paper-writer-skill

Paper Writer Skill

Not a manuscript factory. A research engine.

A Claude Code / agent skill for medical and scientific manuscripts: Discovery gates, stage-gated IMRAD drafting, EN+JP humanizing, adversarial review, and submission/revision loops. Humans stay sovereign on the 💡 IDEA and the 📊 DATA.

English | 日本語

Live intro: https://kgraph57.github.io/paper-writer-skill/

Why This Gets Starred

• Discovery before drafting. Phase −1 forges and ranks research questions, checks novelty against live literature, and designs/powers the study — then stops until a human locks the 💡 IDEA and confirms 📊 DATA (real, IRB-aware, never model-invented).
• Stage-gates that auto-fix. Eight quality gates. FAIL → structured feedback → fixer agent → re-check (up to 3 loops). No silent proceed to the next phase.
• Adversarial review can KILL. Before submission, the skill red-teams its own central claim. Fatal problems return to Discovery instead of polishing a bad paper.
• EN + JP humanizer. Phase 4 removes academic AI-tell patterns (18 English + 13 Japanese) with before/after examples and section priorities — see references/humanizer-academic.md.
• Six paper types, 20+ guidelines. Original, Case Report, Review, Systematic Review, Letter/Short, Study Protocol — CARE / CONSORT / STROBE / PRISMA / SPIRIT and more wired into templates.
• Team mode. Seven parallel specialist agents for literature, drafting, humanize, and review (v3.0+).
• Tested utilities. table1.py, SR helpers, compile/word-count scripts — CI runs unittest + shell syntax checks.

60-Second Start

# 1. Install as an agent skill (Claude Code, Cursor, Codex, and other skills hosts)
npx skills add kgraph57/paper-writer-skill

Or clone into Claude Code’s skills directory:

git clone https://github.com/kgraph57/paper-writer-skill.git ~/.claude/skills/paper-writer

Then ask in plain language:

Use the paper-writer skill to start a CARE case report from these de-identified notes…

論文を書く。症例報告。CARE。データは匿名化済み。

Triggers: write paper / start manuscript / research paper / 論文を書く / 論文執筆 / 原稿作成 — or /paper-writer in Claude Code.

The Pipeline

graph LR
    P0["−1. Discovery\n(question · novelty · design · pre-reg)"] --> P1[1. Literature Search]
    P1 --> P2[2. Outline]
    P2 --> P25[2.5 Tables/Figures]
    P25 --> P3[3. Draft]
    P3 --> P4[4. Humanize]
    P4 --> P5[5. References]
    P5 --> P6[6. Quality Review]
    P6 --> P65["6.5 Adversarial Review"]
    P65 --> P7[7. Pre-Submission]
    P65 -.->|KILL| P0
    P7 --> P8["8. Revision"]
    P8 --> P9["9. Post-Acceptance"]
    P7 -.-> P10["10. Rejection → Resubmit"]
    P10 -.-> P1

Optional Python packages for analysis/PDF utilities: python -m pip install -r requirements.txt. Literature work uses WebSearch/WebFetch and public literature APIs — not a zero-network skill (by design).

Case Study (2-minute read)

Discovery → draft → humanize → adversarial catch — an illustrative CARE spark (PHI-free) showing why gates beat “just write the abstract.”

Supported Paper Types

Architecture (deep dive)

Autonomous Stage-Gate System (v3.1)

Every phase is guarded by a quality gate. If the gate returns FAIL, the system generates structured feedback, dispatches a fixer agent in revision_mode, and re-checks — up to 3 iterations before escalating to the user.

8 Quality Gates

Literature (≥10 papers, valid DOIs) → Outline (IMRAD + citations mapped) → Tables/Figures → Section draft score → Humanize (high-priority AI patterns = 0) → References (no fabrication / orphans) → Cross-section consistency → Submission package.

Team Mode: 7 Parallel Agents (v3.0)

Repository Map

Full file tree and phase tables remain in SKILL.md and the templates/references trees (37 templates · 30 reference docs · 8 scripts).

Language Support

Requirements

• Claude Code CLI or another agent that loads SKILL.md
• WebSearch / WebFetch (literature)
• Python 3 for optional utility scripts
• python -m pip install -r requirements.txt for analysis/PDF helpers

Development

python -m py_compile scripts/*.py
bash -n scripts/*.sh
python -m unittest discover -s tests -v

License

MIT — Copyright (c) 2026 KEN.

Versions

• v3.2.0 — Research project folder management
• v3.1.0 — Autonomous Stage-Gate System
• v3.0.0 — Team Mode (7 parallel agents)

See CHANGELOG.md for details.

If you want manuscripts that survive review — and refuse to lie about data — ★ Star the repo and install it on your agent.