YangCazz / CazzPatent

English | 中文

---

What is this?

CazzPatent is a DeepSeek Harness plugin that acts as an AI patent agent. It walks a raw technical proposal through an 8-stage pipeline — environment check, material parsing, patent-point mining, prior-art search, drafting + diagram generation, review & iteration, reflection & learning, and multi-format export — producing a disclosure ready for firm review.

It ships as two complementary parts:

Part	Role	Install
cazz-patent Skill	Instruction brain: 8-stage workflow + self-improving memory + 6 Python tools	Drop into .dsh/skills/ or customSkillDirs
plugin/ Cordis plugin	Deterministic tools: 5 schema-validated Python tool wrappers	dsh plugin add github:YangCazz/CazzPatent

Pipeline

flowchart LR
    A[Technical proposal] --> B[Stage 0<br/>Env check]
    B --> C[Stage 1<br/>Material parsing]
    C --> D[Stage 2<br/>Patent mining]
    D --> E[Stage 3<br/>Prior-art search]
    E --> F[Stage 4<br/>Draft + diagrams]
    F --> G[Stage 5<br/>Review & iterate]
    G --> H[Stage 5.5<br/>Reflect & learn]
    H --> I[Stage 6<br/>Export]
    I --> J[(MD / DOCX / PDF)]
    H -. inject rules .-> F
    H -. inject rules .-> G

Stage 5.5 (Reflect & learn) extracts reusable rules from iteration history, dual-scores them, and regenerates injection fragments — "fix once, never repeat" — so later proposals benefit automatically.

Highlights

• LaTeX → OMML native equations: 350+ symbols + 96 structural commands — fractions, sub/superscripts, sums, integrals, matrices, accents, Greek, and fonts (\mathbb/\mathfrak/\mathsf/\mathtt/\mathscr) all render as editable Word equations — no MathML, no images.
• Four-phase diagram pipeline: Mermaid logic sketch → 6-point validation → HTML+SVG polish → 3× DPI PNG, three formats per figure.
• Self-improving memory: ledger.json with dual scoring (confirm_count + effectiveness) + domain routing + injection fragments, seeded with 9 rules distilled from 15 real proposals.
• Triple-engine PDF export: Word COM / LibreOffice / weasyprint fallback, cross-platform.
• Deterministic tools: the Cordis plugin registers the 6 scripts as schema-validated tools with configurable pythonPath / scriptsDir / defaultTimeoutMs.

Installation

Skill (instruction brain)

# User-level: available to every project
git clone https://github.com/YangCazz/CazzPatent ~/.dsh/skills/cazz-patent

# Or project-level: copy into a single project
#   cp -r cazz-patent/ <project>/.dsh/skills/

# Or a custom root: point customSkillDirs at the repo in cordis.yml
#   customSkillDirs: [/absolute/path/to/CazzPatent]

Plugin (deterministic tools)

dsh plugin --profile demo add github:YangCazz/CazzPatent#<sha>

Quick start

# 1. Python environment (the Skill's toolchain)
pip install -r requirements.txt
playwright install chromium

# 2. Invoke the Skill in a DSH session
#    /skill:cazz-patent <your technical proposal>

# 3. Or use the CLI tools directly
python cazz-patent/scripts/md_to_docx.py disclosure.md -o disclosure.docx
python cazz-patent/scripts/docx_to_pdf.py disclosure.docx -o disclosure.pdf
python cazz-patent/scripts/batch_diagrams.py --base outputs

Configuration

The Skill is identity-neutral; placeholders resolve from your material at run time:

• {提案人} — the proposer's name, taken from the 基本信息 table or your material's file-name prefix; the model asks when it is missing.

Plugin tool options (override in the profile's cordis.patch.yml):

Field	Default	Meaning
pythonPath	python	Python interpreter to invoke the scripts
scriptsDir	bundled scripts/	Directory containing the Python scripts
defaultTimeoutMs	300000	Per-run timeout

Directory layout

CazzPatent/
├── cazz-patent/          # DSH Skill (instruction brain)
│   ├── SKILL.md          # 8-stage entry point
│   ├── prompts/          # 6 stage prompts
│   ├── scripts/          # 6 Python tools
│   ├── templates/        # disclosure template + diagram HTML template
│   └── memory/           # self-improving memory (ledger + corrections + injections)
├── plugin/               # Cordis plugin (deterministic tools)
│   ├── src/index.ts      # 5 tools + config
│   ├── scripts/          # bundled Python scripts
│   ├── esbuild.config.mjs
│   └── cordis.patch.yml
├── tests/                # pytest suite (11 cases)
├── README.md / README_zh.md
├── requirements.txt
├── LICENSE
└── .github/workflows/ci.yml

Tests

pip install python-docx pytest
python -m pytest tests -q

CI runs the tests on every push and pull request.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md.

License

MIT © 2026 YangCazz