Paper Review

Paper Review

Overview

This skill runs a terminal, cross-cutting audit before submission. It is not a deep prose critic (that is paper-writing), not a statistical robustness reviewer (that is robustness-checks), not a reproducibility verifier (that is replication-driven-research). Instead, it detects the seams between those skills — numerical inconsistency across the manuscript, narrative incoherence between sections, missing literature dialogue, LaTeX overflow, script errors in data pipelines, stylistic tells of AI writing — in one pass, and writes a single consolidated report with severity-ranked findings and a go/no-go verdict.

The skill is standalone. It works on any paper folder, whether produced by the superpapers pipeline or not. When the canonical layout is absent, it asks the user for the paths it needs.

When to Use

• The paper is drafted, tables and figures are in place, and the user wants a final check before submission.
• An external collaborator delivered a manuscript and you need a structured audit.
• The execute-plan skill has finished and suggested this step.
• After a major revision cycle, before re-submitting.
• The user asks for "a full paper review", "pre-submission audit", or "audit my paper".

Do not use this skill:

• Before the paper is substantially drafted — without text there is nothing to audit. Use paper-writing instead.
• For deep prose rewriting. That is paper-writing.
• To design or run additional robustness checks. That is robustness-checks.
• To verify end-to-end reproducibility execution. That is replication-driven-research.

Mandatory Steps

1. Discover the paper structure. Locate: paper.tex (or the equivalent main TeX file), references.bib, output/tables/, output/figures/, output/logs/, the code directory, run_all.sh. Walk up from the working directory. If any essential path is missing, ask the user via AskUserQuestion with concrete options (include "Other" for custom paths). Record the paths for reuse. Do not assume CLAUDE.superpapers.md or docs/superpapers/plans/ exist — the skill must work on externally authored papers.

2. Inventory artifacts. List every file relevant to the audit by category: prose files (.tex), bibliography (.bib), table files (output/tables/*.tex), figure files (output/figures/*.pdf), code (.R, .py, .do, .jl, etc.), logs, and the data manifest. Store the inventory for reuse across subsequent steps.

3. Numerical consistency. Extract numbers from the paper prose — percentages, point estimates, standard errors, sample sizes, R², test statistics, p-values. Cross-check each against its source: the output/tables/*.tex file it came from, or the log in output/logs/ that produced it. Any discrepancy beyond a reasonable rounding tolerance (default: last reported digit of the paper figure) is a Critical finding. Report both paper location (file:line) and source location.

4. Narrative coherence and structure.

   • Coherence: abstract ↔ introduction ↔ results ↔ conclusion must agree on research question, main result with magnitude, identification strategy, and key caveats. Divergence — abstract says "12% effect", conclusion says "around 10%" — is a Major finding.
   • Literature dialogue: Results and Discussion sections must cite prior literature, comparing the paper's findings against cited papers rather than merely reporting numbers in isolation. A Discussion section with zero \cite{} is a Major finding. Fewer than one citation per page of prose in Discussion or Results is a Major gap.
   • Structural flow: flag any \section, \subsection, or \subsubsection whose body is under 80 words. Flag pages where the density of subheadings suggests over-fragmentation (roughly ≥3 subheadings in one page's worth of text). Over-fragmentation is Minor with a suggested merge; severely choppy writing — multiple short sections in a row — escalates to Major.

5. Citation integrity. Every \cite{key} must have a matching entry in .bib. Every .bib entry must be cited at least once. No duplicate keys. For malformed entries or missing required fields, delegate to citation-management.

6. Code and reproducibility hygiene. Scan scripts in the code directory for:

   • Hard-coded paths: /home/..., C:\\..., absolute URLs to local disks.