ai-asset-pricing: Skill for Claude Code

Audit Captions Skill

Systematic audit of all \caption{} blocks, tablenotes, and figure notes for cross-caption consistency in language, notation, terminology, number formatting, and structure.

Examples

• /audit-captions -- audit all captions in main.tex
• /audit-captions tables -- audit only table captions
• /audit-captions figures -- audit only figure captions
• /audit-captions appendix -- audit captions in appendix sections only

Input

The user provides an optional scope filter. If omitted, all captions in main.tex are audited.

Workflow

Step 1: Load Context

1. Read .claude/rules/notation-protocol.md for mathematical notation conventions (if it exists)
2. Read .claude/rules/academic-writing.md for terminology rules and banned words
3. Read guidance/paper-context.md for canonical numbers, sample description, and project-specific terminology (if it exists)
4. Read main.tex in full

Step 2: Extract All Captions

For each \caption{...} in main.tex, extract:

• Line number in main.tex
• Environment type: table or figure
• Label: the \label{...} associated with this float
• Caption title: the text inside \caption{...}
• Caption notes: any footnote-size block that follows the caption. Common patterns:
  • Tables: \caption{Title.} followed by \begin{spacing}{1}{\footnotesize ... }\end{spacing}
  • Figures: \caption{Title.} followed by \begin{justify}\begin{spacing}{1}\footnotesize{...}\end{spacing}\end{justify}
  • Or \begin{tablenotes}...\end{tablenotes}
• Panel descriptions: any \textbf{Panel A:} or \multicolumn{...}{...}{\textbf{Panel A:} ...} text inside the float

Build a numbered registry of all captions for cross-comparison. Record the section each caption belongs to (using %% BEGIN/END markers).

Step 3: Language Consistency

Compare the phrasing of recurring concepts across ALL captions. Flag inconsistencies in:

3a. Sample period description

• Extract every mention of the sample period
• Flag: different date formats for the same sample (e.g., "YYYY-YYYY" vs "YYYY-MM to YYYY-MM" vs "Month YYYY to Month YYYY")
• Flag: inconsistent use of observation counts (some captions include $T$=N, others omit it)
• Note: different sample start dates are legitimate when different datasets start at different points. Only flag genuinely inconsistent formatting, not substantively different samples.