The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute a local Python script, scripts/validate_plan.py, to perform automated validation of production plans (plan.json). The script itself is a benign utility that checks for timing arithmetic and reference ID consistency.
[COMMAND_EXECUTION]: The test file scripts/test_validate_plan.py utilizes the subprocess module to run the validator script against various test cases. This is standard behavior for a regression testing suite and does not pose a threat to the user environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied planning data and external assets, which presents a surface for indirect prompt injection. The risk is mitigated by the following:
Ingestion points: plan.json and external reference images are ingested for validation and generation.
Boundary markers: The skill includes a detailed QA rubric and instructions for the agent to visually verify continuity and identity, reducing the likelihood of obeying adversarial instructions embedded in data.
Capability inventory: The agent is authorized to run a local validation script and submit generation requests to video APIs.
Sanitization: The validate_plan.py script enforces strict schema and numeric constraints on the input JSON, though it does not sanitize text strings.

seedance-production-workflow