English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

nature-skills: An Open-Source Project Engineering Nature-Level Academic Writing into Reusable AI Skills

Forum topic · 小凯 · 2026-05-28

Summary

This article profiles nature-skills, an open-source project by Shanghai Jiao Tong University PhD candidate Yuan Yizhe, which converts Nature-level academic writing expertise into a modular pipeline of nine AI Skills. While most AI writing tools only polish grammar or generate paragraphs, nature-skills treats scholarly writing as a system: it enforces journal-specific rules, separates language from argumentation, and delivers submission-ready outputs. The pipeline covers language polishing, argument scaffolding, publication-grade figures, citation management, data availability statements, bilingual paper reading, point-by-point reviewer responses, paper-to-PPT conversion, and multi-source literature search. Core design principles include evidence-first authoring, section-aware templates, British English with sentence-length limits, hedging calibration, and a clear AI traffic-light policy. The article explains architecture, installation via Claude Code or Codex Desktop, current limitations, and the author's vision of turning implicit writing experience into explicit, auditable rules.

Key points

  • Project origin: nature-skills is an open-source repository by Yuan Yizhe, a PhD candidate at Shanghai Jiao Tong University working on medical AI. It has 151 commits and breaks Nature-level academic writing into 9 reusable Skills rather than a single tool.
  • Nine-Skill pipeline:
  • nature-polishing (Stable): language polishing from Chinese drafts to Nature-style English
  • nature-writing (Draft): argument scaffolding from claim to evidence
  • nature-figure (Stable): publication-grade figures via Python/R
  • nature-citation (Beta): CNS-grade retrieval with RIS/ENW/Zotero export
  • nature-data (Draft): Data Availability statements and FAIR checks
  • nature-reader (Beta): bilingual Markdown reader with figure grounding
  • nature-response (Beta): point-by-point reviewer reply workflow with risk checks
  • nature-paper2ppt (Beta): paper-to-Chinese-PPT conversion
  • nature-academic-search (Beta): parallel PubMed + CrossRef + arXiv search
  • Polishing is rewriting, not editing: a 6-layer diagnostic (paper-type logic, section responsibility, argument gaps, sentence stacking, tense/voice mismatch, vocabulary/citation format) runs before a 12-step sentence workflow that enforces ≤30-word sentences, British English, hedging calibration (demonstrate / suggest / may reflect), and overclaim detection.
  • Evidence-first writing order: Results → Introduction + Conclusion → Title → Discussion → Materials and Methods → Authors → Abstract. Section templates specify exact structural elements, e.g., field scale → bottleneck → prior attempts → unresolved gap → present study for Introductions.
  • Figures as visual argument: a mandatory "figure contract" requires stating the one-sentence claim, mapping each panel to the claim, selecting an archetype (quantitative grid, schematic-led composite, image plate + quant, asymmetric mixed-modality), choosing Python or R backend, and defining journal/export specs. Provided matplotlib settings target Arial, 7pt body text, editable SVG/PDF fonts, and triple export to .svg / .pdf / .tiff.
  • Citation integrity: only sources the author has personally read are cited; primary vs. secondary sources are distinguished; four attribution types (idea, data, method, wording/structure/image) are tracked; DOIs/PMIDs are resolved before export.
  • Reviewer response workflow: triage (major/minor/ambiguous) → action mapping → risk check → point-by-point response (acknowledgement → reply → modified location → revised text).
  • AI traffic-light policy:
  • Green: grammar fixes, outline options, translation, literature summary
  • Yellow: explaining methods/results, drafting reply frames (requires line-by-line review)
  • Red: writing the core argument, inserting AI-generated references or data, uploading unpublished manuscripts to public models
  • Rules are grounded: every rule derives from published Nature/Nature Communications papers or official Author Guidelines, not personal preference.
  • Output-first design: each Skill returns immediately usable artifacts (prose, .svg/.pdf/.tiff, .ris, .pptx), not intermediate planning documents.
  • Installation:
  • Claude Code: /plugin marketplace add https://github.com/Yuan1z0825/nature-skills, then /plugin install nature-skills, then /reload-plugins
  • Codex Desktop: add the repo URL as a custom plugin marketplace on main and install
  • Manual: git clone https://github.com/Yuan1z0825/nature-skills.git and copy each skills/nature-* directory into ~/.claude/skills/ or ~/.codex/skills/
  • Limitations: only polishing and figure are Stable; the rest are Draft/Beta; rules target the Nature family so Cell/Science may need adjustment; plugin installation requires Claude Code or Codex Desktop; the pipeline does not analyze raw experimental data, so statistical errors remain the author's responsibility.
  • Roadmap candidates: nature-stats, nature-methods, nature-cover, nature-review.
  • References

  • GitHub: github.com/Yuan1z0825/nature-skills
  • Author: Yuan Yizhe, PhD candidate, Shanghai Jiao Tong University (medical AI)
  • Source article: Juejin community, 2026-05-10

Tags

#academic-writing#open-source#claude-code#ai-tools#nature-journal#scientific-figures#citation-management#reviewer-response

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980441