The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external URLs, including X (formerly Twitter) and general web pages. This creates a vulnerability surface for indirect prompt injection, where malicious instructions hidden in external content (such as web page metadata or social media posts) could attempt to influence the agent's downstream behavior.
Ingestion points: The scripts/capture_web_reference.py script fetches raw HTML and JSON from external URLs, and references/web-reference-ingestion.md provides instructions for broader web ingestion.
Boundary markers: The instructions include behavioral guidance to focus on mechanisms rather than labels, but lack strong technical delimiters for the ingested content.
Capability inventory: The skill can perform network operations (via requests), file system writes (writing reference files), and command execution (running the capture script).
Sanitization: The capture script uses BeautifulSoup and trafilatura for content extraction, providing some structural validation, but does not explicitly sanitize for natural language instructions.
[EXTERNAL_DOWNLOADS]: The skill utilizes a custom Python script (scripts/capture_web_reference.py) and references external tools such as yt-dlp, gallery-dl, and playwright to download media and text content from the internet. While these are intended for research purposes, they involve connecting to and fetching data from arbitrary, user-supplied or discovered URLs.
[COMMAND_EXECUTION]: The workflow involves the agent executing a local Python script (scripts/capture_web_reference.py) to perform its content capture tasks. The script takes user-supplied URLs as arguments, though it implements basic slugification for output file naming.

reference-hunting-board