nice-figures

Soft Research Figures

A matplotlib visual register for AI research figures — alignment plots, training curves, eval bar charts, scaling scatter plots. Bold sans-serif titles, soft pastel + warm earth palettes, smoothed trends with shaded bands, minimal axes. Conference-ready (white background) by default.

Use this skill for posters, slide decks, full-page appendix figures, blog posts, and conferences that lean toward modern ML aesthetics (NeurIPS, ICML, ICLR talks/posters). It favors a relaxed, presentation-friendly look rather than the tight single-column widths of a strict print journal — but the white-background default is clean enough for most paper figures too.

First steps

1. Copy scripts/soft_style.py to the working directory.
2. Read references/chart_recipes.md and pick the archetype closest to what the user described.
3. Plan: how many panels, what data, what colors, which "better" direction.
4. Write the script using the recipe; adapt only what's necessary.
5. Export both PDF and PNG via save_figure().

When to use this skill

Trigger on phrases like:

• "alignment plot," "training curve," "eval comparison"
• "Anthropic-style figure," "research-blog style," "soft pastel"
• "conference figure," "poster figure," "talk slide chart"
• "make me a [bar chart / ROC / heatmap / box plot / parity plot / scaling law / distribution] in our soft style"
• "match the figures I showed you" (when reference images are present)
• "make it pretty for the appendix"

Don't trigger on:

• Pure data exploration (Jupyter scratch plots) — overkill
• Schematics, diagrams, flowcharts — wrong tool entirely

The sixteen archetypes

Anthropic-blog originals (1–5): the soft-pastel research-figure register, smoothing bands, rounded bars.

#	Recipe	Use for
1	Multi-panel trend with band	Training curves, multi-benchmark comparison
2	Scatter with error bars + baseline	Scaling plots, ablation comparisons
3	Grouped bars + error bars (multi-condition)	Model comparison across metrics
4	Grouped bars (3 conditions × N scenarios)	Intervention comparison
5	Multi-line with markers	Single-parameter sweeps

Standard publication (6–11): ROC, heatmaps, distributions, etc.

#	Recipe	Use for
6	Heatmap / confusion matrix	Similarity matrices, attention, ablation grids
7	ROC / PR curve	Classification diagnostics
8	Distribution comparison (hist + KDE)	Property distributions, score shifts
9	Box / violin plot	Seed stability, dataset comparison
10	Scaling-law plot	Log-log scatter with power-law fit
11	Parity / calibration plot	Predicted vs. actual, regression diagnostics

Domain-specific (12–16): embedding scatter, ECDFs, forest plots, horizontal rankings, Pareto trade-offs.

#	Recipe	Use for
12	2D embedding scatter	t-SNE / UMAP / PCA visualizations
13	ECDF / CDF plot	Distribution comparison without binning, tail behavior
14	Forest plot / dot-and-whisker	Effect sizes with CIs across many interventions
15	Horizontal bar chart (sorted)	Many-method comparisons (>8 categories)
16	Pareto front / trade-off plot	Accuracy-vs-cost, quality-vs-latency trade-offs

Full code for each lives in references/chart_recipes.md. Copy the nearest one and adapt.

Style at a glance

from soft_style import (
    configure_style, figure_title, panel_subtitle, better_badge,
    smooth_curve, rolling_band, rounded_bars, rounded_hbars, save_figure,
    LINE_PALETTE, BAR_PALETTE, MULTILINE_PALETTE,
    CMAP_SEQUENTIAL, CMAP_DIVERGING,
)

configure_style()           # white background (default, conference-ready)
# configure_style(cream_bg=True)  # original warm cream blog background

Palettes

• LINE_PALETTE — blue, mustard, sage, pink. Use for line plots with shaded bands. Always assign colors in this order for visual consistency across figures in the same paper.
• BAR_PALETTE — coral, peach, dark_gray, light. Warm-earth grouped bar charts. The coral and peach pair particularly well to show "baseline vs. intervention."
• MULTILINE_PALETTE — teal, pink, dark_green, mustard, blue. Up to 5 categorical lines. Use white-edged markers on every point.
• CMAP_SEQUENTIAL — white → cream → peach → coral → dark coral. Use for unsigned data: similarity matrices, confusion matrices, attention weights, density heatmaps.
• CMAP_DIVERGING — blue → cream → coral. Use for centered data with symmetric vmin/vmax: correlation matrices, log-fold-change, signed deviations from a baseline.

Typography

• Figure title: 18 pt, bold, near-black (figure_title())
• Panel subtitle: 12 pt, normal, gray (panel_subtitle())
• Axis labels: 11 pt, gray (#6B6960) — set automatically
• Tick labels: 9.5 pt, lighter gray (#8C8A82) — set automatically
• Legend: 9.5 pt, white-filled rounded box with thin gray edge

Inter is the preferred font; the skill falls back to Helvetica, Arial, then DejaVu Sans. On systems with Inter installed, figures pick it up automatically.

Axes

• Only bottom and left spines, in soft gray (#B8B5AC)
• No grid lines
• Tick direction outward, short
• Labels in muted gray, never pure black

Rounded bars (the signature look)

The Anthropic alignment-figure bar charts have a subtle but distinctive feature: bars sit square on the baseline but the top corners are softly rounded. Use rounded_bars() instead of ax.bar():

from matplotlib.patches import Patch

fig, ax = plt.subplots(figsize=(11, 5.5))
# IMPORTANT: set ylim BEFORE rounded_bars so corners are circular in
# display space — not afterwards
ax.set_xlim(-0.5, n_categories - 0.5)
ax.set_ylim(0, y_max)

for i, (col, lab) in enumerate(zip(colors, conditions)):
    offset = (i - (n - 1) / 2) * width
    rounded_bars(ax, x + offset, heights, width=width, color=col,
                 radius_frac=0.20, label=lab)
    ax.errorbar(x + offset, heights, yerr=errs, fmt="none",
                ecolor="#3a3a37", elinewidth=1.1, capsize=3, zorder=3)

# Use Patch handles for the legend — PathPatch auto-handles look wrong
handles = [Patch(facecolor=c, label=l) for c, l in zip(colors, conditions)]
ax.legend(handles=handles, loc="upper right")

radius_frac is corner radius as a fraction of bar width:

• 0.10–0.15: barely rounded (subtle)
• 0.18–0.22: matches the Anthropic look (default sweet spot)
• 0.30–0.40: pronounced rounded caps
• 0.50: half-pill cap (fully rounded top)

The corner radius is computed in display pixels then converted back to data units, so corners look consistently circular regardless of axis aspect ratio. This requires axis limits to be set before calling rounded_bars(). Forgetting this is the most common gotcha — corners will look stretched if you set ylim after.

For very short bars (height < radius), the corner caps at half the height, giving a nice rounded-button look. This is what you want.

Horizontal bars use rounded_hbars() with the same conventions mirrored 90°: bars grow from x=0 to the right, with a square left edge (flush against the baseline) and softly rounded right corners. The signature is rounded_hbars(ax, y, widths, height, color, radius_frac=0.20, ...). Use this for sorted leaderboards, method rankings, and any other many-category comparison (recipe 15).

Background choice

The skill defaults to white because that's what conferences and journals want. The original Anthropic blog figures use a warm cream outer + lighter cream panel — opt in with configure_style(cream_bg=True) only when generating blog-style assets, never for paper submissions.

When in doubt: white.

Smoothing and confidence bands

For training-trace data:

# Smoothed trend line
xs, ys = smooth_curve(steps, scores, frac=0.30)
ax.plot(xs, ys, color=c, linewidth=2.0)

# Shaded ± 0.85·σ band (visually tight, not statistically formal)
xs, lo, hi, _ = rolling_band(steps, scores, frac=0.30, k=0.85)
ax.fill_between(xs, lo, hi, color=c, alpha=0.18, linewidth=0)

# Underlying scatter
ax.scatter(steps, scores, color=c, s=8, alpha=0.55, linewidth=0)

frac controls smoothing bandwidth (fraction of points in the kernel). 0.30 is the sweet spot for typical training curves. Drop to 0.15–0.20 if you want to preserve sharp transitions; push up to 0.40+ for very noisy data. k=0.85 gives a visually tight band that reads as "trend uncertainty"; raise to k=1.0 for ±1σ or k=1.96 for ~95% CI.

For real LOWESS with handling for outliers and non-uniform x, prefer statsmodels.nonparametric.smoothers_lowess — soft_style's smoother is a lightweight Gaussian-kernel approximation that depends only on numpy.

"↓better" / "↑better" badge

Place a small rounded badge in any corner of an axis to make the "direction of better" unambiguous:

better_badge(ax, direction="down", loc="upper right")

Skip this badge when:

• The metric is already obviously directional (accuracy, F1)
• The plot has only one series (the trend is its own narrative)
• The axis label already says "(lower=better)" or "(↑ higher is better)"

Use it when:

• Multiple panels share the same y-axis semantics
• The metric is unusual or ambiguous (e.g., "misalignment score")
• A reader skimming the figure should not have to read the caption

Common mistakes

1. Mixing palettes. Don't combine LINE_PALETTE and MULTILINE_PALETTE in the same figure; their blues and pinks differ slightly and look like printing errors. Pick one palette per figure.
2. Too many series per panel. This register handles ~4 series cleanly. Beyond 5, split into small multiples.
3. Forgetting figure_title(). The bold display title is a defining feature of this register. Don't substitute fig.suptitle(text) with default weight — it looks like a different style.
4. Sharing y-axis when scales differ. In the multi-panel archetype, independent y-scales communicate "these are separate benchmarks." Only share when the panels are directly comparable.
5. Cream background for paper figures. Reviewers will flag it as off-spec for nearly every conference. White by default; cream only for blog/marketing assets.
6. Pure-black text. Use #1B1A18 for titles, #6B6960 for labels, #8C8A82 for ticks. The skill sets these automatically — don't override with color="black".
7. Using ax.bar() instead of rounded_bars(). Square-topped bars read as "different aesthetic" — they don't belong in this register. Always use rounded_bars() for bar charts.
8. Setting ylim after rounded_bars(). Corner radii are computed at call time using current axis limits. Always set limits first.

Export

save_figure(fig, "figure_1")
# Saves figure_1.pdf and figure_1.png at 300 dpi with bbox_inches="tight"
# and facecolor matching the figure (white or cream)

For very large figures or print-quality posters:

save_figure(fig, "poster_fig", formats=("pdf", "png"), dpi=600)

Checklist before delivering

• configure_style() called before any plotting
• Title applied via figure_title() (bold, large, sentence case)
• Each panel labeled via panel_subtitle() (multi-panel only)
• Y-axis units / direction-of-better clear (badge or label suffix)
• Confidence band, smoothing, and scatter all in same color per series
• No grid, only bottom + left spines
• White background unless cream_bg=True was explicitly requested
• Exported as both PDF and PNG via save_figure()
• If used in a paper: figure also legible at 50% scale (sanity check)

If the user has actual data

When real arrays or a CSV are provided:

1. Load the data first; print shapes and a head so the user can confirm.
2. Map columns to the recipe's variables. Don't invent placeholder data alongside theirs — replace synthetic data entirely.
3. If multiple seeds are present, prefer fill_between(mean ± std) over rolling_band(), which uses kernel smoothing instead.

When the user describes a figure conceptually with no data, generate realistic synthetic data via np.random.seed(...) for reproducibility, and clearly mark it as a placeholder in a comment.