The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages a broad range of cloud infrastructure through natural language instructions, creating a surface for potential injection via untrusted resource names or configurations.
Ingestion points: User-provided resource names, JMESPath queries, and configuration values in SKILL.md and reference files.
Boundary markers: Implements a specific 'Agent Safety Model' in SKILL.md that classifies operations by risk (READ, WRITE, DESTRUCTIVE) and requires explicit user confirmation for non-read actions.
Capability inventory: High-privilege operations including resource deletion, RBAC management, and network security configuration across all service files.
Sanitization: Instructions mandate the use of -o json for parseable output and explicitly forbid passing secrets as CLI arguments.
[EXTERNAL_DOWNLOADS]: The skill references official installation sources for the Azure ecosystem.
Directs the agent to download the Azure CLI from Microsoft's official aka.ms and winget sources.
Includes commands for installing the Bicep compiler and CLI extensions from Microsoft's managed infrastructure.
[COMMAND_EXECUTION]: The primary purpose of the skill is to execute az shell commands.
Provides detailed syntax for resource lifecycle management (create, update, delete).
Mitigates risk by providing clear warnings against suppressing confirmation prompts (e.g., advising against --yes for destructive tasks).
[CREDENTIALS_UNSAFE]: The skill includes instructions for managing authentication and secrets.
Guides the agent to use Azure Key Vault and Managed Identities, which are industry best practices for secure identity management.
Explicitly warns against committing credentials to code or exposing passwords in shell history.

azcli