Paper Agent

---

✨ Overview

Most LLM writing tools are general-purpose and produce text that reviewers immediately recognise as AI-generated. paper-agent is purpose-built for hydrology and water-resources peer review: it enforces journal-specific structure, refuses to fabricate citations, and surfaces consistency failures — gap-vs-conclusions mismatches, sample-size drift, terminology inconsistency — that authors and reviewers routinely miss.

paper-agent runs in your local Claude Code session and produces journal-submission-quality output. It is not a writing assistant in the usual sense — it follows strict mode boundaries, refuses to invent values or fabricate citations, and pauses for confirmation after every section.

Use it when you have:

• A workspace of project data (CSVs, source code, metadata) and need a complete first manuscript draft.
• An existing .docx manuscript and need reviewer-style feedback, section-by-section revision suggestions, or a language-level polish.
• Reviewer comments from a journal decision letter and need help producing a point-by-point response and revised text.

🎬 See it in action

Two bundled, synthetic demos let you try the skill in ~60 seconds with no data of your own — see examples/.

Audit mode reads a manuscript and surfaces inconsistencies authors miss. Given this (synthetic) draft:

Abstract: "We analysed six years of monitoring data (January 2018–December 2022)… a network of 33 wells…" Results: "Across all 35 wells, the model reproduced the seasonal cycle (Figure 1)… the spatial pattern is shown in Figure 3… the correlation with the tidal signal was not significant (rho = 0.02, p = 0.94)."

Audit flags, among others:

• 🔴 Temporal error — "six years" but January 2018–December 2022 is five years.
• 🔴 Sample-size drift — 33 wells in the Abstract/Methods, 35 in the Results.
• 🔴 Argument-honesty conflict — the Abstract calls the tidal correlation "strong", but the Results report it as not significant.
• 🟠 Broken cross-reference — the text cites Figure 3, but only Figures 1–2 are captioned.

The full demo (with answer key) and a Draft-mode demo that turns synthetic CSVs into manuscript prose are in examples/.

🎯 Modes

	Mode	Input	Output
✍️	Draft	Project data files (CSVs, source code, metadata)	Complete manuscript sections with inline citations, plus a .docx export
👀	Review	Existing .docx manuscript	Structured reviewer report in chat (no file edits)
🔄	Revise	Existing .docx + optional reviewer comments	Section-by-section BEFORE / AFTER / RATIONALE suggestions in chat (you apply them), plus a response-to-reviewers letter when reviewer comments are provided
🔍	Proofread	Existing .docx manuscript	Revised .docx with language-level fixes only — no scientific changes, no restructuring, no new citations
🧭	Audit	Existing .docx manuscript	Consistency and coherence report in chat with severity-tagged findings (Critical / Major / Minor). Catches gap-vs-conclusions mismatches, sample-size inconsistencies, broken cross-references, terminology drift, argument honesty conflicts

The user selects one mode at session start. Each mode has its own reference file under references/ with the exact protocol, allowed/forbidden operations, and output format.

📦 Installation

paper-agent is a Claude Code skill. Install it as a plugin (recommended) or clone it into your skills directory.

Option A — Install as a plugin (recommended, one command each):

/plugin marketplace add Rekin226/paper-agent
/plugin install paper-agent@paper-agent

This registers the repo as a marketplace and installs the skill. The bundled .mcp.json also wires up the Semantic Scholar MCP server automatically (see Dependencies).

Option B — Clone directly into the skills directory:

git clone https://github.com/Rekin226/paper-agent.git ~/.claude/skills/paper-agent