Medical/scientific paper writing workflow skill. Manages the full pipeline from literature search to submission-ready manuscript. Creates and manages a project directory with IMRAD-format section files, literature matrix, reference management, and quality checklists. Supports both English and Japanese papers. Triggers: 'write paper', 'paper-write', 'start manuscript', '論文を書く', '論文執筆', '論文プロジェクト', 'manuscript', 'research paper', '原稿作成'.
Medical/scientific paper writing workflow skill. Manages the full pipeline from literature search to submission-ready manuscript. Creates and manages a project directory with IMRAD-format section files, literature matrix, reference management, and quality checklists. Supports both English and Japanese papers. Triggers: 'write paper', 'paper-write', 'start manuscript', '論文を書く', '論文執筆', '論文プロジェクト', 'manuscript', 'research paper', '原稿作成'.
context
fork
Paper Writer Skill
Full-pipeline academic paper writing assistant. From literature search to submission-ready manuscript.
Overview
This skill manages the entire paper writing workflow:
Each paper is a project directory containing structured Markdown files for every section, a literature matrix, and quality checklists.
AI-for-Science Operating Model
This skill is not only a manuscript factory (write → format → submit). It is a
research engine that wraps the writing pipeline in a discovery loop and names the
two things only a human can supply. Before doing anything else, read
~/.claude/skills/paper-writer/references/ai-for-science-model.md — it defines:
The two human-sovereign inputs. 💡 IDEA (what is worth asking, what it
means, what is ethical) and 📊 DATA (real, IRB-approved, never
machine-originated). AI proposes and executes everything else at full power; the
human owns exactly these two gates. AI must never originate a data point,
participant, or result.
The loop. Phase −1 Discovery (hypothesis → novelty → design → pre-registration
lock) feeds the existing pipeline; Phase 6.5 Adversarial Review red-teams the
central claim before any journal sees it. A red-team KILL sends the project back
to Discovery — that is the system working.
The three integrity guardrails that make AI-accelerated research more
rigorous, not less: pre-registration (anti-HARKing), novelty check
(anti-reinvention/inflation), adversarial self-review (anti-slop). Each
prevents a documented frontier failure mode.
The autonomy dial (Manual / Co-pilot / Autopilot). Hard rule for clinical
work: the 💡 IDEA gate, the 📊 DATA gate, and the pre-registration lock are
never autopilot.
The rest of this document is the execution detail. When a phase touches a sovereign
gate, stop and get the human; everywhere else, run at full power.
Supported Paper Types
Type
Structure
Reporting Guideline
Notes
Original Article
Full IMRAD
STROBE / CONSORT
Default
Case Report
Intro / Case / Discussion
CARE
Separate templates
Review Article
Thematic sections
-
Flexible structure
Letter / Short Communication
Condensed IMRAD
Same as original
Word limit focus
Systematic Review
PRISMA-compliant
PRISMA 2020
With PRISMA checklist
Study Protocol
SPIRIT-compliant
SPIRIT 2025
For trial registration papers
Workflow
Phase −1: Discovery (the research engine)
This phase is what separates a research engine from a manuscript factory. The
rest of the skill assumes the research question and the data already exist. Phase −1
produces them — a novelty-checked, powered, pre-registered study plan — before
Project Init. Read ~/.claude/skills/paper-writer/references/ai-for-science-model.md
first for the operating model.
Phase −1 is re-enterable — enter at the first guardrail not yet passed. It is
not all-or-nothing: a study that already has a sharpened question (but no novelty
check, power, or pre-registration) enters mid-chain, not at the forge. Route by
the Phase −1 entry matrix:
What the user arrives with
Enter at
How
(a) A raw clinical observation
−1.1 Forge
Run templates/research-question.md in Mode A (forge a question from the spark), then continue −1.2 → −1.3 → −1.4 in order.
(b) An existing question / advanced protocol, pre-data
−1.2 Novelty
Run templates/research-question.md in Mode B (resume/refine — back-fill PECO, single Attack pass, FINER) first, then −1.2 novelty, then −1.3 design as an AUDIT of the existing protocol (not a fresh draft — check it against templates/study-design.md, fix gaps), then −1.4 prereg, then run references/adversarial-review.md in design-stage mode (§0, pre-data) BEFORE the pre-registration lock so cheap design fixes land before freezing.
(c) Question + design + data all locked
Skip to Phase 0
Pure writing-up. Still confirm the 💡 IDEA and 📊 DATA gates are human-owned and that a pre-registration exists or is consciously waived (and disclosed as such).
−1.2 novelty is the mandatory minimum entry for any unpublished study — novelty
cannot be assumed from the fact that a protocol is already being written. Only path
(c) (already locked + data in hand) may skip it.
Start the project's accountability ledger now: create
log/human-loop-ledger.md from ~/.claude/skills/paper-writer/templates/human-loop-ledger.md
and declare the autonomy mode (Manual / Co-pilot / Autopilot). Record every gate
decision in it from here on.
Step −1.1: Forge the research question (💡 IDEA gate)
Read ~/.claude/skills/paper-writer/templates/research-question.md. From the user's
clinical observation, generate 5–15 candidate questions, debate and rank them by
FINER, evolve the top 2–3 — then stop and have the human select. AI never
auto-selects the question. Output: one sharpened research question with its PICO.
Read ~/.claude/skills/paper-writer/references/novelty-check.md. Run a live-literature
sweep on the selected question using the real literature tools (PubMed MCP,
OpenAlex, Europe PMC, Semantic Scholar — see Phase 1 plumbing). Classify the gap:
genuinely novel / incremental / already-answered / contested. An already-answered
question is killed here at near-zero cost. Do not inflate novelty — that is the
Sakana v2 failure mode.
Step −1.3: Design the study & power it
Read ~/.claude/skills/paper-writer/templates/study-design.md. Choose the design,
operationalize every PICO element into a measured variable, define the single
primary outcome, map confounders with a DAG, and run a sample-size/power
calculation (justify the effect size from the novelty-check literature, not from
hope). Check feasibility against the clinic's real volume. This design becomes both
the pre-registration and, later, the Methods section.
Read ~/.claude/skills/paper-writer/templates/preregistration.md. Freeze the
hypotheses and the primary analysis plan (OSF / UMIN-CTR / jRCT / PROSPERO) before
the 📊 DATA gate. After the lock: pre-registered analyses are confirmatory;
everything else is exploratory and labeled as such. This is the integrity backbone
for publishing under your own name. For retrospective data, register before
examining outcome data and disclose the data's pre-existence honestly.
The 📊 DATA gate: only after the plan is locked does the human supply real,
IRB-approved data. AI never originates data. Proceed to Phase 0.
Phase 0: Project Initialization
When the user invokes this skill, ask for:
Working title (can change later)
Paper type (Original Article / Case Report / Review / Letter / Systematic Review)
Target journal (optional but recommended)
Language (English / Japanese / Both)
Research question in one sentence
Key data available (what Tables/Figures already exist?)
Step 0.1: Capture Journal Requirements
If a target journal is specified, look up and document:
Word limits: total manuscript, abstract, each section (if specified)
Citation style: Vancouver, APA, NLM, or other
Required sections: some journals require separate Conclusion, others don't
Abstract format: structured or unstructured, word limit
Figure/Table limits: maximum number allowed
Reporting guideline: which checklist the journal requires
Special requirements: cover page format, line numbering, etc.
AI disclosure: whether the journal requires AI usage disclosure, and where (Methods, Acknowledgments, or dedicated section). See references/ai-disclosure.md.
Keywords: number required, MeSH preferred or free-text. See references/keywords-guide.md.
Graphical abstract: required or optional. See templates/graphical-abstract.md.
Use WebSearch to look up the journal's "Instructions for Authors" page.
Record all requirements in the README.md under a "Journal Requirements" section.
Step 0.2: Select Reporting Guideline
Based on paper type and study design, select the appropriate reporting guideline:
Study Type
Guideline
Reference
Randomized Controlled Trial
CONSORT 2025
references/reporting-guidelines-full.md
Observational study (cohort, case-control, cross-sectional)
STROBE
references/reporting-guidelines-full.md
Systematic review / meta-analysis
PRISMA 2020
references/reporting-guidelines-full.md
Case report
CARE
references/reporting-guidelines-full.md
Diagnostic accuracy study
STARD 2015
references/reporting-guidelines-full.md
Quality improvement study
SQUIRE 2.0
references/reporting-guidelines-full.md
Study protocol (clinical trial)
SPIRIT 2025
references/reporting-guidelines-full.md
Prediction model (incl. AI/ML)
TRIPOD+AI 2024
references/reporting-guidelines-full.md
Animal research
ARRIVE 2.0
references/reporting-guidelines-full.md
Health economics
CHEERS 2022
references/reporting-guidelines-full.md
Read ~/.claude/skills/paper-writer/references/reporting-guidelines.md (summary) or references/reporting-guidelines-full.md (comprehensive) and note the key checklist items for the selected guideline. These items will be checked throughout the writing process.
Step 0.3: Create Project Directory
For Original Article / Review / Letter / Systematic Review:
Read ~/.claude/skills/paper-writer/templates/project-init.md with the Read tool and use it to generate README.md. For Case Reports, use project-init-case.md instead.
File numbering follows the recommended writing order, not the reading order. This is intentional.
Step 0.4: Organize Research Data
If the user has existing research data (clinical records, CSV files, statistical output, etc.):
Read ~/.claude/skills/paper-writer/templates/data-management.md for the full template
Ask the user to place raw data files in data/raw/ — these files are READ-ONLY from this point
Create data/raw/README.md documenting the data source, extraction date, and IRB information
Create data/data-dictionary.md listing all variables with types, ranges, and labels
Confirm de-identification status — if not yet de-identified, create a processing plan in data/processed/README.md
Security rules:
NEVER commit patient-identifiable data to git
Add data/raw/*.csv, data/raw/*.xlsx etc. to .gitignore if the repository is shared
Always confirm IRB approval number before proceeding with data analysis
Data flow:raw/ (never modify) → processed/ (clean, de-identify) → analysis/ (statistical output) → tables/ and figures/ (manuscript-ready)
Step 0.5: Data Analysis
If the user has quantitative data ready for analysis, Claude Code can execute Python scripts directly. Read ~/.claude/skills/paper-writer/templates/analysis-workflow.md for the full workflow.
See references/statistical-reporting-full.md for detailed SAMPL guidelines and templates/analysis-workflow.md for step-by-step commands.
Phase 1: Literature Search & Organization
Step 1.1: Define Search Strategy
Create 00_literature/search-strategy.md with:
Databases: PubMed, Google Scholar (always available); Scopus, CiNii (if user has institutional access)
Search terms: MeSH terms + free-text keywords
Inclusion/exclusion criteria for papers
Date range
How to search — use REAL literature tools, not plain web search.
This skill runs in an environment with a real PubMed MCP and research APIs. These
return structured, verifiable records (PMID, DOI, authors, abstract) — use them as
the primary path. Plain WebSearch is a fallback, not the default.
Primary: PubMed MCP (biomedical, authoritative). Build the query with
references/pubmed-query-builder.md, then:
mcp__claude_ai_PubMed__search_articles — run the MeSH + free-text query
mcp__claude_ai_PubMed__get_article_metadata — pull structured metadata per PMID
mcp__claude_ai_PubMed__find_related_articles — snowball from a key seed paper
mcp__claude_ai_PubMed__lookup_article_by_citation — resolve a citation to a PMID/DOI
mcp__claude_ai_PubMed__get_full_text_article — fetch full text where available
Cochrane / PROSPERO / Epistemonikos — check for existing or in-progress systematic reviews
Why this matters: structured-record retrieval means every paper carries a real
PMID/DOI, so the "is this citation fabricated?" risk drops sharply versus
free-text web search. Still verify per references/citation-verification.md.
Workflow:
Ask the user for their 3–5 key papers (they usually know them) — use these as snowball seeds for find_related_articles
Run the PubMed MCP query; supplement with OpenAlex / Europe PMC / Semantic Scholar for non-PubMed and preprint coverage
De-duplicate by DOI; have the user validate the final list for completeness
Verify every citation resolves to a real record (references/citation-verification.md)
Step 1.2: Build Literature Matrix
Read ~/.claude/skills/paper-writer/templates/literature-matrix.md with the Read tool.
For each relevant paper found, extract and organize:
Author (Year)
Design
N
Population
Key Finding
Limitation
Relevance
Aim for 15-30 papers for an original article, 8-15 for a case report, 30-50 for a systematic review.
Step 1.3: Identify Key Papers
For the 3-5 most important papers, create individual notes in 00_literature/key-papers/ with:
Applies only to Systematic Reviews. Skip for all other paper types.
Phase 1 builds a search; Phase 3-D writes the PRISMA Methods/Results. Between
them sits the actual study selection — dedup, dual screening, and the record
counts that fill the PRISMA flow diagram. This phase runs that pipeline.
Read ~/.claude/skills/paper-writer/templates/sr-screening-pipeline.md with the
Read tool for the full procedure. In brief:
Prerequisite — registered protocol. Eligibility criteria must exist in
00_literature/protocol.md (from templates/sr-prospero.md) and the
protocol must be registered (PROSPERO) BEFORE screening. Do not start
otherwise.
Stage 1 — De-duplicate (deterministic). Place raw DB exports in
00_literature/screening/00_imported/ (one file per database), then run:
Stage 2 — Title/Abstract screening (DUAL). Spawn two independent
screener passes (Agent tool, or team mode) that cannot see each other's
decisions; each judges include/exclude/unclear against protocol.md only.
Reconcile into 02_title_abstract_screen.csv; surface every conflict to the
user. An LLM is one arm of a dual review, never the sole arbiter.
Stage 3 — Full-text screening (DUAL). Link PDFs to records, then run two
independent full-text passes:
Every full-text exclude carries a PRISMA reason category. Write
03_fulltext_screen.csv; the human resolves all conflicts.
Stage 4 — Extraction hand-off. For each included study, create one
extraction/{record_id}.md from templates/sr-data-extraction.md (dual,
no guessing — NR/N/A only).
Copy the counts into templates/sr-prisma-flow.md and the Cohen's κ values
into the Methods selection-process paragraph (Phase 3-D, item 5).
Team mode: the two screening passes per stage are naturally parallel — run
them as two concurrent agents, each given only protocol.md + the records, then
reconcile. κ < 0.6 means the criteria are ambiguous: revise protocol.md and
re-screen rather than proceeding.
Phase 2: Outline
Create 01_outline.md with the paper skeleton.
Read ~/.claude/skills/paper-writer/references/imrad-guide.md with the Read tool for the detailed IMRAD structure. For Case Reports, this guide does not apply directly — use the CARE structure instead.
The outline should specify:
Each section's key points (bullet list)
Which papers support which points
Which Tables/Figures go where
The story arc: Background Problem → Gap → Our Approach → Findings → Implications
For Case Reports: Background → Why Reportable → Case Details → Clinical Lesson
Get user approval on outline before proceeding to drafting.
Phase 2.5: Tables & Figures
Read ~/.claude/skills/paper-writer/references/tables-figures-guide.md with the Read tool.
Tables and figures are the backbone of a paper — many reviewers look at the abstract, then the tables/figures, before reading the text. Design them before writing prose so the text can reference them naturally.
Step 2.5.1: Plan Tables & Figures
Based on the outline, determine:
Which data belongs in a table vs. a figure vs. the text
Table 1 is almost always "Baseline Characteristics" (use the template in references/tables-figures-guide.md)
How many tables/figures are allowed by the journal (check Phase 0 requirements)
table2_*.md — Additional tables as needed (regression results, outcomes, etc.)
Rules:
Title above the table
No vertical lines (horizontal lines only)
Consistent decimal places within each column
Footnotes for abbreviations and statistical tests
Total sample size in the header row
Step 2.5.3: Plan Figures
Create caption files in figures/ directory:
fig1_caption.md — Often a flow diagram (CONSORT/PRISMA) or study design
fig2_caption.md — Key result visualization
Rules:
Captions must be self-explanatory without reading the main text
Include key statistics in captions
Specify resolution requirements (300+ DPI for print, 600+ for line art)
Use colorblind-friendly palettes
Step 2.5.4: Graphical Abstract (if required)
If the journal requires or encourages a graphical abstract, read ~/.claude/skills/paper-writer/templates/graphical-abstract.md and plan the visual summary.
Get user review on table/figure plan before proceeding to drafting.
Phase 3: Drafting
The writing order is intentional and produces better papers. Follow it strictly.
3-A: Original Article Workflow
Step 3.1: Methods & Results (Write as a pair)
Read ~/.claude/skills/paper-writer/templates/methods.md and ~/.claude/skills/paper-writer/templates/results.md with the Read tool.
Methods rules:
Reproducibility is everything
Include: study design, patients/subjects, data collection, statistical analysis, ethics
Every method must have a corresponding result
Results rules:
Facts only, no interpretation
No references to other studies
Every Table/Figure must be mentioned in text
Methods ↔ Results must correspond 1:1
Write sections/02_methods.md and sections/03_results.md together, ensuring perfect correspondence. Cross-check: every subsection in Methods must map to a corresponding subsection in Results, and vice versa.
Step 3.2: Introduction (Paragraph 3) & Conclusion (Write as a pair)
Read ~/.claude/skills/paper-writer/templates/introduction.md and ~/.claude/skills/paper-writer/templates/conclusion.md with the Read tool.
Why write Paragraph 3 first? The study objective (Introduction P3) and the conclusion must mirror each other. Writing them together guarantees alignment. Paragraphs 1-2 provide background that funnels toward the objective — they are easier to write once the objective is locked.
Introduction structure (3 paragraphs):
General background (everyone agrees with this)
Clinical question / knowledge gap (but we don't know X)
Study objective (therefore, we investigated...)
Conclusion rules:
Must directly answer the objective stated in Introduction paragraph 3
One core message
Brief and direct
Write the final paragraph of sections/04_introduction.md and sections/06_conclusion.md together to ensure they mirror each other.
Step 3.3: Discussion
Read ~/.claude/skills/paper-writer/templates/discussion.md with the Read tool.
Discussion structure:
Summary of main findings
2-N. Comparison with prior literature (use 00_literature/literature-matrix.md)
N+1. Limitations — read ~/.claude/skills/paper-writer/templates/limitations-guide.md for categories, templates, and bilingual examples
N+2. Clinical implications / future directions
Discussion rules:
No new results
No excessive speculation
Support every claim with a reference
Keep it focused
Limitations subsection is mandatory — be specific about direction of bias and mitigation
Step 3.4: Introduction (Paragraphs 1-2)
Now write paragraphs 1-2 of sections/04_introduction.md. The background should funnel toward the research question already written in paragraph 3.
Step 3.5: Abstract
Read ~/.claude/skills/paper-writer/templates/abstract.md with the Read tool.
Write sections/07_abstract.md as a structured abstract:
Background/Objective (1-2 sentences)
Methods (2-3 sentences)
Results (3-4 sentences)
Conclusions (1-2 sentences)
Check the journal-specific word limit captured in Phase 0. The Abstract must be consistent with the full text. Cross-check all numbers.
Step 3.6: Title
Write sections/08_title.md with 3-5 title candidates. Evaluate each against:
Specific (what was studied?)
Concise (< 15 words ideal)
Contains keywords (searchable)
No conclusion spoilers
Get user approval on final title.
3-B: Case Report Workflow
Step 3.1-CR: Case Presentation
Read ~/.claude/skills/paper-writer/templates/case-report.md with the Read tool.
Write 02_case.md following the CARE structure:
Patient information (demographics, history)
Clinical findings
Timeline (consider a timeline figure)
Diagnostic assessment
Therapeutic intervention
Follow-up and outcomes
Patient perspective (CARE item 10) — when possible, include the patient's own experience in their words
Rules:
Chronological order
Only clinically relevant details
Document informed consent for publication
Report both positive AND negative findings
Patient perspective strengthens the report and is recommended by CARE guidelines
Step 3.2-CR: Discussion
Read ~/.claude/skills/paper-writer/templates/discussion.md with the Read tool.
Write 04_discussion.md:
Why this case is significant (clinical lesson)
Comparison with published literature
Limitations of the case
Clinical implications
Keep it focused and shorter than in an Original Article.
Step 3.3-CR: Introduction
Read ~/.claude/skills/paper-writer/templates/case-introduction.md with the Read tool.
Write 03_introduction.md:
Brief background on the condition
Why this case is reportable (rarity, novelty, instructive value)
Optional: "We report a case of... to highlight..."
Write the Introduction AFTER the Case section — you need to know the full case to justify its reporting.
Step 3.4-CR: Abstract
Read ~/.claude/skills/paper-writer/templates/case-abstract.md with the Read tool.
Write 05_abstract.md using the CARE abstract structure:
Background (1-2 sentences: why this case is worth reporting)
Case Presentation (3-5 sentences: demographics, findings, diagnosis, treatment, outcome)
Conclusions (1-2 sentences: clinical lesson)
Do NOT use Methods/Results structure for Case Report abstracts.
Step 3.5-CR: Title
Write 06_title.md with 3-5 title candidates. For case reports:
Title MUST contain "case report" (CARE requirement)
Include the diagnosis or key finding
Example: "Successful treatment of severe pediatric asthma with dupilumab: a case report"
Get user approval on final title.
3-C: Review Article Workflow
Review articles synthesize existing literature on a topic. The structure is thematic rather than IMRAD.
Step 3.1-RA: Thematic Sections
Read ~/.claude/skills/paper-writer/templates/discussion.md for general writing guidance.
Organize the body into thematic sections based on the outline. Common structures:
Chronological: Evolution of understanding over time
Thematic: Grouped by subtopic (most common)
Methodological: Grouped by study approach
Each section should:
Synthesize findings across studies (not just summarize one at a time)
Identify areas of consensus and controversy
Highlight gaps in the literature
Use the literature matrix to ensure comprehensive coverage
Step 3.2-RA: Introduction
Write the introduction:
Scope and importance of the topic
Why a review is needed now (new evidence, controversy, emerging field)
Objectives and scope of this review
Step 3.3-RA: Conclusion & Future Directions
Write the conclusion:
Synthesize the key themes identified
Current state of knowledge
Gaps and future research directions
Clinical implications (if applicable)
Step 3.4-RA: Abstract
Write an unstructured abstract (unless journal requires structured format):
Purpose of the review
Methods (databases searched, date range, selection criteria)
Key findings synthesized across themes
Conclusions
Step 3.5-RA: Title
Write title candidates. For review articles:
Include "review", "narrative review", or "scoping review" in the title
Clearly state the topic
Example: "Artificial intelligence in diagnostic radiology: a narrative review"
Get user approval on final title.
3-D: Systematic Review Workflow
Read ~/.claude/skills/paper-writer/templates/sr-outline.md with the Read tool for the complete PRISMA 2020-compliant template.
Systematic reviews follow a strict, pre-registered protocol. The template provides the full structure with PRISMA 2020 checklist item numbers.
Step 3.1-SR: Methods
The Methods section is the most critical part. Write it following PRISMA items P-5 through P-18:
Protocol and registration (PROSPERO ID)
Eligibility criteria (PICO/PECO)
Information sources (databases, dates)
Search strategy (full strategy in supplementary)
Selection process (screening, inter-rater reliability)
Scientific content unchanged (no data or citations lost)
No "Additionally" / "Furthermore" at sentence start (max 1 per section)
No "pivotal" / "crucial" / "landscape" / "delve"
No "-ing" phrases tacked on for fake depth
No "serves as" / "stands as" (use "is")
Em dashes used sparingly (< 2 per page)
Consistent terminology (no synonym cycling)
Sentence rhythm varies (short and long sentences mixed)
No generic conclusions remaining
Hedging proportionate to evidence strength
日本語:
「さらに」「また」「加えて」の連発がない（各セクション最大1回）
同じ語尾が3回以上続いていない
根拠なき「非常に」「大きな」がない
受動態の過剰使用がない（能動態に直す）
定型的な締めの句がない（「参考になれば幸いである」等）
抽象語だけで押し切っていない
カギ括弧を多用していない
Phase 5: References
Read ~/.claude/skills/paper-writer/references/citation-guide.md with the Read tool.
Build references/09_references.md (or references/07_references.md for Case Reports):
Collect all cited papers from all sections
Format according to target journal style captured in Phase 0 (Vancouver, APA, etc.)
Number sequentially as cited
Verify completeness: every reference is cited in text, every citation has a reference entry
Verify authenticity: For EVERY AI-suggested reference, confirm the paper exists against a structured record — mcp__claude_ai_PubMed__lookup_article_by_citation or mcp__claude_ai_PubMed__get_article_metadata (by PMID/DOI), falling back to CrossRef (https://api.crossref.org/works/{DOI}) and WebSearch on the exact title for non-PubMed sources. AI frequently fabricates plausible-sounding citations; a citation that does not resolve to a real PMID/DOI is removed or replaced. See references/citation-verification.md.
Phase 6: Quality Review
Read ~/.claude/skills/paper-writer/references/section-checklist.md with the Read tool. For Case Reports, also check the CARE-specific items in templates/case-report.md.
Run the quality checklist against each section. Update checklists/section-quality.md with results.
No interpretation in Results (Original Article only)
No new results in Discussion
Abstract numbers match full text
All references cited and formatted
Word count within target journal limits (check Phase 0 requirements)
Reporting guideline followed (check Phase 0 selected guideline)
AI writing patterns removed (Phase 4 verification passed)
Consistent terminology throughout all sections
Ethics approval and informed consent documented
Phase 6.5: Adversarial Review (red-team the claim)
Before any journal's red team sees the paper, run your own. Quality Review
(Phase 6) checks that the manuscript is internally consistent and well-formatted.
This phase checks something different and harder: is the central claim actually
true and supported? It is the corrective to the Sakana v2 failure mode (nothing in
that loop tried to make the paper fail, so hallucinations and novelty inflation
shipped).
Read ~/.claude/skills/paper-writer/references/adversarial-review.md. Run a hostile
internal panel that attacks the central claim from four angles, plus a
"steelman the null" pass:
Statistical reviewer — p-hacking, multiplicity, power, post-hoc subgroups; does every confirmatory claim trace to the pre-registered plan (templates/preregistration.md)?
Methodological reviewer — confounding, selection/information bias, the DAG; is each causal-sounding claim supported by the design?
Novelty reviewer — re-attack the novelty claim against references/novelty-check.md; has a larger/better study already shown this?
Integrity / clinical reviewer — every number traces to raw data (📊 gate), citations are real, no overclaiming, no clinical harm if a reader acts on the conclusion.
Steelman the null — argue as hard as possible that the finding is chance, confounding, or bias; does the paper already answer that argument?
Verdict: KILL (claim not supported → return to Phase −1 Discovery; this is the
system working, not a failure) / MAJOR / MINOR / PASS. The human owns the final
KILL/PASS adjudication (💡 gate) — the AI panel advises. Record the verdict in
log/human-loop-ledger.md. Only a PASS proceeds to Pre-Submission.
In team mode, run the four reviewers as four parallel paper-red-team agents (opus,
distinct lenses) and synthesize the verdict. A KILL or MAJOR loops back: fix the
affected sections, then re-run Phase 4 (Humanize) and Phase 6 (Quality) before
re-entering Phase 6.5.
Phase 7: Pre-Submission
Read ~/.claude/skills/paper-writer/templates/cover-letter.md and ~/.claude/skills/paper-writer/templates/submission-ready.md with the Read tool.
Create:
Title page — read ~/.claude/skills/paper-writer/templates/title-page.md for the template (running head, all authors with ORCID, affiliations, word counts, corresponding author, clinical trial registration)
Highlights / Key Points — read ~/.claude/skills/paper-writer/templates/highlights.md and create the appropriate summary box for the target journal (JAMA Key Points, BMJ "What is known", Elsevier Highlights, Lay Summary, etc.)
Acknowledgments — read ~/.claude/skills/paper-writer/templates/acknowledgments.md and draft (non-author contributions, AI tool disclosure, patient acknowledgment)
Declarations — read ~/.claude/skills/paper-writer/templates/declarations.md and complete (Ethics, COI using references/coi-detailed.md, Funding, Data Availability, AI Disclosure, CRediT)
Cover letter using the template
checklists/submission-ready.md using the template — fill in journal-specific limits from Phase 0
Compile all sections into a single reading-order Markdown file → submissions/v1_{journal}/compiled-manuscript.md
Create submissions/v1_{journal}/submission-log.md with submission date, portal, manuscript ID
Title → Abstract → Introduction → Case Presentation → Discussion → References
The compiled file should include all section content in sequence. Tables and Figures should be referenced but kept in their separate folders. All submission documents go into submissions/v1_{journal}/.
Phase 8: Revision (Post-Review)
When the user receives reviewer comments (peer review, editorial decision letter):
Step 8.1: Organize Reviewer Comments
Create revisions/r1/reviewer-comments.md:
Parse the decision letter and reviewer comments
Number each comment sequentially (R1-1, R1-2, R2-1, R2-2, etc.)
Categorize each comment:
Must fix: Factual errors, missing data, methodological concerns
Should fix: Reasonable suggestions that improve the paper
Rebut: Comments based on misunderstanding (requires polite explanation)
Step 8.2: Create Response Letter
Create revisions/r1/response-letter.md:
For each comment, use this format:
**Comment R1-1:** [Quote the reviewer's comment]
**Response:** [Your response]
**Changes made:** [Specific changes with page/line numbers, or explanation if no change]
Rules for response letters:
Thank the reviewer for constructive feedback (once at the beginning, not per comment)
Be specific about what was changed and where
For rebuttals, acknowledge the reviewer's perspective, then explain with evidence
Never be defensive or dismissive
If a change was not made, explain why with references or data
Step 8.3: Implement Revisions
Track which sections need modification based on reviewer comments
Make changes in the relevant section files
Mark changed text (many journals require highlighted changes or a diff)
Roll back affected phases: re-run Humanize (Phase 4) and Quality Review (Phase 6) on modified sections
Update word counts and verify journal limits are still met
Step 8.4: Verify Revision Completeness
Every reviewer comment has a response
Every "Must fix" and "Should fix" item has been addressed
Rebuttals are supported by evidence
Changed text is marked/highlighted
References updated if new citations added
Abstract updated if results or conclusions changed
Cover letter for resubmission drafted
Phase 9: Post-Acceptance
Read ~/.claude/skills/paper-writer/templates/proof-correction.md with the Read tool.
After acceptance, the corresponding author receives galley proofs. This is the LAST opportunity to correct errors.
Step 9.1: Proof Review
When proofs arrive (typically 2-8 weeks after acceptance, turnaround: 24-72 hours):
Critical checks:
Author names, affiliations, and ORCID — correct?
Abstract numbers match main text?
All tables — data values correct, no transposition errors?
All figures — correct images, acceptable quality?
Reference list — complete, correct numbering?
Corresponding author email — correct?
Funding and COI statements — accurate?
Clinical trial registration number — present?
NOT allowed at proof stage:
Rewriting sentences or paragraphs
Adding new data, references, or authors
Changing conclusions
Step 9.2: Submit Corrections
Use the journal's proofing system (Proof Central, CATS, eProofing, or direct PDF return). For each correction: state page, column, line, and exact change.
Step 9.3: Post-Publication
After publication:
Verify the final published version matches the accepted manuscript
Share via institutional repository (Green OA) if applicable — see references/open-access-guide.md
Update clinical trial registry with results (if applicable) — see references/clinical-trial-registration.md
Share with co-authors and collaborators
Phase 10: Rejection & Resubmission
Read ~/.claude/skills/paper-writer/references/desk-rejection-prevention.md and references/journal-reformatting.md with the Read tool.
Step 10.1: Assess the Rejection
Decision
Action
Desk rejection (scope)
Reformat and submit to next journal immediately
Desk rejection (quality)
Revise manuscript, then reformat and submit
Peer review rejection
Read reviews carefully; major revision before next journal
Reject with encouragement to resubmit
Treat as major revision; address all comments
Step 10.2: Quick Reformat
Use references/journal-reformatting.md checklist:
Change reference format (use reference manager)
Restructure abstract with new headings — see references/abstract-formats.md
Adjust word count — see references/word-count-limits.md
Add/remove special sections (Key Points, Highlights)
Reformat title page
Write new cover letter (address new editor by name)
Verify no mention of previous journal name in manuscript
Search, organize, summarize — user validates relevance
Methods
Draft based on user's data description — user verifies accuracy
Results
Structure and format — user provides the actual data
Case (Case Report)
Structure chronologically — user provides clinical details
Introduction
Draft background from literature — user refines narrative
Discussion
Suggest comparisons with literature — user controls interpretation
Abstract
Generate from full text — user ensures accuracy
References
Format and organize — user verifies completeness and authenticity
What AI Should NOT Do
Fabricate data or statistics
Invent citations (always verify with WebSearch)
Write Results without user-provided data
Write Case Presentation without user-provided clinical details
Make clinical recommendations beyond the data
Skip the user approval step at outline and title phases
Status Tracking
Update README.md status after each phase. Use these status values:
Not Started: Phase not begun
In Progress: Phase actively being worked on (add details in Notes)
Draft Complete: First draft finished, pending review
Done: Phase completed and reviewed
Use the appropriate status tracker based on paper type:
Original Article:
Phase
Status
Last Updated
Literature Search
Not Started
-
Outline
Not Started
-
Tables & Figures
Not Started
-
Methods & Results
Not Started
-
Introduction & Conclusion
Not Started
-
Discussion
Not Started
-
Abstract
Not Started
-
Title & Keywords
Not Started
-
Humanize
Not Started
-
References
Not Started
-
Declarations
Not Started
-
Quality Review
Not Started
-
Pre-Submission
Not Started
-
Case Report:
Phase
Status
Last Updated
Literature Search
Not Started
-
Outline
Not Started
-
Tables & Figures
Not Started
-
Case Presentation
Not Started
-
Discussion
Not Started
-
Introduction
Not Started
-
Abstract
Not Started
-
Title & Keywords
Not Started
-
Humanize
Not Started
-
References
Not Started
-
Declarations
Not Started
-
Quality Review
Not Started
-
Pre-Submission
Not Started
-
Review Article:
Phase
Status
Last Updated
Literature Search
Not Started
-
Outline
Not Started
-
Tables & Figures
Not Started
-
Thematic Sections
Not Started
-
Introduction
Not Started
-
Conclusion & Future Directions
Not Started
-
Abstract
Not Started
-
Title & Keywords
Not Started
-
Humanize
Not Started
-
References
Not Started
-
Declarations
Not Started
-
Quality Review
Not Started
-
Pre-Submission
Not Started
-
Systematic Review:
Phase
Status
Last Updated
Literature Search
Not Started
-
Outline
Not Started
-
Tables & Figures
Not Started
-
Methods (PRISMA)
Not Started
-
Results (PRISMA)
Not Started
-
Discussion
Not Started
-
Introduction
Not Started
-
Abstract
Not Started
-
Title & Keywords
Not Started
-
Humanize
Not Started
-
References
Not Started
-
Declarations
Not Started
-
Quality Review
Not Started
-
Pre-Submission
Not Started
-
Letter / Short Communication:
Phase
Status
Last Updated
Literature Search
Not Started
-
Outline
Not Started
-
Tables & Figures
Not Started
-
Condensed Draft
Not Started
-
Abstract
Not Started
-
Title & Keywords
Not Started
-
Humanize
Not Started
-
References
Not Started
-
Quality Review
Not Started
-
Pre-Submission
Not Started
-
Resuming a Project
When the user invokes this skill on an existing project directory:
Read README.md to understand current status, paper type, target journal, and research question
Scan section files to assess actual content state:
Read each section file that shows "In Progress" or "Draft Complete"
Check word count and completeness (empty sections, TODO markers, partial drafts)
Compare actual file state with the status tracker — the files are the source of truth
Present a summary to the user: "Here is where we left off: [status]. The next step is [phase]. Shall I continue?"
Check for workflow updates: Compare the README status table against the canonical phase list above. If phases are missing (e.g., old project created before "Humanize" was added), add them with "Not Started" status and inform the user
Resume from the next incomplete phase
Update status tracker
Handling Mid-Project Changes
Changing target journal: If the user wants to change the target journal:
Update README.md Paper Info and Journal Requirements
Re-check: citation style, word limits, abstract format, reporting guideline
Reformat references if citation style changed
Check word counts against new limits
Update cover letter
Adding data or revisions: If the user has new data or reviewer feedback:
Identify which sections are affected
Roll back affected phases to "In Progress"
Re-run from that phase forward (including Humanize and Quality Review)
references/ai-for-science-model.md - Read first. Operating model: the two human-sovereign inputs (💡 IDEA, 📊 DATA), the discovery loop, the three integrity guardrails, the autonomy dial
references/novelty-check.md - Live-literature novelty sweep via real APIs (PubMed MCP, OpenAlex, Europe PMC, Semantic Scholar); four-verdict gap classification
templates/study-design.md - Design selection, PICO→variables, DAG/confounding, sample-size/power, pre-specified analysis plan
