patent-pipeline（专利管道）

Patent Pipeline: From Invention to Filing

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-5.4 — 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)

Utility Model (实用新型) — CN 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-5.4 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

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 mcp__codex__codex is not available (no OpenAI API key), skip external cross-model review and note it in the output. The pipeline must not fail due to missing reviewer access.

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).