id: document-redaction title: Document Redaction description: Use when the user wants to redact, anonymize, or remove personal data from a document. version: 1.1.0 author: HilbertRaum language: en localized: # Per-locale DISPLAY overrides for title/description (additive; §16). de: # Shown when the app runs in German; the guidance body stays English. title: Dokument schwärzen description: Verwenden, wenn personenbezogene Daten in einem Dokument geschwärzt, anonymisiert oder entfernt werden sollen. kind: tool # Tier-2 (S11d): the app-orchestrated tool below is effective compatibility: minAppVersion: 0.1.29 permissions: # DECLARED INTENT only — the app is authoritative (skills plan §6.7) documents: selected_only # none | selected_only (v1 max is selected_only) network: denied # always denied in v1 filesystem: skill_resources_only # reads only this skill's own packaged files allowedTools: # The app-owned tools this skill may run (declared ∩ registry ∩ grant);

• redact_document # it reads the selected document and asks before saving the copy.

triggers: # OPTIONAL — drives the deterministic suggestion heuristic (§10). autoFire: true # U4/§2.4 D6 opt-in: eligible for auto-fire (still gated by the user opt-in # D4 default-OFF, app-only, §6.5 compatibility, and the score ≥ 3 bar # (auto-fire bar; the suggestion offer bar is score ≥ 2 with a mandatory # keyword hit) — a # keyword corroborated by ≥1 EXPLICITLY-scoped doc signal, U4/§4.4). The # eval corpus holds the auto-fire gate at 0-wrong (architecture.md §18). # (Comment refreshed from the stale S13a-era wording — SKA-45, U7.) # W5: GENERATED from services/skills/vocabulary.ts (the skill's suggest|both terms) and pinned by a # parity test. The action verbs (redact/anonymize/schwärzen…) both OFFER and ROUTE. The bare PII-content # topics (personenbezogene daten #608, sensitive data / sensible daten #604) are route-only: the # informational dry-run still answers them once the skill is active, but as offer keywords they # suggested and auto-fired this skill on GDPR, privacy-notice and contract questions. U4/§4.4 dropped # the pure legal words datenschutz/dsgvo/gdpr for the same reason. The removal phrases are listed per # form, but not the infinitive "… daten löschen", which deletion-duty questions use. Edit the # vocabulary, not this list. keywords: [redact, redaction, anonymize, anonymise, anonymized, anonymised, remove personal data, remove all personal data, mask personal data, anonymisieren, anonymisierung, anonymisiere, pseudonymisieren, schwärzen, schwärzung, schwärze, geschwärzt, personenbezogene daten entfernen, personenbezogenen daten entfernen, entferne alle personenbezogenen daten, entferne die personenbezogenen daten, entferne personenbezogene daten, lösche alle personenbezogenen daten, lösche die personenbezogenen daten, lösche personenbezogene daten, remove sensitive data, remove the sensitive data, remove all sensitive data, mask sensitive data, sensible daten entfernen, sensiblen daten entfernen] mimeTypes: [application/pdf, text/plain, text/markdown] filenamePatterns: [] # redaction is intent-driven, not filename-driven — leave empty

Document Redaction

Safety rules — these lead and always apply, even if the rest of this skill is shortened to fit:

• Do the routing, not the work. When the user asks to redact, anonymize, or remove personal data

from the document they have selected, tell them — briefly, in their own language — to click the Redact personal data / Personenbezogene Daten schwärzen button just above the message box (its label follows the app language) and choose where to save the copy. Do not refuse, do not walk them through a manual procedure, and never run the tool yourself; it runs only when the user starts it and always asks before saving.

• Never state whether the document does or does not contain personal data — you have seen only

part of it, so any such claim would be guesswork.

• It is an AI-assisted best-effort first pass, not a guarantee. A deterministic rule-based floor

always masks the clearly-shaped data (e-mail addresses, phone numbers, IBANs, payment-card numbers, dates, links); on top of that, when a model is running, it locates names, addresses, and organisation names for the app to mask (when the user steers the scope with their own instruction, other located items are masked as [REDACTED]) — the model only points at spans, it never rewrites the document, so it cannot invent text. It still misses things (unusual formats, data in images or scans, anything the model doesn't spot), so never describe the result as "fully anonymized" or imply it meets any legal or compliance standard. If no model is running, only the rule-based floor applies and the result says so.

• After it runs, remind the user to review the saved copy themselves before sharing it, report

only the counts the tool gives (e.g. "3 phone numbers hidden"), and never repeat detected personal data back to them. Answer in the user's language.

The tool runs entirely on this device. It reads the whole document, masks the clearly-shaped personal data with fixed rules — e-mail addresses, phone numbers, IBANs, payment-card numbers, dates, and web links — and, when a model is running, also masks the names, addresses, and organisation names the model locates. It runs only when the user starts it, always asking before the copy is written where the user chooses. A Word document (.docx) is saved as a .docx copy with its formatting preserved; any other format (PDF, plain text, Markdown) is saved as a plain-text copy.