arxiv-prep

arxiv-prep — arXiv Submission Preparation

Automate the tedious steps of preparing a LaTeX paper for arXiv upload. Creates a clean copy, removes cruft, verifies compilation, and produces a ready-to-upload tarball.

When to Use

• "prep for arxiv", "make an arxiv package", "I need to upload this to arxiv"
• After the paper is finalized and ready for submission
• Complements presubmit-checks (content quality) — this skill handles packaging

Workflow

Phase 1: Pre-clean (judgment calls)

1. Find the paper

Same search logic as presubmit-checks:

1. User-provided path
2. paper/current/main.tex
3. paper/main.tex
4. main.tex
5. Glob for **/*.tex with \begin{document}

Identify the paper directory (parent of the main .tex file). All subsequent operations are relative to this directory.

2. Optionally run presubmit-checks

Ask the user: "Want me to run the presubmit-checks first?" Skip if they say they already did.

3. Check for supplements

Scan for separate supplement/appendix .tex files (e.g., supp.tex, appendix.tex, si.tex). If found, ask the user whether to:

• Merge into the main file via \appendix (arXiv prefers single-PDF submissions)
• Keep separate (will be included as ancillary files)

4. Review style files

Grep for journal-specific language that should be removed for arXiv:

• "Submitted to", "Under review at", "Accepted by"
• Journal-specific class options (e.g., review, preprint mode toggles)
• Copyright/license statements that conflict with arXiv's license

Flag these and ask before changing.

Phase 2: Automated cleaning

5. Optimize bibliography

Run bib_optimizer to remove unused citations and reorder entries to match their order of appearance in the text:

uvx bib_optimizer bibopt <main.tex> <references.bib> <references_cleaned.bib>

• If the paper has a supplement with its own .bib (e.g., si.tex/supp.tex using si_references.bib), run bibopt separately for each:

  uvx bib_optimizer bibopt <si.tex> <si_references.bib> <si_references_cleaned.bib>
  

• If the supplement shares the main .bib, run bibopt on each .tex file separately against the same .bib, then merge the two outputs (concatenate and deduplicate entries by cite key):

  uvx bib_optimizer bibopt <main.tex> <references.bib> <references_main.bib>
  uvx bib_optimizer bibopt <si.tex> <references.bib> <references_si.bib>
  

  Then merge with bibtool (deduplicates by cite key):

  bibtool -d references_main.bib references_si.bib -o references_cleaned.bib
  

  If bibtool is not available, concatenate both files and manually remove any duplicate @type{key, entries (keep the first occurrence).
• Update the \bibliography{...} command in each .tex file to point to the cleaned file.
• The original .bib is never modified.

If bib_optimizer is not available, skip this step — the .bib will still be handled in step 7.

6. Run arxiv-latex-cleaner

uvx arxiv-latex-cleaner <paper-dir> --resize_images --im_size 500 --compress_pdf

This creates a <paper-dir>_arXiv/ directory with a cleaned copy. The original is untouched. The --compress_pdf flag reduces embedded PDF figure sizes, helping stay under arXiv's 50 MB limit.

If arxiv-latex-cleaner is not available, fall back to manual cleaning:

• Copy the paper directory
• Remove .git/, __pycache__/, .DS_Store
• Remove commented-out text blocks (lines starting with % that aren't TeX directives)
• Remove unused .tex files not referenced by \input or \include

7. Post-cleaner fixes in _arXiv/

Apply these fixes to the cleaned copy:

• 4-pass trick: Add \typeout{get arXiv to do 4 passes} on the line after \end{document} — this ensures arXiv runs enough LaTeX passes to resolve all references
• Ensure .bbl exists: Check if a .bbl file exists. If not, compile with pdflatex + bibtex/biber to generate it. arXiv needs the .bbl, not the .bib
• Ask before deleting .bib: If .bbl exists, ask the user whether to remove .bib files (reduces package size; arXiv uses .bbl directly)
• Clean aux files: Remove .aux, .log, .out, .blg, .fls, .fdb_latexmk, .synctex.gz, .toc, .lof, .lot, .nav, .snm, .vrb. Do not use latexmk -CA here — it also removes .bbl, which arXiv needs
• Remove .git/ if it exists in the copy

Phase 3: Verify & package

8. Test compilation

Run in the _arXiv/ directory:

pdflatex -interaction=nonstopmode main.tex
pdflatex -interaction=nonstopmode main.tex
pdflatex -interaction=nonstopmode main.tex

(Three passes to resolve references.) Report:

• Errors: any lines with ! — these are blockers
• Warnings: undefined references, missing citations, overfull hboxes
• If compilation fails, report the error but continue to the next steps

9. Extract clean metadata

Parse the .tex file to extract metadata for copy-paste into the arXiv submission form:

• Title: from \title{...} — strip LaTeX commands, math mode, line breaks
• Abstract: from \begin{abstract}...\end{abstract} — strip LaTeX commands
• Authors: from \author{...} — extract names, strip affiliations/footnotes
• Comments: suggest standard format. If the paper has supplementary material, use: "Main text: X pages, X figures. Supplementary Information: X pages, X figures". Otherwise: "X pages, X figures"

Present this in a clean, copy-pasteable format.

10. Create tarball

cd <paper-dir>_arXiv && tar -cvf ../arxiv-submission.tar *

• Warn if the tarball exceeds 50 MB (arXiv's limit is ~50 MB for source)
• Report the file count and size
• Note: arXiv also accepts .tar.gz — use gzip if close to the limit

11. Final summary

Present:

• Package location and size
• Clean metadata (title, abstract, authors, comments)
• Compilation status (clean / warnings / errors)
• Remaining manual steps:
  • Upload arxiv-submission.tar at https://arxiv.org/submit
  • Select primary subject area and cross-list categories
  • Paste metadata into the form
  • Set license (usually CC BY 4.0 or arXiv perpetual non-exclusive)
  • Share the submission password with co-authors

Rules

• Never modify the original paper directory — all changes happen in the _arXiv/ copy
• Ask before destructive operations (deleting .bib, merging supplements)
• If any step fails, continue with the remaining steps and report all issues at the end
• Keep the final summary concise and actionable