Patent Writing Chinese

Patent Writing - Chinese (中国专利说明书生成)

This skill helps you produce professional Chinese patent specifications as Word .docx files. It encapsulates years of trial-and-error around the messy parts: OMML inline math, Chinese-friendly fonts, Word's blue heading default, parallelogram data shapes, and clean orthogonal flowchart routing.

When to use this skill

Trigger on requests like:

• "帮我写一份专利说明书"
• "把这篇论文改写成专利"
• "Write a Chinese patent for ..."
• "Convert this methodology into 中国专利"
• "Generate a 专利申请书 / 发明专利"
• Any request to produce a Chinese patent docx with figures and equations

What this skill provides

patent-writing-chinese/
├── SKILL.md                       (this file)
├── scripts/
│   ├── patent_helpers.py          PatentDoc class + OMML builders + table/figure helpers
│   ├── figure_helpers.py          B&W matplotlib primitives (rect, para, diamond, ...)
│   └── patent_template.py         Skeleton script you can copy and customize
└── examples/
    └── example_patent.py          Tiny working example

The two helper modules are pure infrastructure — they contain no patent-specific content. You import them and call their functions to assemble any patent you want.

Workflow

Step 1: Understand the user's request

Ask (or infer) the following before writing code:

• Title of the invention (中文)
• Technical field (one paragraph for 技术领域)
• Background: what existing methods exist, what's wrong with them
• Core innovations: 2–4 bullet points the patent will claim
• Detailed steps (S10/S20/...): the actual algorithm/method to describe
• Equations needed: list each formula
• Figures needed: usually 1 overview + 2–4 module-level diagrams
• Tables: if any, what columns/rows

If the user provides a paper or document, read it first and extract these elements yourself.

Step 2: Set up the build script

Copy scripts/patent_helpers.py and scripts/figure_helpers.py into the user's working directory (or have the build script add them to sys.path). Then create a build_patent.py that imports them.

Use scripts/patent_template.py as a starting skeleton — it contains the standard 5 sections and H(), MP(), EQ(), FIG(), TABLE() calls you can fill in.

Step 3: Define equations early

Use the OMML builders (mr, m_sub, m_sup, m_subsup, m_frac, m_delim, m_nary, m_norm, m_abs, m_bar) to build display equations as XML strings. Wrap with wrap() for display blocks (used by EQ()) or with imath() for inline equations (used by MP() and TABLE()).

Always define a small library of inline symbol constants up front, e.g.:

S_xt    = imath(m_sub(mr("x"), mr("t")))
S_phis  = imath(m_sub(mr("\u03c6"), mr("s")))
S_LBC   = imath(m_sub(mr("L"), mr("BC")))

Then reuse them across paragraphs and tables.

Step 4: Write paragraphs with MP(), never plain P() for math

CRITICAL: never write '风格标签 s_t' or 'L_BC' as plain text in P(). The literal _ will appear as an underscore character in Word, NOT as a subscript.

Instead, split the text and embed the inline equation:

doc.MP('其中，', S_xt, '为自车特征向量，', S_phis, '为编码器的可学习参数。')

MP() accepts an arbitrary mix of text strings and inline OMML XML strings; it routes each segment to either a w:r run or an m:oMath element automatically.

Step 5: Build figures

Use figure_helpers.py for diagrams. Conventions:

• Rectangles (rect): processing modules
• Parallelograms (para): data items / data flow boundaries
• Diamonds (diamond): decision nodes
• Dashed rectangles (dashed_rect): module groups (entire pipelines or sub-systems)
• All B&W: white fill, black border, no color
• Orthogonal arrows only: never diagonal — use vline/hline for path segments and arrow_to/connect_h/connect_v only at the FINAL destination
• Arrow heads belong only at destinations, never in the middle of a multi-segment path

When a path goes through a junction (fork or merge), use plain vline/hline for the segments and put the arrow head only on the final box edge.

Step 6: Validate before delivery

After running the build script, unzip the docx and check:

1. word/document.xml parses as valid XML
2. `` count is 0 (those would cause Word to error if mixed with text)
3. All ` elements live in w:p paragraphs (some inline alongside w:r`, some standalone for display)
4. No w:color element has themeColor="..." blue references (headings must be RGB 000000)

Open the docx and visually verify the figures fit inside their dashed groups, no text overflows boxes, and equations render as italic math.

Patent document structure (Chinese conventions)

Every patent has these five Heading 1 sections, in order:

1. 技术领域 — 1 paragraph naming the technical field and the specific aspect this invention addresses
2. 背景技术 — 3–5 paragraphs reviewing prior art, identifying its limitations
3. 发明内容 — overview of the invention + numbered steps (步骤一/步骤二/...) + benefits comparison
4. 附图说明 — list of figures, one line each ("图1 本发明总体结构框图")
5. 具体实施方式 — detailed implementation, indexed by S10/S11/S12/S20/... with all equations, tables, and embedded figures

Title is centered, 16pt, bold, NOT a Heading style. Body paragraphs use a 0.74cm first-line indent. Headings should be black (the helper overrides Word's default blue).

Common pitfalls

1. oMathPara mixed with w:r → Word reports "this file contains content that can't be read". Fix: never put a display math block in the same paragraph as a text run. Use imath() (returns oMath only) for inline math, wrap() (returns oMathPara) ONLY when nothing else is in the paragraph.

1. Plain text math symbols (x_t, c_t^*, L_BC) → render as literal underscored text. Always use inline OMML via MP().

1. Blue headings → Word's default Heading 1 style has a blue theme color. The helper explicitly sets RGBColor(0, 0, 0) and removes themeColor attributes. Don't bypass this.

1. Diagonal arrows in figures → look messy and cross other elements. Always use orthogonal routing: vline then hline then arrow_to only at the final destination box edge.

1. Text overflowing rectangles → use multi-line labels ('驾驶风格\n聚类与识别') and increase box height accordingly. After applying the 2/3 width scaling, very long Chinese labels need 2 lines.

1. Parallelograms overflowing dashed groups → the skew adds ~0.15 × effective_width to the leftmost/rightmost vertex beyond the bounding box. Account for this when placing parallelograms near a dashed border. Move them inward or reduce width.

1. Figure caption right under a FIG() call is set automatically — don't add a separate caption paragraph.

Tips

• Read scripts/patent_template.py for the canonical structure of a patent script
• Read examples/example_patent.py for a tiny end-to-end example
• All helpers are documented with docstrings — read them before improvising
• For figures, study the make_box return dict — every box returns its edge coordinates so arrows can snap precisely
• When in doubt, prefer vline/hline (plain segments) and a single arrow_to at the destination over multiple chained arrows

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: StellaQIU27
• Source: StellaQIU27/claude-skill-patent-writing-chinese
• License: MIT

Install and usage instructions live in the source repository linked above.