| name | biomed-paper-writing-skill |
| description | Write publication-ready biomedical research papers with IMRAD structure, CNS-style figures, LaTeX generation, and PubMed-verified citations. Use when user mentions writing a paper, drafting a manuscript, IMRAD, journal names (Nature, Science, Cell, NEJM, Lancet, JAMA), reporting guidelines (CONSORT, STROBE, PRISMA, ARRIVE), reviewer comments, cover letters, or converting lab results into manuscript form. Also triggers for generating publication figures, LaTeX compilation, multi-omics paper checklists, and citation verification. |
| version | 1.4.0 |
| metadata | {"author":"cafi","license":"MIT"} |
Biomedical Paper Writing
Expert-level guidance for writing publication-ready biomedical and pharmaceutical research papers. Combines Glasman-Deal's IMRAD section models, Gopen & Swan's reader expectation principles, PubMed citation verification, reporting guideline checklists, and journal-specific requirements.
Core Philosophy
Be proactive. Scientists are busy โ deliver concrete drafts they can react to.
Paper writing is collaborative, but Claude should lead with output, not questions. If the data and contribution are clear, produce a full draft. Reserve questions for genuine blockers (unclear study type, missing data, ethical concerns).
| Confidence Level | Action |
|---|
| High (clear data, obvious finding) | Write full draft, deliver, iterate on feedback |
| Medium (some ambiguity) | Write draft with flagged uncertainties, continue |
| Low (major unknowns) | Ask 1-2 targeted questions, then draft |
Only block for input when: target journal is unclear, study type is ambiguous, results seem incomplete, or ethical concerns exist (missing approvals, consent issues).
Citation Safety
AI-generated citations have high error rates โ studies report 20-40% fabrication depending on model and domain (Walczak & Cellary, JMIR 2024; Gravel et al., JAMIA 2023). In biomedical contexts, fabricated clinical evidence can have patient safety implications โ a hallucinated RCT reference could influence treatment decisions.
The rule is simple: never generate BibTeX from memory. Every citation must be verified through PubMed or CrossRef before inclusion.
| Situation | Do This | Not This |
|---|
| Adding a citation | Search PubMed โ verify PMID โ fetch BibTeX | Write BibTeX from memory |
| Uncertain about a paper | Mark as [CITATION NEEDED] | Guess the reference |
| Can't find exact paper | Note "placeholder - verify" and tell scientist | Invent similar-sounding paper |
When citations can't be verified, use explicit placeholders and always inform the scientist how many remain unverified.
See references/citation-workflow.md for PubMed/CrossRef API code and the full verification workflow.
Paper Writing Workflow
Step 0: Understand the Scientific Questions
Before writing, explore what exists:
- Data files, analysis notebooks, statistical output
- Existing drafts, lab notebooks, preliminary reports
- Figures, tables, supplementary materials
- Ethics approvals, trial registration documents
Then clarify the contribution:
"Based on my understanding of the data, the main finding appears to be [X]. Is this the framing you want, or should we emphasize different aspects?"
Step 1: Determine Paper Type and Route
Different paper types need different structures and guidelines. Route accordingly:
Step 2: Draft the Paper
Write sections in this order (the order scientists typically refine them):
Paper Writing Progress:
- [ ] Define the one-sentence finding (confirm with scientist)
- [ ] Title and Running Title
- [ ] Abstract (structured or unstructured per journal)
- [ ] Introduction (4-component model)
- [ ] Methods/Materials (with ethical approvals)
- [ ] Results (with figures/tables)
- [ ] Discussion (mirror of Introduction)
- [ ] Reporting guideline checklist
- [ ] Verify all citations via PubMed
- [ ] Final review and submission prep
Step 3: Verify Citations and Submit
Follow the workflow in references/citation-workflow.md:
- Search PubMed for each citation
- Verify PMID exists and cross-check DOI via CrossRef
- Retrieve BibTeX programmatically
- Mark unverifiable citations as
[CITATION NEEDED]
IMRAD Section Guidance
Each section follows Glasman-Deal's 4-component model โ a framework for what to DO in each section. Full details with vocabulary, signalling language, and examples are in references/writing-guide.md.
Introduction
| Component | Function | Tense |
|---|
| 1. Establish importance | Significance of field; background; define terms | Present Simple, Present Perfect |
| 2. Literature review | Previous research (organized thematically, not paper-by-paper) | Past Simple, Present Perfect |
| 3. Identify the gap | Gap in knowledge; problem to address | Present Perfect |
| 4. Present work | Describe this paper; state aims | Present Simple, Past Simple |
Start broad, narrow progressively. Jumping straight to your problem without context loses the reader โ they need to understand why the field matters before understanding why the gap matters.
Gap language signals the transition to your contribution. The tense switch from Past Simple to Present Perfect is the key signal: "Groups investigated X. However, Y has not been fully elucidated." The Present Perfect tells the reader this gap exists now.
Methods
Include all applicable elements:
- Ethics: IRB/IACUC approval number, informed consent, trial registration
- Study design: Type, setting, dates
- Participants: Eligibility criteria, demographics, sample size justification
- Interventions: Drug doses, formulations, routes (enough for replication)
- Materials: Cell lines (source, authentication, passage), reagents, antibodies
- Outcomes: Primary and secondary, how measured
- Statistics: Tests, software, significance threshold, missing data handling
Reviewers check Methods first to decide if findings are trustworthy. Missing ethical approvals or unclear statistical methods are common reasons for desk rejection.
Results
- Present in logical order (not necessarily chronological)
- Report exact p-values (P = 0.003, not P < 0.05) โ exact values let readers judge significance themselves
- Include confidence intervals and effect sizes โ p-values alone don't convey clinical relevance
- Reference each figure/table explicitly in text
- Separate results from interpretation โ mixing them obscures what you actually found
Discussion
| Component | Function |
|---|
| 1. Revisit | Summarise key findings (1-2 paragraphs) |
| 2. Map | How findings relate to existing research (agree/disagree) |
| 3. Achievement | Clinical significance, mechanistic insights, translational potential |
| 4. Limitations & Future | Limitations (always include), future directions |
The Discussion mirrors the Introduction in reverse: specific findings โ broader implications. This symmetry is deliberate โ elements introduced in the Introduction are revisited in reverse order.
Limitations matter because reviewers are instructed not to penalize honesty. Omitting limitations signals either naivety or evasion, both of which undermine credibility.
Abstract
Most biomedical journals require structured abstracts (BACKGROUND, METHODS, RESULTS, CONCLUSIONS). Word limits: NEJM 250, Lancet 300, JAMA 350, Nature Medicine 150 (unstructured), Cell 150 (unstructured "Summary").
Avoid generic openings โ editors and reviewers read hundreds of abstracts that start with "X is a leading cause of..." Start with what's specific to your study.
Writing Quality
Gopen & Swan's Principles (abbreviated)
| Principle | Why It Matters |
|---|
| Subject-verb proximity | Long interruptions between subject and verb force readers to hold context in working memory |
| Stress position | Readers naturally emphasize what comes last in a sentence โ put your key point there |
| Old before new | Starting with familiar info anchors the reader before introducing new concepts |
| Action in verb | "We analyzed" beats "We performed an analysis" โ nominalizations hide the action |
Tense Usage
| Tense | Signals | Example |
|---|
| Present Simple | Accepted fact | "Aspirin inhibits cyclooxygenase" |
| Present Perfect | Current relevance, gap still open | "Several studies have shown..." |
| Past Simple | Specific completed study | "Smith et al. demonstrated..." |
The Past Simple โ Present Perfect switch is how you signal a research gap to the reader. Getting this wrong can imply a gap has already been filled.
Word Choice
- Be specific: "sensitivity" not "performance"; "reduces mortality by 15%" not "improves outcomes"
- Eliminate hedging (unless genuinely uncertain): "suggests" not "may potentially suggest"
- Consistent terminology: pick one term per concept ("compound" vs "drug" vs "agent" โ choose one)
- "Significant": always specify โ statistically (with P value) or clinically (with relevance explained)
Full vocabulary lists and connector language: references/writing-guide.md
Journal Requirements
High-Impact General Medical
| Journal | Word Limit | Abstract | Key Requirement |
|---|
| NEJM | ~2700 | 250 words, structured | ICMJE disclosure, data sharing |
| Lancet | 3000 (RCT up to 4500) | 300 words, structured (Background/Methods/Findings/Interpretation/Funding) | Research in Context panel |
| JAMA | 3000 | 350 words, structured | ICMJE, 3 Key Points required |
| BMJ | ~4400 (aim for 2000) | 400 words, structured | ICMJE, patient involvement |
High-Impact Life Science
| Journal | Word Limit | Abstract | Key Requirement |
|---|
| Nature | 2500 (6pp) / 4300 (8pp) + Methods | 200 words, "summary paragraph" (structured: context โ background โ "Here we show" โ implications) | Online Methods (โค3000 words), Reporting Summary, โค50 refs |
| Science | ~3000 + SOM | 250 words (ideally ~200), single paragraph | Supplementary Online Material, one-sentence summary (โค135 chars) |
| Cell | ~7000 + STAR Methods | 150 words ("Summary") | STAR Methods, Key Resources Table, Graphical Abstract (1200ร1200px) |
| Nature Medicine | 4000 + Methods | 150 words, unreferenced | Reporting Summary, data availability, โค60 refs |
Universal Requirements
- ICMJE compliance: Authorship criteria, conflict of interest disclosure
- CRediT: Author contribution roles (14 standardized roles) โ required by Nature, Cell, Lancet, JAMA, BMJ, PLOS. See reporting-guidelines.md
- Ethics: IRB/IACUC approval, informed consent, trial registration
- SAGER guidelines: Sex and gender reporting โ distinguish sex (biological, male/female) from gender (social, men/women). Many journals now require compliance. See reporting-guidelines.md
- Data availability: Statement on data sharing (increasingly mandatory)
- References: Vancouver style (numbered) for most medical journals; author-year for some life science
Figures and Tables
- Resolution: 300 DPI (photos), 600 DPI (line art)
- Colour: Colourblind-safe palettes (Okabe-Ito) โ ~8% of male readers are colourblind
- Scale bars: Required for all microscopy images
- Legends: Self-contained โ reader should understand without reading main text
- Table 1: Baseline demographics (standard for clinical studies) with units, n (%), mean ยฑ SD
- P values: Report exact values and define error bars (SD vs SEM vs 95% CI)
Common figure types and their requirements: references/writing-guide.md
LaTeX Manuscript Generation
Generate clean LaTeX directly โ never use pandoc conversion (produces {[} artifacts). Key rules: \cite{} not {[}, \emph{GENE} for gene names, figure legends in \caption{} only (no separate section), [H] for draft positioning. In Chinese academia, last corresponding author = highest seniority.
Full details: references/latex-workflow.md
Figure Generation for CNS Journals
CNS journals require low-saturation, professional colors. Critical rules: no single-bar charts, no placeholder data, no text overlap, check z-index, ample panel spacing (hspaceโฅ0.65), always self-check figures after generation.
Full color palette, design rules, and self-check list: references/figure-generation.md
Required Tables
Place tables in Results (not Methods) where the data is discussed:
- Table 1: Clinical demographics โ required when study involves patient cohorts
- Gene/feature tables: After the paragraph describing the finding (e.g., IR genes, MI genes)
- Validation summary: After external validation paragraph
QC Section in Methods (MANDATORY)
Every omics paper MUST include a dedicated QC subsection. Include per-modality thresholds, retention rates, batch correction metrics, and statement of uniform thresholds across groups. Missing QC is a common desk-rejection reason.
Common Pitfalls
| Pitfall | Why It's a Problem | Fix |
|---|
| Introduction jumps to problem | Reader lacks context to understand why gap matters | Follow 4-component model: importance โ literature โ gap โ present work |
| Methods missing ethics | Automatic desk rejection at most journals | Always state IRB/IACUC approval number and consent process |
| Methods missing QC | Reviewers question data quality immediately | Add dedicated QC subsection with per-modality thresholds and retention rates |
| Results mixed with interpretation | Obscures actual findings, makes replication harder | Keep Results factual; save interpretation for Discussion |
| No limitations paragraph | Signals naivety or evasion to reviewers | Be honest โ reviewers respect transparency |
| Tables in wrong section | Tables with results data shouldn't be buried in Methods | Place gene/feature tables in Results, right after the paragraph that describes them |
| Figure legends duplicated | Caption in figure AND separate legend section = redundant | Use \caption{} only; delete standalone Figure Legends section |
| "Significant" without qualifier | Ambiguous โ could mean statistically or clinically | Always specify: "statistically significant (P = 0.003)" or "clinically significant (NNT = 12)" |
| Overclaiming | Reviewers will downgrade if conclusions exceed evidence | "suggests" not "proves"; observational data shows "association" not "causation" |
| Delta plot with baseline = 0 | A bar at zero looks like missing data | Show absolute values instead of deltas, or use dot plots |
| Bibliography format | Pandoc-converted refs use {[}1{]} not \cite{} | Always generate clean LaTeX with proper \cite and \bibitem |
Multi-omics Paper Checklist
For spatial multi-omics studies: pseudobulk DE (avoid pseudoreplication), sample-level statistics, cross-modality validation, and honest limitation reporting (cross-sectional design, annotation confidence, missing QC samples).
Full checklist: references/multi-omics-checklist.md
Reference Files
| Document | When to Read |
|---|
| writing-guide.md | Drafting any section โ contains Glasman-Deal models, Gopen & Swan principles, tense rules, vocabulary lists, figure standards |
| citation-workflow.md | Adding or verifying citations โ contains PubMed/CrossRef API code, BibTeX formats, verification checklist |
| reporting-guidelines.md | Any study requiring reporting compliance โ contains CONSORT, STROBE, PRISMA, ARRIVE, STARD, CARE, TRIPOD/TRIPOD+AI checklists, plus CRediT author roles and SAGER sex/gender guidelines |
| sources.md | Looking up original sources โ complete bibliography and journal author guideline links |
| figure-generation.md | Generating CNS-style figures โ color palette, design rules, self-check list |
| latex-workflow.md | LaTeX generation โ workflow, best practices, author conventions, Nature ref format |
| multi-omics-checklist.md | Multi-omics papers โ data description, analysis, validation, limitations checklists |
Examples
Example 1: Draft a manuscript from analysis results
User says: "Write a paper from my spatial transcriptomics analysis targeting Nature Genetics"
Actions:
- Read existing analysis scripts, result files, and figures
- Identify the one-sentence finding and confirm with scientist
- Draft Abstract โ Introduction โ Results โ Methods โ Discussion
- Generate clean LaTeX with embedded figures
- Compile with tectonic, upload to gdrive
Result: Complete manuscript PDF with 7 figures, 3 tables, 59 references
Example 2: Generate publication figures
User says: "Create figures for the manuscript, CNS style"
Actions:
- Read result CSV files for each panel's data
- Generate multi-panel figures with CNS color palette
- Self-check each figure for overlaps, empty panels, z-index issues
- Compile LaTeX and verify figure embedding
Result: 7 publication-ready PNG+PDF figures at 300 DPI
Troubleshooting
Figures have text overlap
Cause: Panel spacing too small or legend inside plot area.
Solution: Increase hspace/wspace to 0.65+; move legend with bbox_to_anchor.
LaTeX has {[} instead of citations
Cause: File was generated by pandoc from markdown.
Solution: Generate clean LaTeX directly, using \cite{} and \begin{thebibliography}.
Figure panel is empty
Cause: Data file path wrong or column name mismatch.
Solution: Always verify data file exists and check column names with head -3 before plotting.
Table appears in wrong section
Cause: Table placed in Methods instead of Results.
Solution: Put gene/feature tables immediately after the Results paragraph that discusses them.
