before-submit
中文说明
A Claude Code skill that audits your academic paper before you submit it.
It catches the things that get papers desk-rejected or embarrassed in review:
fake / hallucinated / retracted citations, broken LaTeX, double-blind identity
leaks, and violations of the target venue's rules (page limits, mandatory
sections, required checklists). It produces a single, ordered report — and an
optional clean HTML view — and it never edits your files without asking.
Built for the LLM-agent era: AI writing tools invent plausible-but-nonexistent
references and leave behind conversational artifacts. before-submit verifies
every reference against real databases (and, for anything it can't index, the
agent's own web search), and sweeps the source for the rest.
What it checks
Bibliography
- ✅ Existence & metadata — every entry verified against CrossRef, Semantic
Scholar, OpenAlex, DBLP, and arXiv in parallel, with cross-source
corroboration. Anything unverifiable is escalated to the agent's web search,
so genuinely-fake references get flagged loudly.
- 🚫 Retractions — DOIs checked for retracted / withdrawn / concern status.
- 📄 Cite the published version, not arXiv — preprint entries that already
have a peer-reviewed version (conference/journal) are flagged, web-confirmed,
and you're pointed at the version of record to cite instead.
- 🔁 Duplicates, unused entries, missing
\cite keys, required
fields per entry type, preprint ratio, optional dead-URL checks.
- 🎯 Citation relevance (optional) — does the cited work actually support the
claim at that spot in the text? Judged by the agent itself (no extra API key).
Every reference the verifier script flags is re-confirmed by the agent with a
live web search before it reaches the report — the script triages, the agent
decides.
LaTeX quality
- Caption placement (venue-aware), cross-references, citation formatting,
equations, AI-text artifacts, grammar & mechanics, weak/hedging
writing, terminology consistency, acronym hygiene (defined-before-use,
not redefined or defined inconsistently), reference-noun consistency
("Figure 3" vs "fig. 3",
Fig./Sec./Eq. style), number formatting and
en-dash ranges, straight-vs-curly quotes, encoding/mojibake.
- The writing checks are done by the agent reading the assembled prose like a
reviewer (no hardcoded linter) — grep only locates the mechanical ones; the
semantic calls (grammar, terminology, acronym sense, citation fit) are judged.
- Anonymization for double-blind: author leaks, identifying URLs,
acknowledgments, self-revealing phrasing — checked only for the review version.
Internal faithfulness (the paper checked against itself)
- 🔢 Numbers match the tables — numeric claims in the prose/abstract are
cross-checked against the actual
tabular cells (with rounding/subset/metric
tolerances, and the surrounding LaTeX — headers, row labels, shared macros — read
to tell a real mismatch from an apparent one, so it doesn't cry wolf); a wrong
headline number is escalated.
- 📊 Tables are self-consistent — the bolded "best" really is the best,
totals add up, ablation deltas are right.
- 🖼️ Figures match the prose — it reads the figure files (PNG/JPG/PDF)
and compares the plotted trend/axes/legend to what the text claims about them;
unreadable plots (e.g. EPS) are noted, never guessed.
- 🔗 Cross-section consistency — the same metric reported the same value
everywhere, and claims point at the right table/figure/appendix.
Venue compliance (ACL · EMNLP · NAACL · CVPR · ICCV · ECCV · NeurIPS · ICML · ICLR)
- Page limits, mandatory sections (e.g. ACL Limitations), required
deliverables (e.g. NeurIPS Paper Checklist, ICML Impact Statement /
Type-1 fonts), style file, paper size, caption convention.
- Live-first: fetches the venue's current Call for Papers when online, and
falls back to a bundled snapshot otherwise — always telling you which it used.
Optional compile (pdflatex only) — if a TeX Live pdflatex toolchain is
available (or you let it install the lightweight, Overleaf-identical TinyTeX),
it compiles and reads the logs for authoritative undefined references/citations,
multiply-defined labels, overfull boxes, and the real page count vs the venue
limit. If there's no pdflatex and you decline the install, the compile and
page-count checks are skipped and clearly noted — it will not substitute a
different engine (e.g. Tectonic/XeTeX), since that would give a misleading page
count vs Overleaf.
