nature-figure (Yuan1z0825): a Claude skill for journal-ready scientific figures in Python or R

• Conclusion, evidence chain, figure type and journal rules before any code

  Loaded for every job from the manifest. Sets a restrained palette, a hero panel and statistics as part of the figure.

  Try: “Make a Nature-style multi-panel figure from this CSV, in Python.”

  Read · Figure contract and default stance

• Asks once and remembers your choice

  Uses scripts/nature_figure_backend.py to save the choice in ~/.config/nature-skills/nature-figure.json, or the path in NATURE_FIGURE_CONFIG.

  Write · Backend gate: Python or R

• matplotlib or seaborn, or ggplot2 and patchwork

  Once chosen, the backend is exclusive for drawing, previews, export and visual QA.

  Runs code · Python and R backend fragments

• One Results-level question per figure

  Gives each panel a distinct evidence role instead of redrawing the same metric.

  Read · Multi-panel evidence architecture

• Static checks on the plotting source

  Reproducible QA for Python and R plotting source.

  Runs code · scripts/validate_figure.py

• Flags glyphs below the 5 pt floor in the exported PDF

  Scans the PDF for text runs that are too small.

  Runs code · scripts/audit_pdf_text.py

• 1.5 pt tolerance on comparable panels

  audit_panel_alignment.py for Matplotlib, panel_alignment.R for patchwork. A FIX BEFORE DELIVERY result blocks export.

  Runs code · Panel alignment gate

• Finds overlapping text and marks in the final PDF

  Writes a JSON report and an optional marked diagnostic PDF. Needs PyMuPDF.

  Runs code · scripts/audit_figure_collisions.py

• Flagship Nature and Nature Machine Intelligence rules

  Covers initial versus production files, display budgets and legend limits.

  Read · Journal requirement references

• Brief, composition, policy gate, human review, provenance

  Treats AI images as draft concepts, not data panels, and tells you to check the target journal's current policy.

  Read · AI graphical abstract route

• Draft a schematic with OpenRouter's Images API

  Only when you explicitly ask. The script builds a prompt from your title, abstract text and panel flow, adds any reference images (a local file is base64-encoded into the request), and posts it with OPENROUTER_API_KEY. It saves the images and request_metadata.json. --dry-run prints the request without calling the API.

  Runs code · scripts/generate_openrouter_schematic.py

• Hand-built editable SVG diagrams

  Uses references/vector-schematic-workflow.md and scripts/audit_svg_schematic.py instead of an AI image.

  Write · Vector schematic route

Watch out for

• The AI schematic route sends your manuscript text out

  generate_openrouter_schematic.py builds its prompt from your --title, --abstract or --abstract-file and --panel-map, adds any --reference-image (local files are base64-encoded, URLs are passed on) and posts it with a bearer key from OPENROUTER_API_KEY. The skill's reference says not to send confidential manuscript content to OpenRouter without permission. Use --dry-run first and leave unpublished results out of the prompt. github.com

• It stores a preference file outside the project

  nature_figure_backend.py writes your Python or R choice to ~/.config/nature-skills/nature-figure.json unless NATURE_FIGURE_CONFIG points elsewhere. github.com

• A package install is part of the setup

  The README tells you to run python -m pip install -r skills/nature-figure/requirements.txt, which installs PyMuPDF, for the collision audit. github.com

• Bundled third-party figures have no clear license

  The figures4papers files carry a notice that the repository license does not automatically apply. Do not republish them. github.com

• The endpoint and key variable can be redirected

  The script takes --api-url or SCHEMATIC_IMAGE_API_URL and --api-key-env, so the prompt and the key named there go to whatever address you set. It also downloads image URLs returned in the response. github.com