The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch its operational logic, including phase definitions, search channels, and API configurations, from an external GitHub repository (https://github.com/Socialpranker/claude-deep-research). This repository is not associated with a trusted organization or well-known service, making the skill dependent on unverifiable remote content.
[CREDENTIALS_UNSAFE]: Phase 3.5 in SKILL.md directs the agent to "Audit available API keys/channels in the environment." This encourages the agent to discover sensitive credentials, which could lead to unauthorized access or exfiltration if the agent's research process is compromised.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process massive amounts of untrusted data from the web (29+ search channels and 460+ statistical sources), creating a large attack surface for indirect prompt injection.
Ingestion points: Untrusted web content fetched during the parallel search phase (Phase 4).
Boundary markers: The skill lacks explicit instructions to ignore or sanitize embedded commands within the research data, although it uses markdown quotes for source-grounding.
Capability inventory: File system write access (creating report folders and source files) and network access (performing searches and fetching external catalogs).
Sanitization: The skill attempts to mitigate misinformation through a "triangulation" strategy (requiring 3+ independent sources) and a mandatory "adversarial pass" (self-critique), but these do not explicitly prevent the agent from executing instructions found in the text.

meta-research