The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and audit all files within a project directory, including scripts, vision documents, and outputs produced by other specialist skills. This creates an attack surface for indirect prompt injection, where malicious instructions hidden in these files could override the supervisor's auditing logic or influence its downstream revision requests.
Ingestion points: The skill reads all files located in the project/ directory as specified in SKILL.md and README.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands within audited files were identified.
Capability inventory: The skill possesses filesystem write access, command execution via subprocess, and the ability to trigger other agent skills.
Sanitization: No sanitization or escaping of external content before interpolation into reports or revision manifests was observed.
[COMMAND_EXECUTION]: The included helper script scripts/media_registry.py utilizes subprocess.run to execute the ffprobe system utility. This is used to extract technical metadata (duration, FPS, resolution) from media files within the project. While the script correctly uses argument lists to mitigate shell injection risks, it involves running a complex external binary on untrusted media files, which carries a residual risk of exploitation via binary-level vulnerabilities.
[DYNAMIC_EXECUTION]: The skill documentation (references/graph-workflow.md) describes a 'Graph Engineering' mode that requires the execution of a validate_graph.py script and the installation of external dependencies via requirements-graph.txt. This indicates a complex execution environment where the agent is instructed to run local scripts to validate project integrity.

creator-pipeline-supervisor