好,这一步我帮你直接拆到“可以交给工程师开工”的程度:Hermes Agent 外贸邮件自动获客系统(开发级设计)。我会分三部分给你:
- 工作流设计图(Agent Orchestration)
- Prompt 模板库(可直接用)
- API / 数据结构设计(可开发)
一、Hermes Agent 工作流设计图(可开发版)
可以理解为:Hermes 是“调度大脑”,Open Claw + LLM + Email系统是“执行器”。
┌──────────────────────────┐
│ 1. Trigger Scheduler │
│ - 每天/每小时触发任务 │
└──────────┬───────────────┘
↓
┌────────────────────────────────────────────┐
│ 2. Open Claw Data Harvester │
│ - 抓取公司(Google / B2B / Maps) │
│ - 输出 raw leads │
└──────────────────────┬─────────────────────┘
↓
┌────────────────────────────────────────────┐
│ 3. Lead Enrichment Service │
│ - 查网站 │
│ - 找邮箱(Hunter / pattern guess) │
│ - 邮箱验证 │
└──────────────────────┬─────────────────────┘
↓
┌────────────────────────────────────────────┐
│ 4. Hermes Agent: Lead Intelligence Engine │
│ - 行业识别 │
│ - 采购意图分析 │
│ - Lead Scoring (A/B/C) │
└──────────────────────┬─────────────────────┘
↓
┌────────────────────────────────────────────┐
│ 5. Message Generator (LLM Layer) │
│ - Email 1 (cold outreach) │
│ - Follow-ups (3–4 sequences) │
│ - Multi-language versions │
└──────────────────────┬─────────────────────┘
↓
┌────────────────────────────────────────────┐
│ 6. Email Dispatch System │
│ - Multi-inbox rotation │
│ - Warm-up control │
│ - Rate limiting │
└──────────────────────┬─────────────────────┘
↓
┌────────────────────────────────────────────┐
│ 7. Engagement Tracking Layer │
│ - open / click / reply tracking │
│ - webhook events │
└──────────────────────┬─────────────────────┘
↓
┌────────────────────────────────────────────┐
│ 8. CRM + Reply Agent │
│ - 自动分类回复 │
│ - lead status update │
│ - push to sales pipeline │
└────────────────────────────────────────────┘
二、Hermes Agent Prompt 模板库(核心资产)
1. Lead Intelligence Prompt(客户分析)
You are a B2B sales intelligence analyst.
Analyze the following company:
Company: {company_name}
Website: {website}
Country: {country}
Industry: {industry}
Tasks:
1. Identify the company type:
- manufacturer / distributor / trader / end user
2. Estimate purchasing intent (0–100)
3. Identify likely pain points in sourcing suppliers
4. Assign lead score:
- A (80–100): high buying intent
- B (50–79): potential buyer
- C (<50): low priority
5. Output JSON:
{
"type": "",
"intent_score": "",
"lead_class": "",
"pain_points": [],
"recommended_angle": ""
}
2. Cold Email Generator Prompt(开发信)
You are a senior B2B export sales copywriter.
Write a cold email for:
Product: {product}
Target company: {company_name}
Country: {country}
Industry: {industry_type}
Pain points: {pain_points}
Rules:
- Keep email under 120 words
- No spam language (no: best price, cheap, discount)
- Sound like a real person, not marketing
- Include one soft question
- One clear CTA only
Generate 3 versions:
1. Direct style
2. Relationship-building style
3. Value-focused style
Return JSON:
{
"subject": "",
"email_body": ""
}
3. Follow-up Email Prompt
Write a follow-up email for a cold outreach.
Context:
- Previous email sent 3–7 days ago
- No response yet
- Industry: {industry}
- Product: {product}
Rules:
- No repetition of first email
- Short and polite
- Add value or clarification
- Max 80–100 words
Return JSON:
{
"subject": "",
"email_body": ""
}
4. Reply Classification Prompt(自动回复识别)
Classify this email reply:
Reply:
{email_text}
Categories:
1. Hot lead (ready to buy / ask for quote)
2. Warm lead (interested, not urgent)
3. Cold lead (not interested / irrelevant)
4. Spam / unsubscribe
Also extract:
- buying intent
- requested product
- urgency
Return JSON:
{
"category": "",
"intent": "",
"next_action": ""
}
三、API & 数据结构设计(工程实现级)
1. Lead 数据结构(核心对象)
{
"lead_id": "uuid",
"company_name": "ABC Logistics",
"website": "https://abc.com",
"country": "USA",
"industry": "trailer logistics",
"contacts": [
{
"name": "John Smith",
"email": "john@abc.com",
"role": "Procurement Manager"
}
],
"lead_score": 87,
"lead_class": "A",
"pain_points": [
"high procurement cost",
"supplier instability"
],
"status": "new",
"last_contacted_at": null
}
2. Email Campaign 数据结构
{
"campaign_id": "uuid",
"product": "semi trailer",
"target_country": "USA",
"sequence": [
{
"step": 1,
"type": "cold_email",
"delay_days": 0
},
{
"step": 2,
"type": "follow_up",
"delay_days": 3
},
{
"step": 3,
"type": "follow_up",
"delay_days": 7
}
]
}
3. Email Sending API(核心接口)
POST /send-email
{
"from_inbox": "sales1@domain.com",
"to": "john@abc.com",
"subject": "Quick question about your trailer sourcing",
"body": "...",
"campaign_id": "uuid",
"lead_id": "uuid"
}
4. Event Tracking API(行为反馈)
POST /email-event
{
"lead_id": "uuid",
"event_type": "open | click | reply | bounce",
"timestamp": "2026-05-07T10:00:00Z"
}
5. CRM 状态更新 API
POST /lead-update
{
"lead_id": "uuid",
"status": "contacted | engaged | quoted | won | lost"
}
四、Hermes Agent 核心逻辑(最关键部分)
你可以理解 Hermes 在做 4 件事:
1. Decide(决策)
- 谁值得发邮件?
- 谁不值得浪费资源?
2. Personalize(个性化)
- 每个客户不同邮件
- 不同国家不同语气
3. Orchestrate(调度)
- 控制 Open Claw
- 控制 Email sender
- 控制 follow-up timing
4. Learn(反馈学习)
- 哪种邮件回复率高
- 哪个国家更容易成交
- 自动优化策略
五、系统升级路径(非常重要)
如果你做完这个 MVP,可以继续升级:
Level 1(现在)
- 自动找客户
- 自动发邮件
- 基础 follow-up
Level 2(增长)
- Lead scoring AI
- reply auto-classification
- CRM自动流转
Level 3(高级)
- Intent engine(预测客户什么时候买)
- multi-channel(Email + LinkedIn + WhatsApp)
- sales closing assistant
Level 4(终极形态)
Autonomous Sales Agent(自动外贸销售系统)
- AI自己找客户
- 自己判断
- 自己发邮件
- 自己跟进
- 人只负责报价成交
六、如果你下一步要继续推进
我可以帮你再往下拆一层非常关键的东西:
👉「Hermes Agent 代码级架构(Python / Node版)」
包括:
- Agent state machine
- task queue设计(Redis / Celery)
- email warm-up algorithm
- multi-inbox scheduler
或者:
👉「专门针对卡车 / 工程机械行业的邮件模板系统(可直接用来成交)」
只要你说一句:
“我要进入工程实现阶段”
我可以直接帮你把这套系统推进到“能开发上线”的级别。

Leave a Reply