figure-generation

figure-generation is a workflow skill for producing publication-quality scientific figures for research papers using matplotlib and seaborn. It solves the problem of iteratively getting a plot right by structuring generation as a three-phase pipeline adapted from MatPlotAgent: query expansion, code generation with an execution loop, and visual refinement using vision-language-model (VLM) feedback.

In phase one it expands a natural-language figure description into detailed step-by-step coding specifications, determining figure type and data mapping. In phase two it generates a self-contained Python script from a provided template, writes it to a temp file and executes it, retrying up to four times by feeding back tracebacks or adding explicit save instructions until a PNG is produced. In phase three it reads the rendered PNG and inspects it against the request, checking figure type, labels, titles, legends, color scheme, and axis scales, then regenerates corrective code as needed. It supports many figure types (bar, training-curve, heatmap, ablation, line, scatter, radar, violin, tsne, attention), outputs both 300 DPI PNG and vector PDF plus LaTeX include code, and enforces quality requirements like colorblind-friendly palettes, minimum text size, and consistent styling.

It targets researchers, ML practitioners, and academic authors preparing figures for papers, and fits into a larger research-skill chain alongside data-analysis, experiment-code, paper-writing, and table-generation skills. The skill runs Python locally to execute generated plotting scripts, which is a standard and expected part of figure generation; the bundled prompts and template are for plotting only and contain no networking, credentials, or destructive operations.