AI IP Document Generation#

Generate Chinese patent applications, software copyright registration materials, or technical disclosure reports from AI project code, research papers, and design docs. Direct Word (.docx) + PPT (.pptx) output. Built-in 7 AI domain coverage with 11 claim templates.

Three output paths:

• Patent: CNIPA-format invention patent (claims + specification + abstract), with technical disclosure as intermediate deliverable
• Software Copyright: CPCC registration materials (manual + source code document)
• Technical Disclosure: From papers/notes to agent-deliverable disclosure

7 AI Domains x 22 Sub-Directions#

DomainSub-DirectionsClaim TemplatePerception AI2D CV · 3D Vision & Graphics · Audio/Video · Sensor Fusion2.1 Architecture · 2.2 3D VisionCognitive & LanguageLLM · Multimodal LLM · RAG · Knowledge Graph2.3 Training · 2.4 MLLM · 2.5 RAGGenerative AIDiffusion & Controlled Gen · AIGC Watermark · Style Transfer2.6 DiffusionDecision & InteractionAgent · Embodied AI · RL2.7 Agent · 2.8 VLAAI EngineeringTraining/Fine-tuning · Inference Opt · Data Engineering · Edge/IoT2.9 Inference · 2.10 DataAI Safety & GovernanceAdversarial Defense · Watermark · Federated Learning · Alignment2.11 WatermarkIndustry AppsAutonomous Driving · Industry QC · Healthcare · Finance · AI4Science · Digital ContentDomain-adapted

Triggers#

patent / claims / specification / software copyright / disclosure / IP application / paper-to-patent / /ai-copyright / /AI知产

Iteration: When user modifies existing output, enter iterative correction flow directly.

Overall Flow#

text

Phase 0  Patentability Pre-Assessment (patent path only)
Phase A  Requirement Diagnosis → path + domain + risk level
Phase B  Project Analysis → auto-detect (11 types + 6 industries) + extract key points
Phase C  Generation (branch by path)
  C1 Patent: search → layout → disclosure → claims (11 templates) → specification → abstract → self-check
  C2 Software Copyright: manual (4 templates) → source code doc → self-check
  C3 Technical Disclosure: mapping (6 general + 7 domain-specific) → drafting → self-check
Phase D  Confirmation Gate
Phase E  Iterative Correction
Phase F  Word Output (docx-js, auto)
Phase G  Briefing PPT (python-pptx, patent default)

Phase 0: Patentability Pre-Assessment#

Per references/ai-patent-special.md §1, check three elements:

ElementCheckPass StandardTechnical ProblemAnchored to specific scenarioNot "low efficiency" / "poor accuracy"Technical MeansAlgorithm steps bound to system architectureEach step linked to HW/SW componentTechnical EffectQuantifiableConcrete numbers or comparison baselines

Domain Risk Assessment (§1.3)#

RiskDomainPitfallCountermeasureHIGHGenerative AI (pure content gen)Classified as "rules of mental activities"Must bind to tech scenario + conditional control meansHIGHFinancial risk controlClassified as "business method"Must bind to data processing meansHIGHAI alignment / explainabilityClassified as "rules of mental activities"Must bind to safety assurance in specific app scenarioMEDEmbodied AIPure motion control = mental activitiesBind each step to sensor input + actuator outputMEDReinforcement LearningPure strategy optimization = math methodBind to physical system + physics-constrained rewardMEDRAGPure information retrieval = info expressionMust show complete technical pipelineMEDAIGC watermarkPure info marking = info expressionWatermark embedding must bind to model internal layers

Decision: All pass + low risk → proceed; All pass + high risk → proceed with mandatory domain countermeasures; Means fail → switch to disclosure path; Effect fail → supplement quantitative comparison.

Phase A: Requirement Diagnosis#

Confirm with user: path selection, tech topic, AI domain (7 domains / 22 directions, auto-detect assisted), applicant info, inventor info, existing materials.

Gate: Output 3-5 line diagnosis summary (including domain attribution and risk level).

Phase B: Project Analysis#

B.1 Auto-Detection (§1.1 Decision Tree)#

text

Entry file → Project type
app.py/main.py/serve.py → AI Service
train.py/trainer.py → AI Training
inference.py/predict.py → AI Inference
render.py/gaussian.py/nerf.py → 3D Vision
generate.py/diffusion.py → Generative AI
robot.py/vla.py → Embodied AI
pipeline.py + langchain → Agent
pipeline.py + rag/faiss → RAG System
package.json → Frontend/Fullstack

Also detect 6 industry categories (§1.3): autonomous driving / industry / healthcare / finance / AI4Science / digital content.

B.2 Key Points Extraction#

Priority: model definition → training/inference scripts → rendering/generation/control scripts → papers → design docs → README.

Output: Key Points List (architecture/algorithm/engineering-level innovations, scheme skeleton, key params, distinctions from prior art, quantifiable effects, domain attribution).

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. Suggest CPC classification (§6, 6 groups / 30+ codes).

C1.2 Layout Advisory (§2)#

Innovation DistributionLayout StrategyCore innovation in architectureSingle application (arch + training + inference)Architecture + training independent2 applicationsIndependent deployment optimization3 applications (arch + training + deployment)3D reconstruction + rendering independent2 applicationsGeneration model + control independent2 applicationsEmbodied perception + control independent2 applications (perception-decision + control-execution)

C1.3 Technical Disclosure (Intermediate Deliverable)#

7-chapter structure: Invention Name → Tech Field → Background & Prior Art Defects → Invention Content (purpose + scheme + effects) → Figure Description → Specific Embodiments (optimal + alternative) → Key Innovation Summary (3-5 points).

Data source: Phase B key points + C1.1 search results + C1.2 layout advice.

C1.4 Claims#

Per references/ai-patent-claims-guide.md. Structure: Method (1 independent + 3-8 dependent) + System (1 independent + 3-8 dependent, step-by-step correspondence) + Storage Medium (1 independent).

11 Claim Templates by Domain (§2):

#TypeDomainKey Structure2.1Model ArchitecturePerception (2D)Sub-module → processing action → intermediate result chain2.23D Vision & GraphicsPerception (3D)Capture → sparse reconstruction → dense optimization → rendering (4-stage) + rendering computation sub-steps2.3Training MethodAI EngineeringData construct → init → forward → loss → optimize → converge (6-step)2.4Multimodal LLMCognitive & LanguageMulti-encoder → projection layer → fusion → decode2.5RAGCognitive & LanguageQuery parse → retrieve → rerank → context reconstruct → generate (5-stage)2.6Diffusion & Controlled GenGenerative AICondition encode → inject → denoise → decode + condition injection sub-steps2.7AgentDecision & InteractionIntent → plan → tool execute → observe/reflect → respond2.8Embodied AI (VLA)Decision & InteractionSensor perception → encode+fuse → policy → actuator drive + feedback2.9Inference OptimizationAI EngineeringOriginal model → optimize → inference strategy → result2.10Data ProcessingAI EngineeringAcquire → preprocess → augment → downstream2.11AI Watermark & TraceabilityAI SafetyEmbed phase (position → generate → inject) + Verify phase (extract → match)

Domain-Specific Requirements (§1.3):

1. Every step must link to system component ("executed via GPU parallel computing unit")
2. 3D Vision: Must include full 4-stage pipeline; rendering step must expand rendering formula
3. Generative AI: Condition injection step must specify injection method (cross-attention/adapter/ControlNet) to avoid "pure content generation" rejection
4. Embodied AI: Every step must bind to sensor input + actuator output; include safety constraint dependent claim
5. RAG: Must show complete 5-stage technical pipeline to avoid "pure information retrieval" rejection
6. AI Watermark: Embed step must specify injection layer/position/encoding; verify step must quantify robustness metrics
7. Training: Must include training data construction, loss function, optimization strategy
8. Multimodal: Must include modality encoding → cross-modal alignment → fusion → decoding

Dependent Claim Expansion: General 5-layer (module structure → computation method → param range → preferred embodiment → training features) + 5 domain-specific expansion tables (§3.2).

C1.5 Specification#

5-chapter structure: Tech Field → Background (prior art + defects) → Invention Content (problem + scheme + beneficial effects, must be quantified) → Figure Description (13-type figure requirements in §4) → Specific Embodiments.

Desensitization Rules (§5): 6 general replacements + 14 industry-specific + 4 3DGS/NeRF-specific.

Figures: Use fenced mermaid ( flowchart TB/ LR). 13-type AI system figure requirements per §4.

C1.6 Abstract#

≤300 chars. Covers: tech domain + core scheme + main effect. No commercial terms. Replace algorithm names with generic expressions.

C1.7 Quality Self-Check (100-point)#

ItemPointsStandardIndependent claim completeness15-5 per missing featureDependent claim back-reference correctness10-3 per errorMethod + System + Medium triple complete10-5 per missing typeSpecification sufficient disclosure15Enabling for skilled personEmbodiments cover all claim features15-3 per uncovered featureBeneficial effects quantified10-3 per vague assertionTerminology consistency5-1 per synonymAbstract corresponds to claim 15-5 if notDesensitization completeness10-3 per leakFigure numbering consistency5-1 per inconsistency

≥80: deliver. 60-80: auto-fix and deliver. <60: rewrite.

Phase C2: Software Copyright#

Per references/ai-software-copyright-guide.md.

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

4 templates by project type:

Project TypeTemplateKey SectionsGeneral AI service§3.1 GeneralStandard 5 chapters3D Vision/Reconstruction§3.2 3D VisionCapture → reconstruct → render → exportGenerative AI§3.3 Gen-AICondition input → generation engine → safety managementEmbodied AI§3.4 EmbodiedPerception → decision → execution → simulation → monitoring

Key notes: Target non-technical reviewers; use [Screenshot: feature name] placeholders; describe deployment/config/monitoring for HCI requirement; declare open-source pre-trained weights outside protection scope.

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

12-level file priority (§2.1): model.py(train required) → train.py(required) → inference.py(required) → render.py(3D vision required) → dataset.py → loss.py → generate.py(gen-AI required) → control.py(embodied required) → retriever.py(RAG required) → config.yaml(optional).

Desensitization (§2.2): Remove API keys, absolute paths, internal addresses, personal info, hardware models, cloud URLs, DB passwords. Retain algorithm comments.

C2.3 Quality Self-Check#

Pages 15 + Screenshots 10 + Feature coverage 15 + Non-tech description 10 + GPU/CUDA info 10 + Code pages 15 + Lines per page 10 + Name consistency 5 + No secret leaks 10.

Phase C3: Technical Disclosure#

C3.1 Paper-to-Disclosure Mapping (§3)#

General 6-type mapping: Research Problem → Technical Problem | Contribution → Technical Solution | Method Modules → Implementation Units | Algorithm Flow → Implementation Flow | Experimental Outcome → Expected Effects | Ablation Study → Alternative Embodiments.

Domain-specific mapping (§3.2): 3D Vision (rendering formula → claim step) | Generative AI (noise schedule → claim step, condition injection → dependent claim) | Embodied AI (Sim2Real → adaptation step) | RL (reward function → computation step) | RAG (retrieval strategies → dependent claims) | AI Watermark (robustness experiments → verification step) | AI4Science (physics constraints → necessary feature).

C3.2 Drafting#

7-chapter structure (same as C1.3).

C3.3 Quality Self-Check#

Problem anchoring 15 + Implementability 20 + Quantified comparison 15 + Clear distinction 15 + Figure coverage 15 + Agent comprehensibility 10 + Completeness 10.

Phase D: Confirmation Gate#

After each phase: present summary, options: Confirm / Modify / Switch path / Save progress (.temp/ai-ip-progress.md).

Phase E: Iterative Correction#

Identify → Locate → Targeted fix → Diff annotation ( <!-- Revision -->) → Save as v{N} → Re-run affected self-check items only. Do NOT re-run full pipeline.

Phase F: Word Output#

Auto-executed after Phase C (unless user requests md only). Uses docx-js workflow: .temp/{case}-docx.js.

PathOutputPatentdisclosure.docx + patent.docx (cover + claims + spec + abstract) + briefing.pptxSoftware Copyrightmanual.docx + source_code.docxTechnical Disclosuresingle disclosure.docx

Format: Song Ti 24pt body / Hei Ti 32-28pt headings / 1.5x line spacing / 480DXA indent / 1-inch margins. Source code: Courier New 18pt, ≥50 lines/page.

Alternatives: md / docx (default) / both.

Phase G: Briefing PPT#

Auto for patent path; --ppt for others. python-pptx workflow: .temp/{case}-ppt.py.

5-8 slides: Cover → Background & Pain Points → Core Innovation → System Architecture → Effects Comparison → Patent Layout → Implementation Plan. 16:9, blue theme (#1B3A5C + #3B82F6), Microsoft YaHei font.

Output Specification#

text

outputs/{case-id}/
├── patent/          disclosure.docx + {case}.docx + briefing.pptx
├── software-copyright/  manual.docx + source_code.docx
└── disclosure/      disclosure.docx

Each path also retains Markdown intermediates. Self-check report: {case}_selfcheck.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.

Knowledge Index#

Reference FileContentsreferences/ai-patent-special.md§1 Patentability + 8 domain risks §2 Layout (incl 3D/embodied) §3 6-type + 10 domain mapping §4 13-type figures §5 6+14+4 desensitization §6 6-group CPC §7 Quick referencereferences/ai-patent-claims-guide.md§1 Triple claim principle §2 11 claim templates §3 5-layer + 5 domain-dependent expansionreferences/ai-software-copyright-guide.md§1 11-type + 6 industry detection §2 12-level priority + desensitization §3 4 manual templates §4 10 pitfalls