Business Analysis

Business Analysis Learning Agent Skill

1. Purpose

Help business beginners learn data analysis while completing real analysis tasks.

The Agent should not only produce outputs, but also explain the reasoning behind the analysis.

1.1 Output Language Policy

• Repository artifacts are written in English.
• Default user-facing outputs are Chinese (unless the user requests English):
• the analysis text report
• Excel pivot/dashboard workbook
• PDF report
• PPT/slide deliverables (if generated)
• If the user requests, the Agent can output English (or bilingual).

1.2 Platform Compatibility

This skill is platform-agnostic. It is not limited to Claude Code.

Any AI agent can use this skill if it can:

• read instruction files
• inspect or receive a business dataset
• follow a structured workflow
• create files when deliverables are required

Claude Code-specific files under .claude/ are optional adapters. The reusable behavior is defined by this SKILL.md, the templates, examples, and test cases.

2. Target User

• beginners in business, management, marketing, operations, or data analysis
• users who can ask an AI Agent to analyze data but do not yet understand the analytical logic
• users who need explanations of metrics, business concepts, and abbreviations

3. When to Use This Skill

• CSV or Excel business datasets
• retail sales data
• customer behavior data
• product performance data
• marketing campaign data
• store operation data
• revenue/profit data
• public business datasets

4. Core Principle and Positioning

4.1 What This Skill Is

Business Analysis Skill is a data analysis workflow skill designed for beginners and business scenarios.

It guides the Agent to complete business data analysis in a stable, explainable, and teachable way.

Its core is NOT to replace SPSS, Stata, Excel, machine learning platforms, or professional data science tools.

4.2 Primary Goals

1. Help users understand what is in the dataset.
2. Help users understand which metrics can be computed.
3. Help users understand why a particular analysis method is chosen.
4. Help users understand the business meaning of analysis results.
5. Help users know what further advanced analyses the current data could support.
6. Prevent the Agent from forcibly applying advanced methods when data conditions are insufficient.
7. Prevent over-interpreting correlation, statistical significance, or prediction results as definitive conclusions.

4.3 Core Teaching Principle — "Learn by Doing"

Business Analysis Skill is not just a report generator. It is a "learn by doing" (做中学) data analysis teaching Skill designed for business beginners.

Its output goal is not only to tell users what the results are, but to teach them:

• Why the analysis is done this way
• How the metrics are computed from the data
• What the results mean in business language
• What the results CANNOT prove
• How to reuse the thinking next time

The output logic upgrades from:

Finding → Suggestion

To:

Finding → Method Explanation → Metric Explanation → Business Interpretation → Risk Boundary → Learning Transfer

For each important analysis step, the Agent must explain these 8 elements:

1. Analysis result — what the data shows
2. Method explanation — why this method was chosen for this data
3. Metric explanation — what the metric means and how it's calculated
4. Formula explanation — the calculation with field sources
5. Field sources — which columns were used
6. Business meaning — what the result means for business decisions
7. Risk boundary — what the result CANNOT prove or what could go wrong
8. Beginner reuse — how to apply this method when seeing similar data next time

See docs/methodology/teaching_output_rules.md for the complete teaching output specification. See docs/methodology/report_structure.md for the standard report structure with teaching annotations.

5. Analysis Method Layering

All analysis methods are organized into three tiers. The Agent must follow this layering when choosing what to execute.

Tier 1: Default Basic Analysis (Always Execute)

These methods run by default for every dataset. They answer "what is in the data?"

• Data profiling — rows, columns, field types
• Field semantics recognition — what each field likely means
• Missing value check
• Duplicate value check
• Outlier / abnormal value check
• Basic metric computation — matching fields to the metric registry
• Grouped aggregation — sum, mean, count by category
• Ranking analysis — top/bottom by key numeric metrics
• Trend analysis — month-over-month changes (when a date field exists)
• Descriptive statistics — mean, median, std, quartiles
• Data quality risk warnings
• Business interpretation and beginner-friendly explanation

These answer:

• What is in the dataset?
• Which fields are important?
• Which metrics are highest or lowest?
• Which segments show large differences?
• Which time periods show growth or decline?
• Are there data quality issues?

Tier 2: Advanced Analysis Recommendations (Suggest, Do Not Auto-Execute)

When data conditions are met, the Agent recommends these methods but does NOT automatically execute them unless the user confirms or the output mode is audit_report with explicit user intent.

• Correlation analysis
• RFM analysis
• t-test
• Chi-square test
• Linear regression
• Logistic regression / classification
• Clustering analysis
• Cohort analysis
• Seasonality analysis
• Time series forecasting
• Sentiment analysis
• Causal inference

For each recommendation, the Agent must explain:

• Why this method is suitable for the current data
• Which fields are required
• Which fields are already present
• Which fields are missing
• What business question this method can answer
• What business question this method CANNOT answer
• Whether auto-execution is suggested
• Whether user confirmation is needed
• What the risks are

Tier 3: Advanced Analysis Execution (Conditional)

A Tier 2 method is executed ONLY when ALL of the following are met:

• The user explicitly requests this method, OR the user has a clear business question that maps to this method
• The target variable or analysis object is clearly identified
• Data fields meet the minimum requirements
• Sample size is adequate
• The Agent can explain the method's boundaries and risks
• The output MUST include risk warnings

If conditions are NOT met, the Agent MUST skip the method and explain why.

See docs/methodology/advanced_method_triggers.md for detailed trigger rules for each method.

6. Standard Workflow

Step 0: Confirm Inputs and Output Location (Required)

Before starting analysis, ask the user three questions in order:

0a. Confirm the input dataset:

Ask where the data file or data folder is located.

0b. Ask for output mode:

• beginner_summary: short beginner-friendly output (default)
• standard_report: medium report with key reasoning
• audit_report: full detailed output for testing or portfolio review

Do not proceed until the user selects a mode (or explicitly chooses the default).

0c. Ask for output folder:

Ask where to save Excel, HTML, PDF, PPT, and audit log files.

Example prompt:

Where should I save the analysis output files (Excel, HTML, PDF, audit log)?

If you're not sure, I can use:
  /analysis_outputs//

The Agent must confirm the final output directory with the user before executing analysis.

• Never silently write output files into the Skill installation directory.
• Never default to the repository's output/ folder without user confirmation.
• If the output directory already contains files with the same names, create a timestamped subdirectory (e.g., analysis_2026-06-03_143052/) to avoid silent overwrites.

Step 1: Understand the User's Goal

If the user goal is unclear, infer possible business questions from the dataset and clearly state assumptions.

Step 2: Inspect the Dataset

Check:

• rows and columns
• field names
• field types
• missing values
• duplicate values
• time fields
• category fields
• numerical fields
• possible business meaning of fields

Before producing conclusions, perform a dataset risk check:

• cancelled orders
• negative or zero quantity
• zero or abnormal unit price
• missing customer identifiers
• non-product transaction items
• incomplete time periods
• dominant market or customer concentration
• duplicate rows

For each relevant issue, explain why it matters, which analysis it affects, and whether the data should be removed, separated, or kept for another analysis.

Step 3: Translate Data Fields into Business Questions

Example mapping:

• sales_amount -> revenue performance
• customer_id -> repeat purchase or customer behavior
• order_date -> trend or seasonality
• category -> product/category comparison
• store_id -> store performance

Use a field-to-analysis mapping table when the dataset has enough structure:

| Field | Metric | Business Question | Analysis Method | |---|---|---|---| | InvoiceNo | Order Count, AOV | How large and efficient are orders? | descriptive statistics | | Quantity * UnitPrice | Revenue | What is total sales value? | metric aggregation | | InvoiceDate | Monthly Revenue | Is there trend or seasonality? | trend analysis | | Country | Revenue Contribution | Which markets contribute most? | contribution analysis | | CustomerID | Repeat Rate, RFM | What is customer quality? | customer analysis | | StockCode / Description | Product Revenue | Which products drive revenue? | product analysis |

Step 4: Choose Analysis Methods

Methods are organized into three tiers per Section 5 (Analysis Method Layering).

Tier 1 — Always execute:

• descriptive statistics
• trend analysis
• ranking analysis
• contribution analysis
• data quality checks
• basic metric computation

Tier 2 — Recommend with explanation (do not auto-execute):

• correlation analysis
• RFM analysis
• t-test
• chi-square test
• linear regression
• logistic regression
• clustering analysis
• cohort analysis
• seasonality analysis
• time series forecasting
• sentiment analysis

Tier 3 — Execute only when conditions met and user confirms:

• causal inference
• any method requiring user-specified target variables

The Agent must explain why each method is chosen or skipped. For Tier 2 methods, see docs/methodology/advanced_method_triggers.md for detailed trigger rules.

Step 5: Execute the Analysis

Output should include:

• tables
• charts if useful
• key findings
• business interpretation
• beginner learning notes

For every major step, include learning-oriented reasoning:

1. What business question this step answers
2. Why this method is appropriate
3. Which dataset fields are used
4. How the metric is calculated
5. What the result means in business language
6. What a beginner should learn from this step
7. What limitations or risks may affect the conclusion

When cleaning data, include a cleaning impact assessment:

Cleaning Rule: Rule: Business Reason: Rows Removed: Percentage Removed: Affected Analysis: Treatment:

Step 6: Explain Metrics and Abbreviations

For each metric or abbreviation, explain:

• full name
• simple meaning
• formula
• current calculation when data is available
• business use
• common misunderstanding
• limitation

Examples include:

• KPI
• GMV
• AOV
• CAC
• ROI
• retention rate
• conversion rate
• contribution rate
• profit margin

Step 7: Generate Business Insights

Each insight should include:

• data evidence
• business interpretation
• possible reason
• suggested action
• risk or limitation

Recommendations must be actionable and tied to evidence. Use this structure:

Finding -> Business Meaning -> Recommended Action -> Metric to Track -> Data Needed

Step 8: Provide Beginner Learning Feedback

At the end, summarize:

• business concepts learned
• analysis methods learned
• metrics learned
• reusable thinking pattern
• suggested next learning step

End with a structured Recommended Next Analysis section rather than a conversational closing.

Step 9: PPT Generation (Optional)

After the analysis report and deliverables are complete, ask the user if they would like to generate a presentation:

Ask in the default output language (Chinese unless the user requested English).

Mandatory presentation-generation boundary:

• Every presentation must be generated by the Agent after reading the matching bundled presentation Skill under vendor/agent_presentation_skills/.
• Do not generate a presentation with scripts/run_analysis.py, a generic python-pptx template, or any other shortcut that bypasses the bundled presentation Skill.
• The universal runner intentionally generates only Excel, HTML, PDF, and audit log files.
• Save the presentation into the same user-confirmed output folder unless the user chooses another location.
• Before delivery, follow the selected presentation Skill's QA/checklist requirements and verify titles, page numbers, repeated slides, output language, and layout.

Offer two style options with a brief description:

1. Editorial magazine style (Guizang Style A)

• Two-page flipping HTML deck, WebGL motion background, narrative layouts
• 5 themes available (see vendor/agent_presentation_skills/guizang-ppt-skill/references/themes.md)
• Deliverable: standalone HTML file

1. Swiss / International Typographic Style (Guizang Style B)

• Minimal grid system, strong hierarchy, structured layouts
• 4 accent colors available (see vendor/agent_presentation_skills/guizang-ppt-skill/references/themes-swiss.md)
• Deliverable: standalone HTML file

Implementation instructions — presentation templates are bundled in vendor/agent_presentation_skills/ (guizang-ppt-skill for Styles A/B):

• If the user selects Editorial magazine style (Guizang Style A):

1. Read vendor/agent_presentation_skills/guizang-ppt-skill/assets/template.html to understand the slide structure
2. Read vendor/agent_presentation_skills/guizang-ppt-skill/references/themes.md for theme reference (5 themes listed in the file)
3. Read vendor/agent_presentation_skills/guizang-ppt-skill/references/layouts.md for available slide layouts
4. Read vendor/agent_presentation_skills/guizang-ppt-skill/references/checklist.md for QA requirements
5. Generate an HTML file with all slides, using the template structure and chosen theme

• If the user selects Swiss / International Typographic Style (Guizang Style B):

1. Read vendor/agent_presentation_skills/guizang-ppt-skill/assets/template-swiss.html to understand the slide structure
2. Read vendor/agent_presentation_skills/guizang-ppt-skill/references/themes-swiss.md for accent color reference (options listed in the file)
3. Read vendor/agent_presentation_skills/guizang-ppt-skill/references/layouts-swiss.md for available S layouts
4. Read vendor/agent_presentation_skills/guizang-ppt-skill/references/checklist.md for QA requirements
5. Generate an HTML file with all slides, using the template structure and chosen theme

• If the user declines PPT generation, skip this step.

Required Deliverables (All Modes)

Every analysis task must generate two deliverable files in addition to the text report:

1. Excel Pivot Table File (*_analysis.xlsx)

• Sheet 1: Overall comparison (key metric averages by group)
• Sheet 2+: One sheet per analysis dimension (by time, category, region, segment, etc.)
• Cross-analysis sheet for 2D breakdowns
• Dashboard sheet with KPI boxes, thinking model summary, and recommendation table
• Charts (bar, pie) anchored to data ranges
• All values stored as float/int, not text

1. PDF Report (*_report.pdf)

• Cover page with dataset description and key metrics
• Executive summary
• Thinking models applied (data evidence, business insig

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Ysx12138
• Source: Ysx12138/business-analysis-learning-agent-skill
• License: MIT

Install and usage instructions live in the source repository linked above.