pii-redaction

SKILL: PII Redaction

What This Skill Does

Teaches Claude Code how to write a production-safe PII redaction service for any project handling sensitive personal data — medical bills, insurance forms, financial documents, HR records, etc.

When To Use This Skill

Invoke this skill when a project needs to strip personally identifiable information from text before that text is sent to any external API, stored in a database, or logged anywhere.

Trigger phrases:

• "add PII redaction"
• "strip patient data before the API call"
• "redact sensitive information"
• "implement PHI removal"

---

Core Principles (read before writing any code)

1. Redact early, redact once PII redaction must happen as the FIRST step in any pipeline that handles user-submitted text. Never pass raw text to downstream services and rely on them to handle it safely.

2. Regex over ML models for this task Do not use spaCy, Presidio, or any NLP model for basic PII redaction unless the project explicitly requires entity-level classification. Reasons:

• Regex is auditable — you can read exactly what it catches
• ML models have false negatives that are invisible and unpredictable
• Regex has zero cold start time, no model download, no GPU needed
• For structured PII (SSN, phone, email, dates) regex is more reliable than NLP anyway

3. Replace don't delete Always replace PII with a labeled placeholder like [SSN], [NAME], [DATE_OF_BIRTH]. Deleting makes the redacted text lose structure and confuses downstream LLM reasoning. Labeled placeholders let the LLM understand what was there without seeing the actual value.

4. Log redaction counts, never redacted values Log how many items were redacted per category for debugging. Never log the original PII values even in debug mode.

5. Validate redaction before sending to any API After redacting, run a second-pass check to catch any PII that slipped through. Fail loudly if the second pass finds anything.

---

Implementation

Patterns to cover (in order of priority)