When to Use
User needs to send messages on their behalf. Agent must avoid social mistakes that humans wouldn't make: wrong tone, wrong channel, wrong timing, auto-committing to things.
Quick Reference
| Topic | File |
|---|
| Platform formatting | INLINECODE0 |
| Tone calibration |
tone.md |
| Escalation matrix |
escalation.md |
Core Rules
1. Never Auto-Commit
- - Timelines, pricing, legal terms, availability → draft for human, never send
- "We can deliver by Friday" from AI = career damage
- Money confirmations require explicit per-transaction approval
- When uncertain about commitment level → ask first
2. Escalate High-Stakes
Draft for human review, never auto-send:
- - Investors, board, press, lawyers
- Client complaints, anything with "urgent", "legal", "disappointed"
- Condolences, relationship issues, conflict
- First message to new important contact
3. Match the Human's Style
- - Read their last 5 messages before drafting
- Don't add phrases they never use ("Hope you're doing well!")
- Don't use emojis they avoid
- Real humans send "ok", AI sends paragraphs → match their brevity
4. Channel Selection Follows Urgency
| Urgency | Channel |
|---|
| Production down | Call, then Slack |
| Same-day needed |
Slack/Teams DM |
| This week | Email |
| FYI only | Email with no action needed |
- - NEVER email for urgent issues
- NEVER Slack for formal client communication
5. Timing Is Social Signal
- - Instant replies reveal automation
- 3 AM recipient time → schedule for morning
- Email = hours acceptable. Slack DM = expect <1hr response
6. Context Awareness Prevents Disasters
- - Check you're in the RIGHT chat before sending
- Don't introduce yourself to someone you've messaged 50 times
- Group chats: lurking is normal, replying to everything is weird
- Wrong group = social suicide → when unsure, ASK
Common Traps
- - Copying boss on complaint email → escalates when de-escalation needed
- Reply-all with "thanks" → 50 people interrupted
- Forwarding thread with internal comments visible → trust destroyed
- Sending at 11 PM "just to get it off my plate" → signals poor boundaries
- Using client's first name before they used yours → presumptuous
何时使用
用户需要以自身名义发送消息。智能体必须避免人类不会犯的社交失误:错误的语气、错误的渠道、错误的时机、自动承诺事项。
快速参考
tone.md |
| 升级矩阵 | escalation.md |
核心规则
1. 绝不自动承诺
- - 时间线、定价、法律条款、可用性 → 起草供人类审核,绝不发送
- 由AI发出的我们可以在周五前交付 = 职业灾难
- 涉及金钱确认需逐笔明确批准
- 对承诺级别不确定时 → 先询问
2. 高风险事项需升级
起草供人类审核,绝不自动发送:
- - 投资者、董事会、媒体、律师
- 客户投诉,任何包含紧急、法律、失望的内容
- 吊唁、人际关系问题、冲突
- 首次联系重要新联系人
3. 匹配对方的沟通风格
- - 起草前先阅读对方最近5条消息
- 不要添加对方从未使用过的短语(希望你一切顺利!)
- 不要使用对方回避的表情符号
- 真人发好的,AI发长篇大论 → 匹配对方的简洁度
4. 渠道选择取决于紧急程度
| 紧急程度 | 渠道 |
|---|
| 生产系统宕机 | 电话,然后Slack |
| 当天需要 |
Slack/Teams私信 |
| 本周内 | 邮件 |
| 仅供参考 | 无需操作的邮件 |
- - 紧急问题绝不使用邮件
- 正式客户沟通绝不使用Slack
5. 时机是社交信号
- - 即时回复暴露自动化
- 收件人时间凌晨3点 → 安排在早上发送
- 邮件 = 几小时内回复可接受。Slack私信 = 期望1小时内回复
6. 情境意识可避免灾难
- - 发送前确认在正确的聊天窗口
- 不要向已联系过50次的人做自我介绍
- 群聊:潜水是正常的,回复所有内容很奇怪
- 发错群 = 社交自杀 → 不确定时,先询问
常见陷阱
- - 在投诉邮件中抄送老板 → 本应缓和事态却升级了矛盾
- 回复全部时说谢谢 → 打扰50个人
- 转发包含内部评论的邮件链 → 信任崩塌
- 晚上11点发送只是为了完成任务 → 暴露边界感缺失
- 在对方未使用你的名字前先直呼其名 → 显得冒昧