nature-paper-skills

🧬 Nature-Paper-Skills

From first draft to submission and rebuttal, build a clear, evidence-grounded manuscript.

27 skills for Codex and Claude Code, connecting structure, scientific writing, figures, citations and reviewer responses. Built for Nature-series life-science, computational-biology and methods papers.

🧠 Structure · 📊 Figures · ✍️ Writing · 📚 Citations · 📨 Rebuttals

🌐 English · 简体中文

🌐 Website · 🗺️ Workflow · 🚀 Quick start · 🧩 Choose a task · 🪄 Examples · 🧩 27 Skills

✨ Why this workflow

• 🎯 Argument before style. Define the scientific question, contribution and claim–evidence chain before rewriting sentences.
• 🔒 Evidence-bound edits. Preserve measured values, independent units, uncertainty, negative results and necessary scope conditions.
• 📊 Figure-led Results. Give each main figure a claim, align its panels and legends with Results, and use optional figure-production and QA tools when needed.
• 📐 Section-specific writing. Keep the Abstract, Introduction, Results, Discussion and Methods focused on their distinct roles, with consistent scientific claims.
• 🌍 Written for the venue's readers. What to keep or cut follows from who reads the paper at the target venue: the editor, the reviewer and the broad peer. Front matter states the reader's problem by the second sentence and keeps the numbers they can use.
• 📚 Citations and rebuttals. Check whether a cited source supports a claim; connect each reviewer response to completed evidence and actual manuscript changes.
• 🧩 Task-sized revisions. Improve a whole manuscript, one section or one paragraph without reopening settled material.

Section patterns are informed by 28 computational-method Articles and 12 Nature method papers. They are descriptive guidance, not journal formatting rules.

🗺️ Workflow at a glance

One connected workflow, from the research question to the reviewer response.

flowchart LR
    A("🎯 Frame the paper<br/>Question · venue")
    B("🧠 Build the argument<br/>Claims · evidence")
    C("📊 Figures & Results<br/>Panels · legends")
    D("✍️ Revise the prose<br/>Logic · style")
    E("📨 Submit & respond<br/>Preflight · rebuttal")
    I(["🔎 Check throughout<br/>Statistics · citations · data"])
    A --> B --> C --> D --> E
    I -.-> B
    I -.-> C
    I -.-> E
    classDef frame fill:#eef2ff,stroke:#818cf8,color:#312e81;
    classDef argument fill:#f5f0ff,stroke:#a78bfa,color:#4c1d95;
    classDef figures fill:#ecfdf5,stroke:#34d399,color:#064e3b;
    classDef prose fill:#fff7ed,stroke:#fb923c,color:#7c2d12;
    classDef response fill:#eff6ff,stroke:#60a5fa,color:#1e3a8a;
    classDef integrity fill:#f8fafc,stroke:#94a3b8,color:#334155,stroke-dasharray:4 3;
    class A frame;
    class B argument;
    class C figures;
    class D prose;
    class E response;
    class I integrity;

Structure and evidence come before sentence polish. Statistics, citation support and data availability are checked at the relevant stages. Enter at the stage your task needs; a paragraph edit does not require the entire chain. Reviews, surveys and Perspectives have their own architecture route.

The diagram includes figure production, available with --figure; the default installation already covers figure planning, manuscript revision and reviewer responses.

🗺️ Full workflow · 📐 Section contracts · 📚 All skills · ✍️ Writing principles · 📊 Figure workflow

🪄 See what changes

A synthetic teaching example: a draft claims improved survival although the supplied survival interval does not establish a benefit. The reference revision corrects the claim rather than merely smoothing the prose.

Evidence retained in the full revision: 12 independent cultures; marker expression +18% (95% CI 10–26%; Fig. 1a); survival +1 percentage point (95% CI −3 to 5; Fig. 1b). The survival interval does not establish either a benefit or no effect.

The reference edit also removes defensive narration while keeping every measurement and figure reference. It is an illustrative expected output, not a recorded agent run or a real study.

See the interactive example · Read the full input, evidence and reference revision · More editing examples

🚀 Quick start

1. Choose a skill profile and install

Copy this instruction into Codex or Claude Code for the full 27-skill profile:

Read and follow https://github.com/Boom5426/Nature-Paper-Skills/blob/main/INSTALL.md to install all 27 Nature Paper Skills for this environment, using the all profile, and verify the result.
If the page cannot be read, use an available GitHub tool or git clone to retrieve the main branch and read INSTALL.md from the repository root; do not substitute search snippets for the file.

For the 19- or 21-skill profile, choose a ready-to-copy instruction from the website's installation builder, or explicitly request the chosen profile when asking the agent to follow INSTALL.md. The installer otherwise defaults to 19 skills.

Local installation needs Bash and Python 3.9+; remote source retrieval also needs curl and tar. The agent should report the actual installation location, profile, integrity check and whether live client discovery was tested. File integrity does not establish that a new agent session can invoke the skill.

curl -fsSL https://raw.githubusercontent.com/Boom5426/Nature-Paper-Skills/main/install.sh | bash -s -- --agent codex --set all

Use the flags in the profile table for smaller sets. Install into the agent's actual environment, then refresh/reopen Codex and invoke $paper-workflow in CLI/IDE. Native Windows and WSL installations need different target paths; follow the Codex setup guide. The repository root is a collection, not a single skill.

curl -fsSL https://raw.githubusercontent.com/Boom5426/Nature-Paper-Skills/main/install.sh | bash -s -- --agent claude --set all

Use the flags in the profile table for smaller sets. Refresh/reopen Claude Code and invoke /paper-workflow. See the Claude Code setup guide.

Web installation is not confirmed. Downloading skills or generating a directory in a web task does not register them for later chats. Persistent use requires an available account/workspace registration flow; report it as unavailable if the flow is absent. See web distribution and verification.

2. Complete your first revision

Once paper-workflow is discoverable in the agent, place your manuscript excerpt and supporting evidence in its working folder, then ask:

Use paper-workflow. Revise only the Results paragraph in draft.md using evidence.md.
Preserve measured values, confidence intervals and figure references.
Save a revised copy and explain the material changes.

Try the self-contained first-run example before working with a private manuscript. You can then ask in plain language to improve a manuscript, review a section or check a citation; you do not need to memorize skill names.

🧩 Choose a task

A request to review or suggest produces findings. A request to revise produces edits within the requested scope. Missing evidence is surfaced; it is never written into existence.

For complete prompts and task setups, see task examples. The optional --set all selection includes a recomputable statistics walkthrough; its example helper needs SciPy.

⚙️ Installation, updates and compatibility

The local installer supports --dry-run, --doctor, --on-conflict, --restore and --ref <full-commit-sha>. It records provenance and hashes, and backs up files when replacing an existing installation. Existing modified or linked skills require the appropriate conflict policy; do not overwrite them blindly.

Installation management, update and restore · Codex setup (including Windows) · Claude Code setup · Supported environments and file formats