939 lawve-ai

lawve-agentic-delegation-audit-ignacio-adrian-lerer Skill

当律师、法务团队或客户需要评估能代表他人行动的 AI 智能体时使用:发送消息、搜索、起草、提交、付款、删除、连接账户、使用工具或依赖外部数据。输出一份面向法务运营的实用授权、监督、问责与控制审计。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

Lawve Agentic Delegation Audit

Use this skill to help lawyers evaluate whether an AI agent workflow is safe enough to use in legal or business operations.

Core idea:

When AI only answers, the user evaluates an output.
When AI acts, the user delegates authority.
Delegated authority needs controls, logs, revocation, and accountability.

Use When

  • A client wants to use an AI agent for legal, compliance, business, or administrative work.
  • A product, law firm, or legal ops team is connecting an agent to email, files, CRM, court portals, payments, messaging, calendars, databases, browser tools, or code execution.
  • A lawyer needs to explain the risks of agentic AI in practical governance terms.
  • A workflow may send, file, delete, approve, purchase, publish, sign, deploy, or change access.
  • A system claims to be "autonomous", "agentic", "assistant with tools", "AI employee", "legal copilot", or "workflow agent".

Do Not Use For

  • Pure doctrinal legal research with no agentic action.
  • Final legal opinions without jurisdiction-specific legal review.
  • Technical implementation of an agent runtime.
  • Approving production use of an agent without evidence, logs, and controls.

Intake

Collect only what is necessary:

  • What task does the agent perform?
  • Who is the human principal?
  • Who deploys or operates the agent?
  • What systems, accounts, files, channels, or tools can it access?
  • What actions can it take without step-by-step approval?
  • Which actions are irreversible, external, financial, legal, confidential, or reputation-sensitive?
  • What logs exist and who can read them?
  • How can the user pause, revoke, appeal, or correct the agent?
  • What data can influence the agent, including emails, webpages, chats, files, tickets, and prompts?

Audit Steps

  1. Classify the autonomy level
  • answer_only: produces information only.
  • draft_only: drafts but does not send or change records.
  • approval_gated_actor: acts only after explicit approval.
  • policy_bounded_actor: acts within predefined limits.
  • long_running_actor: continues across time, sessions, or triggers.
  1. Map delegated authority
  • Name the principal.
  • Name the deployer/operator.
  • List what authority the agent has.
  • Separate read, draft, internal write, external write, financial, legal-sensitive, and privileged actions.
  1. Check observability
  • Can the user see what the agent did?
  • Are tool calls and external actions logged?
  • Are logs understandable by a non-developer?
  • Are source data and model inference separated?
  1. Check control and revocation
  • Can the user pause the agent?
  • Can permissions be narrowed?
  • Can access be revoked quickly?
  • Are irreversible actions previewed before execution?
  1. Check accountability
  • Who is responsible if harm occurs?
  • Is responsibility split across vendor, deployer, user, professional, and client?
  • Is there a human review point before legal reliance or external action?
  1. Check attack surface
  • Can untrusted content influence the agent?
  • Does the agent process emails, webpages, chats, documents, tickets, or social posts as instructions?
  • Are external content and system instructions separated?
  • Are prompt injection, data exfiltration, and tool misuse considered?
  1. Apply legal uncertainty gate
  • Use PASS, ESCALATE, or BLOCK for downstream legal/business reliance.
  • Do not hide material uncertainty in disclaimer text.

Output

Use this compact format:

## Agentic Delegation Audit

### Verdict
PASS | NEEDS CONTROLS | BLOCK

### Why
[2-5 sentences]

### Delegated Authority
- Principal:
- Deployer/operator:
- Autonomy level:
- Systems/tools:
- Highest-risk action:

### Control Checklist
- Permission scope: adequate | weak | missing
- Human approval before external/legal/financial action: yes | partial | no
- User-readable logs: yes | partial | no
- Revocation/pause: yes | partial | no
- Prompt-injection/data-boundary controls: yes | partial | no
- Accountability owner: clear | partial | unclear

### Required Controls
- [control 1]
- [control 2]
- [control 3]

### Legal Reliance Gate
PASS | ESCALATE | BLOCK

### Next Step
[smallest practical next step]

Decision Rules

  • PASS: agent is draft-only or tightly approval-gated, logs are clear, permissions are scoped, and no material legal/client risk remains unmanaged.
  • NEEDS CONTROLS: agent may be useful, but missing controls prevent safe operational reliance.
  • BLOCK: agent can perform external, legal, financial, confidential, destructive, or privileged actions without adequate approval, logging, revocation, or accountability.