qwen_agent/skills/japanese-pii-redactor/SKILL.md

name	description
japanese-pii-redactor	Redact, anonymize, and de-identify personal information in Japanese-language or mixed-language text and tabular data while preserving analytical usefulness. Use this whenever users ask for PII redaction, PII scrub, de-identification, 個人情報匿名化, 匿名加工, 仮名化, 秘匿化, or マスキング; use it for executing anonymization rules, not for legal interpretation or general writing polish.

Japanese PII Redactor

Overview

Detect and redact personal information in Japanese-language or mixed-language content for safer sharing and analysis.

Common targets:

• Person names
• Phone numbers
• Email addresses
• Home/work addresses
• Account/member/employee identifiers
• Free-text notes containing identifiable details

Triggering Cues

Use this skill when user messages include:

• Chinese cues: 脱敏、匿名化、个人信息、隐私遮蔽、PII处理、数据清洗
• Japanese cues: 個人情報、匿名化、マスキング、伏字、漏えい対策、PII
• English cues: redact PII, anonymize Japanese data, privacy masking

Input Requirements

Ask for or infer:

1. Source text/table
2. Target output format (text/table/json)
3. Redaction strength (light/standard/strict)
4. Whether reversible pseudonyms are needed

Output Format

Always output:

1. Redacted Data
2. Redaction Rules Applied
3. Fields Preserved vs Masked
4. Residual Risk Notes

For rules section, use this schema:

Field Type	Detection Pattern	Redaction Method	Example

Workflow

1. Detect direct identifiers first (email/phone/account IDs).
2. Detect contextual identifiers (address/detail combinations).
3. Apply consistent masking policy across the dataset.
4. Keep analytical utility while minimizing re-identification risk.
5. Report what was masked and why.

Examples

Example 1

Input:

• 日文客服对话日志，需要共享给外部分析团队。

Output style:

• Replace identifiers with neutral tokens
• Preserve issue semantics and timeline

Example 2

Input:

• 员工名单（姓名、邮箱、电话、住址、员工编号）。

Output style:

• Table output with masked fields and preserved non-sensitive columns
• Explicit rule list for audit traceability

Guidelines

• Prefer consistency: same entity should map to same token within one output.
• Never expose raw originals in final output.
• Mark uncertain detections as "Needs Manual Review".
• State that redaction reduces risk but does not guarantee zero re-identification risk.