メニュー
Daily email classifier. Categorizes inbox into 4 types (act/read/track/trash), sends summary for confirmation, learns from corrections. Use when user says 'check email', 'daily email', '查邮件', '看邮件'.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | email-classifier |
| description | Daily email classifier. Categorizes inbox into 4 types (act/read/track/trash), sends summary for confirmation, learns from corrections. Use when user says 'check email', 'daily email', '查邮件', '看邮件'. |
| user-invocable | true |
支持多账户：
himalaya envelope list (默认)himalaya envelope list -a Zohohimalaya envelope list # Gmail
himalaya envelope list -a Zoho # Zoho
memory/{account}.json)| 类别 | 含义 | 判断依据 |
|---|---|---|
| 🔔办 (act) | 需要我做什么 | 验证码、安全提醒、催款、登录通知、部署失败、需要回复的邮件 |
| 📬读 (read) | 值得看一眼 | Newsletter、楼管通知、个人邮件、工作沟通 |
| 💰记 (track) | 钱相关的记录 | 收据、账单、发票、支付确认、运单送达、订单确认 |
| 🗑️扔 (trash) | 不用看 | 促销广告、条款更新、营销、系统报告、已取消事件 |
文件: memory/{account}.json
memory/default.jsonmemory/zoho.json结构:
{
"senders": {
"sender@example.com": {
"category": "track",
"confidence": 0.95,
"count": 10,
"mixed": false,
"breakdown": {"track": 10}
}
}
}
查记忆流程:
mixed: false + count >= 3 → 高信心，直接用mixed: false + count < 3 → 参考但也看 subjectmixed: true 或 category: "mixed" → 必须看 subject/正文📊 今日邮件 (Gmail N封 + Zoho M封)
📧 Gmail:
🔔办 (n):
• 主题1
• 主题2
📬读 (n):
• 主题1
💰记 (n):
• 主题1
🗑️扔 (n):
• 主题1
📧 Zoho:
🔔办 (n):
• 主题1
🗑️扔 (n):
• 主题1
确认删🗑️？回复「删」「删 Gmail」「删 Zoho」
纠正？回复「X 不该删」或「X 应该是🔔」
用户回复纠正时:
memory/{account}.json不依赖外部脚本，agent 直接操作 JSON。
| 类别 | 自动动作 |
|---|---|
| 🔔办 | 在摘要中高亮，验证码类立即发送 |
| 📬读 | 列在摘要中 |
| 💰记 | 列在摘要中 |
| 🗑️扔 | 列出等确认，用户说「删」后批量删除 |
🗑️永远不自动删除，必须等用户确认。
# 列出邮件
himalaya envelope list # Gmail (默认)
himalaya envelope list -a Zoho # Zoho
# 读邮件正文
himalaya message read <id>
himalaya message read -a Zoho <id>
# 移动邮件到 Trash（不要用 message delete，会跳过回收站）
himalaya message move <id> "[Gmail]/Trash" # Gmail
himalaya message move -a Zoho <id> "Trash" # Zoho
Full-stack development workflow skill. Guides the complete cycle from understanding requirements to shipping code via PR with Preview validation. USE FOR: - Bug fixes, feature development, UI changes - "修复这个 bug", "帮我实现这个功能", "改一下这个页面" - "fix this", "implement", "build", "add feature", "change the UI" - Any code modification that needs Dev Server → PR → Preview → Merge flow REQUIRES: git, gh (GitHub CLI), pnpm/npm, agent-browser
Apple Silicon local ASR via Qwen3-ASR + MLX. Server-mode for stable memory. USE FOR: - "转文字", "转录", "transcribe this", "speech to text" - 任何音频/视频文件需要转文字时 - SRT/VTT 字幕生成 - 用户发音频/视频文件说"这个说的什么" - douyin-dl 下载后自动转文字 - Agent 需要处理语音内容时 ⚠️ ALWAYS use this skill for any audio/video transcription on Apple Silicon. Do NOT use whisper-transcribe (disabled). Apple Silicon only. No API key needed. No cloud dependency. Uses server-mode (model loaded once, ~4GB memory stable).
Convert Markdown to high-quality PNG/JPG images with zero whitespace, syntax highlighting (Shiki), and optional KaTeX math support. Uses Playwright (Chromium) for rendering with GitHub-style default theme. USE FOR: - Any request to render Markdown as an image - Code snippet screenshots with syntax highlighting - Markdown with math formulas (use --math flag) - Generating styled images for Discord/Telegram/Slack - "markdown转图片", "把这段md变成图片", "render markdown as image", "screenshot this code" - Converting README, changelogs, or docs to shareable images REQUIRES: md2img (npm global), playwright-core, chromium browser
Unified self-reflection and self-improvement. Three-layer memory: Golden Rules (SOUL.md) → Active Lessons (self-review.md) → Raw Log (daily). Auto-captures corrections, reflects at session end, promotes patterns that work. USE FOR: - User corrects agent behavior → auto-log and learn - Session end → "land the plane" reflection - Heartbeat timer → periodic self-check - "what have you learned", "show your lessons", "memory stats" - "remember this rule", "promote this pattern", "demote this" - Agent notices own mistake → self-log REPLACES: self-improving, self-reflection (merged into this skill) COMPANION: soul-keeper (independent — pushes lessons to workspace files) REQUIRES: jq, date
Transcribe audio/video to text using the best available ASR backend. Backend priority: MLX Qwen3-ASR (local, best for Chinese) → Groq API (fast, cloud) → whisper.cpp (local fallback) USE FOR: - "转文字", "转录", "transcribe this", "speech to text" - Any audio file (wav, mp3, m4a, flac, ogg, wma, aac) needs transcription - Video file needs transcription (extracts audio first via ffmpeg) - Douyin/video downloaded → auto transcribe - Need SRT/VTT subtitles from audio/video - "这个视频说的什么", "听一下这个音频" REPLACES: openai-whisper, openai-whisper-api, mlx-whisper, douyin-transcript (for ASR step) REQUIRES: ffmpeg OPTIONAL: mlx-qwen3-asr (pip), GROQ_API_KEY, whisper-cli (brew)
系统化研究 skill — 螺旋收敛模型。把任何问题（模糊或清晰）分解成子问题，迭代搜索，越搜越清晰，直到每个子问题都有答案。 Triggers: "帮我研究", "研究一下", "调研", "分析对比", "research", "investigate", "look up" 也触发: 竞品分析、市场调研、技术选型对比、趋势了解 Does NOT trigger: - 已经知道答案的简单事实问题 - 用户直接给了 URL 要抓取的 - 代码调试、写代码任务 Output: 结构化研究报告（结论 + 子问题答案 + 来源 + 争议点 + 未解决缺口）