The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute arbitrary shell commands for building, launching, and doctoring the application. It also encourages the agent to fix build issues, which may lead to unvetted code modifications and broad system interactions.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes the project's codebase (untrusted data) to determine verification steps. Malicious instructions placed in the project files (e.g., README or documentation) could potentially influence the generated commands that the agent is then instructed to execute.
Ingestion points: Local project files, including documentation and configuration files.
Boundary markers: The instructions do not define delimiters or warnings to prevent the agent from obeying instructions embedded within the project files.
Capability inventory: The generated skill is granted capabilities to launch processes, perform system checks, and run cleanup routines.
Sanitization: There is no requirement for the agent to sanitize or validate the extracted repository content before using it in commands.
[DYNAMIC_EXECUTION]: The skill implements a generate-and-test workflow where a secondary instruction set (the verification skill) is created and immediately executed. This automatic execution of dynamically generated content bypasses the safety of human review before the agent performs potentially impactful actions.

create-verification-skill