paper-audit-skill

                                                  _ _ _
 _ __   __ _ _ __   ___ _ __       __ _ _   _  __| (_) |_
| '_ \ / _` | '_ \ / _ \ '__|____ / _` | | | |/ _` | | __|
| |_) | (_| | |_) |  __/ | |_____| (_| | |_| | (_| | | |_
| .__/ \__,_| .__/ \___|_|        \__,_|\__,_|\__,_|_|\__|
|_|         |_|

Draft papers like a journal editor reviews them.

paper-audit-skill

A Claude Code skill for writing academic papers. Six paper types. Five venues. Tiered voice rules.

Paper-type-aware, venue-aware drafting and revision. The skill asks which paper type you are writing and where you are submitting, then applies the conventions a journal or conference reviewer would expect.

Why This Exists

Generic writing tools produce generic prose. Academic papers are judged by different standards per field and venue: a CHI autoethnography and a NeurIPS methods section have almost nothing in common in voice, structure, or reviewer risk.

This skill encodes those differences. It:

• Identifies the paper type (empirical quantitative, empirical qualitative, autoethnography, theory, systematic review, position).
• Identifies the venue (arxiv, NeurIPS/ICML/ICLR, CHI/CSCW/DIS, journal, thesis).
• Loads the references that apply.
• Applies tiered voice rules (universal hard-rejects and contextual warnings).
• Flags reviewer risks specific to the paper type.
• Points at the right LaTeX template and citation style.

What It Does

You: "I'm writing an empirical HCI paper for CHI. Help me draft the introduction."
  |
  +-- Skill asks clarifying questions if type/venue aren't clear
  +-- Loads references/00-paper-types.md §empirical-qualitative
  +-- Loads references/01-venues.md §CHI
  +-- Loads references/02-section-moves.md §Introduction (CARS model)
  +-- Loads voice rules (03-voice-positive, 04-voice-negative)
  |
  +-- Drafts with:
        - Authorial "we"
        - CARS Introduction (territory → niche → occupy)
        - Four contribution bullets at the end
        - No em dashes, no "In today's world," no "groundbreaking"
        - Hedging where appropriate

The skill does not:

• Generate citations from nothing. You still provide sources.
• Check factual accuracy. The skill structures claims that need citations.
• Run literature searches. It advises on what types of sources to cite.
• Produce figures or run statistics. It advises on specs and reporting.

Installation

One command:

curl -fsSL https://raw.githubusercontent.com/Hiro-Inagawa/paper-audit-skill/main/install.sh | bash

Works on macOS and Linux. On Windows, run in Git Bash or WSL.

Alternative: Manual install

git clone https://github.com/Hiro-Inagawa/paper-audit-skill.git /tmp/paper-audit-skill
cp -r /tmp/paper-audit-skill/skills/paper-audit-skill ~/.claude/skills/
rm -rf /tmp/paper-audit-skill

No dependencies to install. The skill is pure markdown; it does not require Node, Playwright, or any other runtime.

Usage

In Claude Code

The skill auto-activates when you mention an academic paper, manuscript, arxiv, thesis, or related terms. It also activates on .tex and .bib file extensions.

First interaction: the skill asks for paper type and venue. Answer honestly; different combinations load different rules.

Example prompts

"Help me write the introduction to my NeurIPS paper on attention mechanisms."
"Review my Related Work section for a CHI submission."
"I'm writing a systematic review. What sections do I need?"
"My reviewer said n=1 isn't generalizable. How do I respond?"
"Draft the abstract for my arxiv preprint."

Explicit invocation

If the skill does not auto-activate when you want it, reference it directly:

"Using the paper-audit-skill skill, review this paragraph."

Supported Paper Types

Supported Venues

Tiered Voice Rules

The skill enforces two tiers of voice rules so it does not over-enforce across disciplines.

Tier 1 (universal hard-rejects) scanned on every paragraph:

• Em dashes in formal prose
• Contractions in formal prose
• "As an AI," "As a language model"
• Filler openings ("It's worth noting," "In today's world," "Since the dawn of")
• Overclaiming ("groundbreaking," "revolutionary," "very unique")
• "Clearly shows," "obviously" (without hedge)

Tier 2 (contextual warnings) flagged for human review:

• Passive voice outside Methods
• Nominalizations where verb form is clearer
• Hedging where boosters are warranted (and vice versa)
• First-person "I" in a "we" venue

The user can override any Tier 2 flag based on disciplinary convention.

Verification

This skill is auditable. See verify.md in the repository root for 20 ablation test scenarios.

Each test provides a deliberately flawed input and lists the issues the skill should catch, plus false-positive watch items the skill should NOT flag. Run the tests in a fresh Claude Code session to confirm the skill works as documented.

Quick audit

cat verify.md

Read through the tests. Pick three or four. Try them in a fresh Claude Code session. If the skill catches the expected issues and does not over-flag, it passes.

Reference Files

The skill body is in skills/paper-audit-skill/. It contains:

SKILL.md                           # Top-level rules and routing
references/
  00-paper-types.md                # Six types, structure per type
  01-venues.md                     # Five venues, conventions per venue
  02-section-moves.md              # Swales moves per section
  03-voice-positive.md             # What to do: character-subject, old-new, hedging
  04-voice-negative.md             # What to avoid: tiered forbidden phrases
  05-argument-structure.md         # Booth: claim, reason, evidence, warrant, acknowledgment
  06-citation-triangle.md          # Every claim has evidence, evidence has citation
  07-citation-styles.md            # APA, Chicago, IEEE, ACM, Vancouver, natbib, biblatex
  08-latex-conventions.md          # Multi-venue LaTeX, .bbl gotcha
  09-defense-playbooks.md          # Reviewer risks per paper type, preemption
  10-abstract-craft.md             # 150-250 word formula, four moves
  11-revision-discipline.md        # Williams Style six-pass revision
  12-ai-ethics.md                  # Disclosure rules per venue (2024-2026)
assets/
  templates/
    arxiv-generic.tex              # Open-license arxiv scaffold
    journal-generic.tex            # Generic journal scaffold
    POINTERS.md                    # Where to get venue-specific templates
  bibliography-templates.bib       # BibTeX entry templates per source type

Requirements

• Claude Code installed.
• No additional runtime dependencies.
• Target: academic writing in English, Latin-script. Non-Latin-script conventions are out of scope for v1.

Philosophy

Academic writing is a craft. The skill does not replace the writer; it enforces the conventions the writer needs to respect to be taken seriously. Every rule in the skill is derived from:

• Swales & Feak, Academic Writing for Graduate Students.
• Booth, Colomb, & Williams, The Craft of Research.
• Williams & Bizup, Style: Lessons in Clarity and Grace.
• PRISMA 2020 reporting standards.
• Ellis, Adams, & Bochner, autoethnography tradition.
• Yin, Case Study Research.
• APA, Chicago, IEEE, ACM, Vancouver citation standards.
• Current (2024-2026) AI-assisted writing disclosure policies at ACM, ICML, CHI, and major journals.

License

MIT. See LICENSE.

Contributing

Issues and pull requests welcome. Particularly: domain-specific additions (e.g., biomedical reporting standards beyond Vancouver, humanities conventions beyond Chicago), venue-specific templates, and translations of the voice rules for non-English academic contexts.

Related

• web-reader — Read any URL from Claude Code. Same repo pattern; reads websites instead of writing papers.