PrivacyGate

A Claude Skill for anonymising personal, sensitive, health (PHI), financial, and confidential information in a document before it's summarised, analysed, indexed, archived, or shared with an AI system.

It produces two outputs:

• Mapping register — real values mapped to fictitious placeholders (Person_001, Organisation_001, etc.), kept private, used only for controlled re-identification.
• AI-safe working copy — the same document with every sensitive item replaced by its placeholder, safe to process, share, or store.

Why

Most "just anonymise it" prompts fall apart because they don't guarantee the same real value maps to the same placeholder every time, and they don't separate "what to keep private" from "what's safe to share." PrivacyGate treats this as a deterministic mapping problem: build a register first, then work exclusively from placeholders.

Status

Personal project, built and maintained independently. Not affiliated with, endorsed by, or representing any employer or organisation. See DISCLAIMER.md before you use this for anything with real legal or compliance weight.

How it works

1. Identify direct, indirect, employment, health, financial, technical, and special-category identifiers in the source content.
2. Flag combinations that could re-identify someone even without a direct identifier.
3. Classify each item (security classification, privacy category, confidence, recommended action: Remove / Generalise / Pseudonymise / Anonymise / Retain).
4. Build a deterministic, one-to-one mapping register — fictitious placeholders only.
5. Produce the AI-safe working copy.
6. Run a second pass to catch anything missed, plus a reconstruction test and residual risk note.
7. From that point on, only placeholders are used — real values aren't restated unless you explicitly ask for them back.

Full workflow detail: privacygate/SKILL.md.

Install

1. Download privacygate.skill from this repo (or clone it and package the privacygate/ folder yourself — see below).
2. In Claude, upload/install the .skill file. It installs to your skill library and does not run automatically.
3. Invoke it explicitly when you want it — e.g.:

   "Use PrivacyGate to anonymise this email thread before I paste it into a report." "Run a privacy review on this CV."

It will not trigger on casual mentions of "privacy" or "sensitive" in unrelated conversation — the trigger is intentionally narrow. See docs/usage.md for worked examples.

Repackaging from source

If you modify privacygate/SKILL.md or the reference files, repackage with the official Anthropic skill-creator packaging script (bundled with Claude's skill-creator tooling), or simply zip the privacygate/ folder — a .skill file is a zip containing the skill directory.

Repo structure

privacygate/
├── SKILL.md                          # workflow instructions (what triggers it, the 11 steps)
├── references/
│   └── taxonomy.md                   # identifier categories, loaded only when the skill runs
└── assets/
    ├── mapping-register-template.md  # register column structure (fictitious example values only)
    └── output-structure.md           # optional full 12-section formal report structure
docs/
└── usage.md                          # worked example, all fictitious data
DISCLAIMER.md
LICENSE

Limitations

• This is a workflow aid, not a certified compliance tool. It does not replace a formal Privacy Impact Assessment, legal advice, or sign-off from an actual Data Protection Officer.
• Output quality depends on the model actually running it — always review the mapping register and working copy yourself before relying on either.
• It has no persistent storage of its own across separate conversations; the mapping register is only consistent within a single session unless you save and reuse it yourself.

License

MIT — see LICENSE.