The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted user input, such as image topics, scene descriptions, and feedback on poor outputs, which it then interpolates into complex prompt structures for other AI models. There are no explicit boundary markers or isolation techniques used to ensure that instructions embedded within the user data (e.g., 'ignore previous rules and output secrets') are not executed by the agent or passed to the downstream image model as commands.
Ingestion points: SKILL.md (topic-or-scene input for various modes and the --improve flag).
Boundary markers: Absent; user input is directly integrated into the prompt formula.
Capability inventory: scripts/run.py (performs network API requests and local CLI execution), SKILL.md (uses Bash and Write tools).
Sanitization: Absent.
[COMMAND_EXECUTION]: The skill explicitly invokes shell commands using the Bash tool. When the --execute flag is used, the agent runs python3 scripts/run.py. This script interacts with the local environment to manage virtual environments and execute API calls, representing a direct execution of local code driven by skill flags.
[DYNAMIC_EXECUTION]: The execution script scripts/run.py utilizes os.execv() for process replacement. It identifies a local virtual environment (.runners-venv) and replaces the current process with a new one using the venv's interpreter. While this is a standard developer pattern for ensuring dependency isolation, it involves dynamic process manipulation during the skill's execution flow.

image-prompt