Analyze companies and generate investment research for Chinese A-shares, HK stocks, and US stocks. Two output modes:(1) Tear Sheet — concise 3-5 page PDF. (2) Equity Report — in-depth ≥25 page PDF with financial model. Triggered by ANY company analysis request including:'analyze [company]', 'look into [stock]', 'what do you think of [company]', 'tell me about [stock]', 'research [company]', or explicit keywords:'tear sheet', 'one pager', 'deep dive', 'equity report', or any stock code (e.g., 600519.SH, 0700.HK, AAPL). When intent is ambiguous, ask the user whether they want a report or just a quick conversational answer.
Instrucciones de origen · Vista previa de solo lectura
name
equity-research
description
Analyze companies and generate investment research for Chinese A-shares, HK stocks, and US stocks. Two output modes:(1) Tear Sheet — concise 3-5 page PDF. (2) Equity Report — in-depth ≥25 page PDF with financial model. Triggered by ANY company analysis request including:'analyze [company]', 'look into [stock]', 'what do you think of [company]', 'tell me about [stock]', 'research [company]', or explicit keywords:'tear sheet', 'one pager', 'deep dive', 'equity report', or any stock code (e.g., 600519.SH, 0700.HK, AAPL). When intent is ambiguous, ask the user whether they want a report or just a quick conversational answer.
Equity Research Skill
This skill generates institutional-grade investment research in two modes: Tear Sheet (3-5 page PDF, single session) and Equity Report (≥25 page PDF, 3-task architecture with financial model). Both modes share the same analytical philosophy — the difference is depth, scope, and delivery structure.
Your first job: figure out what the user wants. Then carry the Core Principles into the next file.
Phase 0.0: Router — Intent Clarification + Output Type Detection
Step 1: Detect Language
Detect the user's language from their message. Use that language for ALL follow-up questions and the final report.
User Language
report_language
Chinese (any)
zh
English
en
Mixed / unclear
Match the dominant language in user's message
Step 2: Classify Intent (3 Tiers)
Not every company analysis request needs a full report. Before committing resources, determine what the user actually wants.
"帮我分析一下[公司]", "analyze [company]", "帮我看看[股票]", "look into [stock]", "了解一下[公司]", "what do you think of [company]", 个股分析, 公司分析, or just a stock code (e.g. AAPL, 600519)
→ Ask user (Step 2a)
Tier C: Simple question
"XX公司是做什么的", "what's [company]'s market cap", "when is [stock]'s next earnings"
→ Do NOT trigger this skill. Answer conversationally. No report generation.
Step 2a: Clarify Intent (Tier B only)
When the user's request is ambiguous (Tier B), ask them what level of output they want. Do not assume they want a full report — that wastes their time and tokens.
When the user wants an equity report, ask one more question before starting analysis. The report can be built at two valuation depths — this significantly affects time and complexity.
The equity report can be built at two valuation depths:
Full version (with financial model) — Complete 3-statement financial model (IS/BS/CF) + DCF absolute valuation + comparable companies + sensitivity analysis + historical valuation band. Best for in-depth research with precise target price. Approximately 3 steps.
Streamlined version (Level 1 valuation only) — Comparable company valuation (PE/PB/PS multiples) + consensus expectations + scenario analysis, generating a professional report without complex Excel modeling. Faster turnaround. Approximately 2 steps.
L2: Task 2 produces a real Excel model (8+ tabs) with DCF. Task 3 reads numbers from Excel via openpyxl.
L1: Skip Task 2 entirely. Task 1's research document contains all valuation inputs. Task 3 generates valuation tables directly from the research document (no Excel model needed).
If output_type = TEAR_SHEET: Skip this step entirely. Tear sheets always use Level 1.
⚠️ Core Principles — CARRY THESE INTO THE NEXT FILE
These principles apply to BOTH modes. Read them now. They are NOT repeated in the mode-specific files. If you skip them, you will produce a bad report.
Principle
Requirement
Data Authenticity
All data must have real sources; strictly prohibit fabrication. No placeholders, no "TBD".
Data Verification
Critical data cross-verified by 2+ independent sources
Timeliness
Must use latest financial reports and real-time market data
Recent News Weight
News within past 7 days affecting marginal expectations must be included
Authoritative Source Weight
Prioritize official sources and professional financial institutions
Source Attribution
All data attributed. API data labels original source
Deep Analysis
Mandatory six-dimension framework; each data point answers "so what"
Bull/Bear Balance
Both bullish and bearish viewpoints required — no one-sided analysis
Analysis First
Complete Phase 2-3 analysis, THEN Phase 4 begins layout — never skip ahead
Narrative Consistency
All modules develop around Phase 3 core narrative
Default Output
PDF format
File Read Confirmation
Must confirm required files read before each Phase (see Hard Gate Table in mode file)
I'll generate the streamlined equity report in 2 steps:
Step 1: Deep research analysis (data collection + six-dimension analysis + comps-based valuation) → Research Document
Step 2: Generate final PDF report (≥25 pages, comparable-company valuation)
Starting Step 1 now.
Stop sequential reading here. The mode-specific file you read next has all execution instructions you need. The REFERENCE SECTION below (Output Type Comparison, Task architecture, File Index, Common Rules) is a look-up resource — consult specific sub-sections when you need to locate a file or confirm a mode detail. Do not read it cover-to-cover.
REFERENCE SECTION
You do not need to read this section sequentially. It is reference material for when you need to look up mode details, task architecture, or file locations. The mode-specific files will tell you which files to read and when.
An in-depth ≥25 page PDF built across 2 or 3 Tasks depending on valuation depth. Both L1 and L2 share the same Task 1.
L2 (Full Version — 3 Tasks)
Task 1 (SKILL-equity-task1.md): Phase 0 → Phase 1 → Phase 2 → Phase 3 → Research Document (.md)
↓
Task 2 (SKILL-task2-model.md): Financial Model (.xlsx) + Valuation Analysis (.md)
↓
Task 3 (SKILL-task3-report.md): Final PDF Report (≥25 pages)
L1 (Streamlined Version — 2 Tasks)
Task 1 (SKILL-equity-task1.md): Phase 0 → Phase 1 → Phase 2 → Phase 3 → Research Document (.md)
↓
Task 3 (SKILL-task3-report.md): Final PDF Report (≥25 pages, L1 mode)
Task
Entry File
Input
Output
Acceptance Gate
Task 1
SKILL-equity-task1.md
User request + stock code
Research Document (≥9,000 words)
13 completeness + 4 data quality checks
Task 2 (L2 only)
SKILL-task2-model.md
Task 1 Research Document
Excel Model (8+ tabs) + Valuation Analysis
10 model integrity checks
Task 3
SKILL-task3-report.md
Task 1 doc (+ Task 2 Excel + Valuation for L2)
PDF equity report (≥25 pages)
≥10 number cross-checks vs Excel (L2) / research doc cross-check (L1)
Universal Task Rules (Apply to Both L1 and L2)
Never chain Tasks automatically. Each Task ends with delivery + STOP.
User continues with a single word: "下一步", "继续", or "continue" — no file uploads needed.
Session context carries files. The agent maintains the file list internally. When the user says "continue", the agent reads the previously generated files from the session automatically.
Data flows via files, not memory. Task 2 reads Task 1's document. Task 3 reads Task 2's Excel (L2) or Task 1's document (L1). All financial numbers in Task 3's PDF must tie back to either the Excel model (L2) or the research document (L1).
File Index
All paths are relative to this skill's root directory.
Entry Points
File
Mode / Task
Purpose
SKILL.md
Router
Determines output type, routes to mode-specific file
Financial Model + Valuation → produces Excel model + Valuation Analysis
SKILL-task3-report.md
Equity Report Task 3
Report Generation → produces final PDF equity report
Analysis Frameworks (read on-demand per Hard Gate)
File
Content
analysis/six-dimension-analysis.md
六维分析 complete framework
analysis/investment-logic.md
Investment logic + thesis table spec
Moat classification, earnings-quality checks, management assessment, TAM/SAM/SOM, and competitive deep dive are all defined inline in the two handoff templates (references/research-document-template.md for equity reports, references/analysis-brief-template.md for tear sheets). No separate per-framework files.
Deep Research Modules (Equity Report Only — Phase 2.7)
