Turn a raw weekly CRM export (CSV) into a clean regional sales summary: drop duplicate rows, total revenue per region, and emit a structured JSON summary an agent can narrate or hand to a PDF/dashboard step.
This is a bundled example skill — small but real, and used to demonstrate the creator's validation, pipeline, and eval-rollout machinery end to end.
Activates on requests like "clean this CRM export", "weekly sales report", "regional totals from this CSV". Do not activate on general spreadsheet or analytics questions that don't involve a CRM export.
A CSV with at least region and amount columns (extra columns are ignored).
Duplicate rows (identical across all columns) are removed before totalling.
One command produces the summary:
python3 scripts/run_pipeline.py --input <export.csv> --output summary.json
Output JSON shape:
{
"rows_in": 120,
"rows_after_dedup": 118,
"regions": {"West": 40210.5, "East": 38110.0},
"grand_total": 78320.5
}
rows_after_dedup does not reconcile with grand_total. Rows with a blank
or missing region are skipped when totalling but still counted in
rows_after_dedup, so a CSV with unassigned rows reports more rows than it
actually summed. Do not present the two numbers as if one explains the other —
if they must reconcile, filter blank regions before running.amount silently becomes 0.0, it does not raise. A column
of N/A or — produces a region total of 0.0 that looks like a real zero.
Check the input for non-numeric amounts before trusting a suspiciously low total.region and amount. Two rows with
identical sales data but a differing timestamp or record-ID column are both kept.
The "extra columns are ignored" note above applies to totalling, not to dedup.