data-analysis

Stats Compass — Data Analysis Skill

Use this skill whenever working with Stats Compass MCP tools.

Core Rules

1. Call describe_*_tools once before using any new tool category. The schemas are authoritative — do not guess parameter names. Available: describe_eda_tools, describe_data_tools, describe_cleaning_tools, describe_transform_tools, describe_ml_tools, describe_plot_tools.

2. Read errors, don't retry blindly. ValidationError messages list exactly which fields are missing or forbidden. Fix the call based on the error.

3. inspect_data is for scalar expressions only. It does not support value_counts(), groupby(), or any expression returning a Series/DataFrame. Use instead:

   • Value counts / distributions → bar_chart or describe with include: "all"
   • Group aggregations → groupby_aggregate
   • Filtering → filter_dataframe

4. Always present download_url fields from tool results as clickable links to the user.

5. Axis labels: When calling plot tools, always set xlabel and ylabel to descriptive labels. Include units when known (e.g. "Price (USD)"). Never leave axis labels as raw column names.

Hypothesis Tests (t-test, z-test)

Stats Compass uses wide-format input. Both t_test and z_test require two separate columns (column_a, column_b). They do NOT accept a group_column parameter.

Workflow for group comparisons:

Step 1 — Reshape with split_column_by_group:

{
  "tool_name": "split_column_by_group",
  "params": {
    "value_column": "IMDB Score",
    "group_column": "Language",
    "groups": ["English", "Spanish"],
    "save_as": "scores_wide"
  }
}

Step 2 — Run the test on the new columns:

{
  "tool_name": "t_test",
  "dataframe_name": "scores_wide",
  "params": {
    "column_a": "English",
    "column_b": "Spanish",
    "alternative": "two-sided",
    "equal_var": false
  }
}

Use equal_var: false (Welch's t-test) when sample sizes differ. Use z_test only when population standard deviations are known.