The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/youtube.ts executes the yt-dlp utility (or alternatives like uvx or python3 -m yt_dlp) using child_process.spawnSync.
The execution includes the --remote-components ejs:github flag, which allows yt-dlp to fetch and use extractor components from remote GitHub repositories.
It also uses the --js-runtimes bun flag to execute JavaScript-based extractors.
The command can optionally ingest browser cookies through the YOUTUBE_TRANSCRIPT_COOKIES_FROM_BROWSER environment variable to bypass bot detection.
[EXTERNAL_DOWNLOADS]: The skill performs network requests to YouTube's InnerTube API (youtube.com/youtubei/v1/player) and fetches thumbnails from i.ytimg.com. While these are well-known services, the skill downloads external data that is later processed as text.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted transcription data from YouTube and passes it to an LLM for formatting and speaker identification.
Ingestion points: Raw YouTube transcripts and video descriptions are loaded in scripts/main.ts and processed using the instructions in prompts/speaker-transcript.md.
Boundary markers: The skill uses YAML frontmatter and Markdown headers to structure the data, but the transcript content itself lacks explicit boundary markers or instructions for the agent to ignore embedded commands.
Capability inventory: The agent has the ability to write to the local file system (writeFileSync) and execute the yt-dlp command via the provided scripts.
Sanitization: The scripts implement basic HTML entity unescaping and tag stripping in scripts/shared.ts, which mitigates basic XSS-like patterns but not natural language instructions.

baoyu-youtube-transcript