prompt-governance

Prompt Governance

Category: Engineering

Domain: AI Governance

Overview

The Prompt Governance skill provides tools for auditing prompts for security vulnerabilities, bias, and safety issues, plus managing a versioned catalog of approved prompts. Essential for organizations deploying LLM-based applications at scale.

Quick Start

# Audit a prompt for security and safety issues
python scripts/prompt_auditor.py --file system_prompt.txt

# Audit with specific focus
python scripts/prompt_auditor.py --text "You are a helpful assistant..." --checks injection,bias,safety

# Initialize a prompt catalog
python scripts/prompt_catalog_manager.py --init --catalog-dir ./prompts

# Add a prompt to the catalog
python scripts/prompt_catalog_manager.py --add --name "customer-support-v1" --file prompt.txt --catalog-dir ./prompts

# List all prompts in catalog
python scripts/prompt_catalog_manager.py --list --catalog-dir ./prompts

Tools Overview

| Tool | Purpose | Key Flags | |------|---------|-----------| | prompt_auditor.py | Audit prompts for injection, bias, and safety | --file, --text, --checks, --format | | prompt_catalog_manager.py | Manage versioned prompt catalog | --init, --add, --list, --diff, --catalog-dir |

Workflows

Prompt Review Process

1. Author writes or modifies a prompt
2. Run prompt_auditor.py for automated checks
3. Review findings and address critical issues
4. Add approved prompt to catalog with prompt_catalog_manager.py
5. Deploy from catalog (never from ad-hoc sources)

Prompt Versioning

1. Store all prompts in catalog with semantic versioning
2. Use --diff to compare versions before promotion
3. Maintain audit trail of all prompt changes
4. Roll back to previous versions when issues detected

Reference Documentation

• [Prompt Governance Framework](references/prompt-governance-framework.md) - Policies, review processes, and compliance requirements

Common Patterns

Prompt Lifecycle

Draft -> Audit -> Review -> Approve -> Deploy -> Monitor -> Retire

Governance Checklist

• No injection vulnerabilities
• No harmful content generation potential
• Appropriate bias mitigation
• Clear scope boundaries
• Output format constraints
• Error handling instructions