# trade-customer-intel

Adopt this skill: https://proskills.md/skills/other/floydtang-trade-customer-intel

Page: https://proskills.md/skills/other/floydtang-trade-customer-intel
This document is the public skill bundle for that page. It contains SKILL.md and every helper file the catalog carries. No login is required.

## File: SKILL.md
size: 8387
upstream: https://github.com/floydtang/trade-customer-intel/blob/HEAD/SKILL.md

````text
---
name: trade-customer-intel
description: Build bilingual, evidence-backed, SIEGER v2 customer intelligence reports for foreign-trade leads. Use when OpenClaw or a sales operator needs claim-level citations, decision gates, an industry-specific analysis lens, a non-fabricated verdict card, and sales angles that require approval before downstream outreach drafting.
---

# Trade Customer Intel

## Overview

Use this skill to turn sparse lead data into a structured public-web due-diligence report for sales development. This is the core judgment layer of the whole outreach chain: customer search finds candidates, lead screening cleans them, but this Skill decides what is actually worth saying before the email Skill drafts anything.

角色定位：

- `客户情报分析员`
- 负责基于公开网页证据完成客户背调与风险判断
- 负责识别近期客户动态、市场变化、合规/关税/贸易环境信号
- 负责给开发信 Skill 提供清晰、可引用、可复核的销售切入点
- 不负责批量搜客户，也不负责替代人工直接外发邮件
- 必须明确是否允许进入开发信阶段

上下游关系：

- 上游：`线索整理skill/` 输出的 `customer_intel_input`，或人工提供的稀疏 lead 信息
- 下游：`开发信skill/`

当前最稳的是公司级背调主线。联系人、职位和个性化信号只作辅助。

## Workflow

1. Normalize input into the standard shape:

```json
{
  "company_name": "",
  "person_name": "",
  "email": "",
  "company_website": "",
  "country_or_market": "",
  "product_or_offer": "",
  "industry_lens": "auto",
  "seller_context": {
    "company_name": "",
    "product_or_offer": "",
    "product_categories": [],
    "target_customer_types": [],
    "target_industries": [],
    "value_propositions": [],
    "proof_points": [],
    "authorized_materials": [],
    "forbidden_claims": []
  },
  "notes": ""
}
```

2. If only an email is present, extract the email domain and treat it as the first company clue.
3. Search in this order:
   - Official website and domain evidence
   - LinkedIn company page and personal profile
   - Facebook and Instagram
   - X/Twitter and YouTube
   - General web search and news
4. Extract recent and market signals:
   - Recent customer signals: LinkedIn updates, social posts, news, hiring, funding, expansion, new warehouse, channel changes, product launches.
   - Market signals: target-market regulation, certification, tariff, trade agreement, compliance, import/export rule changes.
   - Every signal must keep source, time or period when visible, freshness, confidence, and product relevance.
5. Resolve identities conservatively:
   - Prefer direct identifiers over inference.
   - Do not merge ambiguous people or companies unless multiple signals line up.
   - Mark weak conclusions as inference, not fact.
6. Convert raw sources into an `evidence_ledger`; every item receives an `EV-*` ID and source-quality label.
7. Convert facts, inferences, and hypotheses into a `claim_ledger`; every claim receives a `CL-*` ID and references its supporting evidence.
8. Apply the selected industry lens. Industrial accounts emphasize BOM and component clues; food accounts emphasize SKU, channel, certification, packaging, and cold-chain clues.
9. Run identity, evidence, seller-offer, product-fit, and risk decision gates.
10. Assemble the SIEGER v2 report using [report-template.md](./references/report-template.md).
11. Score only when at least three verdict dimensions have evidence-backed ratings; otherwise return `score = null` and `customer_grade = 未分级`.
12. Generate sales angles as `proposed`; this Skill must never approve its own angle.

## Main Script

Use [build_customer_intel_report.py](./scripts/build_customer_intel_report.py) as the default entrypoint.

Contracts:

- [customer-intel-input.schema.json](./schemas/customer-intel-input.schema.json)
- [customer-intel-report.schema.json](./schemas/customer-intel-report.schema.json)

### Example run

```bash
python3 ./scripts/build_customer_intel_report.py --input-json /path/to/input.json --markdown-out /tmp/customer-intel.md --json-out /tmp/customer-intel.json
```

Or pipe JSON directly:

```bash
cat <<'EOF' | python3 ./scripts/build_customer_intel_report.py
{
  "company_name": "Acme Industrial",
  "person_name": "Jane Smith",
  "email": "jane@acme-industrial.com",
  "country_or_market": "United States"
}
EOF
```

### Script behavior

- If `tvly` is installed, the script uses it first for web search.
- If `tvly` is not installed, the script falls back to DuckDuckGo HTML search.
- If evidence is sparse, the script still produces a report and explicitly flags low confidence.
- If one or more providers fail, preserve the errors in `collection_errors`; do not convert a collection outage into a successful intel decision.
- Reject malformed identity or nested context types instead of coercing them to strings.
- Create missing parent directories for requested JSON and Markdown outputs.

## Output Requirements

- Keep the report structured and CRM-friendly.
- Keep core analysis in Chinese.
- Include English wording in the `Executive Summary` and `Sales Angles` sections.
- Attach source URLs to every material claim when possible.
- Use `Low`, `Medium`, or `High` risk ratings only.
- Include `IntelDecision` with evidence sufficiency and next action.
- Include `decision_brief`, `evidence_ledger`, `claim_ledger`, and `decision_gates`.
- Every factual claim must reference at least one `evidence_id`.
- Preserve `seller_context` and the resolved `industry_lens` in the report.
- Include the SIEGER Verdict Card, but never fabricate a score to make the report look complete.
- Include `recent_signals` and `market_signals` where evidence supports them.
- Include sales angles with `angle_id`, claim/evidence references, authorized-material references, and `approval_status = proposed`.
- Apply BOM, motor/drive, ASRS, or engineering-fit language only when the resolved `industry_lens` is `industrial` and the selected claims plus seller offer support the concept; never leak that copy into food, consumer, or general reports.
- If the person match is weak, say so explicitly instead of inventing a firm personal profile.
- Keep outreach personalization conservative. Do not invent private preferences or present weak inferences as facts.
- Do not replace `客户搜索skill/` or `线索整理skill/` as the lead-entry stage.
- Do not imply the final outreach draft can skip human review.
- Do not imply stable support for precise personal email discovery.

## Openclaw Integration Notes

- Missing fields are allowed; the report header must list what was missing.
- Only use public internet sources. Do not imply private-data access.

## References

- Read [report-template.md](./references/report-template.md) before changing the output shape.
- Read [source-playbook.md](./references/source-playbook.md) before changing search order, confidence rules, or risk scoring.

## Defaults

- Standard depth, not exhaustive crawling.
- Public web only.
- Conservative entity matching.
- Conservative risk scoring.
- This is the flagship judgment layer in the outreach chain; email drafting should depend on this output instead of inventing recent events.
- The email bridge may only consume an angle after a human or authorized workflow changes its status to `approved`.
- Default intel-stage action: `ready_for_email_draft | hold_for_manual_review`

## Table and Rule Capture

- 本 Skill 输出客户背调和风险判断能力，不要求企业使用固定背调表模板。
- 写入企业表格或知识库时，先沿用已有表头和文档归口，再映射主体可信度、证据充分度、近期信号、市场信号、风险评级、推荐销售角度、推荐下一步、报告引用等标准字段。
- 用户没有可用表格时，龙虾可以按企业产品、市场和背调流程新建够用表。
- 当用户确认新的背调规则、风险分级、近期信号判断、市场信号来源、证据来源权重、行业习惯或表头映射后，必须追问：`是否更新到对应 Skill 以便下次自动复用`。真实写入必须得到用户授权。

## Enhancement Entry

如需数据留存、统一编排、多代理协作或飞书落地，请查看飞书文档：

- <https://evenbetter.feishu.cn/wiki/W6GnwTZGFiUdJ0kXZv6cV4PSnpf>
````

## File: README.md
upstream: https://github.com/floydtang/trade-customer-intel/blob/HEAD/README.md

````text
# Trade Customer Intel

中英双语的开源 Codex Skill，用公开网页线索为外贸销售和客户开发生成结构化客户情报报告。

当前状态：可交付

> 唯一开发源是 [trade-skills-collection](https://github.com/FloydTang/trade-skills-collection)；[trade-customer-intel](https://github.com/FloydTang/trade-customer-intel) 是独立发行镜像。

它是整套主动开发链路里的核心判断层：搜索负责找到候选，整理负责标准化，背调负责看清“这个客户最近发生了什么、为什么值得跟、应该从哪里切入”，开发信再调用这些判断写草稿。

## 这个 Skill 解决什么问题

- 只有公司名、联系人、邮箱或官网
- 公开信息分散、真假难辨
- 开发前需要先做保守的公司级背调
- 开发信需要更清晰的近期动态、市场变化和销售切入点，而不是模板式泛写

## 职责边界

- 负责公开网页证据收集
- 负责实体匹配、风险判断和报告生成
- 负责公司级背调主线
- 负责提取近期客户信号：LinkedIn / 社媒动态、新闻、招聘、融资、扩产、新仓库、渠道变化
- 负责提取市场信号：合规变化、关税、贸易协定、认证要求、目标市场采购压力
- 负责把可引用信号整理给开发信 Skill
- 联系人和职位只作辅助匹配
- 不负责批量搜客户入口
- 不负责替代人工直接外发邮件

## 当前默认能力

- 公司级主体核验
- 官网、社媒、公开网页证据整理
- 双语结构化报告
- 保守的风险评级
- 近期客户信号和市场/合规信号
- `EV-*` 证据账本与 `CL-*` 主张账本，区分事实、推断和假设
- SIEGER v2 主体、证据、卖方能力、产品匹配和风险决策门
- 少于 3 个证据化维度时不计算综合分
- 销售角度默认为 `proposed`，不自动批准
- 工业客户可生成 BOM 确认、工程验证和有条件替代角度；其他行业继续沿用对应行业线索，不套工业模板
- 输入类型错误会直接拒绝，不会悄悄转换成文本
- 联网搜索或网页快照失败时仍会输出报告，并通过 `collection_status`、`collection_errors` 和决策门安全停在人工复核
- JSON / Markdown 输出目录不存在时会自动创建

## 当前不默认承诺

- 精准个人邮箱稳定补齐
- 完整人物档案稳定补齐
- 完全自动化个性化触达
- 弱证据下强行生成个性化事实
- 替开发信 Skill 编造“客户最近发生了什么”

## 最小输入输出

- 输入：公司名、联系人名、邮箱、官网、市场信息、可选产品/方案信息
- 输出：执行摘要、身份快照、公司画像、数字足迹、近期客户信号、市场/合规信号、销售切入角度、风险评级、证据清单、放行判断

## 企业表格与规则沉淀

- 这是客户背调能力，不是固定背调表模板。
- 企业已有表头和知识库归口优先沿用；没有可用表格时，再按产品、市场和背调流程新建够用表。
- 用户确认新的背调规则、风险分级、近期信号判断、市场信号来源、证据权重、行业习惯或表头映射后，先追问：`是否更新到对应 Skill 以便下次自动复用`，授权后再写入 Skill。

## 依赖提醒

- 只用公开网络信息
- `已安装` 不等于 `已可用`
- 搜索工具要区分 `已安装`、`已配置`、`已登录`、`已跑通`
- LinkedIn 类能力不是默认开箱即用
- 云端和本地环境可能不同

## Quick Start

```bash
python3 ./scripts/build_customer_intel_report.py \
  --input-json ./examples/sample-input.json \
  --markdown-out /tmp/customer-intel.md \
  --json-out /tmp/customer-intel.json
```

## 增强权益入口

如需数据留存、统一编排、多代理协作或飞书落地，请查看飞书文档：

- <https://evenbetter.feishu.cn/wiki/W6GnwTZGFiUdJ0kXZv6cV4PSnpf>
````
