Paper Reviewer
Proofread it. Then face Reviewer 2.
Proofreading prompt by Hyungtae Lim
All it takes is a PDF. Mistakes are marked right on the page,
and a reviewer from your venue gives a verdict in English and Korean.
Try it
Review your paper
Paper
PDF and/or LaTeX source zip
Shrink it with Ghostscript and upload the result. This one line installs it and compresses the file, saving -small.pdf next to the original.
Venue
Review criteria follow the venue
Reviewer
How tough should it be?
History
Saved in this browser only · last 20 runs · kept until you resetRun it with your Claude account
For each task, copy the prompt, paste it into Claude, then paste the answer back.
- claude.ai Start a new chat, attach the PDF, and paste the prompt.
- Claude Code Run
claudein the PDF's folder and paste the prompt after@paper.pdf.
Proofreading
- 1
- 2
Open claude.ai
claude - 3
Review
- 1
- 2
Open claude.ai
claude - 3
Run a proofreading pass first, or open one from History. This tab checks whether those issues are fixed in your revision.
Upload the revision
Upload a PDF you compiled, or the Overleaf source zip; the bridge compiles it and the Claude check reads the source too.
Compile a LaTeX folder
The Claude Code bridge compiles it on your machine with latexmk (or pdflatex). Give the folder's absolute path and the main file. The build goes to a temp folder; your project folder is left untouched.
How it works
Quote, don't guess coordinates.
The proofreading prompt
Hyungtae Lim's 02_paper_proofreading.md supplies the persona, severity levels, and category A–I rules verbatim. The original assumes it can read LaTeX source, so checks that only work on source are swapped for their visible symptoms in the PDF, and the output comes back as JSON.
Finding the spot
Coordinates from a model can't be trusted, so every issue comes with a page number and a verbatim quote. The page's text layer is searched for that quote and the match becomes the box, ignoring line-break hyphens and spacing differences. Issues that still can't be found stay in the list as not located.
Review, then translate
The "I am a harsh CVPR reviewer…" prompt plus venue-specific criteria produces an English review, which is then translated into Korean as is. Generating the two separately could give different verdicts. In API mode the PDF sits in the prompt cache, so proofreading and review share it.
Proofreading rules © 2026 Hyungtae Lim, MIT License (license, snapshot bec513d) · PDF rendering pdf.js · Annotation export pdf-lib