The Agent Skills Directory

[COMMAND_EXECUTION]: The research instructions in references/research.md permit the use of shell commands such as curl as a fallback mechanism for fetching content when standard web tools fail. While this is used for data retrieval, it represents an invocation of external system tools.
[EXTERNAL_DOWNLOADS]: The skill is designed to fetch content from user-provided seeds and scouted web candidates. This involves downloading data from arbitrary external domains during the 'Scout' and 'Collect' phases defined in SKILL.md and references/research.md.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources (web pages, code, documentation) to build its reports, creating an ingestion surface for indirect prompt injection attacks.
Ingestion points: Research units and candidates identified through web search or user seeds, accessed via web fetch or shell tools (references/research.md).
Boundary markers: The skill mandates a structured 'evidence contract' (references/evidence-contract.md) and a specific claim format, which acts as a delimiter for external content.
Capability inventory: The skill has network access via curl and the ability to delegate tasks to subagents, potentially expanding the attack surface (SKILL.md).
Sanitization: The workflow includes a verification step where the main agent must manually review all records returned by subagents and confirm cited claims against the original primary material before synthesis.

study