| name | paper2patent |
| description | Convert academic papers into complete Chinese invention patent application deliverables, including DOCX/PDF application documents, claims, specifications, abstracts, drawing specifications, and fidelity reviews. Use when drafting patent text, generating Word or PDF patent applications, preparing patent drawings, or checking paper-to-patent fidelity. |
Paper2Patent
Purpose
Use this skill to convert an academic paper into a complete Chinese invention patent application deliverable. The default full-draft output is a Word document (.docx) and, when a converter is available, a PDF copy. Keep the work faithful to the supplied paper: do not invent technical features, embodiments, data, hardware, scenarios, or effects that are not present in the source material.
Workflow
- Determine the output mode. Use
direct by default, human-in-loop when the user asks to review choices before generation, and text-only when the user explicitly wants no files.
- Collect and assess the paper input. Read
references/input-requirements.md when the user provides partial paper content, a PDF path, source paper figures, figure descriptions, or unclear source material.
- Extract the patentable contribution. Build the closed chain of difference point, technical problem, technical solution, and technical effect.
- Draft the structured patent content. Read
references/text-conversion-workflow.md for the five-part content model and execution order. Treat the repository reference_skills/meterial.md prompt as the normative source for the five patent sections, claim rules, specification depth, and fidelity constraints.
- Apply claim and specification rules. Read
references/claims-and-specification-rules.md before writing or revising claims, background, invention content, or embodiments.
- Prepare drawings. Read
references/drawing-generation.md; first derive patent reference drawings from the source paper figures, captions, and method text when available, then align them with the claims and specification. Generate black-and-white SVG reference drawings, converter-compatible PNG fallbacks when possible, and Image2-style refinement prompts.
- Generate files for full drafts. Read
references/document-generation.md, then use scripts/generate_patent_drawings.py to create clean SVG/PNG drawings, scripts/generate_patent_docx.py to embed them in DOCX, and scripts/export_patent_pdf.py for PDF conversion when available.
- Run the internal quality pass. Read
references/quality-checklist.md before finalizing.
Output Rules
- Default to Simplified Chinese for patent deliverables.
- For a full patent application request, default to generating source-faithful SVG reference drawings, PNG fallback drawings, Image2 refinement prompts, a
.docx file with embedded visible drawings, and a .pdf copy when possible.
- Output only the requested artifact summary to the user: file paths, remaining placeholders, and any conversion limitation. Do not expose hidden analysis or internal checklist results.
- In
text-only mode, output plain patent text with no Markdown headings, code fences, checklists, or hidden reasoning.
- Preserve one consistent invention name across the abstract, claims, specification, and drawing descriptions.
- Do not treat text-only drawing descriptions as a complete drawing deliverable when a full DOCX/PDF application is requested; either embed compliant drawings or report the drawing-generation limitation.
- Do not draw figure numbers or figure titles inside the image canvas. Keep figure captions outside the image in the patent document.
- Keep drawing canvases tight around the technical content: avoid excessive blank margins and reject outputs with unrelated titles, labels, watermarks, decorative elements, or unsupported modules.
- Keep claims definite. Avoid uncertain words such as "等", "大约", "优选", "可以", "比如", and "不限于" in claims.
- In
direct mode, if source information is insufficient, generate explicit 【待补充:...】 placeholders and a material gap section instead of filling gaps with assumptions.
- In
human-in-loop mode, ask for missing source material before generating final files when the gap would affect the technical solution or claim scope.
- Do not store user-provided papers, unpublished disclosures, generated drafts, API keys, tokens, local machine paths, or personal contact details in repository files.
Common Requests
- "将这篇论文直接生成中国发明专利申请书 Word 和 PDF": draft the structured content, generate DOCX, and export PDF when available.
- "将这篇论文转换成中国发明专利申请文本": draft the five-part patent document, using text-only mode if no file is requested.
- "帮我写权利要求书": focus on claims and load the claim rules.
- "根据说明书生成专利附图": load the drawing rules and keep drawing terms aligned with claims and specification.
- "检查这份专利草稿是否忠实论文": use the quality checklist and report gaps or risky additions.