The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/download_video.sh interpolates the $QUERY variable directly into a shell command for yt-dlp. Because it is placed within double quotes, the shell will evaluate any command substitutions (e.g., $(command)) contained within the search query provided in scenes.json.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data via the scenes.json configuration file, which serves as a vulnerability surface.
Ingestion points: The scenes.json file is read by scripts/generate_tts.sh, scripts/render_all.sh, and src/index.tsx.
Boundary markers: There are no explicit delimiters or instructions to the agent to ignore embedded commands within the data fields.
Capability inventory: The skill has access to shell execution via yt-dlp, ffmpeg, ffprobe, and npx remotion.
Sanitization: No sanitization or escaping of the searchQuery or ttsText fields is performed before they are passed to shell environments or subprocesses.
[DYNAMIC_EXECUTION]: The script scripts/generate_tts.sh utilizes a bash heredoc to execute an embedded Python block. Bash variables like $SCENES_FILE and $VOICE are interpolated directly into the Python source code. This pattern can lead to syntax breakage or code injection if the variable values contain single quotes or other Python-sensitive characters.

search-video-on-web-and-gen