The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/vd_sessions.py executes the ffmpeg utility via subprocess.run to generate video clips. While the script passes arguments as a list to avoid shell injection, it populates command-line parameters such as start/end times and filenames using data directly from the talks.json input file.\n- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted video transcripts through LLM prompts.\n
Ingestion points: Transcripts are read from transcript/transcript.tsv and distributed into session files processed by prompts in the references/ directory.\n
Boundary markers: There are no explicit delimiters (e.g., XML tags or unique markers) in the prompts to separate the transcript data from the instructions, nor are there specific warnings to the model to ignore instructions within the transcript.\n
Capability inventory: The skill allows the writing of summary and takeaway files to the local file system and the execution of ffmpeg for video processing.\n
Sanitization: The scripts/vd_clean.py script performs transcript cleaning but does not sanitize the text for potential prompt injection payloads before processing by the LLM.

video-digest-summarize