The Agent Skills Directory

[SAFE]: The skill uses specific, built-in platform tools (rp_search, rp_cite, etc.) for all network-facing tasks. This approach is more secure than using raw shell commands or curl directly, as the platform handles the execution environment and credentials internally.
[SAFE]: The skill implements strict operational boundaries, including a maximum traversal depth of 2 levels and mandatory user checkpoints after every 50 papers. These constraints prevent resource exhaustion and uncontrolled automated processes.
[SAFE]: The skill includes a relevance filtering requirement (minimum score of 7 for initial papers and 5 for citations), which acts as a natural defensive layer by limiting the processing of potentially noisy or irrelevant data.
[SAFE]: Indirect prompt injection surface analysis (Category 8):
Ingestion points: The skill ingests academic titles and abstracts via citation tools.
Boundary markers: The relevance scoring rubric acts as a qualitative boundary.
Capability inventory: Capabilities are limited to querying academic databases and writing findings to local JSON files (papers-reviewed.json, citation-graph.json).
Sanitization: Relevance scoring effectively filters out low-signal data that might contain adversarial noise.

traversing-citations