paperFig

paperFig

Create scientifically faithful, publication-quality figures through a repeatable process from evidence inventory to final render QA.

This skill governs the research-figure process only. It does not include anonymization, data perturbation, synthetic replacement, identifier removal, or privacy guarantees. If the user separately requests those operations, treat them as an additional workflow with an explicit data-release contract; do not silently mix them into ordinary scientific plotting.

Mandatory routing

Read the following references before acting:

• references/research-figure-process.md for the phase-by-phase workflow.
• references/reference-deconstruction.md before imitating or adapting a reference figure.
• references/architecture-and-multipanel-design.md for model diagrams or compound figures.
• references/scientific-visual-qa.md before final generation and delivery.

Use the PDF skill whenever a PDF is read, created, or reviewed. Use the spreadsheets skill when the main source is an XLSX workbook requiring inspection or transformation.

Trigger conditions

Use this skill when the user asks to:

• create or improve scientific, academic, or paper figures;
• locate the plotting code and data behind figures in a PDF or directory;
• reproduce the visual logic of public reference figures with project data;
• build a model architecture, Transformer diagram, mechanism schematic, or evidence-backed workflow panel;
• redesign benchmark, ablation, transfer, interpretability, or mechanism figures;
• assemble a consistent multi-figure suite or reference-vs-redraw atlas;
• standardize an existing plotting project into a reusable process.

Do not use it for generic illustration, ordinary photo editing, UI design, or privacy/anonymization as the primary objective.

Core principles

1. Scientific truth before aesthetics. Preserve actual data, statistics, units, pairing, uncertainty, sample size, ordering, and analysis logic.
2. Reference mechanism, not superficial copying. Identify why the reference persuades and transfer that mechanism to the new scientific claim.
3. Source traceability. Every final panel must map to plotting code and data inputs, or be explicitly marked as a schematic.
4. One panel, one job. Each panel answers a distinct scientific question.
5. Complexity must be earned. Dense figures need a clear reading path and evidence hierarchy; decorative complexity is not rigor.
6. Render verification is mandatory. A script that runs is not a finished figure until the exported artifact has been visually inspected.