The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands to interact with external tools, incorporating user-provided input directly into the command string.
Evidence: In SKILL.md, the workflow executes yt-dlp using the user-supplied URL: /Users/marc/Library/Python/3.9/bin/yt-dlp "[URL]" .... While the URL is wrapped in quotes, this pattern can be susceptible to command injection if the input is not strictly validated.
[PROMPT_INJECTION]: The skill processes untrusted external data (video transcriptions), creating a surface for indirect prompt injection.
Ingestion points: The skill transcribes external audio content from Instagram reels via SKILL.md (Step 2) and saves the results to a local database.
Boundary markers: Hooks are appended to hooks-database.md using blockquotes (>), which provide minimal protection against instructions that may be embedded within the reel's audio transcript.
Capability inventory: The skill has shell execution capabilities (yt-dlp, whisper) and file system access.
Sanitization: No sanitization or validation of the transcribed text is implemented before it is processed or saved to the swipe file.
[EXTERNAL_DOWNLOADS]: The installation process involves downloading software from well-known package managers and repositories.
Evidence: install.sh uses pip3 to install yt-dlp and openai-whisper from PyPI, and brew for ffmpeg. These tools are necessary for the skill's core functionality and are sourced from standard registries.
Evidence: The README and installation script reference the official Homebrew installation command using curl | bash from GitHub's raw content domain.

hooks