scientific-writing – Data & Science skill for AI agents

Scientific Writing

Purpose

Produce clear scientific prose without inventing evidence or concealing uncertainty.

Keep drafting, evidence verification, and submission approval as separate stages.

The accountable human authors control scientific decisions and final approval. AI is

not an author, and generated fluency is never evidence [SW-S01, SW-S03].

Non-negotiable safety rules

Confidentiality

Do not send unpublished manuscripts, peer-review or editorial material, sensitive or

restricted data, PHI or other personal data, proprietary content, or source documents

to an external service without:

1. explicit authorization from a person or body empowered to grant it; and

2. a documented review of journal, institutional, funder, consent, ethics, contractual,

legal, and data-use policy.

When authorization or policy is unclear, keep processing local and use only the minimum

metadata needed. De-identification requires expert review; removing obvious names is

not sufficient. See references/authorship_ai_confidentiality.md.

No fabrication

Never invent or complete:

• citations, references, DOI, PMID, PMCID, ISBN, URLs, or quotations;
• results, data values, denominators, sample sizes, units, effect estimates,

uncertainty, statistical tests, or significance claims;

• methods, materials, protocol details, software versions, analysis choices, or

deviations;

• registrations, approvals, consent, ethics statements, participant details, or dates;
• authors, author order, CRediT roles, acknowledgments, or permissions;
• funding, sponsor roles, conflicts, data or code availability, or AI disclosures.

Use an explicit missing, unverified, or not-applicable state. Do not substitute plausible

boilerplate.

Evidence binding

Every factual or numeric manuscript claim must map to verified evidence IDs. A human

verifier must open the source, confirm the proposition and locator, verify bibliographic

metadata, and record who verified it and when.

Search snippets, generated summaries, memory, and another work's bibliography may aid

discovery but do not verify a claim. See references/evidence_workflow.md.

Scientific fidelity

• Preserve uncertainty and alternative explanations.
• Distinguish confirmatory, exploratory, descriptive, and post hoc work.
• Keep methods and results consistent.
• Reconcile units, denominators, sample sizes, populations, time points, and labels.
• For binary trial outcomes, report group event counts/denominators plus both absolute and relative effects with uncertainty when the reporting guideline requires them. Distinguish risk difference (percentage points), relative risk, and odds ratio; do not rewrite one as another or infer an absolute effect without the baseline risk. See [CONSORT 2025 explanation](https://www.bmj.com/content/389/bmj-2024-081124).
• Report negative, null, adverse, unexpected, failed, and inconclusive findings when

they belong to the study record.

• State concrete limitations and bound generalizability.
• Do not convert association into causation or non-significance into equivalence.

Intake

Before drafting, obtain or mark unresolved:

• document type, study design, stage, audience, and target venue;
• current author instructions and policy access date;
• protocol, registration, analysis plan, amendments, and reporting guideline;
• manuscript or section scope;
• verified source manifest and claim registry;
• methods, results, tables, figures, and supplements;
• authorship, CRediT, declarations, and approval records;
• confidentiality classification and authorized processing boundary;
• data, code, materials, and repository constraints.

Do not ask for restricted source material if metadata or a local user-run audit is

sufficient.

Workflow

1. Establish the local workspace

For a new draft, optionally generate fail-closed Markdown, JSON, and CSV scaffolds:

python3 scripts/scaffold_manuscript.py \
  --output-dir ./draft-workspace \
  --document-id local-draft \
  --study-design randomized_trial \
  --guideline consort-2025

The generator never overwrites files. Its output is explicitly not submission-ready and

contains placeholders that the linter rejects.

2. Select reporting guidance

Choose by actual design and article type, then open the current official statement,

checklist, explanation document, extensions, and target-journal instructions.

python3 scripts/select_reporting_guidelines.py select \
  --study-design randomized_trial

Current major routes rechecked on 2026-10-01 include CONSORT 2025, SPIRIT 2025,

PRISMA 2020, STROBE, STARD and STARD-AI, TRIPOD+AI, CARE, ARRIVE 2.0, SQUIRE 2.0,

and CHEERS 2022 [SW-S06–SW-S18].

The selector is non-scoring. It does not certify quality, compliance, completeness, or

acceptance. See references/reporting_guidelines.md.

3. Build the evidence record

Assign:

• E IDs to sources in source_manifest.json;
• C IDs to claims in claims.csv;
• N, M, O, and R IDs to numeric facts, methods, outcomes, and results in

consistency_manifest.json.

Store a hash of claim text in CSV rather than raw claim text. Use one claim per

physical Markdown line; the hash binds the entire line after removing claim/evidence

markers and collapsing whitespace. During drafting, append:

[claim:C001] [evidence:E001,E002]

Do not mark a source verified until an accountable human has opened it and confirmed

the exact support. After changing wording, re-verify the claim before updating its

hash; see the exact normalization in references/evidence_workflow.md.

4. Create an evidence outline

Outline only from recorded evidence:

• objective or question;
• section purpose;
• claim IDs and evidence IDs;
• methods and result IDs;
• analysis intent and uncertainty;
• unresolved conflicts or missing information;
• applicable reporting topics.

Keep unsupported content in an unresolved-issues list, not manuscript prose.

5. Draft without adding facts

Transform the verified outline into venue-appropriate prose. Preserve all IDs during

drafting.

• Match title and abstract to the completed main text.
• Describe methods as performed.
• Present results in the declared order and analysis population.
• Separate result from interpretation unless the venue combines them.
• Compare with prior evidence only after verifying it.
• Keep conclusions within the observed design, population, and uncertainty.

Use IMRAD only when appropriate. Structured abstracts, lists, combined sections, and

alternative structures depend on study design and venue. See

references/imrad_structure.md and references/writing_principles.md.

6. Reconcile methods and results

Record repeated numeric facts and method-result mappings, then run:

python3 scripts/check_consistency.py consistency_manifest.json

Resolve every mismatch manually. A changed value may be a legitimate analysis-set

difference, but that difference must be named rather than silently normalized.

7. Verify citations and claims

python3 scripts/validate_manifest.py source_manifest.json \
  --kind source --require-verified
python3 scripts/audit_claims.py manuscript.md claims.csv source_manifest.json
python3 scripts/check_references.py source_manifest.json

The reference checker validates syntax and duplicate identifiers without network

resolution. A human must still compare every identifier and quotation with the opened

source. Follow NLM *Citing Medicine* or the current official style required by the

venue [SW-S20, SW-S21].

8. Validate authorship and disclosure

Use journal criteria for authorship. Record the standardized CRediT roles as

contribution metadata; CRediT does not itself define authorship [SW-S19].

If AI was used, humans must verify all affected content and disclose the tool and

purpose according to current journal and publisher policy. ICMJE's January 2026

Recommendations require transparency and retain human accountability [SW-S01, SW-S02].

python3 scripts/validate_authorship.py authorship.json

The local validator implements ICMJE-style authorship gates and a local guarantor

record; it does not implement every venue policy. Do not generate a disclosure from

assumptions. See

references/authorship_ai_confidentiality.md.

9. Review declarations and open-science statements

Verify each statement independently:

• ethics and consent;
• registration and protocol;
• funding and sponsor role;
• conflicts and relationships;
• author contributions and acknowledgments;
• data, code, materials, and protocol availability;
• AI use.

Be as open as rights and responsibilities permit, but do not expose confidential,

personal, proprietary, licensed, or protected information. Record actual access

conditions. See references/research_integrity_open_science.md.

10. Use figures and tables only when warranted

Figures and tables are optional and provenance-bound. This skill does not generate

… (full content is in the download package)