Write Paper

Write-Paper Skill

8-Phase Pipeline

Phase 0: Init

Gather essential information from the user before any writing begins.

Required inputs:

1. Title (working title is fine)
2. Paper type: original article, AI validation, case report, case series, meta-analysis, technical note, animal study, NHIS cohort, cross-national
3. Target journal: load profile from ${CLAUDE_SKILL_DIR}/references/journal_profiles/
4. Research question / hypothesis
5. Available data: what datasets, tables, analyses already exist

Optional flags:

• --no-llm-disclosure: skip the LLM writing-assistance disclosure. Default is ON.
• --autonomous: run Phases 0–7 without user gates (outline approval, T&F plan, discussion planning, section reviews all skipped). Default OFF.

Actions:

1. Load the journal profile. If none exists, ask for word limits, abstract format, citation style, figure/table limits, and special requirements.

2. Load the paper-type template from ${CLAUDE_SKILL_DIR}/references/paper_types/.

3. Select the reporting guideline: diagnostic accuracy → STARD / STARD-AI · prediction model → TRIPOD+AI · radiology AI → CLAIM 2024 · RCT → CONSORT / CONSORT-AI · systematic review → PRISMA 2020 · observational → STROBE · educational → SQUIRE if applicable.

4. AI/LLM design-stage reporting map (AI validation, LLM/MLLM, NLP extraction, report generation): map every required AI-reporting item to a manuscript section before drafting — model/version/access date, input fields, prompt or fine-tuning protocol, same-backbone zero-shot/few-shot baseline if an adaptation claim is made, test-data independence/contamination, repeatability, and the Methods subsection each will land in. If any item cannot be placed, halt for design clarification rather than burying it as a Phase 7 limitation.

5. Create or confirm the project scaffold directory.

6. Record the --no-llm-disclosure and --autonomous flag states for Phase 1–7 gate logic.

7. Identify a backbone article — scan manuscript/_src/refs.bib first and propose proactively (ranking: Backbone ranking below); ask only as a fallback. Record the chosen citekey in project.yaml::backbone_article. Then gate on its full text — a backbone whose full text is not extracted is a backbone in name only; the draft would follow an abstract:

   python3 ${CLAUDE_SKILL_DIR}/scripts/gate_backbone_fulltext.py \
     --project project.yaml --refs manuscript/_src/refs.bib \
     --fulltext-dir pdfs/ --strict

   A file in --fulltext-dir counts as the backbone when it is named <citekey>.md or names the backbone DOI/citekey before its References heading; a paper that only cites the backbone in its reference list does not count. Known limit: another paper that names the DOI in its own body still resolves — name the file <citekey>.md or pass --fulltext to remove the doubt.

   BACKBONE_FULLTEXT_MISSING / BACKBONE_FULLTEXT_THIN → stop and retrieve it (/lit-sync Phase 2.7, then /fulltext-retrieval pdf_to_md.py). Do not begin Methods drafting until this passes. If the article is genuinely unavailable in full text, record that limitation and get user confirmation before proceeding on the abstract alone.

8. Summarize the setup — journal constraints, paper type, reporting guideline, backbone article, directory path, LLM-disclosure status — and confirm before proceeding.

Phase 0 Gate: Citekey-only references

Citekey discipline turns citation fabrication into a visible placeholder the submission gate can block.

1. Every in-text citation MUST be [@citekey], with citekey present in manuscript/_src/refs.bib. Pandoc/Quarto style only — no "(Smith et al., 2024)" free text.

2. For a citation intended but not yet imported, use [@NEW:short-topic] (kebab-case, ≤30 chars, unique in the manuscript).

3. Never fabricate a citekey that "looks real" ([@Smith_2024_AI]) when the entry is not in refs.bib. [@NEW:...] is the only allowed placeholder.

4. All [@NEW:...] placeholders must be resolved before Phase 7 (/search-lit → /lit-sync imports verified entries; Better BibTeX refreshes refs.bib). Cite only references whose DOI or PMID /search-lit has confirmed; mark any other as [UNVERIFIED - NEEDS MANUAL CHECK].

5. Pre-submission check — must return zero matches before /sync-submission may freeze a package:

   grep -E '\[@NEW:[^]]+\]|\[N\]|\[N–N\]' manuscript/index.qmd

   Bare [N] / [N–N] markers mean a manuscript drafted outside this pipeline (no refs.bib) with method-load-bearing citations left unresolved. Block them exactly like [@NEW:...].

If refs.bib is absent, create it empty with the comment % refs.bib managed by /lit-sync via Zotero Better BibTeX. Do not hand-edit., record reference_manager.required_for: project_owner in SSOT.yaml, and proceed — early citations will all be [@NEW:...] until the first /lit-sync run.

Read on demand — once the paper type is known (step 2), and only the row that matches:

Phase 1: Outline

Build the outline on the paper-type template's section structure. Start with a header — working title, target journal, paper type, total word limit (excluding abstract, references and legends) — then give every section a word budget and a paragraph- or subsection-level plan, with the budgets summing within the journal limits. End with the planned tables, figures and supplemental materials, one line each.

Gate: Present the outline to the user. Do NOT proceed until the user approves or requests changes. Autonomous mode: skip this gate; log the outline to qc/_pipeline_log.md and proceed to Phase 2.

Phase 2: Tables & Figures

Design all tables and figures BEFORE writing prose, so the narrative serves the data.

1. Review available data with the user. Plan Table 1 (demographics / baseline characteristics — always), Table 2+ (primary and secondary outcomes) and supplemental tables; Figure 1 is the study flow diagram (CONSORT/STARD/PRISMA as applicable), followed by performance curves, forest plots, calibration plots, etc.
2. Call /analyze-stats if statistical analysis is needed.
3. Call /make-figures with the full figure set the study type requires — do not ask the user to name each figure. Pass --study-type mapped from the Phase 0 paper type / reporting guideline: diagnostic accuracy → diagnostic-accuracy, prediction model → ai-validation, systematic review → meta-analysis, DTA systematic review → dta-meta-analysis, observational → observational-cohort, RCT → rct, case report → case-report.
4. Visual abstract. If the journal profile's "Visual Abstract" section requires or encourages one, call /make-figures with a visual abstract request: title, Key Points 1 and 3, methodology summary, and the best study figure as the visual element.
5. Figure embedding. Scan analysis/figures/ for every PNG and PDF. For each, insert ![Figure N. Caption](analysis/figures/filename.png){width=80%} at the appropriate place in Results and draft its legend from the figure type and analysis context.
6. Manifest verification (HALT gate). analysis/figures/_figure_manifest.md must exist with at least one figure entry — the Phase 7 DOCX build embeds figures from it, and a missing manifest silently drops every figure. If it is missing or empty: in autonomous mode, HALT with error code MANIFEST_MISSING, log to qc/_pipeline_log.md, and write a recovery note to manuscript/<id>/REPORT.md Tier-3 section ("rerun /make-figures or manually create _figure_manifest.md"); in interactive mode, report the error and ask the user how to proceed.

Gate: Present T&F plan to user. Do NOT proceed until user approves. Autonomous mode: skip this gate; log the T&F plan to qc/_pipeline_log.md and proceed to Phase 3.

Phase 3: Methods

Write Methods first — it is the most objective section and anchors the rest of the paper.

Before writing: Load ${CLAUDE_SKILL_DIR}/references/section_guides/methods.md and skim the matching study type in ${CLAUDE_SKILL_DIR}/references/exemplar_methods/ (diagnostic-accuracy/STARD, AI-validation/TRIPOD+AI·CLAIM, observational-cohort/STROBE, meta-analysis/PRISMA 2020, RCT/CONSORT 2010): what each paragraph must establish, plus the element that type most often omits. Every exemplar used in Phases 3–6 is synthetic, with placeholder specifics — model its structure; it is not prose to copy.

Writing order: (1) Study Design and Setting; (2) Participants / Dataset (inclusion/exclusion, recruitment period); (3) Procedures / Intervention / AI Model description; (4) Outcome Measures (primary and secondary endpoints); (5) Statistical Analysis (${CLAUDE_SKILL_DIR}/references/section_templates/methods_statistical.md); (6) Ethics statement; (7) AI/LLM disclosure (unless --no-llm-disclosure): when the target journal wants it in Methods, add the Methods paragraph from LLM Writing Disclosure.

AI/LLM extraction add-ons (when applicable):

• In Dataset / Inputs, state exactly which text fields the model received and whether clinical history, indication, impression, prior diagnosis, or referral text was masked. If a supplied field can contain the target label, Methods must either exclude it or describe a no-leaky-field sensitivity analysis.
• In AI Model or Statistical Analysis, include a same-backbone zero-shot/few-shot comparator when the claim is that fine-tuning, LoRA, prompt engineering, or a multi-agent wrapper improves performance.

Process: Run the critic-fixer loop (Critic Review Criteria), then present the final Methods to the user.

Phase 4: Results

Write Results aligned to the approved tables and figures. Results = "What did we find?" — nothing more. Every sentence must be a factual statement backed by a number.

Before writing: Load ${CLAUDE_SKILL_DIR}/references/section_guides/results.md (mirror symmetry with Methods, flowchart, missing data, the anti-interpretation self-check) and skim the matching study type in ${CLAUDE_SKILL_DIR}/references/exemplar_results/ (the same five types, each in its Methods sibling's order).

Rules:

• Order: study population (enrollment, exclusions, demographics → Table 1); primary endpoint results (one paragraph per primary outcome); secondary endpoints; subgroup / sensitivity analyses.
• Apply the results.md self-check to every sentence: no "why", no prior literature, no causal language ("caused," "led to," "due to" — use "was associated with"), no interpretive hedges ("suggests," "implies," "consistent with," "as expected"), no evaluative adjective without its number.
• Incremental value must be earned, not asserted. If the paper claims the model/marker adds value beyond / on top of an existing tool (a clinical score, a routine test, a baseline model), Results must report the nested-model comparison — a baseline model from the in-routine-use predictors versus the augmented model — with an incremental metric: ΔC-index / ΔAUC (paired CI, e.g. DeLong), NRI, IDI, or decision-curve net benefit. A standalone discrimination number does not support a "beyond X" claim. If the design did not include the baseline comparator (see /design-study Phase 3), soften the claim to standalone performance rather than implying added value.

Process: critic-fixer loop. Gate: Present final Results to user. Confirm before proceeding to Discussion.

Phase 5: Discussion

Before writing: Load ${CLAUDE_SKILL_DIR}/references/section_guides/discussion.md and skim the matching study type in ${CLAUDE_SKILL_DIR}/references/exemplar_discussion/ (the same five types, each naming the element that type most often omits). For case reports, use ${CLAUDE_SKILL_DIR}/references/exemplar_case_report.md instead (literature-boundary wording, n=1 causal caution, bedside teaching points).

Step 5a: Discussion Planning (interactive)

Ask the user the following questions. Wait for answers before drafting.

Q1. List the 3-5 key findings of this study in order of importance.
Q2. Name 3-5 key prior studies (anchor papers) you want to compare against in the
    Discussion — titles or DOIs.
    - Studies consistent with your results: ?
    - Studies inconsistent with your results: ?
Q3. Are there methodological or population differences that could explain any disagreement?
Q4. State up to 3 limitations of this study.
    (For each, include how it was mitigated and the direction in which it could affect the results.)
Q5. Are there clinical implications you want to emphasize?

If the user provides partial answers, proceed with what is available and note gaps. If the user says "skip", use /search-lit to identify anchor papers from the reference list and proceed with best-effort defaults.

Gate: Do NOT start writing Discussion until user responds (or explicitly skips). Autonomous mode: skip the interactive planning and take the "skip" path.

Step 5b: Discussion Drafting

Write an inverted funnel:

1. Summary (1 paragraph): restate key findings without repeating numbers verbatim, bridging from Results.
2. Anchor-paper comparisons (2-3 paragraphs, one theme or finding each): for each anchor paper, state the prior finding with citation, say whether our result agrees, and explain any discrepancy by methodological or population differences.
3. Clinical implications (1 paragraph): what this means for practice or future research.
4. Limitations (1 paragraph, ordered by severity): for each, (a) what it is, (b) how it was mitigated, (c) direction of residual bias. Do NOT open with "our study has several limitations".
5. Strengths (optional, 1-2 sentences): only for a genuinely novel contribution.
6. Conclusion (1-2 sentences): the single most important finding and its implication, as a citable statement. No "further studies are needed" as final sentence.

Rules:

• Do not introduce data not presented in Results; match language to the evidence level; acknowledge alternative explanations for key findings.
• Endpoint↔conclusion scope. The Clinical-implications and Conclusion sentences must not exceed what the design and endpoint support. A cross-sectional / single-visit / prevalence study cannot license a prognostic or surveillance claim (a rescreen interval, disease progression, predicting future risk) — that requires longitudinal follow-up. A binary surrogate endpoint (present/absent, >0, dichotomized) is risk stratification, not a patient-care directive (defer/withhold/initiate therapy). /self-review §D (check_scope_coherence.py) flags CROSS_SECTIONAL_PROGNOSTIC / SURROGATE_CARE_DIRECTIVE; keep the conclusion verb inside the design's reach.

After the first draft, ask the user: any missing anchor papers or comparisons, any change to the interpretation, and which clinical implications to emphasize or soften. Incorporate the feedback, then run the critic-fixer loop.

Phase 6: Introduction + Abstract

Write these LAST because they frame the paper and depend on knowing what was actually found.

Before writing: Load ${CLAUDE_SKILL_DIR}/references/section_guides/introduction.md and ${CLAUDE_SKILL_DIR}/references/section_guides/title_abstract.md, and skim the structure models ${CLAUDE_SKILL_DIR}/references/exemplar_introduction.md and ${CLAUDE_SKILL_DIR}/references/exemplar_abstract.md. For case reports, use ${CLAUDE_SKILL_DIR}/references/exemplar_case_report.md for the 150-word Introduction / Case Presentation / Conclusion abstract anatomy rather than the IMRAD abstract model.

Introduction (3-4 paragraphs): clinical context establishing importance (prevalence, burden, current practice) → the knowledge gap this study addresses → the study objective, stated precisely, with the hypothesis if applicable. For AI/LLM extraction studies, state the decision-impact path: what clinical or research workflow step changes if the model works, not only that the extracted label is interesting.

Abstract: self-contained; all numbers match the main text and tables; the final sentence is a clinical implication, not "further studies are needed." Lead with the pre-specified primary estimand, not the largest effect — even when a critic or peer-sim pass suggests foregrounding the strongest number. Tightening effect-size language is fine; promoting a secondary, exploratory, or post-hoc estimate to the headline is estimand shopping. The Abstract's primary result must be the registered/protocol primary contrast — the same one Step 7.3b checks. If the primary is null or underpowered, report it as such (see /self-review category C, power-aware null) rather than substituting a more favourable secondary estimate.

Process: critic-fixer loop.

Phase 7: Polish

Final quality pass. Strict sequential execution — each step MUST complete before the next begins. Every HALT stops the pipeline; none is advisory.

7.1 — AI Pattern Scan. Remove AI writing patterns (see AI Pattern Avoidance below), editing manuscript/manuscript.md in place. Then run the deterministic classical-style lint: python3 "${CLAUDE_SKILL_DIR}/../self-review/scripts/check_classical_style.py" --manuscript manuscript/manuscript.md --strict. For an MA / systematic review, or when a senior co-author review is expected, also work the 7-grep checklist in references/section_guides/step7_1_classical_qc.md. Pattern 19–21 body rewrites (§, self-reference, AI-disclosure boilerplate) go to /humanize.

HALT — AI-disclosure meta-applicability. An AI/LLM-use disclosure must itself satisfy the items the manuscript critiques (FLAIR F1.6, TRIPOD-LLM, MI-CLEAR-LLM): version, access channel, date range, responsible party, zero [version]/TODO placeholders. A paper cannot fail a framework item it critiques. (Classical target: title page, not the body.)

7.2 — Reporting Guideline Check. Call /check-reporting. Auto-insert only MISSING items whose fixable_by_ai is true; never invent items needing external facts (IRB / registration numbers). Log every insertion to qc/_pipeline_log.md.

7.3 — Citation Verification. The placeholder gate first — HARD STOP on any hit of grep -nE '\[@NEW:[^]]+\]' manuscript/index.qmd manuscript/manuscript.md, looping back to /search-lit → /lit-sync. Then /verify-refs: parse qc/reference_audit.json and stop the pipeline if submission_safe: false, surfacing every FABRICATED / MISMATCH and any duplicate_findings[].

7.3a / 7.3b / 7.3c — Integrity audits. Run all three between 7.3 and 7.4; each can HALT and route to 7.4a. 7.3a numerical claims (text ↔ Table ↔ extraction CSV + primary-source check; a direction reversal or a p<0.05↔p≥0.05 crossing is a P0 blocker). 7.3b estimand provenance (delegates to /self-review 2.5f; PRIMARY_REASSIGNED, EVALUE_ARITHMETIC, EVALUE_NON_PRIMARY = P0). 7.3c reference adequacy (every named method cited — resolve via /search-lit → /lit-sync → /verify-refs --strict, never fabricate). See phase7_integrity_audits.md.

7.4 — Self-Review + Fix Loop. Call /self-review --json --fix: it reviews, applies fixable_by_ai edits, and re-reviews (≤2 iterations), stopping early on PASS. Log the score, verdict, iteration count, and residual issues. Any surviving severity: "fatal" issue routes to 7.4a — do not proceed to 7.5.

7.4a — Audit Recovery Branch. When a finding is structural — the data, protocol, or analysis script is wrong — polishing yields a clean manuscript on a broken foundation. Inline text fixes are forbidden — recovery means re-extraction, re-analysis, or re-registration. Halt 7.5–7.6, log the branch, invoke the routed skill, re-enter at 7.3. Loop budget: one cycle. Routing: references/section_guides/step7_4a_audit_recovery.md.

7.5 — Generate Deliverables. manuscript/manuscript.md, manuscript/title_page.md, qc/reporting_checklist.md, qc/self_review.md, qc/_pipeline_log.md. Do not hand-number author affiliations — build and verify with scripts/build_title_page_affiliations.py --check title_page.md --strict (a Nature Portfolio / npj technical-check item).

7.6 — DOCX Build. Embed figures from analysis/figures/_figure_manifest.md, then render. Prefer /manage-refs (pandoc + citeproc + journal CSL) for any submission with >5 references, and never hand-type a References list.

7.6a — Cross-Reference QC. After the build, before the final gate:

python3 "${CLAUDE_SKILL_DIR}/../manage-refs/scripts/check_xref.py" \
  --md manuscript/manuscript.md --docx manuscript/manuscript_final.docx \
  --out qc/xref_audit.json --strict

It catches in-text citations resolving to the wrong rendered caption. Any MISSING_DOCX/MISSING_BODY/MISMATCH → submission_safe: false, exit 1, HALT. The body caption is the SSOT — fix the build pipeline, never the reverse.

7.7 — Final Gate. Autonomous: log completion; report word count, figure count, self-review score, reporting-compliance %, FATAL flags. Interactive: present summary, await confirmation.