check-review-alignment_skill

Overview

This skill checks and minimally fixes citation alignment in an existing {topic}_review.tex file. It inspects each \cite entry against provided BibTeX and available PDF excerpts, produces a verifiable alignment report, and can render the corrected review to PDF/Word using the systematic-literature-review renderer when requested. The tool follows a strict "do not change for the sake of changing" policy: only fatal mismatches are edited, and all other issues are reported.

How this skill works

The workflow extracts structured inputs: each citation, its BibTeX metadata, and PDF-derived summary snippets when available. A semantic check classifies issues as P0 (fatal), P1 (warning) or P2 (skip). Only P0 items trigger minimal in-place edits limited to sentences that contain the citation; LaTeX commands and bibkeys are preserved unless a bibkey fix is required. It outputs a single alignment report plus the modified tex and optionally renders PDF/Word via the systematic-literature-review rendering skill.

When to use it

• You already have {topic}_review.tex and its .bib and need to verify that each citation truly matches the cited paper.
• You want an auditable alignment report showing which citations were checked, evidence used, and any edits made.
• You need minimal, constrained fixes for fatal citation mismatches while preserving LaTeX integrity.
• You want automatic generation of PDF and Word after verification and fixes (rendering requires the renderer dependency).
• You need a structured JSON summarizing citations and evidence for external AI review or reproducibility.

Best practices

• Run with --prepare first to generate ai_alignment_input.json and inspect extracted evidence before any edits.
• If rendering is needed, ensure the systematic-literature-review renderer is installed; the skill will abort with a clear message if missing.
• Prefer supplying PDFs in the work_dir so the analyzer can use high-priority PDF excerpts as evidence.
• Do not expect broad stylistic rewrites; only fatal semantic mismatches (P0) are changed and only at sentence level.
• Review the generated ai_alignment_report.md for suggested non-fatal changes (P1) and confirm before committing major wording changes.

Example use cases

• Validate that every \cite in a literature review actually corresponds to the cited paper before submission.
• Audit a collaborator's review draft to find and correct swapped or incorrect bibkeys with minimal edits.
• Produce a verifiable report showing citation alignment for peer review or reproducibility checks.
• Integrate the structured JSON output into an external AI pipeline for deeper semantic validation.
• Quickly re-render the corrected review to PDF/Word after alignment fixes using the renderer dependency.