How to use the Nature Figure Making agent skill (nature-figure)

What is an agent skill and what does nature-figure make the agent do?

An agent skill is a folder of instructions and supporting files that an AI agent loads when a task matches the skill's description. Instead of you explaining every detail of journal figure standards each time, the agent reads the skill's playbook and follows it.

For figures, the skill tells the agent to write down the core conclusion of the figure first, then choose a figure type, set the journal and export contract, and only then write code. It covers multi-panel figures, AI-generated graphical abstracts, and hand-built vector schematics. It is not meant for interactive dashboards, data cleaning, or statistics-only analysis, and the skill says so itself.

• Plans the figure's core message and panel layout
• Plots with Python (matplotlib, seaborn) or R (ggplot2, patchwork, ComplexHeatmap)
• Applies journal-style fonts and export settings
• Checks output with validation scripts, including panel alignment and PDF text checks
• Covers AI-generated graphical abstracts and hand-built vector schematics

Install the nature-figure skill

Download the ZIP of the skill from this page. Unzip it so the nature-figure folder (containing SKILL.md) sits in one of these locations. For a personal install for all your projects, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.

If you use the Claude apps instead of Claude Code, upload the skill under Settings → Capabilities → Skills. Restart or start a new session if the agent does not pick it up straight away.

The skill's licence is Apache-2.0. Note that the example plotting scripts and preview images under assets/figures4papers/ come from a third-party project with no explicit licence in its upstream repository, so do not redistribute or publish them or derived materials without checking the upstream terms and getting permission where required.

• Personal install: ~/.claude/skills/nature-figure/
• Project install: .claude/skills/nature-figure/ inside your project
• Claude apps: Settings → Capabilities → Skills

Set up Python or R for plotting

For plotting, the skill uses one backend at a time. The Python backend needs matplotlib, seaborn, numpy, and scipy. The R backend needs ggplot2, patchwork, and ComplexHeatmap. Install whichever you plan to use before asking for a figure that needs rendering.

The first time you ask for a plot, the agent asks one question: Python or R? It saves your answer as your default using scripts/nature_figure_backend.py. You can also name the backend in your request. Once chosen, the backend stays in use for drawing, previewing, exporting, and checks.

The AI-schematic route is optional. It needs an OpenRouter or image-generation API key for scripts/generate_openrouter_schematic.py. Plotting and vector schematics do not need it.

• Python: matplotlib, seaborn, numpy, scipy
• R: ggplot2, patchwork, ComplexHeatmap
• Optional: OpenRouter or image-generation API key

Trigger the skill with natural requests

You do not need special syntax. Describe the figure in plain language and mention your data, the claim the figure should support, and the target journal if you have one. The skill's description includes phrases like 论文配图, 科研绘图, multi-panel figures, and submission-ready plots, so requests in English or Chinese both work.

For a graphical abstract or mechanism diagram made with an AI image generator, say so explicitly, for example mentioning OpenRouter or an image-generation API. The agent then uses the AI-schematic route and does not ask about Python or R. For a hand-built editable diagram, ask for a vector schematic; the agent will confirm the brief with you before drawing.

• Make a 3-panel figure from results.csv showing the main effect, the control, and the knockdown validation, in Python, for a Nature-style submission.
• Redraw my existing Figure 2 as a cleaner multi-panel figure with consistent fonts.
• Draft a mechanism schematic of this pipeline as an editable SVG.
• Plan the panels for Figure 1 of my manuscript before writing any code.

Check the output before you submit

The skill asks the agent to run its validation scripts on the plotting source and the exported PDF. These include scripts/validate_figure.py, scripts/audit_pdf_text.py, scripts/audit_figure_collisions.py, and, for multi-panel figures, scripts/audit_panel_alignment.py. The agent should also inspect each panel and the whole figure at its final physical size, because automated checks do not replace that review.

The skill tells the agent to verify each target journal's current policy. Passing the skill's checks does not mean your figure is eligible for submission, and that is especially true for AI-generated graphical abstracts. Check the journal's official instructions and disclosure rules yourself.

• Panel alignment: shared edges and gutters must match within the default 1.5 pt tolerance
• PDF text size and text-to-text or text-to-line collisions
• Source data and export format against the target journal
• Never let the agent invent experimental values; supply the real data