The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/ingest-channel.py uses subprocess.run to execute local kb CLI commands and shutil.move for filesystem operations. This is the intended primary purpose of the skill to orchestrate knowledge base management.
The implementation uses a list-based argument format for subprocess.run which prevents shell injection.
Inputs like topic_slug are validated against a strict regex (^[a-z0-9]+(-[a-z0-9]+)*$) before being passed to shell commands.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill explicitly instructs against hardcoding or exposing credentials. It relies on standard environment variables (e.g., YOUTUBE_PROXY, OPENAI_API_KEY) or local configuration files (kb.toml) for sensitive data, which is a safe practice for CLI-integrated tools.
[INDIRECT_PROMPT_INJECTION]: The skill processes YouTube transcripts which are untrusted external data. However, it incorporates these into a structured knowledge base ('raw/' directories) and uses boundary markers () in documentation files, reducing the risk of the agent misinterpreting transcript content as instructions.
Ingestion points: YouTube channel URLs via yt-dlp and kb ingest (scripts/ingest-channel.py).
Boundary markers: Uses HTML comment markers in CLAUDE.md to wrap generated metadata.
Capability inventory: File writes to yt-channels/, raw/youtube/, and execution of the local kb binary.
Sanitization: Slug validation and JSON-safe table cell escaping are present.
[SAFE]: The skill's documentation demonstrates a strong security posture by recommending native backoff over aggressive proxy use and warning against exporting sensitive browser cookies.

kb-yt-channel