🧬 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.
| Draft — unsupported conclusion | Reference revision — evidence-aligned conclusion |
|---|
| “These results demonstrate that the treatment improves both marker expression and survival.” | “The treatment therefore increased marker expression without a demonstrated survival benefit.” |
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
| Profile | Best for | Installer selection |
|---|
| Full research workflow · 27 skills | Manuscript work, figures, literature and analysis extensions | --set all |
| Writing & review · recommended 19-skill stack | Structure, prose, references, submission and rebuttal | --set recommended (installer default) |
| Writing & review + figures · 21 skills | The 19-skill stack plus figure production and checks | --set recommended --figure |
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
| Your task | Provide | Expected result | Entry |
|---|
| 🧬 Improve a whole manuscript | Current draft, relevant results/figures, target venue if decided | Revised draft, key changes, unresolved evidence gaps | paper-workflow |
| 📐 Write or revise one section | The section, current Results and figures, target journal | Revised section held to its contract; claims resting on unsettled results named | paper-workflow |
| ✍️ Fix a paragraph | Passage, surrounding context, requested scope | Replacement passage with a brief explanation where useful | write-scientific-manuscript |
| 🪄 Remove audit/defensive writing | Main text, SI, legend or availability statement | Direct scientific prose retaining numbers and necessary conditions | anti-defensive-writing |
| 🧹 Remove AI tone | A passage or section that reads as generated | Empty intensifiers, frames, recaps and coined terms removed, with numbers and findings kept | paper-workflow |
| 📊 Plan or make figures | Scientific question, result table or existing figure | Panel plan; rendered files when data and tools are available | figure-planner; add --figure for production |
| 🔎 Check before submission | Final draft, SI, bibliography and venue | Prioritized findings with locations; unchecked categories identified | submission-audit |
| 📨 Reply to reviewers | Original comments, manuscript, completed new evidence | Numbered response draft and matching manuscript changes | paper-workflow |
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
The skills provide instructions and optional helper scripts, not the external editors, runtimes, source access or experiments required for every task.
🧭 Scope and limits
This is a focused journal-writing workflow for life sciences, computational biology, methods, benchmarks and resources. Its rules are written for Nature-style journal articles and derived again when the target venue differs. For a named ML conference, what its reviewers check decides what stays, with details in conference-paper-writing (full profile). Explicit project instructions take precedence over all defaults. It is independent of Nature Portfolio and does not predict acceptance.
The typical section shapes come from 28 open-access computational-method papers in three Nature Portfolio journals (evidence and limits). They are checks on a draft, not format rules; the target journal's current guide sets the format.
Skills provide instructions and some helper scripts. They do not themselves supply a Word editor, PDF renderer, LaTeX installation, browsing access or experimental evidence. Compatibility separates readable inputs from editable/exportable outputs.
Repository tests cover scripts, installation and consistency. Behavior cases check task scope, evidence preservation and honest reporting of missing capabilities; their limitations and recorded runs are documented separately.
🤝 Contributing
See CONTRIBUTING.md. Changes to user-facing behavior should include a task case or an example, and both READMEs should stay aligned. Component provenance is in ATTRIBUTION.md.
💙 Acknowledgements
Parts of this repository were inspired by OpenLAIR/dr-claw, Yuan1z0825/nature-skills, and the Claude Science skill pack.
The figure layer also draws on design observations from ChenLiu-1996/figures4papers. None of that repository's code or prose is distributed here; see THIRD_PARTY_NOTICES.
⚖️ License
Original content is MIT. Components carrying Apache-2.0 material retain LICENSE-APACHE and NOTICE, including in installer-managed copies. See ATTRIBUTION for coverage.
⭐ Star History
