The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/ingest.py executes several system commands including yt-dlp, pdftotext, and qmd via subprocess.run. These calls use static or validated arguments (such as the URL) and do not use shell=True, minimizing the risk of shell injection.
[EXTERNAL_DOWNLOADS]: The skill downloads content from external URLs (YouTube transcripts, PDFs, and general web pages) using standard libraries like urllib.request and external tools like trafilatura or yt-dlp. These operations are central to the skill's purpose of archiving web content.
[INDIRECT_PROMPT_INJECTION]: The generate_metadata function in scripts/ingest.py sends up to 3000 characters of extracted web content to an LLM (OpenAI or Anthropic) to generate titles and summaries. While this represents an ingestion point for untrusted data, the skill includes several mitigations:
Ingestion points: fetch_content retrieves data from any user-provided URL.
Boundary markers: The prompt uses a structured template asking for specific JSON keys.
Capability inventory: The LLM output is parsed as JSON and used only for metadata (title, slug, summary, tags) which is then written to a local Markdown file.
Sanitization: The script uses json.loads() and regex to sanitize the LLM response before usage, reducing the risk of a successful injection influencing the host system.

knowledge-base