The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) because it ingests and processes untrusted external data in the form of screenplay files (scene-NN.md), project lore (lore.md), and cast information (cast.json).
Ingestion points: Screenplay scenes (projects/<p>/<ep>/scenes/scene-NN.md), project lore (lore.md), and cast/prop metadata (cast.json, props.json, movie_set.json).
Boundary markers: The instructions do not specify any delimiters or safety markers to isolate the screenplay text from the agent's operational logic, increasing the risk that embedded instructions in the screenplay could be obeyed.
Capability inventory: The agent has the capability to write JSON files to the filesystem (scene-NN.json) and execute local Python scripts (scripts/storyboard.py, scripts/render_shot.py) via the uv tool.
Sanitization: No explicit sanitization, validation, or escaping of the screenplay content is described before it is processed or woven into shot prompts.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands and local Python scripts using uv run for several lifecycle stages, including validation, compilation, and rendering. While these scripts are part of the skill's internal directory (scripts/), they represent a capability to run subprocesses.
Evidence: uv run scripts/storyboard.py validate --scene $N, uv run scripts/storyboard.py compile, and uv run scripts/render_shot.py --shot <id> --force.

spark-video-director