The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests transcription data (captions JSON) which may be derived from external video sources. It extracts hooks and rationales from this text and writes them to a manifest file used by other skills in a processing chain. This creates a surface for indirect prompt injection if the transcript contains instructions intended to manipulate the behavior of downstream agents.
Ingestion points: SKILL.md describes reading captions.json from $SOURCE_EXPORTS/processed/ and capturing output from the transcribe.ts utility.
Boundary markers: None specified. The manifest stores extracted transcript strings directly.
Capability inventory: The skill uses Write to create manifest files and Bash to execute processing commands.
Sanitization: No evidence of sanitization, filtering, or escaping for the transcript content before it is included in the manifest.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to perform video editing tasks via ffmpeg and to run local Python and Node.js scripts. This involves the dynamic construction of shell commands using file paths and generated expressions.
[DYNAMIC_EXECUTION]: In SKILL.md, the agent is instructed to execute Node.js code via the -e flag to handle transcription. This code dynamically imports local modules and interpolates file paths. Additionally, the skill uses a Python script (build_pan.py) to generate complex ffmpeg filter expressions at runtime.

short-form-clipping