The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates data retrieval from well-known academic and scientific services, including PubMed (National Library of Medicine), PMC (PubMed Central), Unpaywall, and OpenCitations. These are established, trusted sources in the research community for fetching abstracts and full-text scientific papers.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources (academic papers and abstracts), which represents a theoretical indirect prompt injection surface.
Ingestion points: Abstracts and full-text articles are fetched via subagents (SKILL.md).
Boundary markers: The instructions explicitly require subagents to return structured JSON, which acts as a data boundary to prevent raw text from being interpreted as instructions by the main agent.
Capability inventory: The skill allows writing to local tracking files like papers-reviewed.json and SUMMARY.md via the main agent.
Sanitization: The workflow relies on the main agent parsing JSON results from subagents, reducing the risk of malicious instructions within paper content influencing the agent's behavior.
[SAFE]: The skill demonstrates security best practices by recommending specific rate-limiting delays (e.g., 500ms to 2.5s) to avoid abusing shared API limits for PubMed. It also correctly advises that subagents should return data to the main agent for consolidation rather than directly modifying tracking files, preventing potential file corruption or race conditions during parallel execution.

Subagent-Driven Literature Review