text-comparison-validator — subagent

What text-comparison-validator does

It runs as a separate agent inside Claude. You hand it two sources: the text after extraction (OCR output, for instance) and the file to compare against. From there it walks the lines and diffs them.

What lands in the report: missing and extra words, substituted characters, typos. Formatting gets its own pass — bullet markers • vs - vs *, numbering styles 1. vs 1) vs (1), heading level mismatches, indentation, stray or dropped line breaks. Then structure: paragraphs merged that should stay apart, paragraphs split that belong together, reordered sections.

Output order is fixed. A short summary with an overall accuracy percentage first. Then findings grouped by category: content discrepancies, character errors, formatting, structure. Each item quotes the relevant lines from both sources, explains the difference, gives the line number, and suggests a likely cause such as an OCR error or a formatting glitch. It closes with correction recommendations.

Findings are ranked by severity: critical for missing content, major for clusters of typos or broken paragraphs, minor for formatting and single characters. The agent carries only Read and Write, so it reads and writes files and runs nothing. When it cannot tell which version is correct, it does not guess — it shows both and asks. Worth knowing before you point it at a document where the right version is not obvious.