Trade Customer Intel for OpenClaw

Overview

This skill is the OpenClaw-native companion to the repository's classic version.

It is designed for cloud OpenClaw environments where search and page retrieval are performed by platform tools first, and a Python report builder then turns the resulting evidence bundle into a structured bilingual report.

This is the core judgment layer in the active outreach chain: search finds candidates, screening normalizes them, customer intel decides what is worth saying, and the email Skill drafts only from these verified signals.

Table Policy

• 优先适配企业已有表头和知识库归口，不强制使用课堂标准表。
• 没有可用表格时，龙虾按企业产品、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、证据来源权重、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。真实写入必须得到用户授权。

Inputs

Normalize operator input into this lead shape:

{
  "company_name": "",
  "person_name": "",
  "email": "",
  "company_website": "",
  "country_or_market": "",
  "product_or_offer": "",
  "industry_lens": "auto",
  "seller_context": {},
  "notes": ""
}

The final report-builder input must wrap that lead together with an evidence bundle:

{
  "lead": {},
  "evidence_bundle": {
    "evidence_items": [],
    "search_results": [],
    "page_snapshots": [],
    "search_runs": [],
    "errors": []
  }
}

Validate this envelope against customer-intel-evidence-input.schema.json. Reject malformed nested objects or arrays instead of coercing them.

evidence_items 可为每条证据显式提供 claims。报告会将结论标为 fact、inference 或 hypothesis，并用 EV-* 和 CL-* 保留证据到结论的对应关系。

Tooling Rules

• Search tool: coze-web-search
• Primary fetch tool: scrapling-official
• Fallback fetch tool: coze-web-fetch

Do not use Tavily, DuckDuckGo HTML scraping, or r.jina.ai in this variant.

Search Order

Search in this fixed order:

1. Official website and domain evidence
2. LinkedIn company page and personal profile
3. Facebook and Instagram
4. X / Twitter and YouTube
5. General web search and news

The evidence bundle should preserve recent-signal clues when available: LinkedIn updates, news, hiring, funding, expansion, new warehouse, channel changes, product launches, target-market regulation, tariff, trade agreements, and compliance changes.

Search Execution Rules

• Keep query budgets small and deterministic
• Prefer direct identifiers over broad inference
• Stop expanding official website candidates once a high-confidence official domain is identified
• Keep only a small number of representative results per platform
• Do not fail the overall task just because one platform is weak or blocked

Fetch Rules

• Use scrapling-official first for page text extraction
• If scrapling-official fails, retry with coze-web-fetch
• If full-page content is unavailable, keep the search result snippet and continue
• Record fetch failures into evidence_bundle.errors

Entity Resolution Rules

• Prefer official website over search guess
• Do not treat public email domains as company-domain proof
• Do not merge ambiguous people and companies unless multiple signals align
• Mark weak conclusions as inference
• If LinkedIn evidence is missing, keep person matching conservative

Output Requirements

Follow report-template.md and source-playbook.md.

• Keep analysis mainly in Chinese
• Keep sales-facing content bilingual
• Keep risk scoring conservative
• Preserve Low, Medium, High ratings only
• Use the SIEGER v2 decision gates: identity, evidence, seller offer, product fit, and risk
• Do not produce a numeric score until at least three SIEGER dimensions have evidence-backed claims
• Include recent customer signals and market/compliance signals when evidence supports them
• Keep every generated sales angle at proposed; only an explicit human or authorized workflow action may mark it approved
• Provide sales angles with valid CL-* and EV-* references so downstream email drafting can consume them without inventing facts
• Generate outreach persona and outreach pack only when public evidence supports them
• If evidence is thin, use limited_evidence instead of forcing personalization

Main Script

Use build_customer_intel_report_from_evidence.py as the report-builder entrypoint.

Example run

python3 ./scripts/build_customer_intel_report_from_evidence.py \
  --input-json /path/to/openclaw-customer-intel.json \
  --markdown-out /tmp/customer-intel-openclaw.md \
  --json-out /tmp/customer-intel-openclaw.json

Notes

• This skill and the classic version share the same v2 contract and judgment core
• This version optimizes for cloud stability and controlled tool orchestration

Feishu 回写规则（经验教训沉淀）

主表回写（Lead Workflow Master）

回挂主表时，至少更新以下 7 个字段：

字段	说明
intel_doc_ref	背调文档链接（URL 类型）
risk_rating	SingleSelect: low / medium / high / critical
entity_confidence	Number: 0-100
current_stage	Text: intel_completed / hold / failed
current_status	SingleSelect: intel_completed 等
recommended_next_action	Text: ready_for_email_draft / hold_for_manual_review
last_updated_at	DateTime: 当前时间戳（毫秒）

背调结果表回写（Customer Intel Results）

回写字段：

字段	说明
线索编号 lead_id	Text
风险等级 risk_rating	SingleSelect: low / medium / high / critical
主体置信度 entity_confidence	Number: 0-100
公司匹配状态 company_match_status	SingleSelect: verified / partial / unverified / mismatch
联系人匹配状态 person_match_status	SingleSelect: verified / partial / unverified / not_applicable
背调阶段结果 intel_result	SingleSelect: intel_completed / intel_insufficient_evidence / intel_failed
背调文档引用 intel_doc_ref	URL: {link, text} 对象
中文摘要 summary_cn	Text
英文摘要 summary_en	Text
销售切入点 key_sales_angles	Text
关键风险 key_risks	Text
证据清单 evidence_list	Text

Feishu API 注意点（踩坑记录）

1. 字段名必须完全匹配 — 飞书 API 字段名是「中文 英文」混合格式（如 线索编号 lead_id），带空格，必须精确一致。少一个空格或顺序颠倒都会报 FieldNameNotFound。
2. URL 类型字段需要对象格式 — 不能用纯字符串 "https://"，必须用 {"link": "https://...", "text": "显示文本"}
3. SingleSelect 字段用选项名 — 传 "low" 不是 "optXXXX" ID
4. Number 字段传数字 — 95 不是 "95"
5. DateTime 字段传毫秒时间戳 — 1745570700000
6. 分两步写入 — URL/SingleSelect 类字段和 Text/Number/DateTime 字段最好分两批写入。第一次 create_record 只写 Text/Number/SingleSelect，第二次 update_record 补 URL 字段。一次性大批量写入容易触发字段类型转换错误。
7. 文本内容不要过长 — 长文本（如 evidence_list）控制在 300 字以内，完整内容写入背调文档而非 table 字段。
8. 创建文档后回写 wikidoc_token — 用 feishu_wiki(action="create") 创建文档后，记录 obj_token（doc_token），再用 feishu_doc(action="write") 写入内容。URL 拼接格式：https://evenbetter.feishu.cn/wiki/{node_token}。

背调文档标题规范

文档创建路径：首页/ 下创建独立页面

命名格式：{lead_id}_{公司简称}_背调报告

文档内必须包含的章节（详见 report-template.md）：背调摘要（中英文）、身份快照、Intel Decision、公司画像、数字足迹、主题信号、销售切入点（≥3个）、风险评级、证据清单、需人工确认事项、推荐下一步。

搜索层注意事项（OpenClaw 环境）

本地脚本（classic 版本）在 OpenClaw 环境的已知问题：

• ddg_search() 依赖 DuckDuckGo HTML 接口，中国境内大概率超时
• fetch_snapshot() 依赖 r.jina.ai 外部服务，不可靠
• 脚本没有网络超时后自动降级机制

正确做法： OpenClaw 环境下按本 Skill 的工具规则使用 coze-web-search + scrapling-official，失败时再用 coze-web-fetch，然后组装证据包。

Enhancement Entry

增强权益不在仓库中展开正文。

如需飞书落地、统一编排或多代理协作，请查看飞书文档：

• https://evenbetter.feishu.cn/wiki/W6GnwTZGFiUdJ0kXZv6cV4PSnpf