The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script executes the yt-dlp command with the --remote-components ejs:github and --js-runtimes bun flags in scripts/youtube.ts. This allows the tool to fetch and execute external parsing logic from GitHub and run JavaScript from video pages using the local Bun runtime.\n- [COMMAND_EXECUTION]: The SKILL.md file contains instructions directing the AI agent to autonomously ensure yt-dlp is available (e.g., by installing it) if it is missing from the environment. This bypasses standard user oversight for software installation and system modifications.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it ingests untrusted YouTube transcript data and passes it to an LLM via the prompts/speaker-transcript.md template for speaker identification and formatting.\n
Ingestion points: Raw transcript data fetched from YouTube InnerTube API or yt-dlp in scripts/main.ts.\n
Boundary markers: The prompt uses a simple # Transcript header but lacks strong instructions to ignore embedded commands within the content.\n
Capability inventory: The skill can perform file system writes (writeFileSync) and execute shell commands (spawnSync) in scripts/main.ts and scripts/youtube.ts.\n
Sanitization: There is no evidence of filtering or sanitizing the transcript text before it is provided to the LLM.

baoyu-youtube-transcript