Full-Link Data Analysis

---\r
name: full-link-data-analysis\r
description: Full-Link Data Analysis Engine: From business议题 to analytical report with complete seven-layer architecture. Built-in 15 analysis methods, supports data-aware routing (议题 semantics + data structure + problem type three dimensions), built-in quality assurance, outputs Feishu doc format analytical reports.\r
---\r
\r
# Full-Link Data Analysis\r
\r
## Overview\r
\r
The seven-layer architecture automatically transforms business议题 into structured analytical reports:\r
\r
```\r
L1 Persona Understanding → L2 Data Scope Acquisition → L3 Analysis Data Scoping → L4 Problem Decomposition → L5 Method Selection → L6 Execution & Computation → L7 Result Output\r
```\r
\r
The Agent writes Python analysis code as needed, no pre-packaged code library required. Each layer acts as a quality gate for the next.\r
\r
## Trigger Scenarios\r
\r
- User requests data/business metric analysis\r
- User wants to understand causes/trends/patterns in data\r
- User needs a data analysis report\r
- User mentions "analyze", "attribute", "predict", "cluster", "trend", "data report"\r
- User presents business议题 requiring data-driven insights.\r
\r
## Core Principles\r
\r
- **Code on Demand**: Use `pandas`/`numpy`/`scipy`/`sklearn`/`statsmodels` to write analysis scripts tailored to the actual data.\r
- **Data-Aware Routing**: Method selection综合考虑议题 semantics, data structure, and problem type.\r
- **Quality Gate**: Every conclusion output must include confidence annotation and cross-validation description.\r
\r
## Seven-Layer Process\r
\r
### L1 — Persona Understanding\r
\r
Identify the questioner's role, decision-making scenario, and success criteria.\r
\r
Key Questions:\r
- **Role**: Executive, analyst, product manager, operations personnel?\r
- **Decision Scenario**: Strategic planning, operational review, special investigation?\r
- **Success Criteria**: What metrics/goals define "good"?\r
- **Data Literacy**: How technical should the report be?\r
\r
**Output**: Structured persona summary (role, decision scenario, success criteria).\r
\r
### L2 — Data Scope Acquisition\r
\r
Discover and validate available data sources.\r
\r
Steps:\r
1. Ask user for provided data (files, databases, APIs)\r
2. Upon receiving files, check schema, sample data, inspect types and missing values\r
3. Identify relevant fields/tables\r
4. Confirm data timeliness (update frequency)\r
\r
**Output**: Data inventory (source, schema, quality notes).\r
\r
### L3 — Analysis Data Scoping\r
\r
Narrow down from "all available data" to "data relevant to the议题".\r
\r
Steps:\r
1. Map business problem keywords to required data dimensions\r
2. Filter only necessary fields/records\r
3. Define time window, aggregation granularity, filter conditions\r
4. Identify potential confounding factors\r
\r
**Output**: Data scope specification (dimensions, time window, filters, aggregation granularity).\r
\r
### L4 — Problem Decomposition\r
\r
Break down the business problem into analyzable sub-problems.\r
\r
Framework Selection:\r
- **MECE** (Mutually Exclusive, Collectively Exhaustive): Revenue = Transaction Value × Customer Flow\r
- **Drill-down**: Decompose by region → channel → product layer by layer\r
- **Before/After**: Pre-change vs. post-change\r
- **Cohort**: Group by time/attributes, compare trajectories\r
- **Funnel**: Step-by-step conversion analysis\r
- **Hypothesis Tree**: Structured hypothesis testing\r
\r
**Output**: Problem tree (decomposition structure with clear hypothesis statements).\r
\r
### L5 — Method Selection\r
\r
Select analysis method based on data-aware routing. See `references/routing.md`.\r
\r
Three-Dimensional Routing:\r
1. **议题 Semantics**: Growth, churn, conversion, risk, attribution...\r
2. **Data Structure**: Time series, cross-sectional, panel, hierarchical...\r
3. **Problem Type**: Descriptive, diagnostic, predictive, prescriptive\r
\r
See `references/methods.md` (details on 15 methods).\r
\r
**Output**: Method plan (primary method + alternative cross-validation method).\r
\r
### L6 — Execution & Computation\r
\r
Execute analysis using Python. Process:\r
\r
1. **Environment Check**: `pip list` to confirm `pandas`, `numpy` availability; install missing packages\r
2. **Data Loading**: Load data according to L3 scope specification\r
3. **Data Cleaning**: Handle missing values, outliers, type conversions\r
4. **Analysis Execution**: Write and execute Python script for selected method(s)\r
5. **Cross-Validation**: Run comparison using alternative method (see `references/quality.md`)\r
6. **Result Interpretation**: Extract key figures, statistics, effect sizes\r
\r
Coding Standards:\r
- Use `pandas` for data manipulation, `scipy.stats` for statistical tests, `statsmodels`/`sklearn` for modeling\r
- Print results with clear labels; script output directly constitutes report content\r
- Handle edge cases (empty data, all-null columns, single-category variables)\r
- Output structured text rather than raw numbers\r
\r
### L7 — Result Output\r
\r
Output analytical report in Feishu document format. See `references/feishu-report.md`.\r
\r
Report Structure:\r
1. **Analysis Overview** — Executive summary (1 paragraph)\r
2. **Key Findings** — Data-backed key insights\r
3. **Analysis Process** — Methods, data scope, key assumptions\r
4. **Detailed Results** — Chart descriptions, statistical results, effect sizes\r
5. **Conclusions & Recommendations** — Actionable recommendations with confidence annotations\r
6. **Appendix** — Method details, caveats, data quality notes\r
\r
## Quick Reference\r
\r
| Layer | Action | Reference |\r
|-------|--------|-----------|\r
| L1 Persona | Identify role and decision context | Built-in question set |\r
| L2 Data | Discover and validate data sources | Built-in check process |\r
| L3 Scope | Narrow to relevant data | Built-in mapping logic |\r
| L4 Decompose | Break down into sub-problems | Built-in frameworks |\r
| L5 Method | Select analysis method | `references/methods.md` + `references/routing.md` |\r
| L6 Execute | Write Python analysis | Built-in coding standards |\r
| L7 Output | Generate report | `references/feishu-report.md` + `references/quality.md` |\r
\r
## Checklist\r
\r
- [ ] L1: Confirm persona before accessing data\r
- [ ] L2: Always check schema and sample before analysis\r
- [ ] L3: Clearly state data scope (time, filters, dimensions)\r
- [ ] L4: Must decompose first, do not jump directly to methods\r
- [ ] L5: Consult routing table and document method selection rationale\r
- [ ] L6: Run at least one cross-validation\r
- [ ] L7: Each conclusion must include confidence annotation\r