The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes untrusted data from external sources such as URLs, RTSP streams, and local video files using vision-language models (VLM).
Ingestion points: Data enters the agent context through coll.upload(), coll.connect_rtstream(), and video.understand(), which extracts transcriptions and visual descriptions from remote media.
Capability inventory: The agent uses Bash to execute generated Python scripts that can perform network operations and file system modifications across the SDK.
Boundary markers: The instructions do not mandate the use of delimiters or warnings to ignore instructions that might be embedded in the processed video metadata.
Sanitization: Extracted metadata from external video sources is interpolated into the agent's context without explicit sanitization or escaping procedures shown in the examples.
[DYNAMIC_EXECUTION]: The skill involves the dynamic generation and execution of Python scripts via shell heredocs and inline commands to interact with the VideoDB SDK.
Evidence: Multiple examples in SKILL.md instruct the agent to construct and run Python blocks to manage video collections, upload media, and configure analyzers at runtime.

videodb