prism

Start from the message, not the chart: trends use lines, category comparisons use bars, distributions use histograms or box plots, relationships use scatter plots, and part-of-whole uses stacked bars. The data shape — categorical vs. continuous, number of series and points — rules chart types in or out.

Use a heatmap normalized per row so it shows recall rather than class frequency. Apply a sequential colormap like viridis, annotate each cell with its value, and order classes so confusable pairs sit adjacent to expose off-diagonal hotspots.

Bar charts encode value through length, so starting the axis above zero exaggerates differences — a 2% change can look like a cliff. Start bars at zero, or switch to a dot plot with a clearly flagged broken axis if the narrow range is genuinely the story.

Use a perceptually uniform sequential colormap such as viridis for ordered magnitude, or a diverging palette only when a meaningful midpoint exists. Avoid rainbow or jet colormaps, which invent banding that is not in the data and fail colorblind-safety checks.

Fifteen overlapping lines are unreadable, so pick the fix by intent: highlight the top performers and gray out the rest, facet into small multiples with a shared y-axis, or replace the curves with a sorted bar of final values if only the endpoint matters.

Avoid pies with more than about three slices, since angles are hard to compare accurately. A sorted horizontal bar chart reads instantly for many categories; keep a pie only when there are few parts that genuinely sum to a whole.