The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection vulnerability surface.
Ingestion points: Video titles, descriptions, and transcripts are fetched from external Loom and YouTube URLs in SKILL.md and scripts/loom_meta.py.
Boundary markers: Absent. The prompt template in scripts/analyze_video.py interpolates external metadata directly into the system prompt without using delimiters or instructions to ignore embedded commands.
Capability inventory: The skill utilizes subprocess calls for yt-dlp and performs network POST requests to Google and OpenAI APIs.
Sanitization: Absent. External content is passed to the model without filtering.
[REMOTE_CODE_EXECUTION]: Automated scanners flagged a pattern where curl is piped to python3.
Evidence: curl -s "https://www.loom.com/v1/oembed?url=LOOM_URL" | python3 -c "..." found in SKILL.md.
Context: This is identified as a false positive. The command uses the -c flag to execute a hardcoded Python string for JSON parsing of the API response, rather than executing the content of the response itself.
[EXTERNAL_DOWNLOADS]: Fetches video data and metadata from external platforms.
Evidence: Uses yt-dlp and curl to interact with loom.com, youtube.com, and Google/OpenAI API endpoints.
Context: These interactions involve well-known services and are strictly aligned with the skill's stated purpose of video analysis.

watch-video