The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes natural language input via the $ARGUMENTS variable to generate cloud automation logic. While the skill includes explicit instructions to 'refuse to bypass validation' and 'push back on unsafe requests,' the mapping of user-supplied text to autonomous actions in a cloud environment represents a potential injection surface.
Ingestion points: The user's flow description is passed directly into the agent context via the $ARGUMENTS parameter in SKILL.md.
Boundary markers: There are no explicit delimiters surrounding the user input in the prompt template to separate it from system instructions.
Capability inventory: The skill has access to a wide array of tools in SKILL.md for creating, modifying, running, and deleting flows (create_flow, delete_flow, run_flow, cancel_all_runs), as well as local shell execution (Bash).
Sanitization: The skill employs mandatory validation calls (validate_flow, preflight_flow) and specific negative constraints to prevent the generation of known unsafe flow configurations (e.g., preventing legacy API connections or unauthorized authentication injection).
High-Privilege Cloud Operations: The toolset includes capabilities for administrative tasks such as environment discovery, account switching, and destructive operations like deleting flows or cancelling all active runs. These tools provide the agent with significant autonomy over the Power Automate environment.
Local Command Execution: The skill is configured to allow the Bash tool, intended for performing CLI-based operations that are not covered by the primary toolset. This enables the agent to interact with the local host environment to support flow development and management workflows.

build-flow