Chart Rendering Skill

Repeatable SOP for turning a result set into a chart that actually communicates the point.

Steps

1. Match chart type to question. Decide before drawing:

   Question shape	Chart
   "How does X change over time?"	Line (single series) or multi-line
   "Composition / share of total"	Stacked bar, or 100% stacked area
   "Comparison across a small set of categories"	Bar (horizontal if labels are long)
   "Relationship between two numeric variables"	Scatter
   "Distribution of a single variable"	Histogram or box plot
   "Cumulative total"	Area

   If none fit, ask the user which view they want — do not default to "any chart will do".

2. Prepare the data. The query result from the [[sql-analysis]] Skill should already be tidy (one row per data point, one column per dimension). If not, reshape first using pandas (pivot, melt, groupby). Save the cleaned frame to knowledge/cache/<topic>.csv so it can be re-rendered.

3. Render with matplotlib. Write the script to a file under scratch/charts/<topic>.py and execute it via shell_run. Conventions:

   • Title — one sentence answering the question (not "Untitled chart").
   • Axes — label both, include units (Active users (count), Date (UTC)).
   • Legend — include only if there is more than one series.
   • Colours — use the matplotlib default palette unless the user has a stated brand colour in knowledge/style.md.
   • Output — save as PNG (bbox_inches='tight', dpi=150) under knowledge/charts/<topic>.png.

4. Show and explain. In the assistant reply, link the saved image with a markdown image embed (![title](knowledge/charts/<topic>.png)) and write 2–3 sentences interpreting it — what the chart shows, the one thing the user should take away, and any caveat (incomplete latest data point, outliers excluded, etc.).

5. Keep the script reproducible. The chart file is the artefact; the script under scratch/charts/<topic>.py is the source of truth. If the user asks for a tweak ("can we use a log y-axis?"), edit the script and re-run — do not regenerate from scratch.

Anti-patterns

• ❌ Pie charts with > 5 slices.
• ❌ Charting raw counts when proportions are what the user actually asked.
• ❌ A chart whose axes have no labels or unit.
• ❌ Embedding the image without 2 sentences of interpretation.

When to delegate

If the user wants a multi-chart dashboard with narrative around it (e.g. "build me a weekly product health deck"), spawn the report-writer sub-agent — this Skill is for individual charts.