Nature-Style Scientific Writing

Use this skill when the user needs help creating or rebuilding manuscript prose, not merely polishing existing sentences.

Core stance

• Author evidence comes first. Do not invent results, mechanisms, references, methods, novelty, sample sizes, statistics or limitations.
• Write the argument before writing the sentences.
• Make the paper easy to judge: relevance, novelty, trust, reuse and meaning.
• Use ambitious but bounded claims.
• If essential evidence is missing, write a placeholder or ask for the missing input instead of filling the gap.

Mined writing memory

Before drafting or restructuring academic prose, check the active installed skills/ml-paper-writing/references/knowledge/paper-miner-writing-memory.md under the current client's skill home. Read only entries relevant to this paper's section, article type, and venue. Use mined patterns for structure and wording ideas, while grounding every claim in the author's evidence and the target journal's requirements. Do not copy source phrasing. If the memory is absent or has no relevant entries, continue with this skill's references.

When to open extra files

Intake

Before drafting, identify:

• manuscript section: title, abstract, introduction, results, discussion, conclusion, significance paragraph or full outline
• paper type: mechanism, method, resource, device, model, clinical, materials, computational or interdisciplinary
• core claim: what the paper actually demonstrates
• evidence: figures, measurements, comparisons, datasets, statistics or examples
• boundary: where the claim stops
• target journal or word limit, if provided

If any of core claim, evidence or boundary is absent, expose the gap before drafting. You may still produce a scaffold with explicit placeholders.

Writing workflow

1. Build a one-sentence argument: In [system/problem], we show [advance] using [approach], supported by [evidence], with [boundary].
2. Choose the section architecture from references/article-architecture.md.
3. Map each paragraph to one job: context, gap, approach, result, comparison, mechanism, implication or limitation.
4. Draft from evidence outward. Keep claims near the data that support them.
5. Calibrate verbs: show, demonstrate, suggest, indicate, enable, may, could.
6. Remove unsupported novelty and universal claims.
7. Run a paragraph-flow check: one paragraph, one message, with a clear first sentence and explicit sentence-to-sentence relation.
8. Return prose plus concise notes on assumptions and missing inputs.

Section defaults

Abstract

Default Nature pattern:

context/problem -> gap -> approach -> key result -> implication -> boundary

For technical AI, ML, CV or method-heavy manuscripts, open references/abstract.md and choose one of:

• challenge -> contribution
• challenge -> insight -> contribution
• multiple contributions

Keep it compact. Include quantitative or comparative detail when the user provided it. End with what the work enables, not generic importance.

Introduction

Use:

field scale -> bottleneck -> prior attempts -> unresolved gap -> present study

For method-heavy papers, open references/introduction.md and reason backward from the technical challenge and contribution before drafting forward.

Do not summarize all results. The final paragraph should state what this paper does and how it addresses the gap.

Results narrative

Use an evidence ladder:

system/workflow -> validation -> main result -> baseline comparison -> mechanism/diagnostic analysis -> application or generalization

Each subsection should have a claim-first opening and then data support.

For ML/conference-style experiment sections, open references/experiments.md and make sure each major claim is backed by comparison, ablation, or stress-test evidence.

Related Work

Use:

topic scope -> representative methods -> limitation tied to this paper -> distinction

Group prior work by technical topic and mechanism, not by publication year.

Discussion

Use:

central advance -> evidence meaning -> relation to prior work -> constraints -> future use

This is where interpretation and limitations belong. Do not repeat the Results section figure by figure.

Conclusion

Use:

contribution -> decisive evidence -> implication -> boundary

No new data. No unsupported promises.

Title

Prefer concrete titles that combine:

system/object + action/capability + application or consequence

Avoid slogan titles, grant-style aims and overbroad field claims.

Output format

Default output:

1. Draft: with the requested prose.
2. Section outline: with 3-7 compact bullets when the task involves a full section.
3. Assumptions or missing inputs: with only material issues.
4. Claim-evidence map: for major claims, using Claim: ... | Evidence: ... | Status: supported/needs evidence.
5. Why this structure: with 2-4 short bullets.

For Chinese author notes, provide polished English first, then brief Chinese notes explaining major structural choices.