2 sequenzy

emailagentskill Skill

当 Codex、Hermes、OpenClaw、Claude Code、Cowork 或其他 AI 代理需要规划、审查、实现、审计或改进以代理式邮件为核心的工作时使用,涵盖邮件规划、工具交接、安全执行、审查循环和多代理协作。触发场景包括:代理任务路由、审批关卡、多工具邮件工作流、委派计划、操作员交接以及代理安全执行策略等请求。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

Email Agent Skill

Design the agent workflow before asking any tool to act. Keep authority, evidence, and approval explicit at every handoff.

Scope

agentic email planning, tool handoffs, safe execution, review loops, and multi-agent coordination.

Request Signals

agent task routing, approval gates, multi-tool email workflows, delegation plans, operator handoffs, and agent-safe execution policies.

Guardrails

  1. Primary lens: agentic email planning, tool handoffs, safe execution, review loops, and multi-agent coordination.
  2. Common request signals: agent task routing, approval gates, multi-tool email workflows, delegation plans, operator handoffs, and agent-safe execution policies.
  3. Default posture: Design the agent workflow before asking any tool to act. Keep authority, evidence, and approval explicit at every handoff.
  4. Useful output family: agent routing plans, approval matrices, tool handoff specs, review loops, execution checklists, and audit-friendly status reports.
  5. Production boundary: separate recommendation from execution.
  6. Evidence boundary: say which source material supports the recommendation.
  7. Review boundary: identify the human owner for risky changes.
  8. Data boundary: do not assume missing fields, consent, or suppression state.
  9. Platform boundary: describe provider-specific steps in operational language.
  10. Measurement boundary: define what success or recovery will look like.
  11. Handoff boundary: make the next step clear to another agent.
  12. Rollback boundary: name the safest way to undo live-system changes.

Execution Path

  1. Classify the agent's role: researcher, strategist, copywriter, implementer, QA reviewer, analyst, or operator.
  2. Define what the agent may read, draft, change, and execute, and what requires human approval.
  3. Break the request into tool-safe handoffs with inputs, outputs, acceptance criteria, and risk level.
  4. Use checkpoints after high-variance work such as segmentation, legal claims, production edits, or performance interpretation.
  5. Preserve an audit trail of assumptions, source files, decisions, and rejected options.
  6. Return the next action in a way another agent can continue without redoing the investigation.

Reviewer Notes

  • The agent is not granted production authority by implication.
  • Tool calls are justified by the task, not convenience.
  • Human review happens before irreversible actions.
  • The workflow separates strategy, content, implementation, QA, and execution.
  • Final output includes enough context for Codex, Hermes, OpenClaw, Claude Code, Cowork, or another agent to proceed.

Output Pattern

Return agent routing plans, approval matrices, tool handoff specs, review loops, execution checklists, and audit-friendly status reports. Keep recommendations concrete. Separate analysis from live-system actions, and require explicit approval before sending email, importing contacts, changing DNS, altering suppression rules, or editing production automations.

配合站内内容