astro-plotting-craft

A figure is an argument, and defaults are where the argument leaks. The reflex is matplotlib defaults and Anthropic-orange — neither is the house style. Plot like an expert astrophysicist: the house theme, perceptually-uniform color, honest axes, and typeset math.

House workflow — jaxstroviz is the source of truth

• Apply the theme, don't hand-set rcParams → set_paper() / set_slides() / set_poster() from jaxstroviz own spines, grid, fonts, and the color cycle.
• Size with the helpers → newfig(width=, aspect=), gridfig(nrows, ncols); save with savefig(fig, path) (tight bbox, 300 dpi). Don't reinvent figure sizing.
• Pull brand colors from PALETTE / COLOR_CYCLE, never literal hexes.

See references/house-style.md for the palette spec and the jaxstroviz API map; seaborn palette/property API is in references/seaborn-plotting-reference.md.

Color (these rules are stricter than jaxstroviz's current cycle — see house-style.md)

• Continuous → seaborn/perceptually-uniform colormaps: mako/crest/viridis/magma (sequential); diverging RdBu/vlag (or sns.diverging_palette(...)). Never jet/rainbow/bwr, and never a categorical palette as a continuous map.
• Categorical → pull house colors from jaxstroviz PALETTE/COLOR_CYCLE, but for ≥3 series use a CVD-verified set (sns.color_palette("colorblind"), Wong/Tol) and compose color with marker/linestyle — never hue alone. (jaxstroviz's current cycle has a red-green CVD gap; see house-style.md.) husl for many categories.
• Verify against a colorblind simulator when color is load-bearing.

Axes & scale

• Choose scale from the data: log for power laws, many decades, or positive multiplicative ranges (loglog, semilogy); linear otherwise. Don't default to linear out of habit.
• Label every axis with units in brackets; add minor ticks on log axes.

Typography & math

• Math and units are LaTeX/mathtext, never unicode: r"$M_\odot$", r"$\mathrm{erg\,s^{-1}}$", r"$\rho$" — not M☉, ρ, erg/s. Always use raw strings (r"...").
• The theme sets the fonts (serif/Computer Modern for paper, sans for slides/poster); for reproducibility prefer a bundled/usetex serif over relying on a system "Computer Modern Roman" install (see house-style.md).

Uncertainty & overlays

• Show uncertainty — error bars or fill_between bands; a point estimate with no error is a claim without a bar.
• Beat overplotting — alpha, 2D histogram / hexbin / datashader for dense scatter; don't dump 10⁵ opaque points.
• Legends frameon=False, placed off the data; annotate sparingly.

seaborn objects (optional — layered/exploratory)

• so.Plot(df, x=, y=).add(so.Dot()).add(so.Line()) with so.Nominal/so.Continuous scales (properties in the reference). Use for faceting/layering; classic matplotlib + the theme for final publication figures.

Related

• plot-craft-reviewer — audits a finished plot for the defects this skill prevents.
• plot-design-inspector — which chart and composition best tell the story.
• figure-interpretation-guard — what a finished figure does and doesn't license you to conclude.
• publication-figure-validator — journal DPI/font/format compliance at submission.