The Agent Skills Directory

[SAFE]: The skill outlines a robust and defensive methodology for deep research, prioritizing evidence verification and source grounding over simple summarization. The process is structured to maintain a high level of epistemic rigor and logical consistency.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data from the web, creating an attack surface for indirect prompt injection. However, the skill explicitly enforces architectural 'invariants' (such as Invariants 1 and 4) that prevent raw content from being directly processed by the synthesis model, and Invariant 2 requires targeted verification by separate agents. These measures serve as strong boundary markers and sanitization steps that mitigate the risks associated with processing external content.
[EXTERNAL_DOWNLOADS]: The skill makes use of external search and retrieval tools (e.g., Exa, Jina, arXiv) and academic sources. These references are appropriate for the skill's primary function and do not involve downloading or executing untrusted code or scripts.
[DATA_EXFILTRATION]: The skill utilizes a session-specific temporary directory for storing research artifacts. The data storage is localized to this workspace, and the workflow is focused on reporting findings to the user, with no evidence of unauthorized data transfer to external domains.

deep-research