China Patent Application Multi-Agent Parallel Reviewer

Problem

Drafting issues in Chinese patent application files are often spread across the abstract, claims, description, figure references, and cross-section consistency. Manual line-by-line review is slow and easy to miss; giving one model all rules at once creates heavy context and can degrade output quality.

How It Works

The skill uses a multi-agent batched parallel architecture: the main agent handles orchestration and key decisions, while subagents perform extraction, review, merging, annotation, and compliance checks.

• It accepts an absolute path to a .docx or .doc file and returns a reviewed_docx with tracked changes and comments, plus a work_dir of intermediate files.
• The 14 review rules are split across 13 review agents and N image subagents, launched in 7 batches. Each agent loads only one rule and reads only its assigned section files.
• It first extracts abstract, claims, description, figures, and section metadata, then reviews the abstract, simple and complex claim rules, description, cross-section consistency, figure metadata, and image-level quality.
• Results are merged and deduplicated, image context anchors are repaired, comments are injected with review_adder.py, content integrity is validated, skipped comments are relocated, JSON is formatted, compliance is checked, and logs are archived.
• .doc conversion requires Windows with Word or WPS and pywin32.

Boundaries

It checks drafting quality only; it does not provide patent strategy, patentability analysis, legal advice, or translation. Documents must pass hard checks such as paragraph count and file structure before delivery. For .doc inputs, ensure the Office/WPS conversion environment is available; the useful output is an annotated document with tracked changes and review logs.