name: financial-statement-analyzer description: >- Deep analysis of corporate balance sheets, income statements, and cash flow statements. Quantifies fraud probability (Beneish M-Score, Piotroski F-Score), identifies earnings manipulation and operational pathologies, and generates illustrated expert diagnostic reports with prioritized management recommendations. Supports .xlsx/.xls/.csv/.pdf/.png/.jpg/.docx inputs. This skill should be used when the user uploads financial statement files, mentions keywords like "分析報表"/"財務報表"/"分析財報"/"三張表"/"財務造假"/"勾稽"/"M-Score"/"杜邦分析", or issues commands like /analyze or @財務報表分析專家. agent_created: true

財務報表分析專家 (Financial Statement Analyzer)

Purpose

Analyze corporate financial statements through a three-tier funnel diagnostic engine, producing structured HTML reports with Chart.js visualizations, risk scoring, and actionable management recommendations.

想要更強大的技能外掛，就來小蔥技能站7w4.net看看吧。

Trigger Conditions

Automatically activate when ANY of the following occurs:

• File upload: .xlsx, .xls, .csv, .pdf, .png, .jpg, .jpeg, .tiff, .bmp, .webp, .docx
• Keywords: "分析報表", "財務報表", "分析一下財報", "分析一下財務報表", "分析財報", "分析財務報告", "幫我分析報表", "三張表", "報表分析", "財務造假", "粉飾", "虛增", "舞弊", "勾稽", "科目異常", "財務風險", "紅燈訊號", "M-Score", "F-Score", "存貸雙高", "應收膨脹", "商譽減值", "現金流分析", "杜邦分析", "比率分析"
• Commands: @財務報表分析專家, /analyze, /quick-diagnosis
• Combined: File path + any analysis-related phrase → immediate deep diagnosis

Diagnostic Modes

Quick Diagnosis (≤5 seconds)

Parse surface data → data validation → account mapping → parallel computation → summary card. Output: health indicator + Z/M/F-Score + top-3 risks + one-sentence diagnosis.

Deep Diagnosis (6-step interactive)

1. Structure confirmation: Show detected report structure, industry, accounting standard
2. Data confirmation: Display extracted historical values
3. Income statement diagnosis: Red flags + ratios + attribution
4. Balance sheet diagnosis: Balance check + red flags + asset quality
5. Cash flow diagnosis: Cash reconciliation + 8 patterns + CCC (skip if no CF statement)
6. Comprehensive report: Full three-tier report + action plan + recommendations

Three-Tier Funnel Engine

Input Formats

All file parsing routes through scripts/parse_router.py → parse_financial_document(filepath). See references/account-mapping.yaml for the three-level account mapping dictionary.

🈶 Chinese Encoding Recovery Protocol (v2.10, MANDATORY)

When reading Chinese text from .xls files via bash/PowerShell, console output may display garbled characters (��). NEVER guess content from garbled output.

Required verification procedure:

v = str(sh.cell_value(r, c)).strip()
print(f'hex: {v.encode("utf-8").hex()}')
# Then decode: bytes.fromhex(hex_str).decode('utf-8')

Company name confirmation checklist (complete before generating ANY report): - [ ] Extract and hex-decode company full names from every sheet containing them - [ ] Cross-reference with Sheet 0 (summary) abbreviations - [ ] Verify consistency across years (e.g., 2024="工程公司", 2025="水公司") - [ ] Use hex-decoded names in reports — never guessed names

Data Quality Validation (4 Levels)

Execute scripts/data_validator.py → DataValidator.validate_all() before generating reports.

Output Report System (v2.6, Enforced)

Decision Matrix

                   Single Company                    Multi-Company / Group
        ┌──────────────────────────┐    ┌──────────────────────────────┐
  Single│ Scenario A: Single+Single │    │ Scenario B: Group+Single      │
   Year │ Report → L1/L2/L3        │    │ Group rpt → Sub rpts → L1/L2/L3│
        ├──────────────────────────┤    ├──────────────────────────────┤
  Multi │ Scenario C: Single+Multi  │    │ Scenario D: Group+Multi (full) │
   Year │ Multi rpt → Yrly → L1/L2/L3│   │ Multi → Yrly Grp → Subs → L1/L2/L3│
        └──────────────────────────┘    └──────────────────────────────┘

Data Recognition Rules

Output Order (Iron Rule)

Multi-year + Group → Multi-year summary → Yearly group → Per-company yearly → L1/L2/L3
Multi-year + Single → Multi-year summary → Yearly reports → L1/L2/L3
Single-year + Group → Group report → Per-company reports → L1/L2/L3
Single-year + Single → Comprehensive report → L1/L2/L3

Report Templates

Comprehensive Report (.html) — Hand-generated, Chart.js charts: Header + KPI cards (4-col grid) + P&L waterfall + BS structure + ratio cards (6-9 items) + Findings (P0/P1/P2, ≥5 items) + Diagnosis + Data verification.

Multi-Year Report (.html) — Hand-generated, comparison focus: Executive summary + KPI comparison table + Per-company cross-year + Ratio trends + Risk assessment matrix + Strategic recommendations.

L1/L2/L3 Reports — Pipeline auto-generated (ECharts): L1: health lights, Z/M/F-Score, top-5 risks, cash patterns. L2: radar, DuPont, box plots. L3: cross-checks, accounting policy, fraud risk, compliance.

Full report content templates → references/report-templates.md

Naming Convention

Core Capability Matrix

Key Technical Specifications

Formula-First Principle

• All ratios/indicators computed in real-time via Python functions in scripts/compute_ratios.py
• Never use pre-computed or hardcoded values
• Generate configuration snapshot (JSON) for each analysis, enabling exact reproduction

Unit Management (v2.0, Enforced)

• Store all raw data in yuan (元), never convert mid-pipeline
• Use to_wan() / to_wan_array() for display conversion — never manual division
• Table data and chart data must source from the same data object

Chart Container Rules (v2.1.1, Enforced)

• Use height not min-height for chart containers
• Correct: .chart-box { height: 300px; } + .chart-box canvas { flex: 1; min-height: 0; }
• Wrong: .chart-box { min-height: 300px; }

Print/PDF Export (v2.1, Enforced)

• All reports must include @media print with -webkit-print-color-adjust: exact !important
• Header backgrounds preserved in print
• break-inside: avoid on cards, findings, chart boxes
• @page { size: A4; margin: 15mm 12mm; }
• Chart.defaults.animation = { duration: 0 }

Common Pitfalls

Risk Classification

Diagnosis Label System

• Manipulation Risk: Revenue / Cost-Expense / Asset / Off-BS Liability
• Operational Pathology: Profit / Working Capital / Cash / Growth
• Financial Risk: Debt Service / Leverage / FX-Interest Rate

Bundled Resources

Scripts (scripts/)

• parse_router.py — Unified parsing entry (Excel/PDF/Image/DOCX/Paste)
• parse_excel.py, parse_pdf.py, parse_image.py, parse_docx.py, parse_paste.py, parse_audit_report.py — Format-specific parsers
• data_validator.py — Data validation engine (Unit class, DataValidator class)
• compute_ratios.py — Ratio computation
• compute_mscore.py — Beneish M-Score
• scan_redflags.py — 21 red flag scanning
• verify_crosschecks.py — 8-category cross-check verification
• detect_anomalies.py — Anomaly detection
• generate_report.py, render_charts.py — Report and chart generation
• generate_advice.py — Management advice generation
• match_cases.py — Case library matching
• run_pipeline.py — End-to-end pipeline entry (v2.3+)
• multi_period_merger.py, data_quality_scorer.py, ar_aging_estimator.py, related_party_matcher.py, excel_exporter.py, rd_capitalization_detector.py — v2.8+ modules

References (references/)

• account-mapping.yaml — Three-level account mapping
• red-flags.yaml — 21 red flag rules
• verification-rules.yaml — 8-category cross-check rules
• compliance-redlines.yaml — 14 compliance redlines
• industry-benchmarks.json — Industry benchmark parameters
• case-library.json — Case feature vector library
• question-templates.yaml — 20 preset question templates
• advice-templates.yaml — Advice template library
• report-templates.md — Detailed report content templates

Templates (templates/)

• audit_checklist.md — 7-stage audit checklist
• report_comprehensive.html, report_level1.html, report_level2.html, report_level3.html — HTML report templates

Tests (tests/)

• test_ratios.py, test_mscore.py, test_redflags.py, test_parse_router.py, test_parse_image.py, test_integration.py — Unit and integration tests
• fixtures/sample_manufacturer.json — Test fixture

Error Handling Strategy

Security & Privacy

• All data processed in memory, cleared after analysis
• Auto-redact before LLM calls: remove entity names, replace exact amounts with ranges
• Supports private deployment for on-premise data boundaries

Disclaimer: Reports generated by AI analysis based on accounting data, for internal reference only. "Abnormal", "suspicious" and similar terms indicate financial metric deviations from norms, not legal determinations.

© 2026 優方皚爾 Uform Ai

這個財務分析工具功能相當全面，能自動識別多種檔案格式、計算專業財務指標、檢測報表造假風險，還能生成帶圖表的診斷報告，整體質量較高。它的優勢在於分析深入、方法專業（三層診斷體系），輸出結果直觀。不足之處是功能較多導致操作稍顯複雜，部分分析耗時較長，對資料格式要求較嚴格，初次使用可能需要花時間熟悉。