China Patent Disclosure Assistant
From project docs to a deliverable technical disclosure, or read existing patents into plain-language notes and a knowledge graph.
Evidence: The skill declares allowed tools including Bash, WebSearch, Read/Write, and explicitly requires reading prompt files step-by-step; it fixes scripts for prior-art search and PDF download (cnipa_epub_search.py, fetch_patent_pdf.py) and forbids writing ad-hoc download scripts in-session; it includes desensitization requirements for sensitive data; iteration saves new timestamped files without overwriting old drafts. Deductions: No explicit disclosure of data flow (e.g., where search results go, whether processing is local-only); no rollback or undo mechanism; third-party dependencies (Playwright) not subject to security review. Main risks visible but not fully addressed, so 12.
Evidence: The skill provides detailed step-by-step instructions and tool tables, but no executable tests or CI configuration is present; test files exist (tests/test_cnipa_epub_chain.py etc.) covering only some tools, but no CI workflow is in the repo; happy path behavior is described, but error handling (e.g., network failures, missing files) is not clearly specified. Deductions: Key path reproducibility cannot be determined statically; limited feedback on failure scenarios. Thus 8.
Evidence: The skill clearly distinguishes two modes (disclosure drafting and patent reading) with trigger conditions and target scenarios; Chinese language support and access to CNIPA (cnipa.gov.cn) from mainland China are core features; boundaries are stated (e.g., when a patent number is provided, it prioritizes reading mode). Deductions: Capability boundaries (e.g., unsupported patent types) not fully detailed; availability of CNIPA service not verified; if core functionality depended on unreachable overseas services, deduction would apply, but here domestic services dominate, so 10.
Evidence: Repository includes README, INSTALL.md, LICENSE (MIT), examples/ directory, docs/, prompt mapping tables; version is v2.0.0 in SKILL.md; no changelog but documentation is comprehensive. Deductions: No explicit maintenance responsibility or update path; no FAQ; version number appears only in SKILL.md, not unified across the repo. Thus 10.
Evidence: The skill claims to produce technical disclosures and patent reading notes, with examples/ directory, but no verifiable output samples or evidence that outputs are directly usable; value proposition is clear (time-saving, deliverable), but comparative benefit lacks evidence. Deductions: Static read cannot verify output quality; capped at 7, so 7.
Evidence: Repository includes multiple test files (tests/test_cnipa_epub_chain.py etc.) covering some tool modules, serving as auditable primary material, but no CI configuration or third-party execution evidence; key claims (e.g., auto-bootstrap, search success rates) lack independent verification. Deductions: Test coverage is limited, and no independently reproducible conclusions, so 4.
- The skill depends on the CNIPA website (epub.cnipa.gov.cn); if unreachable or network-restricted, the prior-art search degrades to WebSearch; verify the reliability of this fallback.
- Obsidian installation and full configuration are strongly recommended, but the degradation path (e.g., outputs/ directory) lacks detailed documentation.
- The toolchain relies on Playwright and Node.js, which are complex to install and large; in constrained environments, core functionality may be unavailable.
- The skill claims automatic bootstrap of Obsidian vault, but provides no verification or rollback mechanism for this process.
What does this skill do, and when should you use it?
This skill targets Chinese patents with two modes: Mode A mines patent points from project materials, performs prior-art search (prioritizing the CNIPA official publication site), drafts anonymized technical disclosures (with Mermaid diagrams and Word export), and runs a self-check loop; Mode B interprets existing patents (by publication number or PDF) into plain-language narratives, claim trees, and an Obsidian knowledge graph. The skill includes step-by-step prompts, Python utilities (e.g., Office-to-Markdown conversion, EPUB search, diagram rendering), and supports iterative revision (merge/correction) with versioning. Installation follows the AgentSkills convention, requiring Python and optional Node.js dependencies.
Reads project documentation (converting .docx/.pptx first), scans code and design docs, and analyzes candidate patent points; calls cnipa_epub_search.py (primary) or falls back to WebSearch for prior art; generates an anonymized technical disclosure template with Mermaid system/flow diagrams, renders PNGs, and exports .docx; runs an internal self-check checklist; supports merging or correcting an existing disclosure and saving a new timestamped file with a revision log; for existing patents, extracts text (auto-downloading PDF if needed), builds a claim tree, and creates a plain-language narrative plus an Obsidian note with automatic bootstrap (CSS/Canvas).
- R&D engineers: have project design docs and code, want to identify patent points and draft a disclosure—just describe it in natural language.
- Patent engineers: need online prior-art search (preferring CNIPA) and a Word disclosure with system/flow diagrams.
- Iterative review: when you add materials or correct an existing disclosure file, the skill auto-detects and saves a new version.
- Tech enthusiasts: have a publication number or PDF, want to quickly understand the patent and generate a knowledge graph note.
- Knowledge managers: link multiple patent interpretations into an Obsidian vault with cross-referenced Canvas graphs.
What are this skill's strengths and limitations?
- Covers the full disclosure pipeline: mining, search, drafting, self-check, iteration.
- Prior-art search prioritizes the official CNIPA site, reducing reliance on search engines.
- Supports Office document conversion and integrates Mermaid diagrams with docx export.
- Iteration mode auto-detects, saves new files, and logs revision history.
- Mode B generates Obsidian knowledge graphs, great for personal knowledge management.
- Primarily for Chinese patents; limited support for international or non-Chinese materials.
- CNIPA search depends on Playwright and network, and may fail, falling back to WebSearch.
- No automated test suite; reliability depends on actual runtime conditions.
- Obsidian integration strongly recommends a configured vault path; otherwise degrades to output files.
- Some features (Node.js rendering) require extra installs, with a non-trivial initial setup.
How do you install this skill?
Place the cloned repository into the skills path of Claude Code or Cursor (e.g., .claude/skills/patent-disclosure-skill). Run pip install -r requirements.txt; optionally pip install -r tools/requirements-cnipa.txt and python -m playwright install chromium for CNIPA search; install Node.js for Mermaid rendering and run npm install in the tools folder or use npx mmdc. See INSTALL.md for details.
How do you use this skill?
Describe your need in natural language to the agent, e.g., “Write a technical disclosure based on the project at /path/to/project” or “Explain patent CN12345678A”. The skill auto-selects Mode A or B based on intent. You can also use slash commands like /patent-disclosure-skill and should specify a project path or technical topic in the prompt.