Audit Your Agent's PII at Scale Without Ever Leaking It

Where this fits

You've been running an AI agent for weeks, and now the compliance team asks: "How much personal data is in those session logs?" You have a folder of hundreds of transcripts, but you can't send them to a cloud API—that would defeat the purpose. You need a quick, local, aggregate-only scan. The audit skill does exactly that: it runs a layered PII detector over every file and returns counts, distributions, and a residual proxy—never a raw value or filename.

Why agents benefit

• Corpus-scale awareness: Agents can answer "how much PII across these transcripts?" without loading every file into context—the skill streams through the folder and aggregates.
• Zero-exposure reporting: The output contains only anonymized IDs and counts, so an agent can safely surface the report in a dashboard, email, or chat without risking a leak.
• Red vs. green auditing: Run it on raw (RED) sessions to size the PII problem, or on redacted (GREEN) sessions to verify your redaction holds at scale.
• Layered detection: Uses regex → Natasha → local LLM, giving a coarse residual proxy that catches what simple patterns miss—without external calls.
• Local by design: Raw text never leaves the machine, so agents can run this in air-gapped or sensitive environments.

Practical scenarios

Pre-launch PII inventory

Before shipping a customer-facing agent, you run audit over your beta test transcripts to see which PII types (emails, phones, addresses) appear most often. The aggregate report tells you where to focus your redaction rules.

Redaction regression check

After updating your redaction pipeline, you run audit on a GREEN corpus. The residual proxy shows if any PII slipped through—without exposing what it is. You can compare distributions across runs to catch regressions.

Compliance audit prep

Your legal team wants evidence that you're handling PII responsibly. Run audit on your production transcripts, generate a stats-only report, and share it—safe to email because it contains no raw data.

Add it to your agent workflow

Install via npx:

npx skills add glebis/claude-skills --skill audit

Then point it at a folder:

User: Audit my sessions in ./transcripts
Agent: I'll run the audit locally. (After processing) Here's the aggregate report: 1,234 PII instances across 87 sessions. Emails: 890, phones: 210, addresses: 134. Per-session redaction rate: median 92%, p95 78%. Residual proxy: 3 likely misses.

No raw values ever appear in the response.

Compare before adopting

Before committing, compare against other privacy tools:

• Quality signals: Does it use multiple detection layers? The regex→Natasha→local LLM stack is more robust than a single regex pass.
• Maintenance freshness: Check the GitHub repo's last commit and open issues. A stale skill may miss new PII patterns.
• Alternatives: Consider presidio (Microsoft) or custom scripts—but they often require cloud calls or expose raw data in logs.
• Workflow fit: Does it integrate with your existing agent framework? The npx install is clean, but verify it works with your local environment.

Why it is worth tracking

With 370 stars and an active author, this skill shows real community traction. It addresses a growing need: agents that handle sensitive data must prove compliance without sacrificing privacy. The stats-only output is a thoughtful design choice that makes it safe to surface reports in shared contexts. If you're building agents that process personal data, evaluate this skill now—it could become a standard part of your security toolkit. For more details, see its OpenAgentSkill listing.