nature-article-writer

What this skill does

# Nature Article Writer

Write and revise primary research manuscripts so they feel editorially mature: precise, proportionate, detailed where detail matters, and genuinely pleasurable for a scientist to read. Beautiful Nature-style prose is not ornate prose. It is clear, load-bearing prose with strong logic, good sentence movement, and no wasted claims.

This skill optimises for editorial quality, reader trust, and human-sounding scientific prose. It does **not** optimise for AI-detector evasion.

Do not imitate a named living author. Emulate journal expectations, the user's own prior writing if supplied, and the specific paper's evidence profile.

## When to activate this skill

Use this skill when the user:
- names Nature or a Nature Portfolio journal
- asks for "Nature-style", "Nature journal", or "high-impact journal" scientific writing
- wants a title, summary paragraph, abstract, introduction, results, discussion, methods, figure legend, presubmission package, cover letter, or reviewer response
- wants a scientific draft to sound more natural, less generic, less formulaic, or less obviously machine-written
- wants to convert notes, figures, bullet points, or a rough draft into a submission-ready manuscript
- wants a diagnostic pass on manuscript structure, claim calibration, prose quality, or compliance

## Success standard

A strong output from this skill should feel like it was written by a careful scientist-editor who understands both the data and the journal:
- the central claim is evident early and never overstated
- adjacent-field readers can follow the logic without drowning in jargon
- each paragraph has a job
- each sentence earns its place
- results progress by question and answer, not lab chronology
- the prose is varied but restrained
- limitations are surfaced before reviewers must drag them out
- end matter and policy-sensitive statements are present or explicitly marked as missing

## Non-negotiables

- Never invent data, methods, figures, ethics approvals, accession numbers, references, software versions, statistical results, or journal-specific limits.
- Never strengthen a claim beyond the evidence actually supplied.
- Never hide uncertainty. Mark missing facts explicitly with `[confirm]`, `[insert ref]`, `[insert accession]`, or a short `Issues to confirm` list.
- Never use AI-generated figures or image content for publication.
- Never copy distinctive phrasing from published papers. Use exemplars for structure, rhythm, and level-setting, not sentence theft.
- If AI did more than copy editing, remind the user to check whether disclosure is required under the target journal's policy. Human authors remain accountable for the final text.

## Default workflow

### 1. Build the manuscript brief

Infer or assemble the minimum brief:
- target journal and content type
- one-sentence central claim
- why it matters outside the immediate subfield
- evidence ladder: 3-6 concrete results, figures, or analyses that support the claim
- strongest prior work and the precise gap
- strongest limitation or boundary condition
- data, code, and materials availability
- ethics or compliance facts if humans, animals, clinical samples, or sensitive data are involved

If the user has scattered notes, use [assets/manuscript-brief-template.md](assets/manuscript-brief-template.md).

### 2. Calibrate before you draft

Use both of these calibration layers when possible.

#### A. Journal calibration
Consult [references/modes.md](references/modes.md) and [references/journal-calibration.md](references/journal-calibration.md).

- Choose the closest bundled mode:
  - `nature-article`
  - `nature-letter`
  - `portfolio-article`
  - `portfolio-letter`
- If the user names a specific journal and web access is available, verify the live guide and inspect 2-4 recent primary research papers from that journal.
- Build a short internal style card: title texture, opening-paragraph shape, heading policy, legend density, end-matter order, and how aggressively claims are hedged.

#### B. Exemplar anchoring
If the user supplies their own accepted papers, lab style guides, or a high-quality draft they want to sound like, use [references/exemplar-anchoring.md](references/exemplar-anchoring.md) and optionally run:

```bash
python3 scripts/prose_fingerprint.py --candidate draft.md --reference exemplar1.md exemplar2.md --format text
```

Imitate **broad habits** such as sentence length range, paragraph density, degree of overt signposting, and tolerance for technical detail. Do not imitate distinctive turns of phrase.

### 3. Build the editorial architecture

Before line-level drafting, create:
- a one-sentence paper promise
- a figure-claim matrix
- a paragraph map for the major sections

Use:
- [assets/editorial-blueprint-template.md](assets/editorial-blueprint-template.md)
- [assets/figure-claim-matrix-template.md](assets/figure-claim-matrix-template.md)
- [assets/paragraph-map-template.md](assets/paragraph-map-template.md)
- [references/editorial-architecture.md](references/editorial-architecture.md)

This is the main upgrade over a generic "write the paper" prompt. The prose improves when the structure is load-bearing before wording starts.

### 4. Draft in evidence order, not display order

Default drafting order:
1. figure plan and one-sentence take-home message for each figure
2. Results
3. Methods
4. Discussion or concluding synthesis
5. opening context paragraph or Introduction
6. summary paragraph or abstract
7. title
8. figure legends
9. availability statements and other end matter
10. cover letter or presubmission material if requested

Starting from figures and claims produces more grounded prose than starting from the title or abstract.

### 5. Shape paragraphs deliberately

Every paragraph needs:
- a topic sentence that names the paragraph's job
- evidence or reasoning that advances the job
- a final stress position that lands the important point or hands the reader to the next paragraph

Use [references/sentence-craft.md](references/sentence-craft.md) and [assets/paragraph-map-template.md](assets/paragraph-map-template.md). Prefer old-to-new information flow, concrete verbs, and sentences that end on the point that matters.

### 6. Run a human-voice pass tuned for scientific prose

Consult [references/voice-and-variation.md](references/voice-and-variation.md).

Target common instruction-tuned LLM artefacts without turning the paper chatty:
- overuse of present-participial clause chains
- noun-heavy nominalized phrasing
- conveyor-belt transitions (`Additionally`, `Moreover`, `Importantly`, `Taken together`)
- inflated significance language
- generic concluding sentences that claim importance without stating the implication
- repeated weak sentence openings (`This`, `These`, `It`, `We`)
- flat sentence rhythm and uniform paragraph shape

Do **not** blindly ban passive voice, repetition, or technical compounds. Scientific prose needs all three sometimes. The aim is selective repair.

### 7. Run integrity and compliance checks

Use [references/integrity-and-compliance.md](references/integrity-and-compliance.md) and, if Python 3 is available:

```bash
python3 scripts/nature_preflight.py --input draft.md --mode nature-article --format text
```

or the relevant mode:

```bash
python3 scripts/nature_preflight.py --input draft.md --mode portfolio-article --format text
```

Use the report to fix:
- title length and title texture
- missing required or expected sections
- opening paragraph length or structure
- missing Data Availability or Code Availability sections
- bracket citations that need conversion
- figure legends missing title sentences or statistical detail
- hype words, generic AI-ish phrases, rhythm flatness, or repeated weak openers
- obvious overclaim or unsupported forward-looking claims

If scripting is unavailable, do the same checks manually.

## Section guidance

Use [references/section-rubric.md](references/section-rubric.md) for detailed section-by-