figura

Read-only visual audit of a rendered figure (PNG, PDF, or SVG). Delegates to the figura-image-auditor subagent so the image bytes do not pollute the main context.

Upgrade a "looks like default" figure to publication style — fonts embedded, palette colorblind-safe, spines/grid cleaned, vector exports. Works on matplotlib (.py) and TikZ (.tex).

Export every PDF figure in a directory to PNG at 300 DPI (for board uploads, Slack, slide decks, README embeds).

Targeted fix for overlap/collision defects — tick labels, legend covering data, axis labels, multi-panel collisions (matplotlib) or arrow-through-text, loop-arrow-crossing-nodes, label-on-node (TikZ).

Run the render → view → fix loop on a figure script (matplotlib .py) or TikZ source (.tex) until print-size defects are gone or two cycles pass.

Apply venue-specific publication style (NeurIPS, ICML, ICLR, IEEE, CVPR, ACM, Nature) to a matplotlib figure script.

Read-only visual audit of a rendered figure (PNG or PDF). Reports defects across six categories — legibility, collision, truncation, encoding, layout, dynamic-range. Use proactively whenever the user asks to "review", "audit", "check", or "look at" a paper figure, or when the parent agent wants a fresh-eyes defect pass without filling its own context with image bytes.

Create publication-quality figures, plots, and diagrams for academic papers, with a render-view-fix iteration loop for catching visual defects. Use this skill whenever the user is preparing visuals for a paper, thesis, preprint, technical report, or conference submission — including data plots (line, bar, scatter, heatmap, violin, etc. via matplotlib/seaborn), system or architecture diagrams, schematics, model illustrations, ablation comparisons, or any figure intended for inclusion in a written research deliverable. Trigger on phrases like "figure for my paper," "plot for the manuscript," "architecture diagram," "schematic," "make this look publication-ready," "submission-quality figure," "iterate on this figure," "fix this plot," or any time the user references LaTeX, NeurIPS, ICML, ICLR, IEEE, ACM, Nature, Science, or arXiv. Also use when the user wants to fix or upgrade a figure that "looks like default matplotlib," asks for a colorblind-safe / vector / print-ready version, or wants Claude to visually inspect a rendered figure and fix legibility / overlap / legend / clipping issues. Output files should be vector when possible (PDF/SVG) with PNG only as a high-DPI raster fallback (the PNG also serves as the print-size preview for visual iteration).