Proofread: A 5-Phase Agent Pipeline for LaTeX

February 26, 2026

2026   ·   AI   agents   latex   proofreading   claude-code     ·   tools  

I built proofread, an automated proofreading pipeline for LaTeX papers and textbooks.

It is not just a single prompt. It is a multi-phase agent workflow that splits documents into sections, proofreads each section, verifies report quality, and then produces an aggregate summary.

What It Is

proofread is a Claude Code based proofreading system for LaTeX source.

The pipeline is designed for long technical writing where errors are distributed across many sections: typos, grammar issues, LaTeX issues, notation inconsistencies, and mathematical mistakes.

What It Does

The system runs in 5 phases:

  1. Split (Phase 0): split each ch<N>.txt chapter into section-level files using \section boundaries.
  2. Context Builder (Phase 1): build a chapter context index (definitions, notation, conventions).
  3. Proofreader (Phase 2): review each section for language, LaTeX, and math issues.
  4. Verification (Phase 3): audit the proofreading report for false positives, missed issues, and format quality.
  5. Verdict + Summary (Phases 4-5): either accept the section report (DONE) or retry (CONTINUE, up to 3 iterations), then generate a global summary report.

This verification loop is the key design choice: the system does not trust the first draft of its own report.

Inputs and Outputs

Input format is simple but strict:

  1. You provide files named ch1.txt, ch2.txt, …
  2. Each chapter file must contain \section{...} or \section*{...} markers.
  3. Content before the first \section is ignored.

Main outputs:

  1. Per-section reports in result/proofread/ch<N>/...
  2. Verification files in result/verification/ch<N>/...
  3. Final aggregate summary in result/summary/proofread_summary.md

Why I Built It

I built proofread because I believe current frontier models are already strong enough to save a large chunk of proofreading labor.

The goal of this pipeline is to automate as much of that repetitive work as possible while keeping final decisions in human hands.

Scope and Limits

This tool is a proofreading and review assistant, not a replacement for author judgment.

I still treat theorem validity, final mathematical correctness, and editorial decisions as human-owned.

Also, because the pipeline uses Claude Code with --dangerously-skip-permissions, it should be run carefully (ideally in a sandboxed environment), and API cost should be monitored for large books.

Repository

If you are writing in LaTeX and want agentic quality control, this project is a practical starting point.