Academic Writing Agents

A Claude Code plugin that brings multi-agent review to academic writing — the same "multiple expert reviewers" model that makes peer review work, built into your writing process.

Why

The gap between doing great research and writing about it clearly costs real impact. Findings get buried in unclear prose, reviewers stumble over notation inconsistencies, and figures fail to carry their weight. Most researchers get feedback from one or two people, often late in the process.

This plugin gives you a team of 12 specialist agents — each focused on a different dimension of writing quality — working in parallel, available at any point in the writing process.

Design Philosophy

Informed by Michael Black's "Writing a Good Scientific Paper" and real thesis/paper writing experience:

• Parallel expert review. Multiple specialized agents review simultaneously, each catching what others miss — like having a writing coach, a technical reviewer, a figure specialist, and a bibliography checker all reading at once.
• Review-then-act. Always diagnose before fixing. The writing-reviewer identifies issues; the prose-polisher implements fixes. Never blindly edit.
• 30 codified principles. Not ad-hoc corrections but a systematic framework organized into 6 categories: Structure & Narrative, Prose & Style, Math & Equations, Figures & Tables, Citations & Bibliography, and Process & Meta.
• Human-in-the-loop. Agents propose; humans decide. The plugin presents a prioritized plan and the author chooses what to act on.
• Goal-Problem-Solution rhythm. From Black's guide — every section should follow the GPS pattern. The plugin itself is structured this way: identify what you want to improve (goal), diagnose what's wrong (problem), deploy the right agents to fix it (solution).

How It Works

1. Auto-triggers when Claude detects academic writing context (.tex files, thesis chapters, paper drafts).
2. The academic skill reads project conventions from your .claude/CLAUDE.md.
3. Deploys the right specialist agents in parallel.
4. Synthesizes findings into a prioritized report (Critical / Important / Minor).
5. Offers to fix issues via action agents or direct edits.
6. Iterates with you until you're satisfied.

You can also invoke it manually: /academic review my introduction.

What's Included

Skill

Agents (12 total)

Review Agents (read-only analysis):

Audit Agents (read + verify):

Research Agents (read + web):

Survey Agents (read + web + write):

Action Agents (read + write):

30 Principles

Organized into 6 categories:

A. Structure & Narrative — A1 Recursive Consistency, A2 Logical Chaining, A3 Definition Order, A4 Paragraph Closers, A5 Claim-First, A6 GPS Rhythm, A7 The Nugget

B. Prose & Style — B1 Enumerations, B2 Negation-Contrast, B3 Colloquial Terms, B4 Thesis Voice, B5 One Idea/Sentence, B6 Calibrated Confidence, B7 Ruthless Conciseness, B8 AI-Tell Detection

C. Math & Equations — C1 Math for Clarity, C2 Triple Explanation, C3 Equation-Code Correspondence

D. Figures & Tables — D1 Active Figures, D2 Cross-Reference Floats, D3 Figure-Text-Caption Consistency, D4 One Message, D5 Interpret Figures, D6 Row Alignment, D7 Caption Self-Sufficiency

E. Citations & Bibliography — E1 Cite Named Entities, E2 Citation Completeness, E3 Bibliography Hygiene

F. Process & Meta — F1 Strategic Limitations, F2 Negation-Contrast Audit

Usage

The plugin auto-triggers when Claude detects academic writing context — .tex files, thesis chapters, paper drafts, bibliography work. Just describe what you need in natural language:

Review chapter 3 for consistency and technical correctness

Draft a transition paragraph from the method section to experiments

Polish the abstract in parts/abstract.tex

Audit the bibliography for missing fields and arXiv updates

How does our approach compare to recent test-time adaptation papers?

Survey recent work on test-time adaptation from top venues 2023–2025

Plan the structure for a related work section on domain generalization

Prepare this paper for NeurIPS submission — full review pipeline

You can also invoke the skill explicitly with /academic <task> if auto-triggering doesn't activate.

Installation

# Install from the marketplace:
claude plugin install andrehuang-academic-writing-agents

# Or install directly from GitHub:
claude plugin install --url https://github.com/andrehuang/academic-writing-agents

Or test locally during development:

claude --plugin-dir /path/to/academic-writing-agents

Review Persistence & Coordination (NEW in v2.1)

The orchestrator now persists review findings and supports incremental review:

• Review results saved to .review/: After synthesis, aggregated findings (Critical/Important/Minor) are written to .review/YYYY-MM-DD-<scope>.md in your project. This means review results survive across sessions.
• Incremental review: Before deploying reviewers, the orchestrator checks if unchanged files were already reviewed. If nothing changed since the last review, it presents the prior findings and asks if you want a fresh review or want to proceed to fixes.
• Cross-agent findings sharing: When action agents (prose-polisher, section-drafter) are deployed to fix issues, they receive the specific reviewer concerns relevant to their scope. This ensures fixes address flagged issues rather than applying generic improvements.

These features work with Claude-Claw for session continuity and task management.

Customization

Project-level agents

Add domain-specific agents to your project's .claude/agents/ directory. The plugin auto-discovers them and includes them in deployment plans.

Project-level conventions

Add a .claude/CLAUDE.md to your project with:

• Document structure (chapters, files, build commands)
• LaTeX conventions (macros, figure paths, citation style)
• Author writing style profile (voice, tone, patterns to match)

All agents read this before starting, so they adapt to your project's specific conventions.

Extending the agent roster

Add new agents to ~/.claude/agents/ (global) or <project>/.claude/agents/ (project-level). Each agent is a Markdown file with YAML frontmatter specifying name, description, tools, and model. The plugin picks them up automatically.

Acknowledgments

The principles and workflows in this plugin draw on:

• Michael Black's "Writing a Good Scientific Paper" — the GPS rhythm, the nugget concept, and the emphasis on figures as first-class citizens
• Real thesis supervision experience — patterns observed across multiple rounds of advisor feedback
• Peer review experience — what reviewers actually flag and what they miss

License

MIT