# Convert Research Papers into Chinese Patent Drafts

Converting a research paper into a Chinese invention patent draft requires tracing every claim feature back to source evidence and preserving formulas as editable math. This skill provides a structured workflow with validation gates that produce a complete, evidence-grounded Chinese patent application package.

## Install

```bash
npx skillstore add yuan1z0825/nature-paper-to-patent
```

## Metadata

- - Slug: yuan1z0825-nature-paper-to-patent
- - Version: 1.0.0
- - Author: Yuan1z0825
- - GitHub username: Yuan1z0825
- - License: MIT
- - Repository: https://github.com/Yuan1z0825/nature-skills/tree/main/skills/nature-paper-to-patent
- - Ref: main
- - Supported tools: Claude, Codex, Claude Code
- - Risk level: low
- - Risk factors: external\_commands, network, filesystem
- - Quality score: 85
- - Quality tier: gold
- - Public page: https://skillstore.pages.dev/skills/yuan1z0825-nature-paper-to-patent
- - Manifest: https://skillstore.pages.dev/api/skills/yuan1z0825-nature-paper-to-patent/manifest

## Capabilities

- Extracts searchable text from PDF papers using pypdf with OCR-needed detection
- Builds a structured draft JSON with source map, evidence ledger, claim-feature map, and formula inventory
- Validates draft completeness with quality thresholds for evidence support, claim architecture, and terminology consistency
- Renders core equations as native editable Office Math \(OMML\) in DOCX output
- Generates claim-aligned SVG flowcharts and methodology figures in black-and-white patent style
- Produces separate Chinese DOCX files for claims \(权利要求书\), specification \(说明书\), abstract \(说明书摘要\), and abstract figure \(摘要附图\)

## Use Cases

- University researcher drafting a Chinese patent: A PhD student has a published paper on a new defect-detection algorithm and wants to file a Chinese invention patent. The skill extracts the paper, maps every claim feature to source pages, and produces a validated DOCX package.
- Patent agent auditing a paper-patent relationship: A patent agent receives a client paper and an existing granted patent. The skill performs a bidirectional paper-patent audit, identifying paper-only, patent-only, broader, and narrower claim features with exact locators.
- R&D engineer preparing a disclosure analysis: A corporate R&D engineer has a technical report and source code. The skill produces a disclosure analysis extracting patentable contributions with formula preservation and evidence-grounded support states.

## Prompt Templates

### Basic PDF to Chinese patent draft

```
Convert the paper at paper/main.pdf into a complete Chinese invention patent application. Use full-draft mode with algorithm-software invention type. Produce separate DOCX files for claims, specification, abstract, and abstract figure.
```

### Scanned PDF with OCR uncertainty

```
The paper is a scanned PDF at paper/scanned.pdf with some formulas unclear after OCR. Generate a claim draft for the confirmed technical solution only. Do not guess uncertain equation symbols and mark them with [TO CONFIRM].
```

### Paper-patent audit comparison

```
Compare paper/new-method.pdf with existing-patent/cn-grant-2023.pdf. For each claim in the existing patent, determine if the paper provides support. Label items as paper-only, patent-only, patent-broader, or patent-narrower with exact locators.
```

### Mixed project with source code and figures

```
The workspace contains a paper PDF, supplementary source code in supplementary/source-code/, and source figures in source-figures/. Use mixed-project mode. Include code-derived evidence with C-prefixed source IDs and preserve algorithm pseudocode in the specification.
```

## Limitations

- Does not provide patentability, infringement, or filing-guarantee opinions \(explicitly stated in SKILL.md\)
- Cannot infer inventorship, ownership, publication dates, or prior-art conclusions without explicit confirmation
- OCR for scanned PDFs is not included; the skill flags low-text PDFs and recommends external OCR
- Output is a drafting aid for inventor and patent-professional review, not a filed application

## Best Practices

- Always state the detected source\_format, task\_mode, and invention\_type axes before loading fragments to avoid loading irrelevant reference files
- Create stable source IDs \(P001, E001, F001, C001\) before drafting so every claim feature can be traced back to evidence
- Run validate\_patent\_draft.py after populating the structured draft and resolve all ERROR findings before building the final package
- Label uncertain content with \[TO CONFIRM: specific question\] outside formal claims to keep claims clean and legally defensible

## Anti Patterns

- Do not draft the patent application directly from the paper abstract or contribution list without building the full source map first
- Do not include unsupported features \(status=unsupported\) in formal claims; they must be excluded or confirmed with the inventor
- Do not use plain LaTeX strings as visible formulas in the DOCX output; always render as native Office Math for editability
- Do not infer inventorship, ownership, or publication dates; these require explicit human confirmation and legal review

## Security Audit

- - Safe to publish: true
- - Audited at: 2026-06-24T06:01:19.621\+00:00
- - Summary: The static analyzer flagged 291 potential issues across 34 files, but the vast majority are false positives from regex pattern matching on documentation and standard library usage. All 122 'weak cryptographic algorithm' alerts are false positives matching on Python \`hashlib\` imports for content hashing and on the word 'hash' in regex patterns. The 154 'external\_commands' alerts are markdown backtick characters \(code formatting\) in README and reference files, not actual command execution. The 5 'shutil operations' are legitimate file copying for project workspace initialization. The 1 \`subprocess.run\` call in build\_patent\_package.py uses hardcoded script paths with no user input injection vector. No malicious intent, data exfiltration, or unauthorized network activity was found.

## Stats

- - Views: 0
- - Downloads: 1
- - Favorites: 0
- - Popularity score: 0