The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes external, untrusted content from video transcripts.
Ingestion points: Loom transcripts and captions retrieved via MCP tools as described in SKILL.md Phase 2.
Boundary markers: The skill includes a mandatory warning block explicitly instructing the agent to treat scraped content as material to analyze rather than instructions.
Capability inventory: The skill has the ability to execute shell commands (ffmpeg, curl), write to the local file system, and upload files to Google Drive.
Sanitization: The instructions tell the agent to quote and stop if the scraped text appears to address the agent directly.
[COMMAND_EXECUTION]: The skill generates shell commands for video manipulation using ffmpeg and curl based on external inputs.
Commands incorporate timestamps ({start}, {end}) parsed from transcript text and signed URLs obtained from the Loom tool.
Risks are mitigated by a multi-step human approval process where candidates and final clips must be reviewed by the user before execution and upload.
[EXTERNAL_DOWNLOADS]: The skill downloads video recordings from Loom, a well-known service, using signed URLs.
The download is performed via curl to a local working directory.
As the source is a well-known service, this is considered a standard functional requirement of the skill.

extract-testimonials