Scientific Writing

Scientific Writing is the core drafting skill in K-Dense's Claude Scientific Writer (MIT). Version 2.0 is less about producing polished prose fast and more about never inventing evidence. Drafting, evidence verification, and submission approval are separate stages, and the accountable human authors keep final control. The skill states plainly that AI is not an author and that fluent text is never evidence.

Key Features

• No fabrication list: the agent must not invent citations, DOIs, quotations, data values, sample sizes, statistics, methods, approvals, author roles, funding, or AI disclosures. Missing items get an explicit missing or unverified state instead of plausible boilerplate.
• Evidence binding: sources get E IDs, claims get C IDs, and drafts carry markers like [claim:C001] [evidence:E001,E002]. A human must open the source and confirm the exact support before a source counts as verified.
• Reporting guidelines: a non-scoring selector routes by study design to guidance researched on 2026-07-24, including CONSORT 2025, SPIRIT 2025, PRISMA 2020, STROBE, STARD, TRIPOD+AI, CARE, ARRIVE 2.0, SQUIRE 2.0, and CHEERS 2022.
• Offline checks: eight dependency-free scripts scaffold a workspace, validate manifests, audit claims against sources, reconcile repeated numbers between methods and results, check reference syntax, validate authorship and CRediT records, and lint the manuscript.
• Confidentiality rules: unpublished manuscripts, reviewer material, and personal data stay local unless someone with authority approves and policy is reviewed.

Use Cases

• Drafting a manuscript section from a verified outline without adding new facts.
• Catching a sample size that differs between Methods and Results before submission.
• Preparing a revision where reviewer comments trigger registry updates before prose edits.

Pricing and Access

Free under MIT. The bundled tools run offline and need no API keys; you pay only for the model session, such as Claude Code.

Getting Started

1. Install the plugin: /plugin marketplace add https://github.com/K-Dense-AI/claude-scientific-writer, then /plugin install claude-scientific-writer.
2. Scaffold a workspace: python3 scripts/scaffold_manuscript.py --output-dir ./draft-workspace --document-id local-draft --study-design randomized_trial --guideline consort-2025.
3. Pick guidance with python3 scripts/select_reporting_guidelines.py select --study-design randomized_trial, draft, then run check_consistency.py and audit_claims.py.

Limitation: the skill no longer generates figures or ships LaTeX templates; K-Dense removed them so polished placeholders could not ship. Passing the linter does not mean a paper is submission-ready.

Frequently Asked Questions

Does it still require bullet-free prose and IMRAD?

IMRAD is used only when it fits the design and venue. The current version focuses on evidence traceability rather than a fixed format.

Can it look up references?

Not by itself. Pair it with Research Lookup or Citation Management and verify every source by hand.

Alternatives

• Doc Co-Authoring Skill: general collaborative document drafting without scientific evidence registries.
• DOCX Skill: produces and edits Word files when a journal requires them.
• Literature Review Skill: builds the evidence you later cite.

Conclusion

Choose Scientific Writing if traceability matters more than speed: each claim is tied to a source a human has checked. For fast first drafts with no audit trail, a general writing skill is simpler. Explore related tools in the skills hub.