yjh051108/mmkit/paper-write-nature

Nature-Style Paper Writing

Draft a Nature-quality LaTeX paper based on: $ARGUMENTS

Constants

• TARGET_VENUE = `Nature` — Override via Additional Parameters. Supported: Nature, Nature Methods, Nature Machine Intelligence, Communications journals.
• MAX_PAGES — Nature Article: ~5 pages main text + Methods. Override via Additional Parameters.
• ANONYMOUS = false — Nature uses non-anonymous submission.
• DBLP_BIBTEX = true — Fetch real BibTeX from DBLP/CrossRef. Never fabricate.
• CUSTOM_REQUIREMENTS — Highest priority.
• REVIEWER_SCRIPT — External reviewer script.

Inputs

1. PAPER_PLAN.md — outline with claims-evidence matrix, figure plan
2. RESULTS.md — structured experiment results
3. figures/ — SVGs/PDFs + latex_includes.tex
4. Existing .bib file (or will create)

If no PAPER_PLAN.md, generate minimal outline from available docs.

Core Architecture

1. Identify Paper Type First

Before writing, determine the paper type:

• Research paper: why the phenomenon matters → what was done → what was found → what it means
• Methods paper: does the method work → is it reproducible → is it better under fair comparison
• Hypothesis-based: establish or rule out a causal explanation
• Algorithmic/device: propose procedure/tool/system → show reliable and advantageous performance

Do not use one narrative logic for all paper types.

2. Reader-First Writing Order

Write for the reader's cognitive sequence:

1. Is this relevant to me? (Introduction hook)
2. What is new here? (Contribution statement)
3. Do I trust it? (Results + Methods)
4. Can I reuse it? (Methods detail + Data Availability)
5. What does it mean? (Discussion + boundaries)

3. Productive Writing Order

For research articles, write in this order:

1. Results — anchor everything in evidence
2. Introduction — frame the gap now that you know the findings
3. Title — crystallize the contribution
4. Discussion — interpret and bound
5. Materials and Methods — reproducibility
6. Abstract — mini-paper summary last

4. Hourglass Structure

• Introduction: open broadly → narrow to specific gap → state question/hypothesis
• Discussion/Conclusion: widen again → connect findings to literature → explain how gap was filled

5. Paragraph Architecture: Claim-Evidence-Boundary

Every paragraph follows:

• Claim: one controlling idea (topic sentence)
• Evidence: data, comparison, explanation, literature support
• Boundary: limitation, scope, or transition to next claim

Section Responsibilities

Title

• ≤ 75 characters including spaces (Nature guideline)
• Searchable, specific, restrained, defensible
• Pattern: [Core entity] in/through/by [mechanism or context]
• No vague hooks, no unverified "first"

Abstract (150–200 words)

Mini-paper structure: context/problem → gap → approach → key result with numbers → implication

Introduction (~600–800 words for Nature)

• Hook: why the topic matters broadly
• Known: what is established
• Gap: what remains unresolved
• Aim: what this study asks/does
• Value: brief indication of approach and significance
• Do NOT summarize Results or Conclusion here
• Short paragraphs OK (Nature style allows 3–4 sentence paragraphs)

Results

• Past tense: report what was observed
• Orient reader to figure/table → state main observation → quantitative detail → patterns
• Results = what happened, NOT what it means
• Each result paragraph tied to a specific figure or table
• Active voice preferred: "We observed..." not "It was observed that..."

Discussion

• Restate main finding → plausible explanations → compare with earlier work → limitations → implications → future work
• Short rule: Results = what we observed; Discussion = how we understand it and when it may fail
• Three-part close: contribution → key evidence → implication with boundary

Materials and Methods

• Specific, complete, transparent, reproducible
• Another group must determine: ethical conformity, materials/conditions, key parameters, data processing, statistical tests, software versions
• Never: "under standard conditions", "using routine methods", "data were analyzed statistically"

Data Availability Statement

Generate using Nature data policy principles:

• Map each dataset to access route: public repository, controlled access, within supplement, reused source, third-party restricted
• Prefer DOI/accession numbers over personal websites
• Pattern: "The [data type] generated in this study have been deposited in [repository] under accession code [XXX]. Source data are provided with this paper."
• Flag "available upon request" as weak unless legally/ethically required

Failure Mode Diagnosis

Before editing any section, diagnose the main problem in priority order:

1. Paper type — wrong narrative logic for this paper type?
2. Section job — section not fulfilling its rhetorical responsibility?
3. Paragraph logic — claim without evidence? evidence without claim? missing boundary?
4. Sentence polish — clutter, passive voice, overclaim?

Fix from top down. Do not polish sentences while reasoning is broken.

Nature-Specific Style Rules

Sentence control

• Each sentence ≤ 30 words
• One core subject-verb proposition per sentence
• Split overloaded sentences rather than polishing cosmetically
• Active voice preferred: "We show..." not "It is shown that..."

Paragraph control

• Short paragraphs OK (3–5 sentences typical for Nature)
• Each paragraph: one controlling idea + support
• Thematic linking, not repetitive "This suggests..." openings

De-AI Polish Rules

Remove or replace these AI-typical words:

• "delve" → "examine", "investigate"
• "pivotal" → "important", "central"
• "landscape" → "field", "area"
• "multifaceted" → "complex"
• "underscores" → "shows", "highlights"
• "leveraging" → "using"
• "novel" (overused) → "new", or remove if claim is clear from context
• "groundbreaking" → remove or use specific evidence
• "paradigm shift" → describe the actual change
• "in conclusion" → just state the conclusion directly

Hedging (Academic Phrasebank patterns)

• "These results suggest that..."
• "A possible explanation is that..."
• "This discrepancy may reflect..."
• "To our knowledge, this is the first..."
• "Further work is needed to determine whether..."

Transitions

• Contrast: "However,", "By contrast,", "Nevertheless,"
• Addition: "Moreover,", "Furthermore,", "In addition,"
• Cause: "Consequently,", "As a result,", "Therefore,"
• Concession: "Although...,", "Despite...,", "Notwithstanding,"

Limitations acknowledgment

• "These results should be interpreted with caution because..."
• "A limitation of this study is that..."
• "The generalizability of these findings is limited by..."

⛔⛔⛔ Output Contract (highest priority)

Mandatory output depends on `params.output_format`:

• PDF mode: paper/main.tex (≥ 5KB) + paper/sections/*.tex + paper/references.bib
• docx mode: paper/main.md (single file, ≥ 5KB). Do NOT create paper/main.tex

⛔ MUST run output verification before ending the step:

MODE=$(grep -q "Word（.docx）\|docx mode" CLAUDE.md 2>/dev/null && echo docx || echo pdf)
PASS=true
if [ "$MODE" = "docx" ]; then
    [ -f paper/main.md ] && SZ=$(wc -c < paper/main.md) || SZ=0
    [ "$SZ" -ge 5120 ] && echo "✅ paper/main.md ($SZ)" || { echo "❌ paper/main.md missing"; PASS=false; }
else
    [ -f paper/main.tex ] && SZ=$(wc -c < paper/main.tex) || SZ=0
    [ "$SZ" -ge 5120 ] && echo "✅ paper/main.tex ($SZ)" || { echo "❌ paper/main.tex missing"; PASS=false; }
fi
[ "$PASS" != true ] && echo "⛔ Output verification FAILED — must complete before ending"

If verification fails, complete the missing files instead of exiting.

docx mode conventions

When params.output_format == 'docx', read references/docx-conventions.md in this skill for the Markdown conventions (headings / abstract / math / figures / tables / citations). In docx mode: write paper/main.md only, never .tex.

Workflow

Step 1: Read inputs and plan

Read PAPER_PLAN.md, RESULTS.md, figures/. Identify paper type and writing order.

Step 2: Draft Results first

Write Results section anchored to figures and data. Each paragraph: orient → observe → quantify → pattern.

Step 3: Draft Introduction

Frame the gap now that findings are known. Hourglass: broad → narrow → specific aim.

Step 4: Draft Title

Crystallize contribution in ≤ 75 characters.

Step 5: Draft Discussion

Interpret findings, compare with literature, acknowledge limitations, state implications.

Step 6: Draft Methods

Complete, reproducible, specific. Include statistical tests, software versions, parameters.

Step 7: Draft Abstract

150–200 words. Context → gap → approach → key result → implication.

Step 8: Data Availability + Author Contributions

Generate Data Availability statement. Add Author Contributions template.

Step 9: Self-review

Run failure mode diagnosis on each section. Check:

• [ ] Hourglass structure intact
• [ ] Each paragraph has claim-evidence-boundary
• [ ] No AI-typical language remaining
• [ ] Active voice dominant
• [ ] Sentences ≤ 30 words
• [ ] Title ≤ 75 characters
• [ ] Abstract 150–200 words
• [ ] No fabricated references
• [ ] Hedging appropriate (no overclaim)
• [ ] Data Availability complete

Step 10: Compile

Save to paper/main.tex with proper Nature formatting. Generate paper/references.bib.

Related Files

Key Rules

• ⛔ Never fabricate references — use DBLP/CrossRef or flag with [VERIFY]
• ⛔ Never upgrade association to causation
• ⛔ Never let AI draft the core scientific argument from scratch
• ⛔ No plt.title() in figures — captions in LaTeX only
• Language serves argument — do not polish sentences while reasoning is broken
• Write with empathy for the reader: relevance → novelty → trust → reuse → meaning
• If the draft is structurally rough, reconstruct logic first, prose second
• Nature allows short paragraphs — do not pad for length
• Direct claims with evidence, not hedged-to-meaninglessness

⛔ FIGURE_MANIFEST audit (run before finishing — must produce + embed every planned figure)