tabular-review-lawvable

表格审阅

从多个文档中提取结构化数据，并整理成带引用的 Excel 矩阵。

必需技能

• pdf - 用于读取 PDF 文档
• docx - 用于读取 Word 文档
• xlsx - 用于创建 Excel 输出文件

工作流

步骤 1：收集用户需求

使用 AskUserQuestion 收集：

1. 文档文件夹路径 - 文档位于何处？
2. 输出文件名 - Excel 文件的名称
3. 要提取的列 - 从每份文档中提取哪些信息

列定义示例：

- Parties: Names of all parties to the agreement
- Effective Date: When the agreement becomes effective
- Term: Duration of the agreement
- Governing Law: Jurisdiction for disputes

步骤 2：查找文档

使用 Glob 查找所有文档：

Glob(pattern: "**/*.pdf", path: "<folder>")
Glob(pattern: "**/*.docx", path: "<folder>")

步骤 3：并行处理文档

启动后台 Agent，并行处理文档。每个 Agent：

• 使用 pdf 或 docx 技能读取分配的文档
• 提取每列的值
• 记录页码/段落引用
• 返回结构化 JSON

启动 Agent：

Task(
  prompt: "<agent_prompt>",
  subagent_type: "general-purpose",
  run_in_background: true
)

Agent 提示词模板：

You are processing documents for a tabular review.

DOCUMENTS TO PROCESS:
<list of document paths>

COLUMNS TO EXTRACT:
<column definitions>

For each document:
1. Read the document using the pdf skill (for .pdf) or docx skill (for .docx)
2. Extract the requested information for each column
3. Note the page number (PDF) or section (DOCX) where you found the information
4. Include a brief quote (30-50 chars) showing the source text

Return your results as JSON:
{
  "results": [
    {
      "document": "<filename>",
      "path": "<absolute_path>",
      "extractions": [
        {
          "column": "<column_name>",
          "value": "<extracted_value>",
          "page": <page_number>,
          "quote": "<brief_context_quote>"
        }
      ]
    }
  ]
}

If you cannot find information for a column, set value to "Not found" and explain in the quote field.

分配策略：

• 对于 N 份文档和 M 个 Agent，每个 Agent 处理 ceil(N/M) 份文档
• 默认最多使用 10 个 Agent
• 根据文档数量调整

步骤 4：收集结果

等待所有后台 Agent 完成：

TaskOutput(task_id: "<agent_id>", block: true)

将所有结果汇总为一个文档提取结果数组。

步骤 5：生成 Excel 输出文件

调用 xlsx 技能 创建输出文件：

Create an Excel workbook at <output_path>:

SHEET 1: "Document Review"
- Header row: Document | <Column1> | <Column2> | ...
- Data rows: One row per document

For each extraction cell:
- Cell value: The extracted text
- Cell hyperlink: file://<document_path>#page=<N> (for PDFs)
- Cell comment: "Page <N>: '<quote>'"

SHEET 2: "Summary"
- Total documents: <count>
- Documents processed: <count>
- Extraction date: <today>

JSON 架构

提取结果格式：

{
  "document": "Contract_ABC.pdf",
  "path": "/path/to/Contract_ABC.pdf",
  "extractions": [
    {
      "column": "Parties",
      "value": "Acme Corp and Beta Inc",
      "page": 1,
      "quote": "entered into between Acme Corp and Beta Inc"
    },
    {
      "column": "Effective Date",
      "value": "January 15, 2025",
      "page": 1,
      "quote": "effective as of January 15, 2025"
    }
  ]
}

Excel 输出格式

带引用的单元格：

• 值："Acme Corp and Beta Inc"
• 超链接：file:///path/to/Contract_ABC.pdf#page=1
• 批注：Page 1: "entered into between Acme Corp and Beta Inc"

颜色编码（可选）：

• 绿色：已找到值，且置信度高
• 黄色：已找到值，但不确定
• 红色：未找到值

错误处理

情况	操作
文档无法读取	记录错误，将该行标记为失败，然后继续
未找到列	将值设为 "Not found"，并在批注中说明
Agent 超时	收集部分结果，并注明未完成
缺少技能	提示用户安装所需技能

使用示例

User: I want to do a tabular review of my contracts

Claude: [Uses AskUserQuestion]
  - What folder contains your documents?
  - What should I name the output Excel file?
  - What columns do you want to extract?

User: ~/Contracts, review.xlsx, Parties/Date/Term/Governing Law

Claude: [Discovers 15 documents via Glob]
Claude: [Launches 5 background agents, 3 docs each]
Claude: [Collects results via TaskOutput]
Claude: [Creates review.xlsx via xlsx skill]

Output: review.xlsx with 15 rows, 4 columns, hyperlinks and citations