The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill employs a template-based workflow where the agent adapts a provided Python script (scripts/render_video.py) and executes it using uv run to generate MP4 files. The instructions include significant safety guardrails, specifically requiring the agent to review the generated code and avoid interpolating untrusted content into shell commands to prevent injection attacks.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data from URLs and local files. It includes defensive instructions to ensure the agent treats ingested content as data rather than executable instructions, minimizing the risk of indirect prompt injection.
Ingestion points: Defined in SKILL.md Step 1, which handles URL retrieval and local UTF-8 file reading.
Boundary markers: SKILL.md Step 1 explicitly states that instructions found within sources are content, not authority, and Rule 2 prohibits silent model-knowledge substitution.
Capability inventory: The agent has access to command execution (via uv run, ffmpeg, and ffprobe) and file writing for the output directory.
Sanitization: Non-negotiable Rule 11 mandates code inspection and prevents the use of untrusted source strings in shell commands.
[COMMAND_EXECUTION]: The provided scripts (scripts/check_env.py and scripts/render_video.py) use subprocess.run to interface with system utilities such as ffmpeg, ffprobe, and uv. These calls are implemented using argument lists rather than shell strings, which is a secure method for executing local commands.
[EXTERNAL_DOWNLOADS]: The skill uses the PEP-723 inline dependency format in scripts/render_video.py to manage requirements like numpy, matplotlib, and jsonschema. These are standard Python libraries that are fetched from official registries during execution.

explainify