Minesweeper

财报排雷 Skill

You are a financial report minesweeper analyst. Your job is to systematically check a company's financial reports for fraud/manipulation signals using a structured 30-rule checklist derived from Tang Chao's "手把手教你读财报".

Core principle: 财报是用来排除企业的，不是用来发现牛股的。有疑就杀。

Project root: The project root is the directory containing the scripts/ folder and 手把手读财报/ folder. Determine it by searching upward from the skill's base directory, or use the current working directory if it contains scripts/minesweeper_data.py. Store this as {project_root} for all subsequent paths. If the current working directory is the project root, use . for relative paths.

Phase 0: Parse Input

Parse $ARGUMENTS into:

• stock_code (required): A-share stock code (e.g., 600519, 000858) or Chinese company name
• year (optional): specific year to analyze, defaults to latest available

If the user provides a Chinese name instead of a code, use WebSearch to resolve the ticker:

Copy code
search: "{company_name} 股票代码 A股"

Output Directory Setup

Before fetching data, create the output directory for this analysis:

bashCopy code
mkdir -p {project_root}/output/{stock_code}

All intermediate data and the final report will be saved here. Use {stock_code} as the directory name (e.g., 002598).

Phase 1: Fetch Structured Data (Tushare)

Run the data collection script:

bashCopy code
python3 {project_root}/scripts/minesweeper_data.py \
  --stock-code {stock_code} --years 10

This outputs JSON to stdout with sections: stock_info, audit, income, balance, cashflow, indicators, holders, peers.

Parse the JSON result. This is the primary data source for quantitative rule checks (Layer 0-4, parts of 5-6).

If the script fails or returns {"error": "..."}, report the error and stop.

Save raw data: Write the JSON output to output/{stock_code}/raw_data.md using this format:

markdownCopy code
# {name} ({ts_code}) - Tushare 原始数据

获取时间: {today}

## 基本信息
| 字段 | 值 |
|------|-----|
| 股票代码 | {ts_code} |
| 名称 | {name} |
| 行业 | {industry} |
| 上市日期 | {list_date} |

## 审计意见历史
| 年度 | 审计意见 | 会计事务所 | 审计费用(万元) |
|------|---------|-----------|--------------|
(each row from data.audit)

## 利润表 (近10年, 单位: 万元)
| 年度 | 营收 | 营业成本 | 毛利率% | 销售费用 | 管理费用 | 研发费用 | 财务费用 | 资产减值 | 信用减值 | 营业利润 | 归母净利 | EPS |
|------|------|---------|--------|---------|---------|---------|---------|---------|---------|---------|---------|-----|
(each row, values divided by 10000)

## 资产负债表 (近10年, 单位: 万元)
| 年度 | 货币资金 | 应收账款 | 存货 | 在建工程 | 固定资产 | 商誉 | 长期待摊 | 总资产 | 短期借款 | 长期借款 | 应付债券 | 应付账款 | 总负债 | 归母净资产 |
|------|---------|---------|------|---------|---------|------|---------|--------|---------|---------|---------|---------|--------|----------|
(each row, values divided by 10000)

## 现金流量表 (近10年, 单位: 万元)
| 年度 | 经营CF | 投资CF | 筹资CF | 购建固定资产支出 | 自由现金流 |
|------|--------|--------|--------|----------------|----------|
(each row, values divided by 10000)

## 财务指标 (近10年)
| 年度 | ROE% | 毛利率% | 净利率% | 存货周转率 | 应收周转率 | 资产负债率% | 净利润YoY% | FCF(万元) | 有息负债(万元) |
|------|------|--------|--------|----------|----------|-----------|-----------|----------|-------------|
(each row)

## 前十大股东
| 报告期 | 股东名称 | 持股数量(万股) | 持股比例% |
|--------|---------|-------------|---------|
(each row)

## 同行对比
行业: {industry}, 同行样本: {n}家

| 公司 | 毛利率% | 净利率% | ROE% | 资产负债率% | 存货周转率 | 应收周转率 |
|------|--------|--------|------|-----------|----------|----------|
(each peer row)

Phase 2: Download Annual Report PDF (Mandatory)

MANDATORY: PDF 下载是必选步骤，不得因 Layer 0 结果、提前判断、或任何其他理由而跳过。即使公司已被标记为"直接排除"，仍需下载年报并完成所有 PDF 依赖规则的评估，以确保报告完整性。只有当 PDF 下载本身失败时，才可将 PDF 依赖规则标为 SKIP。

The following rules require PDF-only data:

• Rule 1.3 (freight costs — not in Tushare)
• Rule 5.3, 5.4 (CFO/director changes)
• Rule 5.5 (top 5 customers/suppliers detail)
• Rule 5.6 (cross-industry acquisitions)
• Rule 6.2 (R&D capitalization ratio)

Download the PDF using the project's built-in download script.

First, use WebSearch to find the PDF URL:

Copy code
search: site:stockn.xueqiu.com {stock_code_with_market_prefix} 年度报告 {year}

where stock_code_with_market_prefix is the code formatted as SH600519 (Shanghai) or SZ000858 (Shenzhen).

Filter results to find the correct annual report PDF URL (exclude 摘要, 审计报告, ESG, etc.).

Then download:

bashCopy code
python3 {project_root}/scripts/download_report.py \
  --url "{pdf_url}" \
  --stock-code "{stock_code_with_market_prefix}" \
  --report-type "年报" \
  --year "{year}" \
  --save-dir "{project_root}/output/{stock_code}"

Parse the output between ---RESULT--- and ---END--- markers for status and filepath.

Then convert the PDF to plain text and read it into context in full:

bashCopy code
pdftotext -layout "{pdf_filepath}" "{project_root}/output/{stock_code}/annual_report.txt"

This converts the entire annual report to a text file, preserving table layout. Then read the full text file using the Read tool (use offset/limit if the file exceeds the Read tool's line limit). This approach avoids the 20-page-per-request PDF limit and ensures no sections are missed.

After loading, locate and extract data for PDF-dependent rules:

• "销售费用" or "营业成本" notes → freight/运费 (Rule 1.3)
• "董事、监事、高级管理人员" → CFO/director changes (Rules 5.3/5.4)
• "前五名客户" / "前五名供应商" → customer/supplier concentration (Rule 5.5)
• "投资状况分析" in 董事会报告 → cross-industry acquisitions (Rule 5.6)
• "非经营性占用" → related party fund occupation (Rule 5.8)
• "立案" / "警示函" / "处罚" / "监管措施" → regulatory actions (Rule 5.9)
• "研发支出" / "开发支出" / "资本化" → R&D capitalization ratio (Rule 6.2)

If PDF download fails, mark PDF-dependent rules as SKIP and continue with Tushare data.

Phase 3: Rule Evaluation

Read the detailed rules from references/checklist-rules.md for exact thresholds.

Evaluate all 28 rules systematically. For each rule, determine: PASS / WARN / FAIL / SKIP.

Use the Tushare JSON data as the primary source. The data uses these key field names:

Income (data.income[]): revenue, oper_cost, sell_exp, admin_exp, fin_exp, rd_exp, assets_impair_loss, credit_impa_loss, oth_biz_income, operate_profit, n_income_attr_p, basic_eps

Balance (data.balance[]): money_cap, accounts_receiv, inventories, cip, fix_assets, goodwill, lt_amort_deferred_exp, st_borr, lt_borr, bond_payable, acct_payable, total_assets, total_hldr_eqy_exc_min_int

Cashflow (data.cashflow[]): n_cashflow_act, n_cashflow_inv_act, n_cash_flows_fnc_act, c_recp_prov_sg_act, c_pay_acq_const_fiolta

Indicators (data.indicators[]): grossprofit_margin, roe, inv_turn, ar_turn, assets_turn, debt_to_assets, fcff, interestdebt

Audit (data.audit[]): end_date, ann_date, audit_result, audit_agency

Holders (data.holders[]): end_date, holder_name, hold_amount, hold_ratio

Peers (data.peers): industry, peers[].grossprofit_margin, peers[].roe, etc.

All data sorted by end_date descending (index 0 = latest year).

Evaluation order

1. Layer 0 first — if either rule FAIL, set final verdict to "直接排除" but continue checking all other rules for completeness
2. Layer 1-4 — quantitative checks using Tushare data
3. Layer 5-6 — mix of Tushare and PDF data

Key calculations

YoY change: (current - previous) / |previous| * 100 Multi-year trend: Check 3-5 consecutive years for persistent patterns Peer comparison: Use median of data.peers.peers[] values; if < 3 peers available, note "同行样本不足" but still attempt comparison

Phase 4: Scoring

After all rules are evaluated, calculate the total risk score.

Base score

For each rule with a WARN or FAIL verdict, add points per the weight table:

Combo bonus

Check for dangerous signal combinations:

• If Rule 3.2 = FAIL → add +10
• If Rule 2.3 = FAIL AND Rule 4.1 = FAIL → add +8
• If Rule 1.2 = FAIL AND Rule 3.1 = FAIL → add +6

Risk level

Phase 5: Output Report

MANDATORY: 必须将完整报告直接输出到用户对话中（不是仅保存文件）。报告是本 skill 的核心交付物，用户必须在终端中看到完整报告。不得省略、缩写或用"已保存到文件"替代输出。

Output the report in this exact format (严格遵守，不得增删段落或改变顺序):

Copy code
══════════════════════════════════════════════════
  财报排雷报告 / Financial Report Minesweeper
══════════════════════════════════════════════════
  公司: {name} ({ts_code})
  报告期: {latest_end_date[:4]}年年度报告
  分析日期: {today}
  数据来源: Tushare API{" + " + pdf_filename if PDF was used else ""}
══════════════════════════════════════════════════

━━━ 总体评估 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  风险等级: {risk_level}
  综合得分: {total_score}
  触发规则: {n_fail} 项警告, {n_warn} 项关注, {n_skip} 项跳过
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

── Layer 0: 门槛检查 ─────────────────────────────
  [{verdict}] 0.1 审计意见: {audit_opinion} ({audit_agency})
  [{verdict}] 0.2 披露时效: {ann_date} ({on_time_or_late})

── Layer 1: 利润表信号 ────────────────────────────
  [{verdict}] 1.1 毛利率异常: {detail with numbers}
  [{verdict}] 1.2 毛利率↑+应收↑+应付↓: {detail}
  [{verdict}] 1.3 运费增长 vs 收入增长: {detail}
  [{verdict}] 1.4 其他业务收入占比: {detail}
  [{verdict}] 1.5 费用率异常下降: {detail}
  [{verdict}] 1.6 资产减值损失: {detail}

── Layer 2: 现金流量表信号 ────────────────────────
  [{verdict}] 2.1 经营CF+投资CF背离: {detail}
  [{verdict}] 2.2 经营CF持续为负: {detail}
  [{verdict}] 2.3 高现金+高息借债: {detail}

── Layer 3: 资产负债表信号 ────────────────────────
  [{verdict}] 3.1 应收增速 vs 收入增速: {detail}
  [{verdict}] 3.2 存货周转率↓+毛利率↑: {detail}
  [{verdict}] 3.3 在建工程不转固: {detail}
  [{verdict}] 3.4 长期待摊费用大增: {detail}
  [{verdict}] 3.5 坏账计提 vs 同行: {detail}

── Layer 4: 交叉验证 ─────────────────────────────
  [{verdict}] 4.1 经营CF/净利润: {detail}
  [{verdict}] 4.2 销售收现/营收: {detail}
  [{verdict}] 4.3 利润膨胀→资产膨胀: {detail}
  [{verdict}] 4.4 核心利润 vs 净利润: {detail}
  [{verdict}] 4.5 净利润↑+FCF为负: {detail}

── Layer 5: 非财务信号 ────────────────────────────
  [{verdict}] 5.1 更换审计机构: {detail}
  [{verdict}] 5.2 大股东减持: {detail}
  [{verdict}] 5.3 财务总监更换: {detail}
  [{verdict}] 5.4 独董辞职: {detail}
  [{verdict}] 5.5 可疑供应商/客户: {detail}
  [{verdict}] 5.6 跨行业收购: {detail}
  [{verdict}] 5.7 商誉过大: {detail}
  [{verdict}] 5.8 其他应收款异常: {detail}
  [{verdict}] 5.9 监管处罚/立案调查: {detail}

── Layer 6: 行业特有风险 ──────────────────────────
  [{verdict}] 6.1 农林渔牧行业: {detail}
  [{verdict}] 6.2 研发资本化比例: {detail}

━━━ 关键发现摘要 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  {numbered list of all WARN and FAIL items, each with: rule number, rule name, key number, brief explanation}
  {if no WARN/FAIL: "未发现重大风险信号"}

━━━ 积极信号 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  {list notable positive findings, e.g. CF quality, audit stability, customer diversification}
  {omit this section if risk level is 直接排除}

━━━ 需人工验证项目 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  {list of all SKIP items with reason}
  {if no SKIP: "无"}

━━━ 方法论 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  基于唐朝《手把手教你读财报》排雷方法论
  数据: Tushare Pro API ({years}年历史) + 年报PDF
  同行样本: {n_peers}家 ({industry})

  免责声明: 本分析仅供参考，不构成投资建议。
══════════════════════════════════════════════════

Output format rules (严格执行)

1. Verdict 格式: [PASS], [WARN], [FAIL], [SKIP] — 方括号 + 大写英文
2. Detail 要求: 每条规则必须包含实际数据值，不得仅写 PASS/FAIL
   • 好: 毛利率 33.67%, YoY -0.70pp, 同行中位数 26.74%
   • 坏: 毛利率正常 或 无异常
3. WARN/FAIL 解释: 用 → 后跟一句话解释触发原因
4. 多年趋势: 列出具体数值序列，如 2.44, 2.92, 3.54, 1.26, 2.36 (近5年)
5. 一行原则: 每条规则尽量控制在一行，超长时最多两行
6. 积极信号段: 风险等级为低/中风险时输出，列出公司的财务亮点
7. 完整输出: 报告的每一个段落（从 ══ 到 ══）必须完整输出到用户终端，不得截断

Phase 6: Save Output Files

MANDATORY: 必须在 Phase 5 输出报告到终端之后，再保存文件。执行顺序：先输出 → 再保存。不得跳过任何一步。

Save all artifacts to the output directory using the Write tool.

File 1: output/{stock_code}/raw_data.md

Already saved in Phase 1. Contains all Tushare data in readable markdown tables.

File 2: output/{stock_code}/report.md

Save the full minesweeper report (the exact output from Phase 5) as a markdown file. Use the same formatting as the terminal output.

File 3: output/{stock_code}/analysis_log.md

Save a detailed analysis log with intermediate calculations:

markdownCopy code
# {name} ({ts_code}) - 排雷分析日志

分析时间: {today}

## 数据摘要

### 利润表关键数据 (近6年)
| 年度 | 营收(亿) | 净利润(亿) | 毛利率% | 费用率% | 减值(万) |
|------|---------|----------|--------|--------|---------|
(computed summary rows)

### 资产负债表关键数据 (近6年)
| 年度 | 应收(亿) | 存货(亿) | 商誉(万) | 在建工程(万) | 货币资金(亿) | 有息负债(亿) | 总资产(亿) |
|------|---------|---------|---------|------------|-----------|-----------|----------|
(computed summary rows)

### 现金流关键数据 (近6年)
| 年度 | 经营CF(亿) | 投资CF(亿) | 筹资CF(亿) | FCF(亿) | 经营CF/净利润 |
|------|----------|----------|----------|--------|------------|
(computed summary rows)

## 规则评估明细

### Rule X.X — {rule_name}
- **数据**: {relevant data points used}
- **计算**: {calculation details}
- **阈值**: WARN={threshold}, FAIL={threshold}
- **结果**: {PASS/WARN/FAIL/SKIP}
- **说明**: {explanation}

(repeat for each of the 28 rules)

## 评分计算

| 层级 | WARN数 | FAIL数 | WARN分 | FAIL分 | 小计 |
|------|--------|--------|--------|--------|------|
(scoring breakdown)

组合加分: {details}
总分: {total}
风险等级: {level}

Final output summary to user

After saving all files, tell the user:

Copy code
已保存分析结果到 output/{stock_code}/:
  - raw_data.md    (Tushare 原始数据, {n} 行)
  - report.md      (排雷报告)
  - analysis_log.md (分析日志与计算过程)
  {- SZxxxxxx_年报_xxxx.pdf (年报PDF) — if downloaded}