gws-modelarmor_skill

Overview

This skill provides Google Model Armor functionality to filter and sanitize user-generated content for safety. It exposes commands to apply templates that check and transform prompts and model responses, and to create custom armor templates. The goal is to help teams enforce content policies before forwarding text to or from models.

How this skill works

The skill runs user text through defined armor templates that detect policy violations, remove or replace unsafe content, and return a sanitized version plus metadata about applied rules. Templates can be created and updated to match your safety requirements, and you can preview method schemas to understand required parameters and expected outputs. Helper commands let you sanitize prompts, sanitize responses, and manage templates programmatically.

When to use it

• Preprocessing user prompts before sending them to a model to reduce toxic, illicit, or disallowed content.
• Postprocessing model outputs to ensure responses meet content policy and do not reveal sensitive data.
• Automating content moderation workflows for chatbots, forums, or UGC platforms.
• Creating repeatable safety filters across services and environments via reusable templates.
• Testing and validating new safety rules before rolling them into production.

Best practices

• Design templates conservatively: prefer blocking clearly disallowed content and flagging ambiguous cases for review.
• Log both original and sanitized text along with rule metadata for auditing and model training feedback.
• Keep templates modular and versioned so you can iterate on rules without breaking integrations.
• Combine automated sanitization with human review for high-risk content or edge cases.
• Use schema inspection to validate required parameters before running sanitization commands.

Example use cases

• Sanitize incoming user prompts to eliminate hate speech and personal data before invoking a large language model.
• Filter model-generated text to redact profanity and policy-violating instructions in customer-facing bots.
• Create a template that enforces regional compliance rules and apply it across multiple services.
• Batch-process archived conversations to remove sensitive information prior to analytics or sharing.
• Build a developer workflow that previews template effects on sample inputs before deployment.