Patent & Software Copyright Generation (AI + Big Data)#

Generate CNIPA invention patent documents or CPCC software copyright materials from AI / big-data project code, design docs, and research papers.

Covers 7 AI domains + Big Data (23 sub-directions), 11 claim templates.

Full version (Chinese, with Word/PPT output): see AI-Copyright-Skill project.

Triggers#

patent / claims / specification / software copyright / disclosure / IP application / paper-to-patent / /patent-software-ip

Overall Flow#

text

Phase A  Requirement Diagnosis -> path + domain classification + risk level
Phase B  Project Analysis -> auto-detect domain + extract key technical points
Phase C  Generation (branch by path)
  C1 Patent: prior art search -> claims (11 templates) -> specification -> abstract -> self-check
  C2 Software Copyright: manual -> source code doc -> self-check
Phase D  Iterative Correction

Phase A: Requirement Diagnosis#

Confirm: path (patent/copyright/both), tech topic, applicant/inventor info, existing materials.

Auto domain classification (see Section "AI Domain Taxonomy" below).

Gate: 3-5 line diagnosis summary including domain + risk level.

AI Domain Taxonomy#

DomainSub-directionsHigh-Risk FlagsD1 Perceptual Intelligence2D vision, 3D vision, multi-sensor fusion3D vision: bind 4-stage pipelineD2 Cognition & LanguageNLP, multimodal LLM, RAG, knowledge graphRAG: show full 5-stage chainD3 Generative AIDiffusion, LLM text gen, cross-modal gen, AIGC watermarkMust bind condition injection method; pure content gen = rejectedD4 Decision & InteractionEmbodied AI, reinforcement learning, multi-agentMust bind sensor + actuator; RL: bind reward to concrete taskD5 AI EngineeringTraining/fine-tuning, inference deployment, data engineering, edge IoTTraining: bind to specific model architecture; inference: bind to hardwareD6 AI Safety & GovernanceAdversarial robustness, watermark/tracing, privacy, alignmentNeed concrete technical measure, not policy-level descriptionD7 Industry ApplicationsAutonomous driving, industrial, medical, financial, AI4ScienceMust bind data processing means; financial: bind to data analysisD8 Big DataDistributed computing, data pipeline, stream processing, data quality, real-time analyticsMust bind to specific application scenario; pure platform = rejected

Phase B: Project Analysis#

B.1 Auto-Detection Decision Tree#

Source files -> domain mapping:

Key fileDetected domainmodel.py, unet.py, vae.pyD3 Generative AItrain.py, finetune.pyD5 AI Engineering (Training)inference.py, triton_serve.py, onnx_export.pyD5 AI Engineering (Inference)render.py, gaussian.py, splat.pyD1 3D Visionllm.py, chat.py, rag_chain.pyD2 NLP / RAGrobot.py, vla.py, env.pyD4 Embodied AIreward.py, ppo.pyD4 Reinforcement Learningwatermark.py, embed_watermark.pyD6 AI Safety / Watermarkspark_job.py, flink_job.py, kafka_consumer.pyD8 Big Dataetl.py, data_pipeline.py, feature_store.pyD8 Big Data (Data Engineering)stream.py, realtime_analytics.pyD8 Big Data (Streaming)dataset.py, dataloader.pyD5 AI Engineering (Data)privacy.py, dp_train.pyD6 AI Safety (Privacy)config.yaml, pipeline.py + langchainD2 RAG / Agent

Also detect 6 industry contexts: medical, financial, autonomous driving, industrial, smart city, education.

B.2 Technical Points Extraction#

Priority: model definition -> training/inference -> domain-specific core -> papers/design docs -> README.

Output: Key Points List (innovations, scheme skeleton, key params, distinctions, quantifiable effects, domain classification).

Gate: Present key points list for user confirmation.

Phase C1: Patent Application#

C1.1 Prior Art Search#

Online search 2-3 rounds: CNIPA patent DB, Google Patents, arXiv. Each result: source ID, scheme summary, limitations.

CPC suggestions by domain:

• D1 3D Vision: G06T 7/50, G06T 17/00
• D2 NLP/RAG: G06F 40/30, G06N 3/08
• D3 Generative AI: G06N 3/045, G06T 13/00
• D4 Embodied: G05B 19/00, B25J 9/16
• D5 AI Engineering: G06N 3/084
• D6 AI Safety: G06F 21/60
• D7 Industry: varies by sector
• D8 Big Data: G06F 16/245, G06F 16/903

C1.2 Claims (11 Templates)#

Structure: Method (1 independent + 3-8 dependent) + System (1 independent + 3-8 dependent) + Storage Medium (1 independent).

Template selection by domain:

TemplateDomainIndependent claim skeletonT1 Model ArchitectureD1/D2/D5Predefined network -> layer composition -> feature extraction -> outputT2 3D VisionD1 3DCapture -> sparse reconstruction -> dense optimization -> **rendering (expand formula)**T3 Training StrategyD5Data construction -> model initialization -> loss design -> optimization -> convergenceT4 Multimodal FusionD1/D2Multi-modal input -> modality-specific encoding -> cross-modal alignment -> fused outputT5 RAG PipelineD2Parse -> retrieve -> rerank -> reconstruct -> generateT6 Diffusion ModelD3Noise scheduling -> condition injection (specify: cross-attention/adapter/ControlNet) -> denoising -> decodeT7 AgentD2/D4Environment perception -> task decomposition -> tool selection -> execution -> feedbackT8 Embodied IntelligenceD4Sensor input -> perception -> planning -> actuator output + **safety constraint (dependent)**T9 Inference OptimizationD5Model loading -> computation graph optimization -> kernel fusion -> outputT10 Big Data ProcessingD8Data ingestion -> distributed processing (specify: Spark/Flink/MapReduce) -> aggregation -> storage/outputT11 Data Engineering & QualityD8Data collection -> quality assessment -> anomaly detection -> cleaning -> feature extraction -> storage

Drafting rules (all domains):

1. Method + System claims in pairs
2. Independent: preamble (prior art) + "characterized by" (essential features)
3. Dependent: "according to claim X..." with further limitation
4. Every step must link to system component
5. Avoid functional limitation; prefer structural/step-based description
6. Quantify effects where possible ("improves accuracy by X%", "reduces latency to Y ms")

C1.3 Specification#

5-chapter: Tech Field -> Background (prior art + defects) -> Invention Content (problem + scheme + effects, quantified) -> Figure Description -> Specific Embodiments.

Desensitization:

• Dataset name -> "preset dataset"
• Parameter count -> "preset-scale model"
• Hardware -> "graphics processor" / "distributed computing node"
• Training duration -> "preset period"
• Framework -> "DL framework" / "distributed computing framework"
• API -> "remote interface"
• Company -> "institution"
• Specific values -> ranges

Figures (mermaid flowchart TB/LR): System architecture + method flow + domain-specific pipeline (training/rendering/data pipeline/stream topology/etc.).

C1.4 Abstract#

<=300 chars. Tech domain + core scheme + main effect. No commercial terms.

C1.5 Self-Check#

• Independent claim contains all essential features
• Dependent claims correctly reference
• Method + System + Medium triple complete
• Specification sufficiently disclosed (enabling)
• Embodiments cover all claim features
• Beneficial effects quantified
• Terminology consistent throughout
• Abstract corresponds to claim 1
• Desensitization complete (no company/person/business name leak)
• Figure numbering consistent
• Domain-specific checks passed (see below)

Domain-specific self-check:

DomainExtra checksD1 3D VisionRendering formula in claim? 4-stage pipeline?D2 NLP/RAGFull 5-stage RAG chain? Specific embedding model?D3 Generative AICondition injection method specified? Not pure content gen?D4 EmbodiedSensor + actuator bound in every step? Safety dependent claim?D5 AI EngineeringSpecific model architecture? Hardware binding for inference?D6 AI SafetyConcrete technical measure? Not policy-level?D7 Financial/MedicalData processing means bound? Not pure business method?D8 Big DataSpecific application scenario bound? Not pure platform? Distributed topology described?

Phase C2: Software Copyright#

C2.1 Software Manual (10-15 pages, >=6 screenshots)#

Structure: Introduction (env + capability) -> Installation (env + weights + config) -> Functions (core + data + API + monitoring) -> Non-functional -> FAQ.

Templates by domain:

• General AI: standard template
• 3D Vision: add rendering/visualization section
• Generative AI: add sampling/inference section
• Embodied AI: add sensor/hardware integration section
• Big Data: add data pipeline/deployment section (distributed topology, cluster config, streaming topology diagram)

C2.2 Source Code Document (front 30 + back 30 pages, >=50 lines/page)#

File priority by domain:

DomainRequired filesDomain-specific requiredD1 3D Visionmodel.py, train.py, inference.py, render.pyrender.pyD2 NLP/RAGmodel.py, train.py, inference.py, retriever.pyretriever.pyD3 Generative AImodel.py, train.py, inference.py, generate.pygenerate.pyD4 Embodiedmodel.py, train.py, inference.py, control.py, env.pycontrol.pyD5 AI Engineeringmodel.py, finetune.py, export.py, deploy.pyfinetune.pyD6 AI Safetymodel.py, watermark.py, adv_train.pywatermark.pyD8 Big Datapipeline.py, etl.py, stream.py, config.yamlpipeline.py

<3000 lines: submit all; >3000: front 1500 + back 1500 by priority.

Desensitization: Remove API keys, absolute paths, internal addresses, personal info, hardware models, cloud URLs, DB passwords. Retain algorithm comments.

C2.3 Self-Check#

• Pages >= 15
• Screenshots >= 6
• Feature coverage complete
• Non-technical description for reviewers
• Code pages with >= 50 lines/page
• Name consistency
• No secret leaks

Knowledge Index#

Deep-dive reference files for domain-specific patent writing rules, claim templates, and software copyright guides.

FileSectionsKey Contenteferences/ai-patent-claims-guide.md11 claim templates (T1-T14)Full legal claim text per template: method/system/medium triples with dependent claims; Big Data T10-T14 includedeferences/ai-patent-special.mdPatentability framework, 8 risk domains, CPC codes, desensitization rulesAI+Big Data patentability risk assessment; domain mapping; figure requirements; industry desensitization; CPC classification (7.1-7.7); 9-domain quick referenceeferences/ai-software-copyright-guide.mdType detection, source file priority, 5 domain templates, FAQDecision tree for 10+ project types; source code priority by domain; Big Data dedicated template (section 3.5); desensitization checklist; common pitfalls

Phase D: Iterative Correction#

Identify -> Locate -> Targeted fix -> Save as v{N} -> Re-run affected self-check items only. Do NOT re-run full pipeline.

Output#

text

outputs/{case-id}/
  patent/          claims.md + specification.md + abstract.md + full.md
  software-copyright/  manual.md + source_code.md

Prohibitions: No skill name/repo path/disclaimers in deliverables. No self-check section in body. No fabricated patent numbers/links. No "approximately" in claims. No commercial terms in abstract.

Quick Reference: 8 High-Risk Rejection Patterns#

PatternWhy rejectedFixPure content generation (no condition injection)"Intellectual activity rules"Specify cross-attention/adapter/ControlNet in claimsFinancial AI without data processing means"Business method"Bind to specific feature engineering + model architectureEmbodied AI without sensor/actuator binding"Pure algorithm"Add "executed via LiDAR module" + "motor controller"RAG without full pipeline"Insufficient disclosure"Show all 5 stages in method claimBig Data platform without application"Abstract idea"Bind to specific scenario (e.g., real-time traffic analytics)RL without reward function"Insufficient disclosure"Include reward computation formulaAI watermark without robustness test"Insufficient technical effect"Add adversarial/noise/compression robustness claimMedical AI without clinical validation"Insufficient enablement"Add evaluation on specific dataset with clinical metrics