PII Scrub
wujiajun4/pii-scrub
Detecta dados pessoais antes de você publicar ou compartilhar
Ver no GitHubScanner defensivo que detecta dados pessoais (PII) antes de publicação — emails, telefones, IDs, chaves API, números de cartão, endereços de carteira cripto e pontuações acadêmicas. Opera localmente sem enviar dados para rede, permitindo triage manual pela severidade detectada. Ideal como step 0 de SOP de publicação segura.
Instalação
npx convez-ai wujiajun4/pii-scrub1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Validar código/conteúdo antes de push para repositórios públicos
- Escanear arquivos antes de compartilhar em chats, fóruns ou com LLMs
- Proteger credenciais e informações sensíveis em documentação técnica
- Implementar step 0 de procedimento de segurança em publicações
Veredito

Ferramenta sólida e minimalista para detectar vazamentos de PII antes da publicação. Abordagem local-first (sem telemetria), padrões configuráveis e triage por severidade a tornam prática; limitação é que apenas reporta, deixando decisão final ao usuário.
README
A tradução para o português deste README ainda está na fila.
🛡️ pii-scrub
A defensive scanner for personal data. Catches real PII (scores, school names, phone numbers, IDs, API keys) before it leaves your machine — to a git push, a public file share, a chat, a forum, or an LLM context.
✨ What it does
Greps your target directory for personal data patterns. Reports matches — never auto-rewrites, never sends data anywhere, never logs telemetry.
Two pattern sources:
- Built-in generic library — email, phone (US / CN / intl), credit card, IBAN, BTC/ETH wallet, US SSN, Chinese resident ID, API key prefixes (sk-, sk-ant-, ghp_, AKIA, AIza...), academic scores (PTE, IELTS, TOEFL, Duolingo, GPA, /100), IPv4/IPv6/MAC, file paths.
- Per-user
user-patterns.txt— your real school, your real phone, your real work email, anything not in the generic library.
See references/pii-patterns.md for the full pattern library.
🛡️ Severity model
| Severity | Meaning | Action |
|---|---|---|
| Critical | Almost always a leak. | Block publish. |
| High | Very likely a leak. | Block publish. |
| Medium | Could be real, could be example. | Triage. |
| Low | Probably example / boilerplate. | Note. |
The skill sorts the report by severity so you deal with Critical and High first.
📦 Install
# Clone
git clone https://github.com/<owner>/pii-scrub.git
# Or copy into your skills dir
cp -r pii-scrub ~/.claude/skills/pii-scrub
# Verify
ls ~/.claude/skills/pii-scrub/
# SKILL.md LICENSE README.md .gitignore references/ examples/
Restart Claude Code. The skill is now active and auto-triggers on words like publish, push, share, 发送, 上传, 发布前.
🚀 Usage
Quick — scan a directory
Just ask Claude anything publish-related and the skill fires:
"I want to publish this skill, scan it for PII first" "Check for any API keys or phone numbers in this folder" "发布前帮我扫一下这个目录"
Add your personal patterns
Edit ~/.claude/skills/pii-scrub/user-patterns.txt (gitignored). One
POSIX extended regex per line:
# My real school (Chinese)
江西师范大学
# My real school (English)
Jiangxi Normal University
# My PTE score
PTE\s*6[0-9]\s*\(\s*[0-9]{1,2}\s*,\s*[0-9]{1,2}\s*,\s*[0-9]{1,2}\s*,\s*[0-9]{1,2}
Re-run pii-scrub and your patterns are added to the scan.
Suppress false positives
For matches you know are intentional (e.g. a security-check command
containing the literal pattern), add to
~/.claude/skills/pii-scrub/user-patterns-whitelist.txt:
# format: <file>:<line> <reason>
SKILL.md:42 example of a generic security-check command
🔍 What it does NOT do
| It does | It does NOT |
|---|---|
| ✅ Grep for patterns | ❌ Auto-rewrite files |
| ✅ Categorise by severity | ❌ Send anything over the network |
| ✅ Show file:line + context | ❌ Remember or log results (no telemetry) |
| ✅ Sort by severity for triage | ❌ Replace deleted content with placeholders |
| ✅ Use generic + per-user patterns | ❌ Decide for you what's a leak |
You stay in control. The skill only reports.
📁 File structure
pii-scrub/
├── SKILL.md # Main entry — frontmatter + workflow
├── README.md # This file (bilingual)
├── LICENSE # MIT
├── .gitignore # Excludes user-patterns.txt + evals
├── references/
│ └── pii-patterns.md # Generic pattern library
├── examples/
│ └── before-after.md # Worked example of a scrub cycle
├── user-patterns.txt # YOUR patterns (gitignored)
└── user-patterns-whitelist.txt # YOUR suppressions (gitignored)
🛡️ Publishing SOP integration
This skill is step 0 of the user's publishing SOP — it runs BEFORE preflight. The full sequence:
0. /pii-scrub ← THIS SKILL — catch PII first
1. /preflight ← 6 content + 3 distribution checks
2. /readme ← README optimisation
3. HUMAN REVIEW ← user explicitly approves
4. ctx_execute push ← git push (via Node, bypasses sandbox)
5. /brain-sync ← sync to Memory MCP
If step 0 reports any Critical / High match, do not proceed to step 1. Scrub first.
🤝 Contributing
PRs welcome. Two specific ways to help:
- Add a new generic pattern — edit
references/pii-patterns.md(the markdown table at the top). Add the ERE, an example, and a note on false-positive rate. - Add a new pattern category — for example, "Medical" (PHN, NHS number), "Education" (NRIC, school IDs), "Travel" (passport numbers by country).
Pre-PR checks
# 1. PII scan (self-aware — see SKILL.md "Self-scan exclusion")
grep -rE "<suspicious-pattern>" ~/.claude/skills/pii-scrub/ \
--exclude-dir=references --exclude=user-patterns.txt \
|| echo "OK no PII leak"
# 2. LICENSE
ls -la LICENSE
head -3 LICENSE
# 3. File structure
ls SKILL.md README.md LICENSE .gitignore references/ examples/
⚠️ Privacy note
The pattern library in references/pii-patterns.md contains realistic
example values to teach the patterns. These are NOT real user data —
they're generic formats (e.g. 110101000000000000 style for ID, sk- proj-AbCdEf1234567890XYZ style for keys). If you scan this skill's
own directory, the patterns will match the example values. That's
expected and harmless — see SKILL.md "Self-scan exclusion" for details.
📜 License
MIT © 2026 pii-scrub contributors
🗺️ Roadmap
-
pii-scrubCLI tool (run without Claude) - GitHub Action wrapper
- Auto-generate per-user
user-patterns.txtfrom past project data - Eval suite with 10+ worked examples
- Locale-specific libraries (CN / IN / EU / US / UK)
🇨🇳 中文版
✨ 做什么
在数据离开你机器之前(git push / 公网分享 / 聊天 / LLM context)抓真实 PII。
两个模式来源:
- 内置通用库 — 邮箱 / 电话 / 信用卡 / IBAN / 加密钱包 / 美国 SSN / 中国身份证 / API key 前缀 / 学术分数 / IPv4 / MAC / 文件路径
user-patterns.txt— 你的真实学校 / 真实电话 / 真实工号 / 任何通用库没覆盖的
详细见 references/pii-patterns.md。
🛡️ 严重度模型
| 严重度 | 含义 | 处理 |
|---|---|---|
| Critical | 几乎肯定是泄漏 | block publish |
| High | 很可能泄漏 | block publish |
| Medium | 可能是真,可能是 example | triage |
| Low | 大概率是 example | 记录一下 |
📦 安装
git clone https://github.com/<owner>/pii-scrub.git
cp -r pii-scrub ~/.claude/skills/pii-scrub
ls ~/.claude/skills/pii-scrub/
重启 Claude Code。Skill 自动激活,触发词:发布 / push / share / 发送 / 上传 / 发布前。
🚀 用法
直接说"扫一下这个目录"或"发布前检查"就行。
加你的个人 pattern: 编辑 ~/.claude/skills/pii-scrub/user-patterns.txt(gitignored)。一行一个 POSIX 扩展正则。
抑制误报: 加到 user-patterns-whitelist.txt。
🛡️ Publishing SOP 集成
这是你发布 SOP 的 step 0(在 preflight 之前):
0. /pii-scrub ← 这个 skill — 先抓 PII
1. /preflight ← 6 项内容 + 3 项分发检查
2. /readme
3. HUMAN REVIEW
4. ctx_execute push
5. /brain-sync
📜 许可
MIT © 2026 pii-scrub contributors