- Read
~/.claude/plugins/config/claude-for-legal-zh/ai-governance-legal/CLAUDE.md→ existing AI system inventory (if any), regulatory registry. - Run the following workflow.
- For each system: describe function → determine provider/user role → assign risk level → map regulatory obligations.
- Output system-level entries + summary table.
/ai-governance-legal:ai-inventory "智能客服系统 v3"
/ai-governance-legal:ai-inventory --full
AI System Inventory Compilation
Purpose
Take stock of every piece of AI you are using—whether as a provider or a user, what the risk tier is, and which regulations govern it. This is the foundation layer for use-case-triage(assessing new things) and aia-generation(in-depth assessment of a single system).
Load Current State
Read ~/.claude/plugins/config/claude-for-legal-zh/ai-governance-legal/CLAUDE.md：
## AI系统清单— existing inventory (if any)## 监管注册表— applicable regulations and obligations## 红线— prohibited use case categories
Single-System Entry
When a system name or description is provided, run single-system entry.
Workflow
Step 1: Collect Basic Information
If the information provided by the user is insufficient to fill in the following fields, ask item by item:
| Field | Description |
|---|---|
| System Name | Unique identifier |
| Function Description | One paragraph—what the system does |
| AI Technology Type | Machine learning/deep learning/rule-based system/large language model/computer vision/other |
| Model Source | Self-developed/fine-tuned based on open-source models/third-party API/purchased commercial product |
| Deployment Method | On-premises deployment/private cloud/public cloud API/SaaS |
| Data Processing | Categories of data involved—whether it includes personal information, sensitive personal information, commercial data, public data |
| Affected Population | Internal employees/business customers/general public users/minors |
| Use Scenario | Internal auxiliary tool/customer-facing functionality/public-facing service |
| Decision Type | Non-substantive (recommendation, ranking)/substantive (affects rights and obligations)/safety-critical |
Step 2: Determine Role
| Role | Determination Criteria |
|---|---|
| Provider | Independently develops and provides AI services to others (including other departments within the company) |
| User | Uses third-party AI services and does not provide the AI service itself to others |
| Dual | Provides services externally after training/fine-tuning based on a third-party model |
If the system is used only internally within the company and is not provided externally, it is generally classified as a user (even if internal data is used). However, if the company develops its own model/system and provides external customers with access to that system, it is classified as a provider.
Gray Area: User-facing functionality built using an LLM API—technically, a third-party model is used, but you built the application layer and provide the service to users. Such cases are usually classified as a "dual" role: to users you are the provider, while at the same time you are a user of the underlying model.
Step 3: Assign Risk Level
| Level | Definition | Trigger Characteristics |
|---|---|---|
| High Risk | Has a material impact on rights and interests, or is aimed at vulnerable groups, or is safety-critical | Automated decision-making affecting credit/employment/education/insurance; involves sensitive personal information; aimed at minors; medical/transportation/infrastructure safety scenarios |
| Medium Risk | Aimed at the public but with no material impact on rights | Content recommendation/personalization; uses personal information but not sensitive information; generates synthetic content |
| Low Risk | Internal use, does not involve personal information, no external impact | Internal data analysis; non-personal information processing; productivity and efficiency tools |
| Not Applicable | There are no AI components in the system | Purely deterministic automation, traditional software |
Step 4: Map Regulatory Obligations
Based on role and risk level, determine obligations:
| Regulation | High Risk + Provider | Medium Risk + Provider | High Risk + User | Medium/Low Risk + User |
|---|---|---|---|---|
Generative AI Security Assessment (Article 17 of the Administrative Measures [法条原文]） |
✅ Required | ✅ Required | ❌ Not directly | ❌ Not directly |
Algorithm Filing (Article 24 of the Provisions on the Administration of Algorithm Recommendations [法条原文]） |
✅ Required (if applicable) | ✅ Required (if applicable) | ❌ | ❌ |
Science and Technology Ethics Review (Ethics Review Measures[法条原文]） |
✅ Required | ⚠️ Depending on the specific scenario | ⚠️ Depending on the specific scenario | ❌ |
Personal Information Protection Impact Assessment (Article 55 of the Personal Information Protection Law [法条原文]） |
✅ Required | ✅ Required | ✅ Required | ⚠️ Depending on the data |
Algorithm Recommendation Transparency (Article 16 of the Provisions on the Administration of Algorithm Recommendations [法条原文]） |
✅ Required | ✅ Required | ❌ | ❌ |
Deep Synthesis Labeling (Article 16 of the Provisions on the Administration of Deep Synthesis [法条原文]） |
✅ Required (if applicable) | ✅ Required (if applicable) | ❌ | ❌ |
Complaint and Reporting Mechanism (Article 15 of the Administrative Measures [法条原文]） |
✅ Required | ✅ Required | ❌ | ❌ |
Step 5: Output the Single-System Entry
### [系统名称]
**角色：** [提供者 / 使用者 / 双重]
**风险等级：** [高风险 / 中风险 / 低风险 / 不适用]
**状态：** [已部署 / 在评估 / 开发中 / 已退役]
| 属性 | 值 |
|------|---|
| 功能描述 | [一段话] |
| AI技术类型 | [类型] |
| 模型来源 | [来源] |
| 部署方式 | [方式] |
| 数据类别 | [类别列表] |
| 受影响人群 | [人群] |
| 面向对象 | [内部/客户/公众] |
| 决策类型 | [类型] |
**监管义务：**
| 义务 | 适用 | 状态 | 截止日期/完成日期 | 证据 |
|------|------|------|-------------------|------|
| [义务] | ✅/❌ | ✅已完成/⚠️进行中/❌未开始 | [日期] | [文档引用] |
Full Review --full
When using --full , run the single-system entry process for each system in ## AI系统清单 , and in addition:
- Cross-check consistency: Ensure that similar systems receive similar classifications. When a system receives a risk level different from other similar systems, flag it and explain.
- Identify inventory gaps: Check whether
## AI系统清单omits AI systems that exist in practice (for example, vendor systems that have been approved in vendor review but are not listed in the inventory). - Regulatory coverage gaps: Which regulations apply to your business but you have no systems classified as needing to comply with that regulation? Is this reasonable?
Full review summary table
## AI系统清单汇总
**审查日期：** [日期]
**系统总数：** [N] | 提供者：[N] | 使用者：[N] | 双重：[N]
**按风险等级：** 高风险：[N] | 中风险：[N] | 低风险：[N]
| 系统 | 角色 | 风险等级 | 适用法规数 | 备案状态 | 上次评估 | 差距数 |
|------|------|----------|-----------|----------|----------|--------|
| [名称] | [角色] | [等级] | [N] | [状态] | [日期] | [N] |
## 监管覆盖缺口
| 法规 | 适用系统数 | 是否有覆盖缺口？ | 说明 |
|------|-----------|----------------|------|
| [法规] | [N] | 是/否 | [说明] |
## 不一致标记
[标记任何类似系统获得不同分类的情况，附说明]
Output
Save to ~/.claude/plugins/config/claude-for-legal-zh/ai-governance-legal/outputs/ai-inventory-[日期].md. Also update the entries in ## AI系统清单 .
[工作成果头 — 按照插件配置 ## 输出]
Interface with use-case-triage .
use-case-triage Evaluate new proposals before the system enters development.ai-inventory is for creating records for systems that already exist (or are about to be deployed). A system that has passed classification typically needs to complete inventory compilation before deployment. When triage gives a "conditional-high" classification, inventory compilation should be triggered automatically.
Wrap-up
Conclude with the next-step decision tree specified in CLAUDE.md ## 输出 . Customization options: complete entry of missing systems, handle inconsistency flags, fill regulatory coverage gaps, escalate to legal counsel.
What this skill does not do
- Does not perform in-depth assessment—that is the responsibility of
aia-generation. This skill classifies systems and assigns obligations; it does not perform clause-by-clause compliance analysis. - Does not make legal judgments that "this risk level is correct." Classification is based on current regulations and technical understanding; when regulations change, reclassify.
- Does not replace the Cyberspace Administration's official classification—regulators may disagree with your risk level determination.