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:

[Discovery] → Literature Search → Outline → Tables/Figures → Draft → Humanize → References → Quality Review → [Adversarial Review] → Pre-Submission → [Revision] → [Post-Acceptance] → [Rejection → Resubmission]

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

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:

−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.

#### Step −1.2: Novelty check (guardrail: anti-reinvention)

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.

#### Step −1.4: Pre-register & lock (guardrail: anti-HARKing)

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:

1. Working title (can change later)
2. Paper type (Original Article / Case Report / Review / Letter / Systematic Review)
3. Target journal (optional but recommended)
4. Language (English / Japanese / Both)
5. Research question in one sentence
6. 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:

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:

{project-dir}/
├── README.md                        # Project dashboard (status, timeline, links)
│
├── 00_literature/                   # Phase 1: Literature
│   ├── search-strategy.md           #   Search terms, databases, dates
│   ├── literature-matrix.md         #   Structured comparison table
│   └── key-papers/                  #   Notes on important papers (1 file per paper)
│
├── 01_outline.md                    # Phase 2: Paper skeleton
│
├── sections/                        # Phase 3: Manuscript sections (writing order)
│   ├── 02_methods.md
│   ├── 03_results.md
│   ├── 04_introduction.md
│   ├── 05_discussion.md
│   ├── 06_conclusion.md
│   ├── 07_abstract.md
│   └── 08_title.md
│
├── tables/                          # Tables (numbered: table1_baseline.md, ...)
├── figures/                         # Figures + captions (fig1_caption.md, ...)
├── supplements/                     # Supplementary materials
│   ├── supplementary-tables/        #   e.g., tableS1_sensitivity.md
│   ├── supplementary-figures/       #   e.g., figS1_subgroup.md
│   └── appendices/                  #   Additional methods, datasets, etc.
│
├── data/                            # Research data (see templates/data-management.md)
│   ├── raw/                         #   Original data (READ-ONLY, gitignored)
│   ├── processed/                   #   Cleaned, de-identified data
│   ├── analysis/                    #   Statistical output, scripts
│   └── data-dictionary.md           #   Variable definitions
│
├── ethics/                          # Ethics & regulatory
│   ├── irb-approval.md              #   IRB approval details, number, dates
│   ├── informed-consent.md          #   Consent process documentation
│   ├── protocol.md                  #   Study protocol (SPIRIT if applicable)
│   └── registration.md              #   Trial/study registration (UMIN, ClinicalTrials.gov)
│
├── submissions/                     # Submission history (1 subfolder per attempt)
│   └── v1_{journal}/                #   e.g., v1_bmj/
│       ├── compiled-manuscript.md   #     Full compiled manuscript snapshot
│       ├── cover-letter.md          #     Cover letter
│       ├── title-page.md            #     Title page
│       ├── declarations.md          #     Ethics, COI, funding, AI disclosure
│       ├── highlights.md            #     Key points (if required)
│       ├── graphical-abstract.md    #     Graphical abstract (if required)
│       └── submission-log.md        #     Date, portal, manuscript #, status
│
├── revisions/                       # Revision rounds (Phase 8)
│   └── r1/                          #   Round 1
│       ├── reviewer-comments.md     #     Original reviewer comments
│       ├── response-letter.md       #     Point-by-point response
│       ├── revision-cover-letter.md #     Revision cover letter
│       ├── diff-summary.md          #     Changes made (section, line, change)
│       └── compiled-manuscript.md   #     Revised manuscript snapshot
│
├── coauthor-review/                 # Co-author feedback tracking
│   ├── review-tracker.md            #   Who reviewed, when, status
│   └── feedback/                    #   Individual feedback files
│
├── correspondence/                  # Editor & reviewer communication log
│   └── YYYY-MM-DD_{subject}.md      #   e.g., 2026-03-05_decision-letter.md
│
├── references/                      # Reference management
│   └── 09_references.md             #   Formatted reference list
│
├── checklists/                      # Quality control
│   ├── section-quality.md           #   Per-section quality scores
│   ├── submission-ready.md          #   Pre-submission checklist
│   ├── reporting-guideline.md       #   CONSORT/STROBE/etc. item tracking
│   ├── gate-state.md                #   Stage-gate iteration state
│   └── feedback-*.md                #   Auto-generated gate feedback
│
└── log/                             # Decision & progress log
    ├── decisions.md                 #   Key decisions with rationale
    ├── meetings.md                  #   Meeting notes (co-authors, supervisor)
    └── timeline.md                  #   Milestone targets & actual dates

For Case Report:

{project-dir}/
├── README.md                        # Project dashboard
│
├── 00_literature/
│   ├── search-strategy.md
│   ├── literature-matrix.md
│   └── key-papers/
│
├── 01_outline.md
│
├── sections/
│   ├── 02_case.md                   # Case presentation (CARE structure)
│   ├── 03_introduction.md           # Introduction (why reportable)
│   ├── 04_discussion.md
│   ├── 05_abstract.md               # Abstract (CARE format)
│   └── 06_title.md                  # Title (must contain "case report")
│
├── tables/
├── figures/
├── supplements/
│   ├── supplementary-tables/
│   ├── supplementary-figures/
│   └── appendices/
│
├── data/
│   ├── raw/
│   ├── processed/
│   ├── analysis/
│   └── data-dictionary.md
│
├── ethics/
│   ├── irb-approval.md
│   ├── informed-consent.md          # Patient consent for publication
│   └── patient-perspective.md       # Patient's perspective (CARE item)
│
├── submissions/
│   └── v1_{journal}/
│       ├── compiled-manuscript.md
│       ├── cover-letter.md
│       ├── title-page.md
│       ├── declarations.md
│       └── submission-log.md
│
├── revisions/
│   └── r1/
│       ├── reviewer-comments.md
│       ├── response-letter.md
│       ├── diff-summary.md
│       └── compiled-manuscript.md
│
├── coauthor-review/
│   ├── review-tracker.md
│   └── feedback/
│
├── correspondence/
│   └── YYYY-MM-DD_{subject}.md
│
├── references/
│   └── 07_references.md
│
├── checklists/
│   ├── section-quality.md
│   ├── submission-ready.md
│   ├── reporting-guideline.md
│   ├── gate-state.md
│   └── feedback-*.md
│
└── log/
    ├── decisions.md
    ├── meetings.md
    └── timeline.md

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.):

1. Create data/raw/, data/processed/, data/analysis/ directories
2. Read ~/.claude/skills/paper-writer/templates/data-management.md for the full template
3. Ask the user to place raw data files in data/raw/ — these files are READ-ONLY from this point
4. Create data/raw/README.md documenting the data source, extraction date, and IRB information
5. Create data/data-dictionary.md listing all variables with types, ranges, and labels
6. 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.

Available analysis scripts:

Workflow:

1. Inspect data: Load data/processed/cohort_final.csv, check shape, dtypes, missing values
2. Table 1: Run scripts/table1.py to generate baseline characteristics table → tables/table1.md
3. Primary analysis: Choose analysis type based on study design:

• Cross-sectional / case-control → logistic regression (OR with 95% CI)
• Cohort with time-to-event → survival analysis (Kaplan-Meier, log-rank)
• Continuous outcome → linear regression
• Group comparison → t-test / Mann-Whitney U

1. Subgroup & sensitivity analyses: By sex, age group, disease severity, etc.
2. Generate figures: Box plots, KM curves, forest plots, ROC curves
3. Link to manuscript: Map analysis output to Results section paragraphs

Analysis output directory: All results go to data/analysis/. Figures for the manuscript go to figures/.

Required Python packages: Install the utility-script dependencies from the skill root:

pip install -r ~/.claude/skills/paper-writer/requirements.txt

Statistical reporting requirements (before writing Results):

• Effect sizes with 95% confidence intervals
• P values to 3 decimal places (P < 0.001 for very small)
• Statistical test names specified
• Software and version documented
• Two-sided tests (unless justified)
• Multiple comparison correction (if >1 primary outcome)
• Missing data handling described

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

Supplementary APIs (broader coverage; fetch via WebFetch / firecrawl_scrape / tavily_search):

• OpenAlex — https://api.openalex.org/works?search=... (filter by year, cited_by_count)
• Europe PMC — https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=...&format=json (full text, preprints)
• Semantic Scholar — https://api.semanticscholar.org/graph/v1/paper/search?query=... (citation graph, influential-citation counts)
• 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:

1. Ask the user for their 3–5 key papers (they usually know them) — use these as snowball seeds for find_related_articles
2. Run the PubMed MCP query; supplement with OpenAlex / Europe PMC / Semantic Scholar for non-PubMed and preprint coverage
3. De-duplicate by DOI; have the user validate the final list for completeness
4. 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:

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:

• Full citation
• Study design and quality assessment
• Key results with exact numbers
• How it relates to the current paper
• What gap it leaves (that our paper addresses)

Phase 1.5: Screening Execution (Systematic Review only)

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:

1. 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.

1. Stage 1 — De-duplicate (deterministic). Place raw DB exports in

00_literature/screening/00_imported/ (one file per database), then run:

   python ~/.claude/skills/paper-writer/scripts/sr-dedup.py \
     --input 00_literature/screening/00_imported \
     --output 00_literature/screening/01_deduplicated.csv \
     --counts 00_literature/screening/counts/identification.json

1. 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.**

1. Stage 3 — Full-text screening (DUAL). Link PDFs to records, then run two

independent full-text passes:

   python ~/.claude/skills/paper-writer/scripts/sr-pdf-link.py \
     --pdfs 00_literature/screening/full-texts \
     --records 00_literature/screening/02_title_abstract_screen.csv \
     --include-only --rename

Every full-text exclude carries a PRISMA reason category. Write 03_fulltext_screen.csv; the human resolves all conflicts.

1. 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).

1. Produce PRISMA numbers (deterministic).

   python ~/.claude/skills/paper-writer/scripts/sr-prisma-count.py \
     --identification 00_literature/screening/counts/identification.json \
     --ta 00_literature/screening/02_title_abstract_screen.csv \
     --ft 00_literature/screening/03_fulltext_screen.csv \
     --output 00_literature/screening/counts/prisma-summary.md

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)

#### Step 2.5.2: Create Tables

Create table files in tables/ directory:

• table1_baseline.md — Baseline characteristics (standard format)
• 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.

Workflow: Write Methods subsection 1 → Results subsection 1 → Methods subsection 2 → Results subsection 2 → ... This interleaving ensures 1:1 correspondence.

##### 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):

1. General background (everyone agrees with this)
2. Clinical question / knowledge gap (but we don't know X)
3. 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:

1. 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:

1. Patient information (demographics, history)
2. Clinical findings
3. Timeline (consider a timeline figure)
4. Diagnostic assessment
5. Therapeutic intervention
6. Follow-up and outcomes
7. 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:

1. Why this case is significant (clinical lesson)
2. Comparison with published literature
3. Limitations of the case
4. 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:

1. Brief background on the condition
2. Why this case is reportable (rarity, novelty, instructive value)
3. 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:

1. Chronological: Evolution of understanding over time
2. Thematic: Grouped by subtopic (most common)
3. 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:

1. Scope and importance of the topic
2. Why a review is needed now (new evidence, controversy, emerging field)
3. Objectives and scope of this review

##### Step 3.3-RA: Conclusion & Future Directions

Write the conclusion:

1. Synthesize the key themes identified
2. Current state of knowledge
3. Gaps and future research directions
4. 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:

1. Protocol and registration (PROSPERO ID)
2. Eligibility criteria (PICO/PECO)
3. Information sources (databases, dates)
4. Search strategy (full strategy in supplementary)
5. Selection process (screening, inter-rater reliability)
6. Data collection process
7. Data items
8. Risk of bias assessment (tool selection)
9. Effect measures
10. Synthesis methods (narrative and/or meta-analysis)
11. Subgroup and sensitivity analyses
12. Reporting bias assessment
13. Certainty of evidence (GRADE)

##### Step 3.2-SR: Results

Write Results following PRISMA items P-19 through P-23:

1. PRISMA flow diagram (Figure 1 — mandatory)
2. Study characteristics table
3. Risk of bias summary
4. Results of individual studies
5. Results of syntheses (forest plots if meta-analysis)
6. Reporting biases (funnel plots if ≥10 studies)
7. Certainty of evidence (GRADE Summary of Findings table)

##### Step 3.3-SR: Discussion

Write Discussion following PRISMA items P-25 through P-27:

1. Summary of evidence with certainty levels
2. Comparison with previous reviews
3. Strengths and limitations (both evidence and review process)
4. Implications for practice and research

##### Step 3.4-SR: Introduction, Abstract, Title

Follow the same principles as Original Article but with SR-specific framing:

• Introduction: justify why this SR is needed (no existing SR, outdated SR, new evidence)
• Abstract: must include number of studies, total participants, key pooled estimates
• Title: must include "systematic review" (and "meta-analysis" if applicable)

Get user approval on final title.

#### 3-E: Letter / Short Communication Workflow

Letters and short communications follow a condensed IMRAD format. The key constraint is the word limit (typically 600-1500 words).

##### Step 3.1-LT: Condensed Draft

Write a single file covering all sections:

1. Introduction (1-2 sentences): State the purpose directly. No lengthy background.
2. Methods (1 paragraph): Essential details only. Reference a fuller description elsewhere if needed.
3. Results (1-2 paragraphs): Key findings only. Usually 1 table OR 1 figure (not both).
4. Discussion (1-2 paragraphs): Main interpretation, 1-2 comparisons with literature, key limitation.

Rules:

• Every word counts — eliminate all filler
• Typically limited to 1 table + 1 figure, or 2 of one type
• References usually limited to 10-15
• No separate Conclusion section (fold into last Discussion paragraph)

##### Step 3.2-LT: Abstract

Write a brief abstract (often 100-150 words, unstructured).

##### Step 3.3-LT: Title

Short, direct titles work best. No need for elaborate structure.

Get user approval on final title.

Phase 4: Humanize

Read ~/.claude/skills/paper-writer/references/humanizer-academic.md with the Read tool.

After drafting, run a humanization pass on every section to remove AI-generated writing patterns.

#### Step 4.1: Scan for AI Patterns

Read each section file and identify:

English papers — check for these 18 patterns:

1. Significance inflation ("pivotal", "evolving landscape", "underscores")
2. Notability claims ("landmark", "renowned", "groundbreaking")
3. Superficial -ing analyses ("highlighting", "underscoring", "showcasing")
4. Promotional language ("profound impact", "remarkable", "dramatic")
5. Vague attributions ("Studies have shown", "Experts argue")
6. Formulaic challenges ("Despite challenges... future outlook")
7. AI vocabulary ("Additionally", "crucial", "delve", "landscape", "pivotal")
8. Copula avoidance ("serves as" instead of "is")
9. Negative parallelisms ("Not only... but also")
10. Rule of three overuse (forcing ideas into groups of three)
11. Synonym cycling ("Patients... Participants... Subjects")
12. False ranges ("from X to Y" on unrelated scales)
13. Em dash overuse
14. Title Case in headings
15. Curly quotation marks
16. Filler phrases ("In order to", "It is important to note", "comprehensive investigation")
17. Excessive hedging ("may suggest... have the potential to")
18. Generic positive conclusions ("The future looks bright")

Japanese papers (日本語) — 13パターン（A〜C）+ AIボキャブラリー一覧（D）をチェック:

A. 記号と表記（3パターン）:

• emダッシュ、カギ括弧多用、丸括弧補足しすぎ

B. 文のリズム（3パターン）:

• 同じ語尾の連続（である。である。である。）
• 接続詞過多（さらに、また、加えて）
• 段落の終わりが毎回きれいに閉じる

C. 学術文特有の問題（7パターン）:

• C-1 保険が多い（逃げ道の常設）
• C-2 根拠なき評価語（非常に有効、大きなメリット）
• C-3 抽象語だけで押し切る
• C-4 AIボキャブラリー（包括的、革新的、シームレス、示唆に富む）
• C-5 同義語の言い換え連打
• C-6 受動態の過剰使用（検討が行われた → 検討した）
• C-7 非学術的な文体の混入（「参考になれば幸いである」「ポイントは以下の通り」→ 削除）

D. AIボキャブラリー一覧: references/humanizer-academic.md の日本語セクションDを参照。C-4のパターン判定に使う語彙リスト。

#### Step 4.2: Rewrite

Consult references/humanizer-academic.md for specific before/after examples. For each identified pattern:

1. Replace with precise, specific academic language
2. Preserve all data, statistics, and citations exactly
3. Use simple constructions ("is" over "serves as")
4. Remove filler and reduce hedging to match evidence strength
5. Ensure consistent terminology throughout
6. If 3+ AI patterns appear in one sentence, rewrite the entire sentence rather than fixing patterns individually

#### Step 4.3: Section-Specific Focus

English:

日本語:

#### Step 4.4: Verify

After humanization:

English:

• [ ] 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):

1. Collect all cited papers from all sections
2. Format according to target journal style captured in Phase 0 (Vancouver, APA, etc.)
3. Number sequentially as cited
4. Verify completeness: every reference is cited in text, every citation has a reference entry
5. 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.

Verification checklist:

• [ ] Methods ↔ Results correspondence (Original Article only)
• [ ] Introduction objective ↔ Conclusion answer
• [ ] All Tables/Figures mentioned in text
• [ ] 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:

1. Statistical reviewer — p-hacking, multiplicity, power, post-hoc subgroups; does every confirmatory claim trace to the pre-registered plan (templates/preregistration.md)?
2. Methodological reviewer — confounding, selection/information bias, the DAG; is each causal-sounding claim supported by the design?
3. Novelty reviewer — re-attack the novelty claim against references/novelty-check.md; has a larger/better study already shown this?
4. 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.
5. 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:

1. 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)
2. 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.)
3. Acknowledgments — read ~/.claude/skills/paper-writer/templates/acknowledgments.md and draft (non-author contributions, AI tool disclosure, patient acknowledgment)
4. Declarations — read ~/.claude/skills/paper-writer/templates/declarations.md and complete (Ethics, COI using references/coi-detailed.md, Funding, Data Availability, AI Disclosure, CRediT)
5. Cover letter using the template
6. checklists/submission-ready.md using the template — fill in journal-specific limits from Phase 0
7. Compile all sections into a single reading-order Markdown file → submissions/v1_{journal}/compiled-manuscript.md
8. Create submissions/v1_{journal}/submission-log.md with submission date, portal, manuscript ID
9. Log the submission in log/timeline.md

Final compilation order (reading order):

For Original Article:

Title → Abstract → Introduction → Methods → Results → Discussion → Conclusion → References

For Case Report:

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:

1. Parse the decision letter and reviewer comments
2. Number each comment sequentially (R1-1, R1-2, R2-1, R2-2, etc.)
3. Categorize each comment:

• Must fix: Factual errors, missing data, methodological concerns
• Should fix: Reasonable suggestions that improve the paper
• Consider: Optional suggestions, stylistic preferences
• 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

1. Track which sections need modification based on reviewer comments
2. Make changes in the relevant section files
3. Mark changed text (many journals require highlighted changes or a diff)
4. Roll back affected phases: re-run Humanize (Phase 4) and Quality Review (Phase 6) on modified sections
5. 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

#### Step 10.2: Quick Reformat

Use references/journal-reformatting.md checklist:

1. Change reference format (use reference manager)
2. Restructure abstract with new headings — see references/abstract-formats.md
3. Adjust word count — see references/word-count-limits.md
4. Add/remove special sections (Key Points, Highlights)
5. Reformat title page
6. Write new cover letter (address new editor by name)
7. Verify no mention of previous journal name in manuscript

#### Step 10.3: Cascading Submission Strategy

Track submissions:

Plan cascade: Reach journal → Target journal → Safety journal → Backup journal.

Section-Specific AI Guidelines

What AI Should Do

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:

Case Report:

Review Article:

Systematic Review:

Letter / Short Communication:

Resuming a Project

When the user invokes this skill on an existing project directory:

1. Read `README.md` to understand current status, paper type, target journal, and research question
2. 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

1. Present a summary to the user: "Here is where we left off: [status]. The next step is [phase]. Shall I continue?"
2. 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
3. Resume from the next incomplete phase
4. Update status tracker

Handling Mid-Project Changes

Changing target journal: If the user wants to change the target journal:

1. Update README.md Paper Info and Journal Requirements
2. Re-check: citation style, word limits, abstract format, reporting guideline
3. Reformat references if citation style changed
4. Check word counts against new limits
5. Update cover letter

Adding data or revisions: If the user has new data or reviewer feedback:

1. Identify which sections are affected
2. Roll back affected phases to "In Progress"
3. Re-run from that phase forward (including Humanize and Quality Review)

Language Support

English Papers

• Use standard academic English
• Follow target journal's style guide
• Flag awkward phrasing for user review

Japanese Papers

• IMRAD形式は英語論文と同じ（Case Reportは例外: CARE形式）
• 「です・ます」ではなく「である」調
• 専門用語は原則として日本語（初出時に英語併記）
• 症例報告では「症例提示」「臨床経過」等の標準的な見出しを使用
• 論文の書き方ガイドが別途ある場合は参照のこと

Team Mode（チームモード）

ユーザーが「チームで」「team mode」「並列で」と指示した場合、各フェーズを並列エージェントで実行し、大幅に高速化する。

チーム構成

Phase別チームワークフロー

#### Phase 0, 2: 逐次実行（変更なし） ユーザーとの対話が必要なため、既存フローのまま実行する。

#### Phase 1: 文献検索（並列 x3）

paper-lit-searcher を3つ並列でAgent toolから起動する：

• Agent A: PubMed検索（MeSH用語使用）
• Agent B: Google Scholar検索（フリーテキスト）
• Agent C: ユーザー提供の重要論文レビュー + ドメイン固有DB（CiNii, EMBASE等）

3エージェント完了後、リードが結果をマージし 00_literature/literature-matrix.md を作成（重複除去）。

#### Phase 2.5: 表・図（並列 x2）

paper-table-figure-planner を2つ並列で起動：

• Agent A: 表の設計（tables/ に出力）
• Agent B: 図の設計（figures/ に出力）

#### Phase 3: ドラフティング（グループ並列）

paper-section-drafter を依存関係に基づいてラウンド実行する。

Original Article の場合:

• Round 1: Methods + Results（ペアリング執筆、1エージェント）
• Round 2: Introduction P3 + Conclusion（並列 x2、ミラー関係）
• Round 3: Discussion + Introduction P1-P2 + Abstract（並列 x3）
• Round 4: Title（1エージェント）

Case Report の場合:

• Round 1: Case Presentation（逐次、ユーザーの臨床情報必要）
• Round 2: Discussion + Introduction（並列 x2）
• Round 3: Abstract + Title（並列 x2）

Systematic Review の場合:

• Round 1: Methods（逐次、最重要セクション）
• Round 2: Results（逐次、Methods構造に依存）
• Round 3: Discussion + Introduction + Abstr