Cookbook: Make the figures

Figures are where a paper is won or lost, and also where LaTeX chooses violence. Researcher generates four kinds of publication-ready visual, each as source you own plus a rendered preview. Everything below is from the worked examples in the repository, on one running scenario: self-supervised pretraining for ECG arrhythmia classification.

Architecture diagram (TikZ)

Make a TikZ diagram of my two-stage pipeline: contrastive pretraining on unlabeled 12-lead ECG, then fine-tuning on PTB-XL. Standalone, colorblind-safe.

The tikz-diagrams skill picks a system-architecture layout, groups the two stages, and emits a standalone .tex that compiles on its own.

Two-stage pipeline: self-supervised contrastive pretraining, then fine-tuning on PTB-XL

Neural-network architecture (PlotNeuralNet style)

Draw a PlotNeuralNet-style 3D diagram of my 1D-CNN ECG encoder. It must compile without the PlotNeuralNet repo.

The plotneuralnet skill inlines every layer macro, so the file is self-contained. Box height shrinks as the signal is pooled; width grows with channel count.

1D-CNN ECG encoder: input, three conv blocks with pooling, global average pooling, dense layer, softmax

Results table (booktabs)

Turn this results CSV into a publication-quality table: group by supervised vs self-supervised, bold the best in each column, add significance markers.

The latex-tables skill follows booktabs rules (no vertical lines, \toprule/\midrule/\bottomrule), bolds the best per column, and puts significance markers in a table note. Synthetic demonstration data is labeled as such.

$Label-efficiency comparison table across 1%, 10%, and 100% labeled fractions$

Chart (matplotlib)

Plot macro-AUROC vs labeled fraction, one line per method, log x-axis, colorblind-safe.

The visualization skill chooses a line plot on a log axis (a trend, not a comparison of categories), keeps the same palette as the diagram, and draws the self-supervised methods as solid lines against dashed supervised baselines.

Label-efficiency line plot: self-supervised curves above supervised, gap largest at 1% labels

Journal style presets

Named presets restyle a figure for a target journal without touching the data. Three ship today, defined in references/figure-styles.md: default, nature, and ieee. Pin one explicitly with a Style: line in your request (Style: nature, Style: ieee), ask for it by name (“restyle this for Nature”), or let the skill pick it up from your manuscript’s target journal. Those three selectors resolve in that order of precedence, and no Style: line at all (or Style: default) is the no-op path.

Here is the same label-efficiency plot under the default and nature presets:

`default` preset	`nature` preset

The Nature variant applies 89 mm single-column sizing, sans-serif type, hairline axes, a muted palette, and a bold lowercase panel letter. Every plotted value is identical: only the styling changes.

Presets are not a matplotlib-only feature. The same two presets restyle the TikZ architecture diagram:

`default` preset	`nature` preset

Here the preset swaps typography, stroke weights, arrowheads, and fills, and adds the bold lowercase panel letter. The topology (nodes, edges, labels, meaning) is untouched. Both palettes are the shipped ones: each variant copies its hexes straight out of the reference file rather than hand-picking them, so what you see here is what the preset produces.

Conceptual art via external generators

For visuals that are not data, the image-prompt-crafting skill crafts prompts for external image generators (ChatGPT/DALL-E, Gemini, Midjourney). It covers conceptual illustrations, graphical abstracts, and cover art only, never data or results figures, and it always requires an AI-disclosure caption on the resulting image.

The full source for each of these lives in examples/visualization-latex/.