LimHyungTae/awesome-claudecode-paper-proofreading

What it solves

Writing a rigorous academic paper or PhD thesis is full of subtle errors—LaTeX infrastructure problems, overclaiming, inconsistent notation, broken cross-references, and leftover review comments. Existing AI tools may jump straight to editing files, making changes you didn’t ask for. This project provides a structured, two-phase prompt workflow for Claude Code and Codex that first detects and lists every issue with a unique number, then waits for your explicit permission before touching any file. It standardizes the checks expected at top robotics and computer-vision venues.

How it works

You use the prompts from inside your paper workspace, not from inside this repository. The two core steps:

1. Workspace audit – reference prompts/01_latex_workspace_review.md (via absolute path or copied into your workspace) and attach your main .tex file. Claude Code or Codex checks the LaTeX infrastructure: preamble config, macro safety, cross-references, citations, figures, hidden TODOs, and academic writing patterns (C1–C9).
2. Content proofreading – reference prompts/02_paper_proofreading.md (or 04_thesis_proofreading.md for PhD theses) plus your root .tex and compiled PDF. The agent reads all \input files and checks language, clarity, structure, figures, notation, and more (categories A–I). The PDF allows cross-checking figure placement and unresolved review annotations.

After the agent lists issues with IDs like [1], [2], you reply with commands like discard 3, 7, 12 or fix all critical before any modification happens. For Codex, you copy the prompts/ directory into your workspace and point Codex at the relevant prompt.

Who it’s for

• Researchers writing papers or theses in LaTeX, especially in robotics and computer vision.
• PhD students preparing a thesis, who need chapter-level checks like architecture, cross-chapter consistency, and abbreviation management.
• Non-native English speakers, since the prompts include patterns common in papers by Korean/Japanese researchers.
• Anyone who wants an AI proofreader that doesn’t edit files without approval and follows top-venue review standards.

Highlights

• Dual compatible with Claude Code and Codex – works with either agent.
• “Detect first, fix later” – the agent never modifies files until the user explicitly confirms which issues to fix.
• Modular checklists – each check is a labeled block (e.g., C1–C9, A–I) that you can reorder or disable.
• PDF cross-check – the compiled PDF lets the agent verify figure placement and spot leftover review comments.
• Thesis-specific extended checklist – adds multi-chapter checks like chapter architecture, cross-chapter consistency, and List of Figures.
• Real-world review patterns – distilled from 100+ paper reviews and editorial work at ICRA, RA-L, CVPR, and other top venues.
• Works in CI/headless environments – just copy the prompts or inline them into AGENTS.md.