The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches content from external sources (RSS feeds, GitHub issues/commits, and JSON APIs) using scripts like watch_rss.py, watch_github.py, and watch_http_json.py. This content is intended to be summarized and delivered to the user by the agent. If these external sources contain malicious instructions (e.g., in a RSS description or GitHub issue body), the agent might execute them upon summarization.
Ingestion points: The skill uses urllib.request to fetch data from user-provided URLs in scripts/watch_rss.py, scripts/watch_github.py, and scripts/watch_http_json.py.
Boundary markers: The output is formatted as Markdown in scripts/_watermark.py, but there are no specific prompt boundary markers or instructions to the LLM to ignore embedded commands within the fetched data.
Capability inventory: The skill utilizes file system writes for state management (scripts/_watermark.py) and standard terminal output. While the scripts themselves are limited, the agent summarized the output may have broader capabilities.
Sanitization: The scripts perform basic XML/JSON parsing and length-based truncation for the 'body' field in format_items_as_markdown, but no safety-oriented sanitization of the content itself is performed.

watchers