Paper Writer Skill ★ Star on GitHub

Claude Code · Agent Skill · v3.2

Not a manuscript factory.
A research engine.

From clinical spark to submission-ready IMRAD — with Discovery gates, stage-gated drafting, EN+JP humanizing, and an adversarial review that can KILL a weak claim before any journal sees it. Two things stay human: the 💡 IDEA and the 📊 DATA.

Install in 60 seconds ★ Star on GitHub

Works with Claude Code and other agents that load SKILL.md. Network used for literature — not for telemetry.

agent · paper-writer
You
Write a case report from this spark: unexplained hypoglycemia after GLP-1 start in a non-diabetic adult. Data is de-identified and IRB-cleared.
Paper Writer
Phase −1 Discovery Forging FINER questions… novelty check against live literature… GATE · IDEA — three candidate questions ranked. Human must pick the 💡 IDEA before any prose is drafted. GATE · DATA — confirming 📊 DATA is human-origin, de-identified, and never invented by the model.
You
Lock Q2. CARE case report. Target: BMJ Case Reports. EN.
Paper Writer
Plan locked · CARE checklist armed Project scaffold ready: 00_literature/ · 01_outline.md · sections/ · ethics/ Next: literature matrix → outline approval → tables before prose. No manuscript factory mode — gates stay on.

Illustrative demo — no patient data. Gates are real.

Non-negotiable

Human stays sovereign on two things

Everything else can be drafted, checked, and red-teamed by the agent. These two cannot be delegated away.

💡 IDEA

What is worth asking, what it means, and what is ethical. Discovery forges candidates; the human locks the question.

📊 DATA

Real, IRB-approved, never machine-originated. Raw clinical data stays local and gitignored. The skill never invents numbers.

Lifecycle

From Discovery to revision — with gates

Ten phases. Eight quality gates with auto-fix loops. Adversarial review can send a claim back to Discovery with a KILL.

  1. −1 DiscoveryQuestion forge · novelty · design · pre-registration
  2. 1–2.5 Literature → Outline → Tables/FiguresMatrix first; design exhibits before prose
  3. 3–5 Draft → Humanize → ReferencesIMRAD writing order · EN 18 + JP 13 AI patterns · DOI checks
  4. 6–6.5 Quality + Adversarial reviewCross-section gates, then red-team the central claim
  5. 7–10 Submit · Revise · Accept · Reject/resubmitCover letter, response letters, cascading strategy

Six paper types: Original · Case Report · Review · Systematic Review · Letter · Protocol. 20+ reporting guidelines (CARE, CONSORT, STROBE, PRISMA…).

Phase 4

Kill the AI tell — before reviewers do

Real patterns from references/humanizer-academic.md. English and Japanese. Section-priority checklists.

Before
represents a pivotal challenge in the evolving landscape
Significance inflation
After
is highly prevalent in patients with diabetes
State the concrete claim
Before
HR 0.65, P = 0.002), highlighting the cardioprotective effects
Superficial -ing analysis
After
HR 0.65; 95% CI 0.50–0.85; P = 0.002.
Numbers over vibes
Why this gets starred

Concrete differentiators — not vibes

Discovery before drafting

Phase −1 forges and ranks questions, checks novelty against live literature, and locks a plan — only after the human picks the IDEA.

Stage-gates that auto-fix

Eight quality gates. FAIL → structured feedback → fixer agent → re-check (up to 3 loops). No silent proceed.

Adversarial review

Red-teams the central claim. A KILL path returns to Discovery instead of polishing a bad paper.

EN + JP humanizer

18 English and 13 Japanese academic AI patterns with before/after examples and section priorities.

Six types, 20+ guidelines

Original, Case Report, Review, SR, Letter, Protocol — CARE / CONSORT / STROBE / PRISMA and more wired in.

Team mode + tested scripts

Seven parallel agents for search/draft/humanize/review. Utilities like table1.py and SR helpers with CI tests.

60-second start

Install it. Ask in plain language.

# Claude Code, Cursor, Codex, and other agents that load skills
$ 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

Natural-language triggers

“write paper” · “start manuscript” · 「論文を書く」 · 「論文執筆」
Proof

One public case study — two minutes

Discovery → draft → humanize → adversarial catch

An illustrative CARE case-report spark (PHI-free) walks through question forge, a naive “just write it” abstract vs. the gated path, two humanizer fixes, and an overclaim the adversarial review kills.

Input → Discovery → Weak path → Gated path → Humanize → Adversarial → Proof

Read the case study