The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple shell commands including ffmpeg for audio extraction and python scripts for video silence removal and subtitle processing. These commands rely on several external scripts (e.g., smart_cut.py, transcribe_groq.py) referenced in the workflow that were not provided for analysis. The workflow also incorporates user-provided strings via CLI arguments in the finalize_subtitles.py script, requiring the agent to handle shell interpolation safely when constructing commands.
[PROMPT_INJECTION]: The skill has a surface for indirect prompt injection because it processes untrusted subtitle data derived from external video files and presents it to the agent.
Ingestion points: Subtitle content is ingested into the agent context via a generated Markdown report at working/<video-id>/_subtitles/dubious_terms.md which the agent is instructed to read.
Boundary markers: The instructions lack delimiters or explicit warnings to treat the processed subtitle content as untrusted data, which could allow adversarial instructions in subtitles to influence agent behavior.
Capability inventory: The agent possesses capabilities to execute shell commands and modify the local filesystem.
Sanitization: Subtitle content is not sanitized or escaped before being included in the reports presented to the agent.

video-editing-and-subtitles