| name | data-analysis |
| description | Guides Claude through correct Stats Compass MCP tool usage — calling describe tools before using a category, reshaping data before hypothesis tests, presenting download links, and following optimal workflows for EDA, ML, and time series analysis. |
Stats Compass — Data Analysis Skill
Use this skill whenever working with Stats Compass MCP tools.
Core Rules
-
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.
-
Read errors, don't retry blindly. ValidationError messages list exactly which fields are missing or forbidden. Fix the call based on the error.
-
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
-
Always present download_url fields from tool results as clickable links to the user.
-
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.
Finding Files
Use the list_files MCP tool to locate files. Never use bash or shell commands — they run in a sandbox without access to the user's filesystem.
list_files is a top-level tool (not a sub-tool of execute_data_tool):
{ "directory": "~/Downloads" }
Note: list_files only works when Stats Compass runs locally. On the hosted version, files must be uploaded via the upload link.
Loading Data
- If loading fails with a codec error, retry with
encoding: "latin-1".
- Use
get_schema after loading to confirm column names and dtypes.
EDA
- Use
run_eda_report_workflow for a full automated report (stats, correlations, missing data, charts) in one call.
- Use
describe with include: "all" for both numeric and categorical column stats.
Preprocessing
- Use
run_preprocessing_workflow for automated cleaning (imputation, outliers, deduplication).
- The cleaned DataFrame is saved under a new name returned in the result. Use that name for subsequent analysis, not the original.
Time Series Forecasting
Always use fast defaults to avoid timeouts. ARIMA grid search and stationarity tests add 20–60 seconds on large datasets.
Step 1 — Pre-filter to recent data:
{
"tool_name": "filter_dataframe",
"params": { "tail": 500, "save_as": "ts_recent" }
}
Step 2 — Run with fast config:
{
"dataframe_name": "ts_recent",
"target_column": "Close",
"date_column": "Date",
"config": {
"auto_find_params": false,
"arima_order": [1, 1, 1],
"check_stationarity": false,
"forecast_periods": 30
}
}
Only enable auto_find_params: true or check_stationarity: true if the user explicitly requests optimised parameters or stationarity diagnostics.