Document Review

The Core Problem It Solves

Document review is a critical bottleneck in professional workflows like contract analysis, technical report publication, and proposal preparation. Manual review is time-consuming and prone to oversight, especially for subtle but critical issues. A document may contain a mix of problems: superficial format inconsistencies and typos, alongside more insidious logical conflicts (contradictory conclusions), factual inaccuracies (erroneous data citations), and logical disarray (jumping arguments). Relying on a single method—either pure scripting or pure AI—fails to address this spectrum efficiently and comprehensively.

How It Works: Dual-Layer Architecture and Key Steps

This skill employs a Dual-Layer Architecture (L1 Quantitative Layer + L2 Semantic Layer) to systematically tackle the challenges.

Core Capabilities:
- L1 Quantitative Layer: Powered by a Python script. It performs fast, deterministic checks: format compliance, typo detection, structural analysis, and data extraction. This provides a reproducible baseline audit.
- L2 Semantic Layer: Powered by an AI model. It handles deep semantic understanding and reasoning: detecting logical conflicts, fact-checking (with optional web verification), and analyzing logical disarray. This layer addresses complex, context-dependent issues.

Key Execution Steps:
1. Content Extraction: Supports .docx, .pdf, .pptx, .txt, .md with a multi-level fallback strategy to retrieve text.
2. Mode Selection: Users choose review intensity based on needs.
- --mode quick: Quick Scan (L1 checks only)
- --mode standard: Standard Review (L1 full + L2 logic analysis)
- --mode deep: Deep Review (L1 full + L2 full + web-based fact-checking)
3. Layered Review:
- First, the script runs L1 checks, outputting structured results and auxiliary scores (covering dimensions like format compliance and content logic).
- Then, the AI performs L2 analysis based on L1 results, focusing on the four core issues (logical conflict, factual conflict, typo, logical disarray).
4. Report Generation: Synthesizes results from both layers into a report that includes issue locations, severity levels (high/medium/low), and actionable modification suggestions. It supports Word export and visualization via show_widget.

Applicability and Important Caveats

This skill is focused on identifying and diagnosing issues, not making them go away automatically.
- Applicable Scenarios: Ideal for formal scenarios requiring rigorous document quality control, supporting single-file, batch, and incremental (version diff) reviews.
- Key Caveats:
- Review must remain objective and neutral; factual conflicts require cited sources for verification.
- All issues must have precise locations (page, section), and suggestions must be concretely actionable.
- For very large documents (500+ pages), using incremental review or chapter-wise processing is recommended to avoid performance issues.
- Its primary role is analysis; final edits still require user action in an editor (though it can integrate with editing skills to write annotations back into the source file).