Description

Chinese-market client-intelligence safety layer for lead-discovery skills. Use for any lead-gen / customer-investigation output targeting a Chinese company (...

README (SKILL.md)

\x3C!-- Derived from anthropics/financial-services-plugins under Apache-2.0. Lead-discovery adaptation by AIGroup, 2026-04-18. -->

CN Lead Safety Skill

Name: Cn Lead Safety
Author: jackdark425

中国大陆客户情报安全层 — 5 条 Rule

Lead-discovery 产出的 markdown intelligence（customer-investigation / key-account-briefing / client-initial-screening 等）如果走中文输出，必须经过这 5 条 Rule。下游 banker workflow (aigroup-financial-services-openclaw/datapack-builder / dcf-model) 读取这份 intelligence 时，只会对自己生成的 MD 跑 provenance_verify — 上游 lead-discovery MD 的质量由本 skill 守门。

Why this skill exists

2026-04-18 实测 MiniMax-M2.7 在 OpenClaw main agent 生成中文 markdown 时存在字符级 escape drift：

公司名典型 typo：寒武纪→宽厭谛79 / 营收→营收（偶发）/ 核心→校虚 / 净利→洁利 / 财务→贜务
硬数字（市值/营收/员工数）偶发缺 source citation → 下游无法 tracing

Lead-discovery 此前无此类防线。升级到 0.2.0 补上。

5 条 Mandatory Rules

Rule 1 — UTF-8 literal over `\uXXXX`

中文字符一律以 UTF-8 literal 写入 markdown，严禁 \uXXXX escape。

✅ ## 贵州茅台经营概览
❌ ## \u8d35\u5dde\u8305\u53f0\u7ecf\u8425\u6982\u89c8

write / edit 工具原生支持中文，不需要 pre-encode。

Rule 2 — Lexicon lookup for recurring terms

公司名 / 行业术语 / 财务指标 / 监管术语等重复中文短语，从 aigroup-financial-services-openclaw 主包的 cn-lexicon.js lookup，不要让模型每次重新打字。

推荐的 cross-skill 查阅路径（主包已装在 macmini 上）：

~/.openclaw/extensions/aigroup-financial-services-openclaw/skills/cn-client-investigation/references/cn-lexicon.js

在 banker intelligence markdown 中，这些字段首选 lexicon 写法：

Target 公司名（全称 / 简称 / ticker）— LEXICON.company.* + LEXICON.industry_terms.consumer_brand.* 等
财务条目（营业收入 / 归母净利润 / 扣非净利 / 毛利率 / 研发费用 / 经营现金流）— LEXICON.finance.*
监管 / 市场术语（A股 / 科创板 / 创业板 / 实际控制人 / 国资背景）— LEXICON.cn_market.*
投资评级（增持 / 中性 / 减持 / 买入 / 卖出）— LEXICON.rating.*

若你只产 markdown（不走 pptxgenjs slide JS），require() 不可用 — 则读 lexicon.js 作为术语白名单在心中校对，不让模型自由造短语。

Rule 3 — Tier-ordered data sources（跟 cn-client-investigation Rule 4 一致）

Tier	源	Lead-discovery 常用入口
T1	巨潮资讯网 cninfo.com.cn / Tushare Pro	`web_fetch` cninfo PDF；`aigroup-market-mcp__basic_info / company_performance / stock_data`
T2	上交所/深交所/港交所官网；天眼查 / 企查查；国家信用 gsxt.gov.cn	`web_fetch` 官网；`aigroup-tianyancha-mcp`（如装）
T3	Wind / 同花顺 / 东方财富；FMP / Finnhub（港股/中概股）	`aigroup-fmp-mcp / finnhub-mcp`
T4	财新 / 21世纪 / 中证 / 上证 / 财联社 / 澎湃 / 第一财经	`brave-web-search` + `web_fetch`

按 tier 依次 try，不要跳级。T1 失败 → 再 T2 / T3 / T4；只有 T1-T3 全空才用 T4 媒体信息且必须标 "单源报道"。

Rule 4 — Inline source citation for every hard number

Lead-discovery intelligence MD 中每个硬数字（digit + 亿/万/%/RMB/USD/元/CNY/HKD/M/B）必须内联source citation，格式形如：

✅ 2024 年前三季度营业收入 **1,088 亿元**（来源：巨潮资讯 2024 年三季报，2024-10-27；Tushare Pro income_all 校验）。

✅ 截至 2026-04-17 收盘总市值 **20,150 亿元**（来源：东方财富 Choice，2026-04-17 T+0；Tushare stock_data 同日对齐）。

❌ 公司 2024 Q3 营收 1,088 亿元，盈利能力持续增强。   ← 数字没 source，下游无法 tracing

verify_intelligence.py （本 skill 附带）会扫 MD，每个硬数字都要有临近 "来源：" / "Source:" 或脚注引用，否则退回 exit 1。

Rule 5 — No fabrication on missing data

T1-T4 都查不到或 MCP 返回权限错误（402/403）时：

DO：标 数据不可得 / N/A（source unavailable） + 简述尝试路径
DO NOT：估算一个"合理数字"顶上

✅ 2024 年员工数量：**数据不可得**（尝试路径：巨潮资讯年报未披露员工表、天眼查 MCP 返回 403）。
❌ 2024 年员工数量：约 5,000 人。   ← 无源凭空估算

Phase 0: pre-flight (触发)

当 user input 或 target context 出现以下任一触发词，先加载本 skill 的 5 条 Rule，再选对应 lead-discovery skill（customer-investigation / client-initial-screening 等）开始工作：

市场/监管：中国 / A股 / 港股 / 科创板 / 创业板 / 北交所 / 中概股 / H股 / 证监会
源系统：巨潮资讯 / cninfo / Tushare / 天眼查 / 企查查 / Wind / 东方财富
Ticker 形态：*.SH / *.SZ / *.BJ / *.HK
显式中文公司名（茅台 / 寒武纪 / 海光 / 华为 / 小米 / 美的 etc.）

Phase 5: 交付前 QA

python3 ~/.openclaw/extensions/aigroup-lead-discovery-openclaw/skills/cn-lead-safety/scripts/verify_intelligence.py \
    path/to/intelligence.md
# exit 0 → clean；exit 1 → 有硬数字缺 citation

补 citation 后重跑。不要 --no-verify 或类似 bypass —— 上游情报没 source，下游 banker 交付 provenance_verify 也救不了。

What this skill does NOT do

不产 pptx（那是 financial-services 的 ppt-deliverable 范畴）
不跑 typo scan（那是 financial-services 的 cn_typo_scan.py 范畴；lead-discovery 只出 markdown，中文错字下游财服 skill 读到后会被 compile 阶段的 typo gate 拦）
不自己带 cn-lexicon 副本（主包有，跨插件引用）

Output contract

每个 lead-discovery skill 的 markdown 输出必须：

文件扩展名 .md，UTF-8 no BOM
标题第一行 # \x3C中文公司名> — \x3Cintelligence 类型>
数据来源 section 标注 tier
硬数字 100% 内联 source citation
数据不可得处明确标 "N/A / 数据不可得"
verify_intelligence.py exit 0

Usage Guidance

This skill is internally consistent: it bundles a small local Python verifier that checks every hard number in a Markdown intelligence file has an inline source citation and documents five human-facing rules for Chinese-market lead outputs. Before installing/using: 1) confirm the referenced aigroup financial package (the cn-lexicon path under ~/.openclaw/extensions/...) actually exists on the host if you plan to follow the lexicon lookup guidance; otherwise follow the guidance manually. 2) Understand the verifier enforces citation rules but does NOT automatically detect Unicode-escape usages or typos — those are described as separate checks in other skills. 3) Running the script requires read access to the target .md file (and to the path if you follow the lexicon guidance); review the script if you want to be certain it meets your policies (it contains no network calls or credential handling). 4) The skill forbids citing certain paid terminals (Wind/同花顺) under strict mode; if you rely on those data sources, adopt the required MCP tools or use acceptable official filings. Overall the skill appears safe and purpose-aligned, but verify the presence and trustworthiness of the cross-plugin lexicon if you will follow that step.

Capability Analysis

Type: OpenClaw Skill Name: cn-lead-safety Version: 0.8.2 The skill bundle is a safety and quality-assurance layer designed to ensure the accuracy and traceability of financial data for the Chinese market. It includes a Python script, `verify_intelligence.py`, which uses regular expressions to validate that numerical data in markdown reports are accompanied by citations from authorized sources (e.g., Tushare, cninfo). The `SKILL.md` file provides clear instructions to the agent to enforce UTF-8 encoding, use specific terminology lexicons, and avoid data fabrication. No evidence of data exfiltration, malicious execution, or harmful prompt injection was found; the logic is strictly focused on data integrity and formatting.

Capability Assessment

✓ Purpose & Capability

Name/description, the included verify_intelligence.py, and the SKILL.md rules all align: the skill's purpose is to enforce citation and source-tier rules for Chinese-market lead intelligence. It does not ask for unrelated credentials or binaries.

ℹ Instruction Scope

SKILL.md prescribes five concrete rules and instructs running the included verification script; it also recommends (but does not require) consulting a lexicon file in another extension under ~/.openclaw/extensions/aigroup-financial-services-openclaw/.... The included script enforces Rule 4 (citation for hard numbers) but does not itself check UTF-8-vs-\uXXXX encoding or force lexicon usage, so some rules are procedural/guidance rather than programmatically enforced. The skill asks the agent to read files from the user's OpenClaw extensions path if the lexicon lookup is used — this is cross-plugin file reading (not network exfiltration) and should be acceptable but relies on that other package being present.

✓ Install Mechanism

Instruction-only skill with a small local Python script bundled; there is no install spec, no downloads, and nothing is written to disk by an installer. Lowest-risk install model.

✓ Credentials

No environment variables, credentials, or network endpoints are requested. The only filesystem access implied is reading markdown files to verify and optionally consulting a lexicon file under the user's ~/.openclaw tree. This is proportionate to a lint/verification/style gate.

✓ Persistence & Privilege

The skill does not request always:true or any elevated persistence. It does not modify other skills or system-wide settings; it runs as an on-demand verification gate and is user-invocable.

Version History

v0.8.2

- Introduces a safety layer for Chinese-market lead-discovery skills to enforce output quality and traceability. - Implements 5 mandatory rules: UTF-8 literal Chinese, lexicon-based recurring terms, tier-ordered data source usage, mandatory inline citations for all hard numbers, and strict no-fabrication on missing data. - Enforces markdown output standards to ensure downstream workflows can verify data provenance. - Provides pre-flight trigger logic and a verification script to ensure compliance before banker deliverables.

Metadata

Slug cn-lead-safety

Version 0.8.2

License MIT-0

All-time Installs 0

Active Installs 0

Total Versions 1

Frequently Asked Questions

What is Cn Lead Safety?

Chinese-market client-intelligence safety layer for lead-discovery skills. Use for any lead-gen / customer-investigation output targeting a Chinese company (... It is an AI Agent Skill for Claude Code / OpenClaw, with 126 downloads so far.

How do I install Cn Lead Safety?

Run "/install cn-lead-safety" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Cn Lead Safety free?

Yes, Cn Lead Safety is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Cn Lead Safety support?

Cn Lead Safety is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Cn Lead Safety?

It is built and maintained by jackdark (@jackdark425); the current version is v0.8.2.

More Skills

Cn Lead Safety