Journal Club Review

Journal-Club Review

Turn any paper (arXiv ID/URL, PDF, text, or a local LaTeX source) into a journal-club presentation: a warm, accurate, discussion-oriented walkthrough in nine sections, with LaTeX math, the paper's real figures (when source is available), and two optional infographic figures. This is a teaching/discussion artifact, not a referee report: no scores, no accept/reject.

When to use

• "Give me a journal-club review of 2401.00001"
• "Review this PDF like a journal club" / "make presentation notes for this paper"
• "이 논문 저널클럽 리뷰 만들어줘" / "발표자료처럼 정리해줘"
• Any text or markdown draft the user wants walked through for a reading group.

If the user wants a peer-review referee report (rating, confidence, weaknesses for OpenReview) use workshop-paper-review. For an adversarial pre-submission audit of their own draft use adversarial-review.

Inputs

One of:

• arXiv id (2401.00001, 2401.00001v2, hep-ph/0101001) or arXiv URL
• a local PDF path
• a local .md / .txt path, or pasted text
• a local LaTeX source: a .tex file or a project directory (e.g. an Overleaf checkout). Figures referenced by the source are harvested automatically.

Workflow

1. Ingest the source

Run the extractor to get a uniform working directory with source.md:

uv run scripts/extract_text.py "<arxiv-id | url | path>"

(scripts/extract_text.py is relative to this skill's base directory; pass an absolute path if your cwd is elsewhere.)

It prints a JSON summary (slug, title, authors, categories, out_dir, source_md, n_chars, plus n_figures, figures_dir, figures_manifest) and writes <out_dir>/source.md (default ./reviews/<slug>/). For pasted text, save it to a .md file first, then pass that path.

For arXiv ids and local PDF paths it also keeps the paper itself at <out_dir>/source.pdf. That file is part of the deliverable: source.md is a lossy text dump with the LaTeX gone, so any verbatim quote you plan to reuse should be checked against source.pdf, and step 6 archives it alongside the review. Text and LaTeX inputs have no PDF to keep, which is fine.

When LaTeX source is available (arXiv e-print tarball, or a local .tex/dir input), the extractor also converts the paper's figures to PNG under <out_dir>/figures/paper/ and writes <out_dir>/figures_manifest.json (n_figures > 0). For PDF-only or plain-text inputs there are no source figures (n_figures is 0); that is fine, just skip step 3.

Read source.md. If extraction yielded little text (scanned PDF, n_chars small), tell the user and proceed with whatever is available (abstract-level).

2. Detect language and write the review

Read references/style-and-math.md, references/section-pipeline.md.

• Detect the source language; write the review in that language unless the user asked otherwise (see style-and-math.md).
• Produce all nine sections in order, grounded in source.md with section / equation / figure citations. Render all math as LaTeX ($...$, $$...$$).
• Build a method figure brief inside section 4 and a results figure brief inside section 5 (schema in references/figure-generation.md).
• Follow the output skeleton in section-pipeline.md. Save the draft to <out_dir>/review.md (leave figure image lines out until steps 3-4 confirm which figures exist).

3. Embed the paper's real figures (when source available)

If n_figures > 0, read references/figure-generation.md ("Real source figures") and <out_dir>/figures_manifest.json. Curate the most relevant figures and embed them into the matching sections with their captions (figures/paper/<name>.png): overview/architecture/schematic into How It Works, result plots into Key Results. Verify each embedded PNG is non-empty before referencing it. Do not dump every figure; note any a reader might expect that you skipped.

Skip if n_figures is 0 (PDF/text input).

4. Generate infographics (optional, on by default)

Read references/figure-generation.md ("Generated infographics"). Check codex login status. If logged in, compose the two friendly-whiteboard prompts from the briefs and launch both codex exec jobs in parallel into <out_dir>/figures/. After they finish, embed ![Method overview](figures/method.png) and ![Key results](figures/results.png) for whichever PNGs are non-empty; note any that were skipped. These coexist with the real figures from step 3.

Skip this step if the user passed --no-figures / "no images" / "text only", or if codex is not logged in (then say so and keep the review text-only).

5. Deliver

• Final file: <out_dir>/review.md (with figures alongside in <out_dir>/figures/: real figures in figures/paper/, infographics in figures/).
• Tell the user the path and give a 1-2 line summary.

6. Export PDF and file into the JournalClub archive (default)

Every finished review is exported to PDF and mirrored into the user's ~/Dropbox/JournalClub archive. Do this automatically, do not wait to be asked.

1. Export the PDF with the md2pdf-typora skill on <out_dir>/review.md, producing <out_dir>/review.pdf. (This runs regardless of the review language; the Whitey theme handles Korean and English both.)
2. Pick the topic folder. The archive is organised as ~/Dropbox/JournalClub/<Topic>/reviews/<slug>/. List the existing topics (ls ~/Dropbox/JournalClub) and choose the one that fits the paper. If none fits, ask the user which topic to use or whether to create a new one (e.g. SpectralGeometry, InverseProblem); create it only after they confirm the name. Do not silently invent a topic.
3. Copy the artifacts into ~/Dropbox/JournalClub/<Topic>/reviews/<slug>/: review.md, review.pdf, source.md, source.pdf (the original paper, when the input produced one), figures_manifest.json (when it exists), and the figures/ directory (both figures/paper/ real figures and the generated figures/*.png infographics). Match the existing layout in sibling review folders. The archive is the long-lived copy, so the original PDF belongs there: a text extraction cannot settle a page reference, an equation number, or a figure panel later.
4. Confirm the destination path and file sizes to the user.

Notes

• This skill is self-contained: it does not require the arXiv Explorer app. It reuses that project's journal-club section design and figure style, but Claude itself does the analysis here.
• For arXiv inputs the extractor uses the PDF for text and the e-print tarball for figures. If you have the LaTeX source already, pass the .tex/dir path instead for cleaner math and section structure plus the same figure harvest.
• Real-figure conversion uses whatever rasterizer is on PATH (pdftoppm, magick/convert, or gs); if none is present, vector figures are skipped and only raster (PNG/JPG) figures survive.
• Generated infographics depend on a logged-in bundled codex runtime (ChatGPT OAuth). Without it the review still renders, just text-only.

Files

• scripts/extract_text.py: arXiv/PDF/text/LaTeX -> source.md + JSON metadata, and source.pdf for arXiv/PDF inputs (arXiv is downloaded, a local PDF is copied); harvests real figures to figures/paper/ + figures_manifest.json when LaTeX source is available (PEP 723 inline deps: pdfplumber, httpx, feedparser; uses system pdftoppm/magick/gs for conversion; run with uv run).
• references/section-pipeline.md: the nine sections and output skeleton.
• references/figure-generation.md: real-figure embedding policy, infographic briefs, style block, codex command.
• references/style-and-math.md: language rule, LaTeX math, tone, anti-patterns.