pii-schema-classifier

What this is for

Every column name is a hint about sensitivity: email, ssn, phone, address, ip_address, credit_card, passport, birth_date. But what about customer_id (public key or PII?), notes (free text with embedded PII?), metadata_json (unknown sensitivity?). This skill classifies schema columns by sensitivity using naming patterns, then the LLM validates and adjusts based on application domain.

The dominant failure mode is the unclassified sensitive column that gets logged, exported, or exposed in an API response without protection.

PROTECTION RULE - never ~/.claude/

Read-only skill. Guard required if write mode added later.

## What You Must Do When Invoked

Step 1

1. If -help is passed, print the ## Usage block below and stop.

Step 2

2. Confirm -ProjectDir is provided and the path exists.

Step 3

3. Run: scripts/pii-scan.ps1 -ProjectDir "<path>"

Step 4

4. LLM reads the JSON output. For each column:

Step 5

• Check the patternClass and confidence from the deterministic scan.

Step 6

• For ambiguous columns: judge sensitivity based on table context and domain.

Step 7

• For all columns: validate pattern classifications, correct false positives/negatives.

Step 8

5. Confidence: proven (exact pattern match), likely (substring match),

Step 9

suspected (ambiguous column needing domain judgment).

Step 10

6. Write pii-classification-report.md to the working directory.

Usage

/pii-scan                         # interactive, prompts for directory
/pii-scan <dir>                   # scan project directory
/pii-scan -help                   # show usage

Returns JSON with columns[]: each entry {table, column, type, patternClass, sensitivity, confidence, ambiguous} plus counts: {scannedFiles, totalColumns, bySensitivity}.

Report Format

pii-classification-report.md with:

• Executive summary (total columns, breakdown by sensitivity level)
• High-sensitivity columns (require masking/encryption)
• Medium-sensitivity columns (review access controls)
• Low-sensitivity columns (monitor for context changes)
• Ambiguous columns (need domain review)
• Confidence column for every finding
• Open questions (suspected, needs human review)