The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches and deeply analyzes untrusted content from multiple external news sources (Hacker News, Weibo, GitHub, 36Kr, etc.) and arbitrary URLs provided in news items via the --deep flag.
Ingestion points: Untrusted data is ingested in scripts/fetch_news.py through various scrapers and APIs, and article text is extracted from remote URLs via the fetch_url_content function.
Boundary markers: No specific delimiters or "ignore embedded instructions" warnings are defined in SKILL.md to prevent the agent from following instructions found within the fetched news content.
Capability inventory: The skill allows the agent to perform network operations (via scripts/fetch_news.py) and instructs the agent to write full reports to the reports/ directory.
Sanitization: While scripts/fetch_news.py performs basic HTML cleanup (removing script and style tags), it does not sanitize the text for potentially malicious instructions designed to influence the agent's behavior during the required "Deep Analysis" phase.
[PROMPT_INJECTION]: The skill contains instructions that explicitly tell the agent to override or ignore user-specified constraints.
Instruction Override: In SKILL.md, the "Smart Fill" logic instructs the agent to disregard user-requested time windows if results are sparse, potentially providing data that the user explicitly tried to filter out.
Keyword Smuggling: The agent is instructed to "automatically expand" user keywords (e.g., expanding "AI" to a broad list including "LLM, GPT, Claude, Generative"). This forces the agent to include results the user did not explicitly request, which can be used to inject additional context or content into the final report.

news-aggregator-skill