The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core functionality involves fetching large amounts of untrusted data from the internet via WebSearch, WebFetch, and external GitHub repositories. This data is then processed by sub-agents and synthesized into a final HTML file. The skill lacks explicit boundary markers to delimit untrusted content and does not specify sanitization procedures to prevent malicious instructions or scripts embedded in the source material from affecting the agent's logic or the final output.
Ingestion points: Untrusted content is ingested from the web and third-party repositories as defined in reference/stage-3-subagent-templates.md.
Boundary markers: There are no specific instructions or delimiters used to separate untrusted data from the agent's internal instructions during the research or synthesis phases.
Capability inventory: The skill uses the Agent tool to invoke sub-tasks and the Write tool to generate persistent HTML and Markdown files.
Sanitization: The instructions do not mandate escaping or filtering of external content before it is written to the final 05-学习材料.html file, which could allow for Cross-Site Scripting (XSS) if the research sources contain malicious code.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to fetch and analyze content from several third-party, community-maintained repositories. These sources are outside the control of the skill author and are not from recognized trusted organizations.
Evidence: reference/stage-3-subagent-templates.md directs the agent to research sources like jujumilk3/leaked-system-prompts and x1xhlol/system-prompts-and-models-of-ai-tools.

system-study