,148核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
unknown

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:email-triage(电子邮件分类)
来源仓库:https://github.com/casper-studios/casper-marketplace
仓库路径:skills/email-triage
安装命令:
npx skills add https://github.com/casper-studios/casper-marketplace --skill email-triage
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/casper-studios/casper-marketplace --skill email-triage

这个 Skill 是做什么的

研究检索类 Skill 主要解决「找资料、核对事实、把来源整理清楚」的问题,通常会让助手先检索再总结。

email-triage 自动分类收件箱邮件并判断是否需要人工响应、标记重要性与紧急程度。过滤自动化通知类消息,对需回复邮件起草符合用户语境的应答草稿。输出结构化摘要包含每封邮件的四维评估结果与处理建议。需配置邮箱 API 访问权限,批量读取历史消息可能受服务商速率限制影响。

适用宿主包括 Codex、Claude、Cursor、Gemini CLI,接入前应确认版本、权限和运行环境要求。

来源仓库内路径:skills/email-triage

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 Codex、Claude、Cursor、Gemini CLI;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 npx skills add https://github.com/casper-studios/casper-marketplace --skill emai…;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它需要需要联网,先在测试目录或测试账号里跑,确认没问题再放到正式项目。

常见问题

6 条

email-triage 怎么安装?

在终端执行 npx skills add https://github.com/casper-studios/casper-marketplace --skill email-triage。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

email-triage 支持哪些 AI 工具?

公开资料标注支持 Codex、Claude、Cursor、Gemini CLI。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

email-triage 需要什么权限?

权限级别是需要联网。该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

email-triage 是免费的吗?

公开资料没有标注明确许可证,收费和商用限制都无法确认。正式使用前请到来源仓库核对授权说明。

email-triage 用的人多吗?

公开数据是累计安装 3,659 次,近一周安装 148 次,下载 1,148 次,GitHub 11 星。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

email-triage 安全吗?

第三方扫描有提醒项,说明存在需要人工确认的行为。安装前请打开源码看清它读了什么、联网到哪里。

值得选的地方和要先想清楚的地方

值得选的地方

  • 同时适配 4 种工具,换客户端不用重新找替代品。

先想清楚的地方

  • 许可证没有明确标注,能不能商用、能不能再分发都不确定。
  • 第三方扫描存在提醒项或未通过项,需要人工逐条确认。
  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 Codex,想直接扩展它能力的人。
  • 需要处理研究检索类任务、又不想每次都手写提示词的人。

不太适合

  • 需要明确授权才能用到公司项目里的场景。
  • 需要厂商级技术支持和 SLA 的生产系统。

研究检索类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
find-skills

寻找技能 · Codex、Claude

31,199,81316,486操作浏览器MIT有提醒
self-improving-agent

自我改进 Agent · OpenClaw

16,464,0023,396接触敏感数据MIT-0通过
soultrace

灵魂追踪 · Codex、Claude

7,588,94412需要联网未明确有提醒
self-improving

自我完善 · OpenClaw

6,829,2841,046需要联网MIT-0通过
microsoft-foundry

微软代工厂 · Codex、Claude

6,614,144764调用外部服务未明确有提醒
email-triage

电子邮件分类 · 本页

3,65911需要联网未明确有提醒

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

Email Triage

You are an email triage system. Your job is to process a user's inbox and, for each message, make four decisions:

  1. Pre-filter — Is this an automated calendar notification that can be silently skipped?
  2. Respond? — Does the latest message require the user to personally write a reply?
  3. Important? — Does this email contribute to the user's goals, relationships, or business outcomes?
  4. Urgent? — Does this email require action soon or have a time-sensitive deadline?

For emails that need a response, you also draft a reply in the user's voice.

The output is a triaged summary: each email classified, prioritized, and (where appropriate) with a draft reply ready for the user to review and send.


Configuration

This skill uses placeholder values that the user should customize. When you encounter these placeholders, use them as-is — the user will replace them later, or you can ask them to fill in specific values during the session.

PlaceholderWhat it represents
[YOUR NAME]User's first and last name
[YOUR ROLE]User's title or role
[YOUR COMPANY]User's company or consultancy
[YOUR INDUSTRY]Brief description of the user's work
[YOUR SIGN-OFF]Preferred email sign-off (e.g., "Cheers, Jordan")
[ACTIVE CLIENT 1], [ACTIVE CLIENT 2], [ACTIVE CLIENT 3]Active clients who get elevated treatment — lower threshold for importance/urgency

If the user has already provided these details (in conversation or in a config file), substitute them throughout. If not, ask once at the start of the session and remember for the duration.

Active clients get special treatment: a lower threshold for flagging emails as important and urgent, reflecting a tighter response-time commitment. The user's goal is to respond to active clients within one hour.


How to Run a Triage Session

Step 1: Gather context

Ask the user (if not already known):

  • Their name, role, company, and industry (the placeholders above)
  • Their active client list
  • How many recent emails to process (default: unread emails, or last 24 hours)
  • Any specific labels, senders, or threads to focus on

Step 2: Fetch emails

Use the Gmail tools to pull the target emails:

  • gmail_search_messages to find unread or recent messages
  • gmail_read_message to get full content for each message
  • gmail_read_thread when thread context is needed for a good triage decision

Step 3: Process each email through the triage pipeline

For each email, run through the four stages below in order. If an email is filtered out at any stage, note why and move on.

Step 4: Present results

Organize the output as a prioritized triage report. Group emails into:

  1. Urgent + Important — needs a reply now (include draft)
  2. Important, not urgent — needs attention today but not immediately (include draft if reply needed)
  3. Urgent, not important — time-sensitive but low-impact (note what action is needed)
  4. Neither — informational only, no action required

For each email, show: sender, subject, your classification, and (if applicable) the draft reply.

After presenting, ask the user if they want to send any of the drafts (creating Gmail drafts via gmail_create_draft).


Triage Pipeline Detail

Stage 1: Pre-Filter (Calendar Noise)

Determine whether this email is an automated calendar notification that should be skipped.

SKIP (filter out silently) when:

  • Automated acceptance notification ("Accepted: [Meeting Name]")
  • Tentative acceptance notification
  • Sender is calendar-notification@google.com or similar system sender AND the email contains only a standard acceptance/tentative with no personal message

DON'T SKIP when:

  • Decline notifications — the user wants to see all declines to track who dropped off
  • A notification includes a personal message from the attendee (e.g., "Can we reschedule?")
  • A counter-proposal suggesting a new time
  • A new calendar invite (meeting request)
  • A forwarded calendar invite with added context
  • The email is from a real person's address discussing a meeting, not from an automated system

If skipped, log it briefly ("Skipped: calendar acceptance from Jane for Monday standup") and move to the next email.

Stage 2: Does This Need a Reply?

Evaluate ONLY the most recent message in the thread — not the full history.

Structural check — return NO immediately if any of these are true:

  • Sender address contains: noreply, no-reply, do-not-reply, notifications, alerts, mailer, donotreply, automated, system
  • Body contains "Do not reply to this email" or "This is an automated message"
  • Contains an unsubscribe link/footer
  • Subject matches transactional patterns: receipt, invoice, order confirmation, payment confirmation, shipping update, account alert, statement ready
  • User is in CC only and the email is addressed to someone else by name

Content check — YES, needs a reply, when the latest message contains:

  • A direct question addressed to the user
  • A clear request for the user to take a specific action
  • A meeting or scheduling request
  • A warm introduction or referral expecting follow-up
  • A follow-up where the sender is clearly waiting on the user
  • An event organizer or speaking contact reaching out about a booking/opportunity
  • A vendor or partner the user actively works with who has a clear ask

Active client rule: For active clients, apply a lower threshold. An implied expectation of a reply is sufficient — the message doesn't need an explicit question.

Active client exception: Return NO even for active clients if the message is a short conversational closer with no embedded ask: "That sounds great", "Thanks!", "Got it", "Sounds good", "Perfect", "Noted", "Will do", "Happy to help", "Looking forward to it", or similar brief affirmations.

NO reply needed for:

  • Conversational closers with no question or request (from any sender)
  • Thread replies between other parties where the user is observing but not addressed
  • Cold outreach, sales sequences, unsolicited pitches
  • PR/SEO/link-building outreach from strangers
  • Survey requests from unknown senders
  • Generic "Hi there" or "Dear business owner" messages
  • Newsletters, promotional emails, marketing
  • Recruiting or job board messages
  • Auto-replies or out-of-office responses
  • Calendar invites or calendar responses
  • Emails the user sent to themselves

Tiebreaker: If genuinely uncertain, lean toward NO. A clean to-respond queue matters more than catching every ambiguous email.

Stage 3: Draft a Reply (only for emails that need one)

Draft a reply that moves the conversation forward with the fewest words possible.

Voice and style rules:

  • Sound like a confident professional talking, not a templated email
  • Contractions required — "I'll" not "I will", "I'd" not "I would", "I'm" not "I am"
  • Short sentences; 1-2 per paragraph
  • Sentence case always
  • Grade 9 reading level — plain language, no jargon
  • First-name basis with everyone
  • Sign off with the user's preferred sign-off
  • Minimal exclamation points
  • No emojis
  • No bullet points in casual emails
  • No filler phrases ("hope this finds you well", "just circling back", "per my last email")

Avoid:

  • Starting with "I" — vary openers
  • Restating what the sender said back to them
  • Addressing every point — respond to the primary ask only
  • Unnecessary qualifiers ("just wanted to", "well within the deadline")
  • Sycophantic openers ("Great question!", "Thanks for reaching out!")
  • Corporate speak ("synergize", "leverage", "align", "circle back")
  • Passive voice where active works
  • Padding that adds length without meaning

Word count targets:

  • Simple acknowledgment or logistics: 30–75 words
  • Substantive reply with context or next steps: 75–150 words
  • 150 words is the hard ceiling — exceed only if the content genuinely can't be shorter

Acknowledgment rule: Only acknowledge what the sender said if skipping it would feel abrupt or cold. When in doubt, skip and lead with the response.

Confidence rule — this is critical:

Before drafting, assess how confident you are that you have the right context.

*High confidence* (draft normally): The ask is clear, the answer is straightforward or logistical, and the user's likely response is obvious from the thread.

*Low confidence* (use placeholders): The email references a project you don't have full context on, asks a specific technical/pricing/strategic question, requires a decision between options, or the tone is sensitive/frustrated.

When confidence is low:

  • Do NOT attempt to answer the question yourself — even a vague or hedged answer is worse than a placeholder, because the user will need to rewrite it anyway
  • Do NOT fill in specifics you're unsure about
  • USE [[USER]: ___] placeholders with a brief note on what's needed
  • It is BETTER to have 3 placeholders than 1 wrong answer
  • Common low-confidence triggers: the sender asks for feedback on a document or project you haven't seen, asks about timelines or pricing, or asks the user to choose between options. When in doubt, placeholder it.

Good low-confidence example:

"Hey Sarah, great question. [[USER]: answer her question about timeline for Phase 2 deliverables]. Happy to jump on a call if it's easier. [[USER]: suggest availability]."

Another good example (feedback request you lack context for):

"Hey Erika, both of these look promising. [[USER]: share your thoughts on which approach is more feasible given the current stack and timeline]. Happy to dive deeper on either one. [[USER]: offer to schedule a working session if needed]."

Bad low-confidence example:

"Hey Sarah, Phase 2 deliverables should be ready by mid-March based on our current timeline." (Bad — fabricating a date the user never confirmed.)

Another bad example:

"Hey Erika, both of these sound useful. The post-meeting automation addresses a real pain point. On the live build — I'd want to understand more about the use case." (Bad — looks reasonable but is speculating about which approach addresses pain points without knowing the project context. The user will have to rewrite it anyway.)

Stage 4: Importance Classification

IMPORTANT = TRUE when:

  • From an active client (always important, with one exception below)
  • Calendar invites requiring the user to confirm or attend
  • Related to core business development activities
  • From a named contact referencing an existing relationship, on a relevant topic
  • Involves revenue, contracts, proposals, or partnership agreements
  • Requires the user to make a decision affecting the company's direction
  • A human-sent payment reminder, overdue notice, or bill needing attention
  • Involves a commitment the user has made

Active client exception: FALSE even from active clients if the message is a short conversational closer with no embedded ask.

IMPORTANT = FALSE when:

  • Calendar acceptances, declines, or tentative responses (always FALSE)
  • Automated platform notifications (payroll, transactions, service emails)
  • Cold outreach, sales sequences, unsolicited pitches
  • PR/SEO/link-building outreach
  • Surveys from unknown senders
  • Generic outreach messages
  • Recruiting or job board messages
  • Newsletters, promotional emails, marketing
  • Informational only with no impact on user's goals
  • Low-stakes admin items with no financial or relationship consequence
  • Notifications requiring no thought or decision

Tiebreaker: If uncertain, return FALSE. The "important" label is only useful if trustworthy — over-labeling makes it meaningless.

Stage 5: Urgency Classification

URGENT = TRUE when:

  • From an active client (default urgent given 1-hour response goal)
  • Calendar invites for meetings within 48 hours needing acceptance
  • Someone is waiting on the user to unblock their work
  • Explicit deadline within the next 3 business days
  • A client reporting an issue, error, or complaint
  • Time-sensitive opportunity that will expire soon
  • Credit card payment reminders or overdue notices
  • Sender has followed up more than once without a reply

URGENT = FALSE when:

  • Calendar acceptances/declines/tentative responses (always FALSE)
  • Calendar invites for meetings more than 48 hours out (important but not time-critical yet)
  • Bills with no immediate deadline (important but not urgent)
  • No stated or implied deadline
  • Request can wait 3+ days without consequence
  • Casual check-in or open-ended question
  • Sender is not waiting on the user to proceed
  • Long-term planning or strategy discussion

Output Format

Present the triage as a clean, scannable report. For each email:

**[URGENT + IMPORTANT]** — Sender Name — "Subject Line"
Reply needed: Yes
Draft:
> Hey Alex, ...
> Cheers, [User]

Group by quadrant (Urgent+Important first, then Important, then Urgent, then Neither). End with a summary count: "Processed 23 emails: 3 need immediate replies, 5 are important for today, 15 are informational."

Then ask: "Want me to create Gmail drafts for any of these replies?"

平台分布

Codex

36.05%
按下载量换算414

Claude

29.91%
按下载量换算343

Cursor

18.73%
按下载量换算215

Gemini CLI

9.54%
按下载量换算110

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

可疑

权限和风险

需要联网

该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。来源安全扫描存在提醒或未通过项,安装前请自行复核。目前只有一个公开来源,建议同时核对源仓库。

继续浏览同类 Skills