A key pain point in Chinese patent filing is the disclosure document (交底书) — the technical brief engineers must prepare for patent agents. Many experienced engineers stall here: not because they can't write, but because they don't know how to mine patentable points, run novelty searches, draw required diagrams, or fill Word templates. handsomestWei/patent-disclosure-skill aims to close this gap with AI, gaining 571 stars in one day on GitHub Trending.
The Problem
Engineers typically struggle with five steps:
1. Patent point mining — deciding which parts of a system are worth filing 2. Novelty search — CNIPA's search system has its own query syntax, unlike a Google search 3. Diagrams — invention patents need mermaid flowcharts; utility models need structural line drawings with part-number callouts; design patents need product outline sketches 4. Word output — the deliverable is a .docx per template, with embedded images and part labels 5. Iteration — agents request revisions to claims and materials, each change requiring tracked edits
Core Capabilities
1. Three patent types, separate templates
- Invention: methods/systems/algorithms, mermaid diagrams
- Utility model: shape, construction, connections — structural line drawings with part numbers
- Design: appearance, pattern, color — product line drawings
- Design line drawings: extract product outlines from photos, or generate via image-to-image/text-to-image if no qualified sketch exists
- Utility model structure drawings: generate outlines and part-number callouts, writing sub-SVGs per
structure_schema.parts - CAD 3D projection: extract isometric and multi-view projections from STEP/STP engineering models
- Policy radar: searches recent CNIPA policy and examination trends (off by default; requires user confirmation)
- Examination response assistance: stores anonymized historical office actions and responses with legal-clause/defect tags; new office actions are matched against this history to draft responses, optionally with a local vector model for semantic retrieval
- Output is a draft: the README states it "must be human-reviewed before filing" — claim wording and scope still need an agent's judgment
- CAD parsing limits: STEP/STP parsing is deferred by default ("drafting continues; asks later"), reflecting format-variability challenges
- Policy compliance: policy radar is opt-in, keeping humans in the decision loop
The skill switches templates automatically based on the declared type.
2. Project scanning: from code to patent points
The skill scans project directories — reading docs, code, and .docx/.pptx files (converted to Markdown first) — to extract patentable technical points. Documentation is scanned before code, matching patent-examination logic: patents protect "technical solutions," not code implementations.
3. Novelty search against CNIPA
A built-in crawler (tools/crawl/cnipa_epub_search.py) queries the CNIPA publication database with the mined patent points. If CNIPA fails or returns nothing, it falls back to WebSearch — a sensible degradation so the government system isn't a single point of failure.
4. Automated diagram generation
5. Obsidian knowledge base
Beyond drafting, the skill extracts claim trees, glossaries, and feature-specification-drawing mappings from published patents into Obsidian. Over time this becomes a personal patent-intelligence layer enabling patent comparison, family analysis, and technology-roadmap mapping.
6. Policy radar + examination response
Why It Matters
The "last mile" problem
AI coding assistants write code well, but converting technical work into deliverable professional documents (patent disclosures, audit reports, bid proposals) has lacked tooling. These documents require domain knowledge + format conventions + multi-round iteration — engineers know the tech, agents know the format, and neither knows both.
Depth of a vertical skill
Unlike general assistants, this skill knows the three patent types' templates, integrates with CNIPA search, extracts views from CAD files, generates compliant line drawings, and tracks how examination-guideline changes affect drafting — work that normally requires an agent, an engineer, and a CAD drafter together.
Obsidian as the knowledge layer
Obsidian's bidirectional links, graph view, and Canvas suit patents' terminology-dense, cross-referential nature. As the library grows, comparison and roadmap analysis become possible.
Caveats
The Bigger Picture
The project signals that vertical AI skills are expanding from "writing code" to "writing professional documents." Code is structured and compiler-verified; professional documents are unstructured, validated only at examination or bidding time — raising the bar for vertical skills, which must encode domain knowledge, formats, and compliance as distinct modes (invention/utility/design/policy/examination-response) with their own toolchains.
---
Project: https://github.com/handsomestWei/patent-disclosure-skill
Audience: engineers writing patent disclosures, patent agents, corporate IP teams
Difficulty: medium (basic patent knowledge needed; Obsidian optional but recommended)