The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script scripts/_common.py contains a get_replicate_token function that scans several hardcoded paths in the user's home directory (e.g., ~/.cursor/skills/avatar-talking-video/config.json) to find a Replicate API token. This involves cross-skill data access which could lead to unintended credential exposure if multiple skills share the environment.
[COMMAND_EXECUTION]: The skill uses subprocess.run in multiple locations. scripts/build_frame.py executes a Python script located at a path provided via the --gpt-image-2-script argument (defaulting to a sibling skill directory). scripts/_common.py executes ffmpeg and ffprobe for media processing tasks such as muting clips and determining duration.
[DYNAMIC_EXECUTION]: scripts/_common.py uses __import__("replicate") to dynamically load the Replicate library at runtime within the _require_replicate function.
[INDIRECT_PROMPT_INJECTION]: The skill accepts user-supplied text for actions and prompts and interpolates them into templates sent to external AI models.
Ingestion points: args.action in scripts/build_frame.py and args.video_prompt in scripts/make_broll_camera.py.
Boundary markers: The prompt templates, such as the TEMPLATE in build_frame.py, lack explicit delimiters or instructions to ignore instructions embedded in the user-provided action text.
Capability inventory: The skill has the capability to execute local commands (subprocess.run), make network API calls (replicate.run), and download files from the internet (urllib.request.urlretrieve).
Sanitization: Input strings are trimmed but not otherwise sanitized or escaped before being placed into the prompt template.

broll-avatar-camera