eda-narrative

name	eda-narrative
description	Use when exploring data without a fixed question — produces a stakeholder-ready narrative backed by numbers and charts

eda-narrative : exploration into a story

Question-answering (ds-star-plus) and exploration are different jobs. Exploration has no single right answer — it surveys the data and reports what's worth knowing. This skill does that and writes it up as a narrative a stakeholder can read: each claim grounded in a computed number or a chart, each caveat stated.

When this applies

A first look at a new dataset; a stakeholder asking "what's interesting here?"; finding trends, correlations, segments, anomalies, or a summary deck's worth of findings. Run after data-profile (don't narrate dirty data as if clean) and instead of ds-star-plus when there is no single precise question.

The cardinal rule

Every finding carries its evidence. No "sales are strong" — instead "Q4 revenue €1.2M, +18% vs Q3 (chart)". An unsupported narrative is a guess dressed as insight. If the data can't support a claim, say what it can support.

How to run it

Stage 1 — Ground in quality

Read the data-profile watch-outs first (or run it). Know the nulls, sentinels, and grain before you compute a single statistic, so findings aren't artifacts of dirty data.

Stage 2 — Survey (breadth before depth)

Compute the standard exploratory passes, each grounded:

• Univariate: distributions of key numerics and categoricals (with the actual numbers).
• Trends: time series of the main metric(s) if a date column exists — level, growth, seasonality.
• Segments: the main metric broken down by the highest-signal categoricals (top/bottom groups).
• Relationships: correlations / cross-tabs among key variables; note strength, not just sign.
• Anomalies: outliers, sudden changes, suspicious gaps — flagged, not hidden.

Stage 3 — Verify before narrating

Each finding you intend to report must be re-derived by code you actually ran (the DS-STAR rule: executable ≠ correct, and narrated ≠ computed). Drop any "finding" you can't reproduce from a real number or chart.

Stage 4 — Write the narrative

Use references/report_structure.md: a 2–3 line executive summary, then 3–7 findings each with its number/chart and a one-line "so what", then caveats (data quality, scope, what was NOT examined). Generate charts with exact titles and axis labels. Order findings by importance, not by computation order. Produce charts following the quality checklist in references/visualization.md. Use whatever plotting library is in the project env — no library is bundled by ds-crew.

Output

A readable narrative report (markdown + chart files) where every claim is backed by evidence and the caveats are explicit. If a finding raises a precise question worth nailing down, hand it to ds-clarify → ds-star-plus.

Offer the narrative as a Markdown report (and optional self-contained HTML) per ../ds-star-plus/references/export.md. Write to the run dir alongside any charts.

Quick reference

Report structure + the exploratory passes: references/report_structure.md. Profile the data first: ../data-profile/SKILL.md.