CNS Plot

Build a compact visual argument, not a decorated default plot. Use cnsplots when its plot types fit; reuse its design language with ordinary Matplotlib or Seaborn when a custom artist is clearer.

Start with a figure contract

Before plotting, establish:

1. The claim each panel must support.
2. The data, unit, grouping, uncertainty, and statistical test behind each mark.
3. The final physical width, panel count, and target formats.
4. Whether downstream editing requires live text and vector objects.
5. The categorical identities and continuous quantities that colors must encode.

Do not infer a journal's current submission limits from the library name. Check the target journal's live guidance when exact limits matter.

Choose the implementation path

• Use cnsplots end to end for its built-in statistical, genomic, survival, heatmap, set, Sankey, and multi-panel plots.
• Use cnsplots styling around Matplotlib/Seaborn when the chart needs custom artists or an API that cnsplots does not expose.
• Use plain Matplotlib/Seaborn with the distilled rules when cnsplots is not installed or would add unnecessary dependencies.
• Use nature-figure first when the figure's evidence structure, panel narrative, or journal strategy is still unresolved.

Install the library into the active project rather than the global interpreter:

uv add cnsplots

The inspected upstream version requires Python 3.10+ and uses Matplotlib 3.10.x. Verify the installed version before relying on version-specific behavior.

Use the core workflow

1. Set style without leaking global state

Prefer a temporary settings context when changing defaults:

import matplotlib as mpl
import cnsplots as cns

with mpl.rc_context(), cns.settings.context(
    palette_qual="OkabeIto",
    palette_seq="YlGnBu_custom",
    title_fontsize=8,
    legend_fontsize=7,
    axes_linewidth=0.5,
):
    cns.figure(width=252, height=180)
    ax = cns.boxplot(data=df, x="group", y="value")
    ax.set(xlabel="", ylabel="Response (a.u.)")
    cns.savefig("figure.svg")

Keep width first and height second. cnsplots treats layout units as pixels at a 72-DPI base, so 252 px is approximately 89 mm.

2. Build on explicit axes

Pass ax= whenever a function accepts it. Return and retain each axes for later label, legend, and limit changes. Use cns.setup_ax(ax) to bring axes produced by another library into the same typographic and spine system.

3. Separate semantic mappings

• Map unordered classes to a qualitative palette with stable class-to-color keys.
• Map ordered magnitudes to a sequential colormap.
• Map signed deviations around a meaningful midpoint to a diverging colormap.
• Keep thresholds, reference values, and nonsignificant observations neutral unless direction itself is the message.
• Add shape, line style, direct labels, or facets when color alone is insufficient.

4. Compose from simple artists

Build unusual plots from a small vocabulary:

• plot, vlines/hlines, and scatter for trajectories and lollipops.
• errorbar for estimates and confidence intervals.
• fill_between for uncertainty bands and flows.
• bar/barh for counts, proportions, and aligned secondary panels.
• text/annotate for direct labels; use collision adjustment for dense labels.
• axvline/axhline for scientifically meaningful references.

Reserve higher zorder for marks and annotations; keep supporting stems, bands, and grids behind them. Rasterize only dense data layers, not labels, axes, or annotations.

5. Lay out panels in final-size units

import cnsplots as cns

mp = cns.multipanel(max_width=518, title="Response overview", loc="left")

ax_a = mp.panel("A", width=150, height=145, pad_left=8, pad_top=6)
cns.boxplot(data=df, x="group", y="value", ax=ax_a)

ax_b = mp.panel(
    "B",
    width=220,
    height=145,
    pad_left=8,
    pad_top=6,
    margin_right=0,
)
cns.scatterplot(data=df2, x="x", y="y", hue="class", ax=ax_b)
cns.take_legend_out(ax=ax_b)

Treat width and height as the axes content area. Treat pad_left/pad_top as the gap between the panel label and rendered decorations. Treat margins as spacing between whole panels. Let the library draw once and remeasure titles, tick labels, and helper axes before export.

6. Export and inspect

Prefer SVG or PDF for plots and high-resolution PNG/TIFF only when raster output is required. Keep SVG text editable with svg.fonttype="none" and PDF fonts as Type 42. cnsplots uses a MuPDF-assisted SVG cleanup path when mutool is available and falls back to standard Matplotlib SVG otherwise.

Export at least one vector file and one raster preview. Inspect the actual outputs at final size for clipping, font substitution, panel-label collisions, colorbar/legend placement, and rasterized text.

Apply the design language

Read design-language.md whenever changing text, colors, graphical objects, panel geometry, or export behavior.

Read recipes.md when choosing a plot family or implementing a cnsplots, Matplotlib, Seaborn, heatmap, volcano, forest, or flow pattern.

Quality gate

Before delivery:

• Verify labels include units and category order is intentional.
• Verify statistical annotations name the test and match the plotted comparison.
• Verify colors retain meaning across panels and remain distinguishable.
• Verify every panel reads at the final physical size.
• Verify legends and colorbars do not obscure data.
• Verify dense layers are rasterized only when useful.
• Verify SVG/PDF text remains editable and Unicode glyphs render correctly.
• Verify the vector and raster exports share the same bounds.
• Keep the plotting code, environment specification, and data transformation reproducible.

Source basis

This skill distills faridrashidi/cnsplots at main commit f0543f2e07d1b602ce5cff18e06c1c9b91fcae15, package version 0.5.0, inspected 2026-07-28. The upstream project is BSD-3-Clause licensed. Reinspect upstream source and docs when the installed version differs.