Csv Data Analysis

Step 1: Extract Data Features (Execute Script)

Use the execute_skill_script_file tool to run csv_analyzer.py, passing in the data file path (supports .csv, .xlsx, .xls, .tsv formats).

Tool call parameter example:

{
  "skill_name": "csv-data-analysis",
  "script_file_name": "csv_analyzer.py",
  "args": {"input_file": "/path/to/data.csv or /path/to/data.xlsx"}
}

Script return explanation: The script returns a large block of text content containing two parts:

1. [Statistical Summary]: For you to read and understand the dataset's basic characteristics, distributions, correlations, and categorical composition.
2. [Marker-wrapped data blocks]: The script output contains marker data blocks in the format ###KEY_START###...###KEY_END###. The backend automatically captures and injects these into the template — you do not need to handle or pass this content.

Step 2: Generate Insights & Display Report (Inject into Template)

Read the "Statistical Summary" obtained in Step 1, and reason about the business significance or patterns behind the data. Then use the html_interpreter tool to load the template and inject data.

Critical Rules (Must Follow):

1. You must set template_path to csv-data-analysis/templates/report_template.html. The template has built-in complete ECharts rendering JavaScript code and all section titles and footer text. You only need to fill in 9 content placeholders via the data parameter. Never write or modify any JavaScript chart rendering code yourself.

2. Marker data blocks are automatically injected by the backend — you must not pass them in data. The backend automatically extracts content from ###KEY_START###...###KEY_END### markers in the script output and injects it into the template; in this skill, this is primarily CHART_DATA_JSON.

3. *_INSIGHTS, EXEC_SUMMARY, and CONCLUSIONS must use HTML formatting (e.g., <p>, <ul>, <li>, <strong>, <ol>) to ensure proper layout. These are deep business insights you write based on the statistical summary.

4. The output language must match the user's input language. You must also pass the LANG placeholder ("en" or "zh") so that the template's hardcoded section titles, labels, and footer text are displayed in the matching language. Detect language from the user's query: if the user writes in English, set LANG to "en"; if the user writes in Chinese, set LANG to "zh". Default to "zh" when uncertain.

5. Pass exactly 9 placeholders — no more, no less. Auto-injected marker fields like CHART_DATA_JSON are handled by the backend and should not be passed by you. The template already hardcodes all section titles (Distribution Analysis, Correlation Analysis, etc.), insight box titles ("Insights"), and footer text — you do not need to pass these (the template will automatically translate them based on the LANG placeholder).

6. Insight content must be substantive. Each insight module should cover 4 layers of information: observation, possible causes, business impact, and action recommendations. Do not merely restate statistical values or write only a few vague conclusions.

7. Foundational analysis first, attribution as an enhancement module. The first half of the report must focus on analyzing the data features of the CSV itself, including numerical distributions, categorical structures, outliers, correlations, ranking patterns, etc., and should incorporate chart interpretations wherever possible. "Data Anomaly Overview," "Attribution Analysis," and "Root Cause Inference" should appear in the second half as enhancement modules — the entire report must not consist solely of attribution content.

html_interpreter call example:

{
  "template_path": "csv-data-analysis/templates/report_template.html",
  "data": {
    "LANG": "en",
    "REPORT_TITLE": "Sales Dataset Deep Analysis Report",
    "REPORT_SUBTITLE": "Multi-dimensional Data Feature & Business Insight Mining",
    "EXEC_SUMMARY": "<p>This dataset contains 1,000 rows and 5 columns with good data completeness. Key findings include:</p><ul><li><strong>Audience Distribution:</strong> Primarily concentrated in the 25-35 age group...</li></ul>",
    "DISTRIBUTION_INSIGHTS": "<p>The numerical distribution chart reveals that Metric A exhibits a pronounced right-skewed distribution, suggesting...</p>",
    "CORRELATION_INSIGHTS": "<p>The heatmap between variables reveals strong positive correlations, particularly between..., which implies...</p>",
    "CATEGORICAL_INSIGHTS": "<p>Category proportions show that Beijing and Shanghai account for over 50% of the 'City' field.</p>",
    "TIME_SERIES_INSIGHTS": "<p>The time series trend indicates a significant seasonal uptick toward year-end.</p>",
    "CONCLUSIONS": "<p>Based on the comprehensive multi-dimensional analysis, the data exhibits clear structural features and patterns.</p><h3>Recommendations</h3><ul><li>Regularly monitor missing value ratios...</li><li>Focus on high-growth market segments...</li></ul>"
  }
}

Strictly Prohibited:

• Do NOT pass CHART_DATA_JSON or any auto-injected marker fields in data (handled automatically by the backend)
• Do NOT add any JavaScript code in data
• Do NOT omit the template_path parameter (omitting template_path will prevent charts from rendering!)
• Do NOT return static PNG images — this tool has been fully upgraded to ECharts dynamic frontend rendering
• Do NOT pass non-existent placeholders (the template only has the following 9 text placeholders + 1 auto-injected CHART_DATA_JSON; other names will be ignored)

Placeholder Reference (9 total, passed by LLM via data)

The placeholders you need to fill in the template are as follows:

Note: csv_analyzer.py includes ###CHART_DATA_JSON_START###...###CHART_DATA_JSON_END### marker data blocks in its output. The backend automatically extracts and injects these into the template — they should not be passed in data. All section titles in the template (e.g., "Distribution Analysis", "Correlation Analysis", "Conclusions & Recommendations"), insight box titles ("Insights"), and footer text are hardcoded in the HTML and are automatically translated based on the LANG placeholder — they do not need to be passed via placeholders.

Why Choose This Tool?

1. Fast & Lightweight: No more slow Python plotting and bulk PNG generation — only core JSON data is transmitted.
2. Modern Interactive Layout: Fully integrated with Tailwind CSS responsive layouts and Apache ECharts smooth animated interactions.
3. Deep Business Insights: By separating machine-driven data extraction from LLM-driven logical reasoning, this tool produces highly valuable data analysis reports.

File Structure

csv-data-analysis/
├── SKILL.md                        # The skill guide you are currently reading
├── scripts/
│   └── csv_analyzer.py             # Python analysis engine (supports CSV/Excel/TSV, lightweight, no graphics dependencies)
└── templates/
    └── report_template.html        # Responsive ECharts report template (with built-in rendering logic and hardcoded titles)