The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes local shell commands to manage the video production pipeline. This includes rendering shots, generating storyboards, and fanning out tasks across chain groups.
Evidence: Execution of uv run scripts/render_shot.py, uv run scripts/storyboard.py graph, and uv run scripts/render_all.py as documented in SKILL.md.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its critique-to-rewrite feedback loop. Malicious instructions or data could be injected via the automated reviews, potentially influencing the subsequent generated video prompts and their execution parameters.
Ingestion points: Critique data provided by the scoring model is read from the out.review.critique field in SKILL.md.
Boundary markers: While rewrite-system.md provides structural rules, it lacks explicit delimiters or instructions to ignore embedded commands within the ingested critique text.
Capability inventory: Rewritten prompts are used as arguments to the render_shot.py script (referenced in SKILL.md), which executes these instructions through a video generation model.
Sanitization: The skill does not perform validation or sanitization on the model-generated critique text before incorporating it into the prompt rewriting process.
[EXTERNAL_DOWNLOADS]: The skill utilizes the uv tool for environment management, which involves downloading and installing Python packages from public registries.
Evidence: The use of uv run in SKILL.md implies standard package management and downloading behaviors.

spark-video-clip-review