Nature Figure

Nature Figure

Build publication figures as visual arguments. Every figure starts from a claim, an evidence hierarchy, and a review-risk check before code or aesthetics.

Boundary

Use this skill for:

• empirical and statistical plots;
• multi-panel results figures;
• static publication maps and spatial evidence figures;
• image plates with quantitative evidence;
• journal-ready SVG, PDF, TIFF, and preview exports.

Use engineering-figure-agent for system architecture, workflows, mechanisms, hardware schematics, and illustration-led diagrams. Use conceptual-figure-workflow first when theory, variables, mechanisms, or claim-evidence boundaries still need review.

Do not use this skill for interactive dashboards, full GIS data-processing workflows, 3D production, or Illustrator/Figma-first infographics. A static map that carries a manuscript claim belongs here even when its source data were prepared in GIS software.

Blocking Backend Gate

If the current request does not specify Python or R and does not provide an unambiguous language-specific workflow, ask Python or R? and stop. Do not write plotting code, create mock data, or choose a backend by default.

After selection:

• use the chosen backend for every plot, preview, export, and visual QA render;
• do not substitute the other backend when a runtime or package is missing;
• report the exact blocker and provide selected-backend installation or run guidance;
• use the other language only for non-visual data inspection or conversion.

Mentioning R code, an R model object, lme4, or the interactions package is an unambiguous R selection. Do not ask the backend question again.

Read references/backend-selection.md when the user asks for a recommendation or backend details.

Workflow

1. Define the contract. Read references/figure-contract.md. State the core conclusion, evidence chain, figure archetype, target size, panel map, statistics, source data and export formats. Reuse the paper's leading advantage and claim/experiment IDs. Check risks internally; surface only material issues that change interpretation or delivery.
2. Select the minimum evidence. Give every panel a unique claim-bearing role. Remove or merge panels that do not change what the reader can conclude.
3. Load the relevant track. Use Python references for Python; use references/r-workflow.md for R. For a static map or spatial figure, also read references/geospatial-figure-contract.md. For an interaction or moderation result, read references/interaction-moderation-track.md and use interactions as the only interaction-analysis plotting interface.
4. Build reproducibly. Keep source data, plotting code, statistics, labels, dimensions, colors, and export settings traceable.
5. Integrate the layout. Use one restrained visual vocabulary, stable method colors, readable hierarchy, and direct labels when they reduce eye travel.
6. Write the legend when in scope. Read references/figure-legend-contract.md and keep the title, panel descriptions, statistics, and source-data wording aligned with the actual evidence and target journal.
7. Export with the selected backend. Produce the requested editable and raster formats at final dimensions.
8. Verify. Read references/qa-contract.md, inspect outputs at final size, and check the data -> transformation -> caption -> manuscript claim trace in the figure contract. Correct rendering alone does not establish support for an interpretation.

Quality Rules

• Never fabricate values, sample sizes, statistics, maps, or source data.
• Preserve the distinction between association, prediction, mechanism, and causation.
• Prefer one dominant evidence panel with subordinate validation or robustness panels when the evidence hierarchy supports it.
• Keep backgrounds white except where image data require a dark field.
• Avoid rainbow palettes and color-only distinctions that fail in grayscale or for color-vision deficiencies.
• Define n, center, spread, tests, corrections, and source-data provenance.
• Preserve editable text in vector outputs when the selected backend supports it.
• Do not disclose private paths, filenames, or internal template provenance unless the user explicitly requests an audit trail.

Reference Router

Load only the references needed for the selected backend and current figure task. Do not open the entire reference collection by default.

For geography, GIScience, health geography, or GeoAI work, load only the relevant files from ../shared/field-context/.

Output

Deliver the figure contract, source-data or input assumptions, selected-backend script, requested exports, and a concise QA verdict. State any missing dependency or unverified visual/statistical condition.