The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently uses shell commands (such as yt-dlp, ffmpeg, ffprobe, and psql) where variables are directly interpolated into the command strings. For instance, the use of $VIDEO_URL, $VIDEO_ID, $START, and $DURATION without sanitization creates a risk of command injection if an attacker provides a malicious URL or if video metadata contains shell metacharacters.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data from external sources.
Ingestion points: Video content is downloaded from arbitrary URLs and processed into text transcripts via Whisper.
Boundary markers: There are no delimiters or instructions used to signal the agent to ignore embedded commands within the processed transcripts.
Capability inventory: The skill possesses extensive capabilities including file system modification, network access via curl, and direct database interaction via psql.
Sanitization: The skill does not sanitize or validate the extracted transcript text before passing it to subsequent processing steps or external API calls (e.g., HeyGen).
[CREDENTIALS_UNSAFE]: The documentation instructs the agent to use a direct database connection string (INSFORGE_DB_URL) within a shell command ($PSQL "$INSFORGE_DB_URL"). Because connection strings often contain plain-text passwords, this practice can lead to credentials being exposed in the system's process monitoring tools (like ps or top) or log files.
[EXTERNAL_DOWNLOADS]: The skill fetches content from external sources including video data via yt-dlp and communicates with the HeyGen and fal.ai APIs. These operations target well-known services and are documented as part of the primary workflow.

short-form-video-clone-edit