visualization

Visualization

Publication-quality data visualization with multiple plotting libraries. Generates actual runnable code that produces journal-ready figures.

Model Routing

This skill runs in the main session: it needs the conversation's data context. For heavy plotting code, you may launch the visualization-agent subagent (pinned to Sonnet in its frontmatter) via the Task/Agent tool; the skill itself does not switch models.

Supported Libraries

matplotlib (Python)

General-purpose plotting. Use for maximum control over every visual element.

• Line plots, bar charts, scatter plots, histograms
• Heatmaps, contour plots, 3D surface plots
• Error bars, annotations, inset axes
• Multi-panel figures with plt.subplots()

seaborn (Python)

Statistical visualization built on matplotlib. Use for statistical plots with minimal code.

• Distribution plots: KDE, violin, box, swarm, strip
• Relationship plots: pair plots, joint plots, regression plots
• Categorical plots: bar with CI, count, point
• Heatmaps with clustering (clustermap)

ggplot2 (R)

Grammar of graphics. Use when the user works in R or wants faceted figures.

• All standard geoms: geom_point, geom_line, geom_bar, geom_boxplot
• Faceting: facet_wrap, facet_grid
• Themes: theme_minimal, theme_classic, theme_bw
• Scale customization: colors, axes, legends

ggpubr (R)

Publication-ready statistical plots with annotations. Use for figures requiring significance brackets.

• ggboxplot, ggbarplot, ggscatter, ggviolin with stat comparisons
• Automatic p-value brackets (stat_compare_means)
• ggarrange for multi-panel layouts
• Ready-made publication themes

plotly (Python / R)

Interactive HTML plots. Use for supplementary materials or web-based presentations.

• Interactive hover, zoom, pan
• 3D scatter and surface plots
• Animated transitions
• Export as self-contained HTML

pgfplots (LaTeX)

Native LaTeX plotting. Use when figures must compile within the manuscript .tex file.

• Integrates directly with LaTeX math notation
• Consistent fonts with document body
• addplot from data files or inline tables
• Grouped bar charts, error bars, fill between

Smart Chart Selection

When the user provides data without specifying chart type, select based on data characteristics:

Data Type	Recommended Chart
One numeric variable	Histogram, KDE, box plot
Two numeric variables	Scatter plot, line plot (if ordered)
One categorical + one numeric	Bar chart, box plot, violin plot
Two categorical	Heatmap, grouped bar chart
Time series	Line plot with date axis
Correlation matrix	Heatmap with annotations
Distribution comparison	Violin plot, ridge plot, overlaid KDE
Part-of-whole	Stacked bar chart (avoid pie charts in academic work)
Spatial data	Contour plot, 2D density, map
Model performance across conditions	Grouped bar chart with error bars

Publication Styling

Journal Requirements

Apply journal-specific figure standards:

• DPI: 300 minimum for raster (PNG), vector preferred (PDF/SVG/EPS)
• Width: Single column (3.3 in / 84 mm) or double column (6.7 in / 170 mm)
• Fonts: Match manuscript font, minimum 6pt for labels
• Color: Use colorblind-safe palettes by default

Colorblind-Safe Palettes

Default to these palettes unless the user specifies otherwise:

• Categorical: Wong palette (#E69F00, #56B4E9, #009E73, #F0E442, #0072B2, #D55E00, #CC79A7)
• Sequential: Viridis, Inferno, Plasma
• Diverging: Coolwarm, RdBu (reversed if needed)
• Always verify grayscale legibility for print journals

Standard Style Template (matplotlib)

import matplotlib.pyplot as plt
plt.rcParams.update({
    'font.family': 'serif',
    'font.size': 10,
    'axes.labelsize': 11,
    'axes.titlesize': 12,
    'xtick.labelsize': 9,
    'ytick.labelsize': 9,
    'legend.fontsize': 9,
    'figure.dpi': 300,
    'savefig.dpi': 300,
    'savefig.bbox': 'tight',
    'axes.spines.top': False,
    'axes.spines.right': False,
})

Style Presets

Named figure style presets: default, nature, ieee. Each preset's rcParams and sizing rules are defined once in references/figure-styles.md; load that file rather than duplicating values here.

Presets resolve in one precedence order, highest first: explicit Style: line > trigger phrase > journal inference from manuscript/config.yaml > default.

Style: line (accepted input)

The invocation may carry an explicit Style: line, which outranks every other selector:

Plot macro-AUROC against labeled fraction, one line per method.
Style: nature

• Style: nature, Style: ieee: apply that preset.
• Style: default, or no Style: line at all: the no-op path, exactly the output this skill produces today. An omitted Style: line is never an error.
• Any other value: do not guess and do not improvise a preset. Say it is not defined, list the presets that are (default, nature, ieee), and ask which to use.

Trigger phrases

With no Style: line, apply a preset when the user asks for it in prose, for example:

• "nature style", "in Nature format", "for submission to Nature" -> nature
• "IEEE two-column figure" -> ieee

Journal inference

If the user gives neither a Style: line nor a trigger phrase, but manuscript/config.yaml specifies a target journal, map that journal to its preset (Nature-family journals -> nature, IEEE venues -> ieee).

Default rule

If nothing above resolves, use the default preset. This is the current behavior, unchanged. Always state which preset was applied and why.

Integrity

Presets restyle only: fonts, sizes, colors, dimensions, DPI. They never change data values or synthetic-data labels.

Statistical Annotations

• Significance brackets: p-value brackets between groups (*, **, ***, ns)
• Correlation coefficients: Pearson r or Spearman rho on scatter plots
• Regression lines: With confidence bands (95% CI shading)
• Error bars: Standard error (SE), standard deviation (SD), or confidence interval (CI): always label which
• Effect sizes: Cohen's d or eta-squared annotations where appropriate

Data Input

From CSV

import pandas as pd
df = pd.read_csv('data/results.csv')

From JSON

df = pd.read_json('data/results.json')

Inline Specification

User provides data directly in the prompt. Parse into a DataFrame or table structure.

From manuscript tables

Extract data from existing LaTeX tables in manuscript/tables/ and visualize.

Multi-Panel Figures

matplotlib/seaborn

fig, axes = plt.subplots(2, 2, figsize=(6.7, 5))
# Label panels: (a), (b), (c), (d)
for ax, label in zip(axes.flat, 'abcd'):
    ax.set_title(f'({label})', loc='left', fontweight='bold')

ggplot2/ggpubr

library(ggpubr)
ggarrange(p1, p2, p3, p4, labels = c("a", "b", "c", "d"), ncol = 2, nrow = 2)

Panel labels follow journal convention: lowercase letters in parentheses, bold, top-left.

Export Formats

Format	Use Case	Command (matplotlib)
PDF	Vector, LaTeX inclusion	plt.savefig('fig.pdf')
PNG	Raster, Word inclusion	plt.savefig('fig.png', dpi=300)
SVG	Web, scalable	plt.savefig('fig.svg')
EPS	Legacy journal requirement	plt.savefig('fig.eps')

Workflow

1. Determine data source: CSV, JSON, inline, or existing manuscript table
2. Select chart type: Based on data characteristics or user request
3. Choose library: Based on user preference, language, or best fit
4. Generate code: Write the plotting code in-session; for heavy plotting code, optionally launch the visualization-agent subagent (Sonnet) via the Task/Agent tool
5. Apply style preset: Load references/figure-styles.md, resolve the preset by precedence (explicit Style: line, then trigger phrase, then journal inference from manuscript/config.yaml, then default), apply its rcParams and sizing, and tell the user which preset was applied
6. Apply publication styling: DPI, fonts, colors, dimensions per journal requirements
7. Add statistical annotations if applicable
8. Generate caption following journal conventions
9. Save to figures/ in appropriate format
10. Generate manuscript inclusion snippet (\includegraphics{} or Word embedding)

Caption Generation

Generate figure captions following academic conventions:

• First sentence: what the figure shows (descriptive)
• Subsequent sentences: key observations, statistical details, panel descriptions
• Example: "Figure 3. Comparison of model accuracy across datasets. (a) Performance on CIFAR-10. (b) Performance on ImageNet. Error bars indicate 95% confidence intervals. The proposed method (blue) significantly outperforms the baseline (orange) on both datasets (p < 0.01)."

Integration Points

• figure-suggestions: Suggests what to plot; visualization generates the actual code
• latex-tables: Data from tables can be visualized; visualizations can be tabulated
• journal-formatting: Provides figure dimension and style requirements
• tikz-diagrams: For schematic/conceptual figures use tikz-diagrams; for data-driven plots use visualization
• word-output: PDF exports for LaTeX; PNG exports for Word, which the user places in the document by hand (automated image embedding is planned, not implemented: templates/word/build-docx.js generates headings, paragraphs, and lists only)
• implementation: Visualization code may be part of experiment pipeline

Alt Text

Alt text is a REQUIRED output of this skill, delivered with every plot (and every preset variant). Write one or two sentences describing the DATA content: name the axes, the series compared, and the direction and size of the trend. Never describe styling (colors, fonts, palette); presets restyle only, so all variants of one plot share ONE data description plus at most a one-clause style note (for example "(Nature preset: 89 mm single column)"). Do not begin with "Image of" or restate the caption verbatim. See the Alt Text section of references/figure-styles.md. When the plot is embedded in Word, templates/word/build-docx.js carries this text into the DOCX image accessibility properties; for LaTeX inclusion, offer \pdftooltip (pdfcomment package) without claiming PDF/UA.