The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted competitor content which presents an indirect prompt injection surface. It incorporates specific safeguards to handle this risk.
Ingestion points: Video transcripts, captions, and metadata fetched from external social media URLs.
Boundary markers: Includes a prominent warning block at the top of the skill instructions explicitly labeling scraped content as untrusted and instructing the agent to ignore any commands within it.
Capability inventory: Capabilities include file system writes via ffmpeg, database operations via Supabase, and record creation in Airtable.
Sanitization: The instructions explicitly direct the agent to quote and stop if any scraped text attempts to address or instruct the agent directly.
[COMMAND_EXECUTION]: The skill invokes several command-line utilities to process media files.
Evidence: Uses ffmpeg for frame extraction and audio conversion, ffprobe for metadata analysis, and yt-dlp for video downloading.
[EXTERNAL_DOWNLOADS]: Fetches media and data from external platforms.
Evidence: Downloads video files from Instagram, TikTok, YouTube, and X using yt-dlp and the Apify MCP. It also downloads transcription results from the OpenAI Whisper API.
[DATA_EXFILTRATION]: Conducts network operations to interact with various cloud services.
Evidence: Performs API requests to OpenAI for transcription, Supabase for data querying, and Airtable for pipeline management. These interactions involve well-known technology providers and are consistent with the skill's primary functional purpose.

remix-reel