The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes yt-dlp using subprocess.run with argument lists. This practice effectively prevents shell command injection as the input is not parsed by a shell.\n- [EXTERNAL_DOWNLOADS]: The skill fetches data from YouTube and transcriptapi.com to perform its intended research functions. These are legitimate operations targeting well-known services.\n- [CREDENTIALS_UNSAFE]: The skill manages API keys via environment variables or specific secret files within the skill's configuration directory. This is a standard and acceptable practice for secret management in agent skills and does not involve hardcoded credentials.\n- [PROMPT_INJECTION]: The skill processes external, untrusted content such as YouTube video titles and transcripts. This creates an inherent surface for indirect prompt injection, where malicious instructions embedded in video metadata or captions could potentially influence the agent's behavior during analysis.\n
Ingestion points: Video titles and channel names in scripts/yt_search.py, transcript text in scripts/yt_transcript.sh.\n
Boundary markers: None present in the output returned to the agent.\n
Capability inventory: Subprocess execution (yt-dlp) and network access (via curl and urllib) across scripts/yt_search.py and scripts/yt_transcript.sh.\n
Sanitization: No specific sanitization or filtering of the retrieved content is performed before returning it to the agent.

yt-research