mcpField
P
patent-pipeline

patent-pipeline

github.com/wanshuiyinDocuments18.2k512 installsUpdated 4d ago
Repository

About patent-pipeline

patent-pipeline coordinates patent drafting from prior-art searches through jurisdiction-formatted application documents. It is for inventors preparing a filing and connects novelty assessment, invention structuring, claims, specification, review, and formatting in a staged workflow with approval checkpoints.

Key features

  1. 01Prior-art search and novelty assessment
  2. 02Invention structuring and claim design
  3. 03Specification and embodiment drafting
  4. 04Examiner-style review cycles
  5. 05CN, US, and EP formatting
  6. 06Checkpointed workflow with saved state

Use cases

  1. 01Assess an invention against prior art
  2. 02Develop a claims hierarchy
  3. 03Draft a patent specification from an invention brief
  4. 04Prepare documents for CN, US, or EP jurisdiction

MCP support

The skill uses Codex MCP for examiner-style review, with reviewer routing that can select other configured routes.

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/wanshuiyin/auto-claude-code-research-in-sleep --skill patent-pipeline

After installing, type/patent-pipelinein Claude Code to confirm it loaded.

Details

From the project README

Patent Pipeline: From Invention to Filing

Before each reviewer slot, follow Selectable Reviewers: resolve invocation → .aris/reviewer.json → Claude Code default (Codex MCP) with reviewer_route.py through the canonical helper resolver, preserving this slot's tier, fresh contexts, source/image inputs and deterministic gates. Codex examples/pins below apply only to backend=codex; other routes use their own first-call/reply contract. Save the route and actual returned identity; unavailable selections are REVIEW_UNAVAILABLE. Classify acceptance from actual families/completion/scope. Forward invocation reviewer/model/effort overrides to reviewer-bearing children. Generation workers are unchanged.

Draft a complete patent application based on: $ARGUMENTS

Overview

This skill orchestrates the full patent drafting lifecycle -- from prior art search through jurisdiction-formatted filing documents. It chains sub-skills into a patent-specific pipeline:

/prior-art-search → /patent-novelty-check → /invention-structuring → /claims-drafting → /specification-writing → /patent-review → /jurisdiction-format
     (search)           (verify)              (structure)             (claims)            (description)          (examiner)         (compile)
                                                                                              ├── /figure-description
                                                                                              └── /embodiment-description

This is a parallel branch, not part of the linear research pipeline. After /idea-discovery produces validated ideas, the user can either:

  • Go to /experiment-bridge → /auto-review-loop → /paper-writing (publish track)
  • Go to /grant-proposal (funding track)
  • Go to /patent-pipeline (patent track) <-- this skill
                    ┌→ /experiment-bridge → /auto-review-loop → /paper-writing  (publish track)
/idea-discovery ────┤
                    ├→ /grant-proposal → [get funded] → ...  (funding track)
                    └→ /patent-pipeline → [file patent]       (patent track)

Patents are about protecting inventions (legal scope), not publishing results (academic contribution). This skill handles the unique requirements of patent drafting: prior art analysis, claims hierarchy design, specification writing with enablement support, embodiment descriptions, and jurisdiction-specific formatting.

Constants

  • JURISDICTION = CN — Target patent jurisdiction. Options: CN (CNIPA), US (USPTO), EP (EPO), ALL (generate all three). Override via argument (e.g., /patent-pipeline "invention — US").
  • PATENT_TYPE = invention — invention (发明专利, 20 year protection) or utility_model (实用新型, CN only, 10 year protection, apparatus claims only). Override via argument.
  • REVIEWER_MODEL = gpt-6-astra — Model used via Codex MCP for examiner-style review.
  • MAX_REVIEW_ROUNDS = 2 — Maximum review-revision cycles.
  • AUTO_PROCEED = false — At each checkpoint, always wait for explicit user confirmation. Patent applications require inventor judgment at every stage. Set true only if user explicitly requests autonomous mode.
  • LANGUAGE = auto — Output language. Auto-detected from jurisdiction: CN->Chinese, US->English, EP->English. Override explicitly if needed.
  • OUTPUT_DIR = patent/ — Directory for generated patent files.
  • OUTPUT_FORMAT = markdown — Draft format. markdown for review, docx for filing-ready.

Override defaults via arguments: /patent-pipeline "invention — US, utility model" or /patent-pipeline "invention — ALL, language: Chinese".

Patent Type Specifications

Invention Patent (发明专利)
FieldDetail
Protection20 years from filing date
Subject matterMethods, systems, products, compositions, processes
ExaminationSubstantive examination required
Inventive stepHigh (must involve an inventive step / 创造性)
Timeline2-4 years to grant (CN); 2-3 years (US); 3-5 years (EP)
ClaimsMethod + system + product claims allowed
Utility Model (实用新型) — CN Only
FieldDetail
Protection10 years from filing date
Subject matterProduct shape, structure, or combination thereof only
ExaminationFormal examination only (no substantive examination)
Inventive stepLower than invention patent
Timeline6-8 months to grant
ClaimsApparatus/device claims only. NO method claims.
RestrictionCN jurisdiction only

State Persistence (Compact Recovery)

Patent drafting is a long task that may trigger context compaction. Persist state to patent/PATENT_STATE.json after each phase:

{
  "phase": 3,
  "jurisdiction": "CN",
  "patent_type": "invention",
  "language": "Chinese",
  "codex_thread_id": "019cfcf4-...",
  "invention_title": "...",
  "claims_count": 15,
  "status": "in_progress",
  "timestamp": "2026-04-10T15:00:00"
}

Write this file at the end of every phase. On invocation, check for this file:

  • If absent or status: "completed" -> fresh start
  • If status: "in_progress" and within 24h -> resume from saved phase (read output files to restore context)
  • If older than 24h -> fresh start (stale state)

On completion, set "status": "completed".

Workflow

Phase 0: Input Parsing & Context Gathering

Parse $ARGUMENTS to extract:

  1. Invention description — may be structured (references INVENTION_BRIEF.md), conversational with figures, or output from IDEA_REPORT.md
  2. Jurisdiction — detect from keywords (e.g., "CN" or "中国" -> CN, "US" or "USPTO" -> US, "EP" or "EPO" -> EP, "ALL")
  3. Patent type — detect from keywords (e.g., "utility model" or "实用新型" -> utility_model, default -> invention)
  4. Overrides — language, output format, review rounds

Then gather context from the project directory:

  1. Read INVENTION_BRIEF.md if it exists (user filled in the template)
  2. Read IDEA_REPORT.md if it exists (from /idea-discovery -- can extract invention from research results)
  3. Read refine-logs/FINAL_PROPOSAL.md if it exists
  4. Read NARRATIVE_REPORT.md if it exists (research results that may be patentable)
  5. Search for user-provided figures (PNG, JPG, SVG, PDF) in the project directory
  6. Check for patent/PATENT_STATE.json (resume from prior interrupted run)

If insufficient context exists:

  • No invention description at all -> suggest user describe the invention or fill in INVENTION_BRIEF.md
  • Has IDEA_REPORT.md -> extract patentable aspects from the research
  • Has figures -> reference them in the invention brief
  • No figures -> note that figures will be needed and plan what drawings are required

If the input is conversational (not a structured brief), parse the description into the invention brief structure and write patent/INVENTION_BRIEF.md for downstream phases.

Phase 1: Prior Art Search & Novelty Assessment
1.1 Prior Art Search

Invoke /prior-art-search:

/prior-art-search "patent/INVENTION_BRIEF.md"

This searches patent databases (Google Patents, Espacenet) and academic literature for relevant prior art.

1.2 Novelty Check

Invoke /patent-novelty-check:

/patent-novelty-check "patent/INVENTION_BRIEF.md"

This assesses novelty and non-obviousness against the prior art found in step 1.1.

🚦 Checkpoint: Present the prior art landscape and novelty assessment:

Prior art search complete:
- [X] patent references found
- [Y] non-patent literature references found
- Closest prior art: [reference] -- [why it's closest]
- Novelty assessment: [PATENTABLE / PATENTABLE WITH AMENDMENTS / NOT PATENTABLE]
- Key risk areas: [list]

Ready to proceed with invention structuring?

⛔ STOP HERE and wait for user response. Do NOT auto-proceed unless AUTO_PROCEED=true.

Options:

  • Reply "go" -> proceed to Phase 2
  • Reply with adjustments -> refine the invention scope and re-check novelty
  • Reply "stop" -> save progress to patent/DRAFT_NOTES.md

State: Write PATENT_STATE.json with phase: 1.

Phase 2: Invention Structuring & Claims Design
2.1 Structure the Invention

Invoke /invention-structuring:

/invention-structuring "patent/INVENTION_BRIEF.md"

This decomposes the invention into core inventive concept, supporting features, and optional features. Produces patent/INVENTION_DISCLOSURE.md.

2.2 Draft Claims

Invoke /claims-drafting:

/claims-drafting "patent/INVENTION_DISCLOSURE.md"

This drafts the claims hierarchy -- the most critical part of the patent. Produces patent/CLAIMS.md.

🚦 Checkpoint: Present the invention structure and claims:

Invention structured:
- Core inventive concept: [summary]
- Claim categories: [method, system, etc.]
- Claims drafted: [X] independent + [Y] dependent = [Z] total
- Independent claim 1 (broadest): [first 50 words of claim 1]
- Examiner review score: [X]/10

The claims define the legal scope of protection. Please review before proceeding to specification.

⛔ STOP HERE and wait for user response. Do NOT auto-proceed unless AUTO_PROCEED=true.

Options:

  • Reply "go" -> proceed to Phase 3
  • Reply with adjustments (e.g., "broaden claim 1", "add more dependent claims") -> revise claims
  • Reply "stop" -> save progress

State: Write PATENT_STATE.json with phase: 2.

Phase 3: Specification Writing

Invoke /specification-writing:

/specification-writing "patent/CLAIMS.md"

This writes the full specification section by section. Internally invokes /figure-description (if user-provided figures exist) and /embodiment-description for the detailed description. The specification-writing skill handles figure processing and embodiment writing as sub-skills.

🚦 Checkpoint: Present the specification overview:

Specification written:
- Title: [title]
- Sections: Technical Field, Background, Summary, Drawings Description, Detailed Description, Abstract
- Embodiments: [X]
- Reference numerals: [Y] components mapped
- Abstract length: [Z] words (limit: [jurisdiction limit])
- Claim support: [all elements covered / X elements missing]

Ready to proceed to review?

⛔ STOP HERE and wait for user response.

State: Write PATENT_STATE.json with phase: 3.

Phase 4: Patent Review

Invoke /patent-review:

/patent-review "patent/"

This runs 2 rounds of examiner-style review via GPT-6-Astra xhigh. The examiner evaluates clarity, written description, enablement, novelty, non-obviousness, and claim scope.

State: Write PATENT_STATE.json with phase: 4 and review score.

Phase 5: Jurisdiction Formatting & Output

Invoke /jurisdiction-format:

/jurisdiction-format "patent/"

This compiles the application into the target jurisdiction format(s).

Final Deliverables
OutputLocationDescription
CN: 权利要求书patent/output/CN/Claims in CNIPA format
CN: 说明书patent/output/CN/Description in CNIPA format
CN: 说明书摘要patent/output/CN/Abstract (CN)
US: Claimspatent/output/US/Claims in USPTO format
US: Specificationpatent/output/US/Description in USPTO format
US: Abstractpatent/output/US/Abstract (US)
EP: Claimspatent/output/EP/Claims in EPO format
EP: Descriptionpatent/output/EP/Description in EPO format
EP: Abstractpatent/output/EP/Abstract (EP)
Final Report
## Patent Pipeline Complete

### Application Summary
- Title: [invention title]
- Jurisdiction: [CN/US/EP/ALL]
- Patent Type: [Invention / Utility Model]
- Language: [Chinese/English]
- Total Claims: [X] independent + [Y] dependent

### Pipeline Scores
| Phase | Score |
|-------|-------|
| Prior Art Search | [completeness assessment] |
| Novelty Assessment | [PATENTABLE/PATENTABLE WITH AMENDMENTS/NOT PATENTABLE] |
| Examiner Review Round 1 | [X]/10 |
| Examiner Review Round 2 | [Y]/10 |
| Final | [Z]/10 |

### Output Files
[Table of all generated files with paths]

### Next Steps
- [ ] Have a patent attorney review the application
- [ ] Conduct professional prior art search (this tool's search is preliminary)
- [ ] Prepare formal drawings (if user figures need professional rendering)
- [ ] File with the patent office
- [ ] For utility model (CN): formal examination typically takes 6-8 months
- [ ] For invention patent: substantive examination may take 2-4 years

State: Write PATENT_STATE.json with phase: 5, status: "completed".

Key Rules

  • Never fabricate prior art references, patent numbers, or citations.
  • Claims must be supported by the specification (written description requirement).
  • Each jurisdiction has strict format requirements -- do not mix formats.
  • Utility model (实用新型) applies ONLY to CN jurisdiction and ONLY covers apparatus/device claims.
  • AUTO_PROCEED defaults to false -- patent applications require human review at every phase. Sub-skills inherit this flag: when AUTO_PROCEED=false, sub-skills present results and wait at their own internal checkpoints too.
  • The patent pipeline produces drafts for attorney review, not final filing documents.
  • Large file handling: if a Write operation fails, retry with Bash cat <<'EOF' heredoc.
  • Never include experimental results or empirical evaluations in the specification.
  • Consistent terminology is mandatory -- same word for the same concept throughout.
  • If the selected reviewer is unavailable, record REVIEW_UNAVAILABLE; the draft pipeline may continue explicitly unreviewed, without switching providers.

Composing with Other Workflows

The patent pipeline can start from multiple entry points:

User describes invention directly ──→ /patent-pipeline

/idea-discovery produces IDEA_REPORT.md ──→ /patent-pipeline (extract patentable aspects)

/research-refine produces FINAL_PROPOSAL.md ──→ /patent-pipeline (from refined research idea)

/auto-review-loop produces strong results ──→ /patent-pipeline (patent the method)

Acknowledgements

Built on the ARIS (Auto-claude-code-research-in-sleep) skill architecture. Patent writing principles adapted from MPEP (US), CN Patent Examination Guidelines (CN), and EPO Guidelines for Examination (EP).

Comments

Markdown supported