eda-report

Skill: EDA Report

Take the executed and annotated notebook at $ARGUMENTS and produce an executive markdown report at reports/eda_report_<dataset_stem>.md. The report is for non-technical stakeholders — managers, product owners, business leads — who need the conclusions without reading 50+ cells.

This skill closes the loop opened by /analyze-dataset (structure with placeholders), /execute-notebook (run the code) and /annotate-findings (write real observations). If the report comes before any of those three, fail with a clear hint.

Principle

A report is synthesis, not transcription. The notebook may have 30 observations across 9 sections; the report has 3-6 findings. The skill must select, rank, and rewrite — never just copy markdown cells verbatim.

Two non-negotiable rules:

1. Every quantitative claim must trace back to a real cell output in the notebook. Never invent numbers (same rule as /annotate-findings).
2. Zero code in the report body. Code lives in the notebook, which is linked from the Appendix.

Step 1 — Validate input

1. Verify the notebook file exists and is readable.
2. Verify it has been executed: at least one code cell with non-empty outputs and a non-null execution_count. If pristine, fail with: Notebook not executed. Run /execute-notebook first.
3. Verify it has been annotated: NO markdown cell should contain the literal marker <!-- annotate-findings: pending -->. If markers remain, fail with: Notebook has unfilled placeholders. Run /annotate-findings first.

Step 2 — Extract content from the notebook

First, ensure the output directory tree exists. Create reports/ and reports/figures/ (relative to the current working directory) if they do not already exist. Use mkdir -p reports/figures via Bash or Path("reports/figures").mkdir(parents=True, exist_ok=True) in the Python script. This is a hard prerequisite for the rest of Step 2 — without these directories, the image extraction below fails silently.